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                 goto out;
 593         }
 594         ptr = ((struct newreno *)tp->ccv->cc_data);
 595         if (CC_ALGO(tp)->ctl_output == NULL)  {
 596                 /* Huh, why does new_reno no longer have a set function? */
 597                 printf("no ctl_output for algo:%s\n", tp->cc_algo->name);
 598                 goto out;
 599         }
 600         if (ptr == NULL) {
 601                 /* Just the default values */
 602                 old.beta = V_newreno_beta_ecn;
 603                 old.beta_ecn = V_newreno_beta_ecn;
 604                 old.newreno_flags = 0;
 605         } else {
 606                 old.beta = ptr->beta;
 607                 old.beta_ecn = ptr->beta_ecn;
 608                 old.newreno_flags = ptr->newreno_flags;
 609         }
 610         sopt.sopt_valsize = sizeof(struct cc_newreno_opts);
 611         sopt.sopt_dir = SOPT_SET;
 612         opt.name = CC_NEWRENO_BETA;
 613         opt.val = rack->r_ctl.rc_saved_beta.beta;
 614         error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
 615         if (error)  {
 616                 printf("Error returned by ctl_output %d\n", error);
 617                 goto out;
 618         }
 619         /*
 620          * Hack alert we need to set in our newreno_flags
 621          * so that Abe behavior is also applied.
 622          */
 623         ((struct newreno *)tp->ccv->cc_data)->newreno_flags = CC_NEWRENO_BETA_ECN;
 624         opt.name = CC_NEWRENO_BETA_ECN;
 625         opt.val = rack->r_ctl.rc_saved_beta.beta_ecn;
 626         error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
 627         if (error) {
 628                 printf("Error returned by ctl_output %d\n", error);
 629                 goto out;
 630         }
 631         /* Save off the original values for restoral */
 632         memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno));
 633 out:
 634         if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
 635                 union tcp_log_stackspecific log;
 636                 struct timeval tv;
 637
 638                 ptr = ((struct newreno *)tp->ccv->cc_data);
 639                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
 640                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
 641                 if (ptr) {
 642                         log.u_bbr.flex1 = ptr->beta;
 643                         log.u_bbr.flex2 = ptr->beta_ecn;
 644                         log.u_bbr.flex3 = ptr->newreno_flags;
 645                 }
 646                 log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta;
 647                 log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn;
 648                 log.u_bbr.flex6 = rack->r_ctl.rc_saved_beta.newreno_flags;
 649                 log.u_bbr.flex7 = rack->gp_ready;
 650                 log.u_bbr.flex7 <<= 1;
 651                 log.u_bbr.flex7 |= rack->use_fixed_rate;
 652                 log.u_bbr.flex7 <<= 1;
 653                 log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
 654                 log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
 655                 log.u_bbr.flex8 = 3;
 656                 tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, error,
 657                                0, &log, false, NULL, NULL, 0, &tv);
 658         }
 659 }
 660
 661 static void
 662 rack_undo_cc_pacing(struct tcp_rack *rack)
 663 {
 664         struct newreno old, *ptr;
 665         struct tcpcb *tp;
 666
 667         if (rack->rc_pacing_cc_set == 0)
 668                 return;
 669         tp = rack->rc_tp;
 670         rack->rc_pacing_cc_set = 0;
 671         if (tp->cc_algo == NULL)
 672                 /* Tcb is leaving */
 673                 return;
 674         if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
 675                 /* Not new-reno nothing to do! */
 676                 return;
 677         }
 678         ptr = ((struct newreno *)tp->ccv->cc_data);
 679         if (ptr == NULL) {
 680                 /*
 681                  * This happens at rack_fini() if the
 682                  * cc module gets freed on us. In that
 683                  * case we loose our "new" settings but
 684                  * thats ok, since the tcb is going away anyway.
 685                  */
 686                 return;
 687         }
 688         /* Grab out our set values */
 689         memcpy(&old, ptr, sizeof(struct newreno));
 690         /* Copy back in the original values */
 691         memcpy(ptr, &rack->r_ctl.rc_saved_beta, sizeof(struct newreno));
 692         /* Now save back the values we had set in (for when pacing is restored) */
 693         memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno));
 694         if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
 695                 union tcp_log_stackspecific log;
 696                 struct timeval tv;
 697
 698                 ptr = ((struct newreno *)tp->ccv->cc_data);
 699                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
 700                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
 701                 log.u_bbr.flex1 = ptr->beta;
 702                 log.u_bbr.flex2 = ptr->beta_ecn;
 703                 log.u_bbr.flex3 = ptr->newreno_flags;
 704                 log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta;
 705                 log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn;
 706                 log.u_bbr.flex6 = rack->r_ctl.rc_saved_beta.newreno_flags;
 707                 log.u_bbr.flex7 = rack->gp_ready;
 708                 log.u_bbr.flex7 <<= 1;
 709                 log.u_bbr.flex7 |= rack->use_fixed_rate;
 710                 log.u_bbr.flex7 <<= 1;
 711                 log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
 712                 log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
 713                 log.u_bbr.flex8 = 4;
 714                 tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
 715                                0, &log, false, NULL, NULL, 0, &tv);
 716         }
 717 }
 718
 719 #ifdef NETFLIX_PEAKRATE
 720 static inline void
 721 rack_update_peakrate_thr(struct tcpcb *tp)
 722 {
 723         /* Keep in mind that t_maxpeakrate is in B/s. */
 724         uint64_t peak;
 725         peak = uqmax((tp->t_maxseg * 2),
 726                      (((uint64_t)tp->t_maxpeakrate * (uint64_t)(tp->t_srtt)) / (uint64_t)HPTS_USEC_IN_SEC));
 727         tp->t_peakrate_thr = (uint32_t)uqmin(peak, UINT32_MAX);
 728 }
 729 #endif
 730
 731 static int
 732 sysctl_rack_clear(SYSCTL_HANDLER_ARGS)
 733 {
 734         uint32_t stat;
 735         int32_t error;
 736         int i;
 737
 738         error = SYSCTL_OUT(req, &rack_clear_counter, sizeof(uint32_t));
 739         if (error || req->newptr == NULL)
 740                 return error;
 741
 742         error = SYSCTL_IN(req, &stat, sizeof(uint32_t));
 743         if (error)
 744                 return (error);
 745         if (stat == 1) {
 746 #ifdef INVARIANTS
 747                 printf("Clearing RACK counters\n");
 748 #endif
 749                 counter_u64_zero(rack_badfr);
 750                 counter_u64_zero(rack_badfr_bytes);
 751                 counter_u64_zero(rack_rtm_prr_retran);
 752                 counter_u64_zero(rack_rtm_prr_newdata);
 753                 counter_u64_zero(rack_timestamp_mismatch);
 754                 counter_u64_zero(rack_reorder_seen);
 755                 counter_u64_zero(rack_tlp_tot);
 756                 counter_u64_zero(rack_tlp_newdata);
 757                 counter_u64_zero(rack_tlp_retran);
 758                 counter_u64_zero(rack_tlp_retran_bytes);
 759                 counter_u64_zero(rack_tlp_retran_fail);
 760                 counter_u64_zero(rack_to_tot);
 761                 counter_u64_zero(rack_to_arm_rack);
 762                 counter_u64_zero(rack_to_arm_tlp);
 763                 counter_u64_zero(rack_paced_segments);
 764                 counter_u64_zero(rack_calc_zero);
 765                 counter_u64_zero(rack_calc_nonzero);
 766                 counter_u64_zero(rack_unpaced_segments);
 767                 counter_u64_zero(rack_saw_enobuf);
 768                 counter_u64_zero(rack_saw_enobuf_hw);
 769                 counter_u64_zero(rack_saw_enetunreach);
 770                 counter_u64_zero(rack_per_timer_hole);
 771                 counter_u64_zero(rack_large_ackcmp);
 772                 counter_u64_zero(rack_small_ackcmp);
 773 #ifdef INVARIANTS
 774                 counter_u64_zero(rack_adjust_map_bw);
 775 #endif
 776                 counter_u64_zero(rack_to_alloc_hard);
 777                 counter_u64_zero(rack_to_alloc_emerg);
 778                 counter_u64_zero(rack_sack_proc_all);
 779                 counter_u64_zero(rack_fto_send);
 780                 counter_u64_zero(rack_fto_rsm_send);
 781                 counter_u64_zero(rack_extended_rfo);
 782                 counter_u64_zero(rack_hw_pace_init_fail);
 783                 counter_u64_zero(rack_hw_pace_lost);
 784                 counter_u64_zero(rack_sbsndptr_wrong);
 785                 counter_u64_zero(rack_sbsndptr_right);
 786                 counter_u64_zero(rack_non_fto_send);
 787                 counter_u64_zero(rack_nfto_resend);
 788                 counter_u64_zero(rack_sack_proc_short);
 789                 counter_u64_zero(rack_sack_proc_restart);
 790                 counter_u64_zero(rack_to_alloc);
 791                 counter_u64_zero(rack_to_alloc_limited);
 792                 counter_u64_zero(rack_alloc_limited_conns);
 793                 counter_u64_zero(rack_split_limited);
 794                 for (i = 0; i < MAX_NUM_OF_CNTS; i++) {
 795                         counter_u64_zero(rack_proc_comp_ack[i]);
 796                 }
 797                 counter_u64_zero(rack_multi_single_eq);
 798                 counter_u64_zero(rack_proc_non_comp_ack);
 799                 counter_u64_zero(rack_find_high);
 800                 counter_u64_zero(rack_sack_attacks_detected);
 801                 counter_u64_zero(rack_sack_attacks_reversed);
 802                 counter_u64_zero(rack_sack_used_next_merge);
 803                 counter_u64_zero(rack_sack_used_prev_merge);
 804                 counter_u64_zero(rack_sack_splits);
 805                 counter_u64_zero(rack_sack_skipped_acked);
 806                 counter_u64_zero(rack_ack_total);
 807                 counter_u64_zero(rack_express_sack);
 808                 counter_u64_zero(rack_sack_total);
 809                 counter_u64_zero(rack_move_none);
 810                 counter_u64_zero(rack_move_some);
 811                 counter_u64_zero(rack_used_tlpmethod);
 812                 counter_u64_zero(rack_used_tlpmethod2);
 813                 counter_u64_zero(rack_enter_tlp_calc);
 814                 counter_u64_zero(rack_progress_drops);
 815                 counter_u64_zero(rack_tlp_does_nada);
 816                 counter_u64_zero(rack_try_scwnd);
 817                 counter_u64_zero(rack_collapsed_win);
 818         }
 819         rack_clear_counter = 0;
 820         return (0);
 821 }
 822
 823 static void
 824 rack_init_sysctls(void)
 825 {
 826         int i;
 827         struct sysctl_oid *rack_counters;
 828         struct sysctl_oid *rack_attack;
 829         struct sysctl_oid *rack_pacing;
 830         struct sysctl_oid *rack_timely;
 831         struct sysctl_oid *rack_timers;
 832         struct sysctl_oid *rack_tlp;
 833         struct sysctl_oid *rack_misc;
 834         struct sysctl_oid *rack_measure;
 835         struct sysctl_oid *rack_probertt;
 836         struct sysctl_oid *rack_hw_pacing;
 837
 838         rack_attack = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 839             SYSCTL_CHILDREN(rack_sysctl_root),
 840             OID_AUTO,
 841             "sack_attack",
 842             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 843             "Rack Sack Attack Counters and Controls");
 844         rack_counters = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 845             SYSCTL_CHILDREN(rack_sysctl_root),
 846             OID_AUTO,
 847             "stats",
 848             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 849             "Rack Counters");
 850         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 851             SYSCTL_CHILDREN(rack_sysctl_root),
 852             OID_AUTO, "rate_sample_method", CTLFLAG_RW,
 853             &rack_rate_sample_method , USE_RTT_LOW,
 854             "What method should we use for rate sampling 0=high, 1=low ");
 855         /* Probe rtt related controls */
 856         rack_probertt = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 857             SYSCTL_CHILDREN(rack_sysctl_root),
 858             OID_AUTO,
 859             "probertt",
 860             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 861             "ProbeRTT related Controls");
 862         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 863             SYSCTL_CHILDREN(rack_probertt),
 864             OID_AUTO, "exit_per_hpb", CTLFLAG_RW,
 865             &rack_atexit_prtt_hbp, 130,
 866             "What percentage above goodput do we clamp CA/SS to at exit on high-BDP path 110%");
 867         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 868             SYSCTL_CHILDREN(rack_probertt),
 869             OID_AUTO, "exit_per_nonhpb", CTLFLAG_RW,
 870             &rack_atexit_prtt, 130,
 871             "What percentage above goodput do we clamp CA/SS to at exit on a non high-BDP path 100%");
 872         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 873             SYSCTL_CHILDREN(rack_probertt),
 874             OID_AUTO, "gp_per_mul", CTLFLAG_RW,
 875             &rack_per_of_gp_probertt, 60,
 876             "What percentage of goodput do we pace at in probertt");
 877         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 878             SYSCTL_CHILDREN(rack_probertt),
 879             OID_AUTO, "gp_per_reduce", CTLFLAG_RW,
 880             &rack_per_of_gp_probertt_reduce, 10,
 881             "What percentage of goodput do we reduce every gp_srtt");
 882         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 883             SYSCTL_CHILDREN(rack_probertt),
 884             OID_AUTO, "gp_per_low", CTLFLAG_RW,
 885             &rack_per_of_gp_lowthresh, 40,
 886             "What percentage of goodput do we allow the multiplier to fall to");
 887         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 888             SYSCTL_CHILDREN(rack_probertt),
 889             OID_AUTO, "time_between", CTLFLAG_RW,
 890             & rack_time_between_probertt, 96000000,
 891             "How many useconds between the lowest rtt falling must past before we enter probertt");
 892         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 893             SYSCTL_CHILDREN(rack_probertt),
 894             OID_AUTO, "safety", CTLFLAG_RW,
 895             &rack_probe_rtt_safety_val, 2000000,
 896             "If not zero, provides a maximum usecond that you can stay in probertt (2sec = 2000000)");
 897         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 898             SYSCTL_CHILDREN(rack_probertt),
 899             OID_AUTO, "sets_cwnd", CTLFLAG_RW,
 900             &rack_probe_rtt_sets_cwnd, 0,
 901             "Do we set the cwnd too (if always_lower is on)");
 902         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 903             SYSCTL_CHILDREN(rack_probertt),
 904             OID_AUTO, "maxdrainsrtts", CTLFLAG_RW,
 905             &rack_max_drain_wait, 2,
 906             "Maximum number of gp_srtt's to hold in drain waiting for flight to reach goal");
 907         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 908             SYSCTL_CHILDREN(rack_probertt),
 909             OID_AUTO, "mustdrainsrtts", CTLFLAG_RW,
 910             &rack_must_drain, 1,
 911             "We must drain this many gp_srtt's waiting for flight to reach goal");
 912         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 913             SYSCTL_CHILDREN(rack_probertt),
 914             OID_AUTO, "goal_use_min_entry", CTLFLAG_RW,
 915             &rack_probertt_use_min_rtt_entry, 1,
 916             "Should we use the min-rtt to calculate the goal rtt (else gp_srtt) at entry");
 917         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 918             SYSCTL_CHILDREN(rack_probertt),
 919             OID_AUTO, "goal_use_min_exit", CTLFLAG_RW,
 920             &rack_probertt_use_min_rtt_exit, 0,
 921             "How to set cwnd at exit, 0 - dynamic, 1 - use min-rtt, 2 - use curgprtt, 3 - entry gp-rtt");
 922         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 923             SYSCTL_CHILDREN(rack_probertt),
 924             OID_AUTO, "length_div", CTLFLAG_RW,
 925             &rack_probertt_gpsrtt_cnt_div, 0,
 926             "How many recent goodput srtt periods plus hold tim does probertt last (bottom of fraction)");
 927         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 928             SYSCTL_CHILDREN(rack_probertt),
 929             OID_AUTO, "length_mul", CTLFLAG_RW,
 930             &rack_probertt_gpsrtt_cnt_mul, 0,
 931             "How many recent goodput srtt periods plus hold tim does probertt last (top of fraction)");
 932         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 933             SYSCTL_CHILDREN(rack_probertt),
 934             OID_AUTO, "holdtim_at_target", CTLFLAG_RW,
 935             &rack_min_probertt_hold, 200000,
 936             "What is the minimum time we hold probertt at target");
 937         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 938             SYSCTL_CHILDREN(rack_probertt),
 939             OID_AUTO, "filter_life", CTLFLAG_RW,
 940             &rack_probertt_filter_life, 10000000,
 941             "What is the time for the filters life in useconds");
 942         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 943             SYSCTL_CHILDREN(rack_probertt),
 944             OID_AUTO, "lower_within", CTLFLAG_RW,
 945             &rack_probertt_lower_within, 10,
 946             "If the rtt goes lower within this percentage of the time, go into probe-rtt");
 947         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 948             SYSCTL_CHILDREN(rack_probertt),
 949             OID_AUTO, "must_move", CTLFLAG_RW,
 950             &rack_min_rtt_movement, 250,
 951             "How much is the minimum movement in rtt to count as a drop for probertt purposes");
 952         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 953             SYSCTL_CHILDREN(rack_probertt),
 954             OID_AUTO, "clear_is_cnts", CTLFLAG_RW,
 955             &rack_probertt_clear_is, 1,
 956             "Do we clear I/S counts on exiting probe-rtt");
 957         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 958             SYSCTL_CHILDREN(rack_probertt),
 959             OID_AUTO, "hbp_extra_drain", CTLFLAG_RW,
 960             &rack_max_drain_hbp, 1,
 961             "How many extra drain gpsrtt's do we get in highly buffered paths");
 962         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 963             SYSCTL_CHILDREN(rack_probertt),
 964             OID_AUTO, "hbp_threshold", CTLFLAG_RW,
 965             &rack_hbp_thresh, 3,
 966             "We are highly buffered if min_rtt_seen / max_rtt_seen > this-threshold");
 967         /* Pacing related sysctls */
 968         rack_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 969             SYSCTL_CHILDREN(rack_sysctl_root),
 970             OID_AUTO,
 971             "pacing",
 972             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 973             "Pacing related Controls");
 974         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 975             SYSCTL_CHILDREN(rack_pacing),
 976             OID_AUTO, "max_pace_over", CTLFLAG_RW,
 977             &rack_max_per_above, 30,
 978             "What is the maximum allowable percentage that we can pace above (so 30 = 130% of our goal)");
 979         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 980             SYSCTL_CHILDREN(rack_pacing),
 981             OID_AUTO, "pace_to_one", CTLFLAG_RW,
 982             &rack_pace_one_seg, 0,
 983             "Do we allow low b/w pacing of 1MSS instead of two");
 984         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 985             SYSCTL_CHILDREN(rack_pacing),
 986             OID_AUTO, "limit_wsrtt", CTLFLAG_RW,
 987             &rack_limit_time_with_srtt, 0,
 988             "Do we limit pacing time based on srtt");
 989         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 990             SYSCTL_CHILDREN(rack_pacing),
 991             OID_AUTO, "init_win", CTLFLAG_RW,
 992             &rack_default_init_window, 0,
 993             "Do we have a rack initial window 0 = system default");
 994         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 995             SYSCTL_CHILDREN(rack_pacing),
 996             OID_AUTO, "gp_per_ss", CTLFLAG_RW,
 997             &rack_per_of_gp_ss, 250,
 998             "If non zero, what percentage of goodput to pace at in slow start");
 999         SYSCTL_ADD_U16(&rack_sysctl_ctx,
1000             SYSCTL_CHILDREN(rack_pacing),
1001             OID_AUTO, "gp_per_ca", CTLFLAG_RW,
1002             &rack_per_of_gp_ca, 150,
1003             "If non zero, what percentage of goodput to pace at in congestion avoidance");
1004         SYSCTL_ADD_U16(&rack_sysctl_ctx,
1005             SYSCTL_CHILDREN(rack_pacing),
1006             OID_AUTO, "gp_per_rec", CTLFLAG_RW,
1007             &rack_per_of_gp_rec, 200,
1008             "If non zero, what percentage of goodput to pace at in recovery");
1009         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1010             SYSCTL_CHILDREN(rack_pacing),
1011             OID_AUTO, "pace_max_seg", CTLFLAG_RW,
1012             &rack_hptsi_segments, 40,
1013             "What size is the max for TSO segments in pacing and burst mitigation");
1014         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1015             SYSCTL_CHILDREN(rack_pacing),
1016             OID_AUTO, "burst_reduces", CTLFLAG_RW,
1017             &rack_slot_reduction, 4,
1018             "When doing only burst mitigation what is the reduce divisor");
1019         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1020             SYSCTL_CHILDREN(rack_sysctl_root),
1021             OID_AUTO, "use_pacing", CTLFLAG_RW,
1022             &rack_pace_every_seg, 0,
1023             "If set we use pacing, if clear we use only the original burst mitigation");
1024         SYSCTL_ADD_U64(&rack_sysctl_ctx,
1025             SYSCTL_CHILDREN(rack_pacing),
1026             OID_AUTO, "rate_cap", CTLFLAG_RW,
1027             &rack_bw_rate_cap, 0,
1028             "If set we apply this value to the absolute rate cap used by pacing");
1029         SYSCTL_ADD_U8(&rack_sysctl_ctx,
1030             SYSCTL_CHILDREN(rack_sysctl_root),
1031             OID_AUTO, "req_measure_cnt", CTLFLAG_RW,
1032             &rack_req_measurements, 1,
1033             "If doing dynamic pacing, how many measurements must be in before we start pacing?");
1034         /* Hardware pacing */
1035         rack_hw_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1036             SYSCTL_CHILDREN(rack_sysctl_root),
1037             OID_AUTO,
1038             "hdwr_pacing",
1039             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1040             "Pacing related Controls");
1041         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1042             SYSCTL_CHILDREN(rack_hw_pacing),
1043             OID_AUTO, "rwnd_factor", CTLFLAG_RW,
1044             &rack_hw_rwnd_factor, 2,
1045             "How many times does snd_wnd need to be bigger than pace_max_seg so we will hold off and get more acks?");
1046         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1047             SYSCTL_CHILDREN(rack_hw_pacing),
1048             OID_AUTO, "pace_enobuf_mult", CTLFLAG_RW,
1049             &rack_enobuf_hw_boost_mult, 2,
1050             "By how many time_betweens should we boost the pacing time if we see a ENOBUFS?");
1051         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1052             SYSCTL_CHILDREN(rack_hw_pacing),
1053             OID_AUTO, "pace_enobuf_max", CTLFLAG_RW,
1054             &rack_enobuf_hw_max, 2,
1055             "What is the max boost the pacing time if we see a ENOBUFS?");
1056         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1057             SYSCTL_CHILDREN(rack_hw_pacing),
1058             OID_AUTO, "pace_enobuf_min", CTLFLAG_RW,
1059             &rack_enobuf_hw_min, 2,
1060             "What is the min boost the pacing time if we see a ENOBUFS?");
1061         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1062             SYSCTL_CHILDREN(rack_hw_pacing),
1063             OID_AUTO, "enable", CTLFLAG_RW,
1064             &rack_enable_hw_pacing, 0,
1065             "Should RACK attempt to use hw pacing?");
1066         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1067             SYSCTL_CHILDREN(rack_hw_pacing),
1068             OID_AUTO, "rate_cap", CTLFLAG_RW,
1069             &rack_hw_rate_caps, 1,
1070             "Does the highest hardware pacing rate cap the rate we will send at??");
1071         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1072             SYSCTL_CHILDREN(rack_hw_pacing),
1073             OID_AUTO, "rate_min", CTLFLAG_RW,
1074             &rack_hw_rate_min, 0,
1075             "Do we need a minimum estimate of this many bytes per second in order to engage hw pacing?");
1076         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1077             SYSCTL_CHILDREN(rack_hw_pacing),
1078             OID_AUTO, "rate_to_low", CTLFLAG_RW,
1079             &rack_hw_rate_to_low, 0,
1080             "If we fall below this rate, dis-engage hw pacing?");
1081         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1082             SYSCTL_CHILDREN(rack_hw_pacing),
1083             OID_AUTO, "up_only", CTLFLAG_RW,
1084             &rack_hw_up_only, 1,
1085             "Do we allow hw pacing to lower the rate selected?");
1086         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1087             SYSCTL_CHILDREN(rack_hw_pacing),
1088             OID_AUTO, "extra_mss_precise", CTLFLAG_RW,
1089             &rack_hw_pace_extra_slots, 2,
1090             "If the rates between software and hardware match precisely how many extra time_betweens do we get?");
1091         rack_timely = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1092             SYSCTL_CHILDREN(rack_sysctl_root),
1093             OID_AUTO,
1094             "timely",
1095             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1096             "Rack Timely RTT Controls");
1097         /* Timely based GP dynmics */
1098         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1099             SYSCTL_CHILDREN(rack_timely),
1100             OID_AUTO, "upper", CTLFLAG_RW,
1101             &rack_gp_per_bw_mul_up, 2,
1102             "Rack timely upper range for equal b/w (in percentage)");
1103         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1104             SYSCTL_CHILDREN(rack_timely),
1105             OID_AUTO, "lower", CTLFLAG_RW,
1106             &rack_gp_per_bw_mul_down, 4,
1107             "Rack timely lower range for equal b/w (in percentage)");
1108         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1109             SYSCTL_CHILDREN(rack_timely),
1110             OID_AUTO, "rtt_max_mul", CTLFLAG_RW,
1111             &rack_gp_rtt_maxmul, 3,
1112             "Rack timely multipler of lowest rtt for rtt_max");
1113         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1114             SYSCTL_CHILDREN(rack_timely),
1115             OID_AUTO, "rtt_min_div", CTLFLAG_RW,
1116             &rack_gp_rtt_mindiv, 4,
1117             "Rack timely divisor used for rtt + (rtt * mul/divisor) for check for lower rtt");
1118         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1119             SYSCTL_CHILDREN(rack_timely),
1120             OID_AUTO, "rtt_min_mul", CTLFLAG_RW,
1121             &rack_gp_rtt_minmul, 1,
1122             "Rack timely multiplier used for rtt + (rtt * mul/divisor) for check for lower rtt");
1123         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1124             SYSCTL_CHILDREN(rack_timely),
1125             OID_AUTO, "decrease", CTLFLAG_RW,
1126             &rack_gp_decrease_per, 20,
1127             "Rack timely decrease percentage of our GP multiplication factor");
1128         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1129             SYSCTL_CHILDREN(rack_timely),
1130             OID_AUTO, "increase", CTLFLAG_RW,
1131             &rack_gp_increase_per, 2,
1132             "Rack timely increase perentage of our GP multiplication factor");
1133         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1134             SYSCTL_CHILDREN(rack_timely),
1135             OID_AUTO, "lowerbound", CTLFLAG_RW,
1136             &rack_per_lower_bound, 50,
1137             "Rack timely lowest percentage we allow GP multiplier to fall to");
1138         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1139             SYSCTL_CHILDREN(rack_timely),
1140             OID_AUTO, "upperboundss", CTLFLAG_RW,
1141             &rack_per_upper_bound_ss, 0,
1142             "Rack timely higest percentage we allow GP multiplier in SS to raise to (0 is no upperbound)");
1143         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1144             SYSCTL_CHILDREN(rack_timely),
1145             OID_AUTO, "upperboundca", CTLFLAG_RW,
1146             &rack_per_upper_bound_ca, 0,
1147             "Rack timely higest percentage we allow GP multiplier to CA raise to (0 is no upperbound)");
1148         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1149             SYSCTL_CHILDREN(rack_timely),
1150             OID_AUTO, "dynamicgp", CTLFLAG_RW,
1151             &rack_do_dyn_mul, 0,
1152             "Rack timely do we enable dynmaic timely goodput by default");
1153         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1154             SYSCTL_CHILDREN(rack_timely),
1155             OID_AUTO, "no_rec_red", CTLFLAG_RW,
1156             &rack_gp_no_rec_chg, 1,
1157             "Rack timely do we prohibit the recovery multiplier from being lowered");
1158         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1159             SYSCTL_CHILDREN(rack_timely),
1160             OID_AUTO, "red_clear_cnt", CTLFLAG_RW,
1161             &rack_timely_dec_clear, 6,
1162             "Rack timely what threshold do we count to before another boost during b/w decent");
1163         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1164             SYSCTL_CHILDREN(rack_timely),
1165             OID_AUTO, "max_push_rise", CTLFLAG_RW,
1166             &rack_timely_max_push_rise, 3,
1167             "Rack timely how many times do we push up with b/w increase");
1168         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1169             SYSCTL_CHILDREN(rack_timely),
1170             OID_AUTO, "max_push_drop", CTLFLAG_RW,
1171             &rack_timely_max_push_drop, 3,
1172             "Rack timely how many times do we push back on b/w decent");
1173         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1174             SYSCTL_CHILDREN(rack_timely),
1175             OID_AUTO, "min_segs", CTLFLAG_RW,
1176             &rack_timely_min_segs, 4,
1177             "Rack timely when setting the cwnd what is the min num segments");
1178         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1179             SYSCTL_CHILDREN(rack_timely),
1180             OID_AUTO, "noback_max", CTLFLAG_RW,
1181             &rack_use_max_for_nobackoff, 0,
1182             "Rack timely when deciding if to backoff on a loss, do we use under max rtt else min");
1183         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1184             SYSCTL_CHILDREN(rack_timely),
1185             OID_AUTO, "interim_timely_only", CTLFLAG_RW,
1186             &rack_timely_int_timely_only, 0,
1187             "Rack timely when doing interim timely's do we only do timely (no b/w consideration)");
1188         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1189             SYSCTL_CHILDREN(rack_timely),
1190             OID_AUTO, "nonstop", CTLFLAG_RW,
1191             &rack_timely_no_stopping, 0,
1192             "Rack timely don't stop increase");
1193         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1194             SYSCTL_CHILDREN(rack_timely),
1195             OID_AUTO, "dec_raise_thresh", CTLFLAG_RW,
1196             &rack_down_raise_thresh, 100,
1197             "If the CA or SS is below this threshold raise on the first 3 b/w lowers (0=always)");
1198         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1199             SYSCTL_CHILDREN(rack_timely),
1200             OID_AUTO, "bottom_drag_segs", CTLFLAG_RW,
1201             &rack_req_segs, 1,
1202             "Bottom dragging if not these many segments outstanding and room");
1203
1204         /* TLP and Rack related parameters */
1205         rack_tlp = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1206             SYSCTL_CHILDREN(rack_sysctl_root),
1207             OID_AUTO,
1208             "tlp",
1209             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1210             "TLP and Rack related Controls");
1211         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1212             SYSCTL_CHILDREN(rack_tlp),
1213             OID_AUTO, "use_rrr", CTLFLAG_RW,
1214             &use_rack_rr, 1,
1215             "Do we use Rack Rapid Recovery");
1216         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1217             SYSCTL_CHILDREN(rack_tlp),
1218             OID_AUTO, "post_rec_labc", CTLFLAG_RW,
1219             &rack_max_abc_post_recovery, 2,
1220             "Since we do early recovery, do we override the l_abc to a value, if so what?");
1221         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1222             SYSCTL_CHILDREN(rack_tlp),
1223             OID_AUTO, "nonrxt_use_cr", CTLFLAG_RW,
1224             &rack_non_rxt_use_cr, 0,
1225             "Do we use ss/ca rate if in recovery we are transmitting a new data chunk");
1226         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1227             SYSCTL_CHILDREN(rack_tlp),
1228             OID_AUTO, "tlpmethod", CTLFLAG_RW,
1229             &rack_tlp_threshold_use, TLP_USE_TWO_ONE,
1230             "What method do we do for TLP time calc 0=no-de-ack-comp, 1=ID, 2=2.1, 3=2.2");
1231         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1232             SYSCTL_CHILDREN(rack_tlp),
1233             OID_AUTO, "limit", CTLFLAG_RW,
1234             &rack_tlp_limit, 2,
1235             "How many TLP's can be sent without sending new data");
1236         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1237             SYSCTL_CHILDREN(rack_tlp),
1238             OID_AUTO, "use_greater", CTLFLAG_RW,
1239             &rack_tlp_use_greater, 1,
1240             "Should we use the rack_rtt time if its greater than srtt");
1241         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1242             SYSCTL_CHILDREN(rack_tlp),
1243             OID_AUTO, "tlpminto", CTLFLAG_RW,
1244             &rack_tlp_min, 10000,
1245             "TLP minimum timeout per the specification (in microseconds)");
1246         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1247             SYSCTL_CHILDREN(rack_tlp),
1248             OID_AUTO, "send_oldest", CTLFLAG_RW,
1249             &rack_always_send_oldest, 0,
1250             "Should we always send the oldest TLP and RACK-TLP");
1251         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1252             SYSCTL_CHILDREN(rack_tlp),
1253             OID_AUTO, "rack_tlimit", CTLFLAG_RW,
1254             &rack_limited_retran, 0,
1255             "How many times can a rack timeout drive out sends");
1256         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1257             SYSCTL_CHILDREN(rack_tlp),
1258             OID_AUTO, "tlp_cwnd_flag", CTLFLAG_RW,
1259             &rack_lower_cwnd_at_tlp, 0,
1260             "When a TLP completes a retran should we enter recovery");
1261         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1262             SYSCTL_CHILDREN(rack_tlp),
1263             OID_AUTO, "reorder_thresh", CTLFLAG_RW,
1264             &rack_reorder_thresh, 2,
1265             "What factor for rack will be added when seeing reordering (shift right)");
1266         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1267             SYSCTL_CHILDREN(rack_tlp),
1268             OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW,
1269             &rack_tlp_thresh, 1,
1270             "What divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)");
1271         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1272             SYSCTL_CHILDREN(rack_tlp),
1273             OID_AUTO, "reorder_fade", CTLFLAG_RW,
1274             &rack_reorder_fade, 60000000,
1275             "Does reorder detection fade, if so how many microseconds (0 means never)");
1276         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1277             SYSCTL_CHILDREN(rack_tlp),
1278             OID_AUTO, "pktdelay", CTLFLAG_RW,
1279             &rack_pkt_delay, 1000,
1280             "Extra RACK time (in microseconds) besides reordering thresh");
1281
1282         /* Timer related controls */
1283         rack_timers = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1284             SYSCTL_CHILDREN(rack_sysctl_root),
1285             OID_AUTO,
1286             "timers",
1287             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1288             "Timer related controls");
1289         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1290             SYSCTL_CHILDREN(rack_timers),
1291             OID_AUTO, "persmin", CTLFLAG_RW,
1292             &rack_persist_min, 250000,
1293             "What is the minimum time in microseconds between persists");
1294         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1295             SYSCTL_CHILDREN(rack_timers),
1296             OID_AUTO, "persmax", CTLFLAG_RW,
1297             &rack_persist_max, 2000000,
1298             "What is the largest delay in microseconds between persists");
1299         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1300             SYSCTL_CHILDREN(rack_timers),
1301             OID_AUTO, "delayed_ack", CTLFLAG_RW,
1302             &rack_delayed_ack_time, 40000,
1303             "Delayed ack time (40ms in microseconds)");
1304         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1305             SYSCTL_CHILDREN(rack_timers),
1306             OID_AUTO, "minrto", CTLFLAG_RW,
1307             &rack_rto_min, 30000,
1308             "Minimum RTO in microseconds -- set with caution below 1000 due to TLP");
1309         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1310             SYSCTL_CHILDREN(rack_timers),
1311             OID_AUTO, "maxrto", CTLFLAG_RW,
1312             &rack_rto_max, 4000000,
1313             "Maxiumum RTO in microseconds -- should be at least as large as min_rto");
1314         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1315             SYSCTL_CHILDREN(rack_timers),
1316             OID_AUTO, "minto", CTLFLAG_RW,
1317             &rack_min_to, 1000,
1318             "Minimum rack timeout in microseconds");
1319         /* Measure controls */
1320         rack_measure = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1321             SYSCTL_CHILDREN(rack_sysctl_root),
1322             OID_AUTO,
1323             "measure",
1324             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1325             "Measure related controls");
1326         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1327             SYSCTL_CHILDREN(rack_measure),
1328             OID_AUTO, "wma_divisor", CTLFLAG_RW,
1329             &rack_wma_divisor, 8,
1330             "When doing b/w calculation what is the  divisor for the WMA");
1331         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1332             SYSCTL_CHILDREN(rack_measure),
1333             OID_AUTO, "end_cwnd", CTLFLAG_RW,
1334             &rack_cwnd_block_ends_measure, 0,
1335             "Does a cwnd just-return end the measurement window (app limited)");
1336         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1337             SYSCTL_CHILDREN(rack_measure),
1338             OID_AUTO, "end_rwnd", CTLFLAG_RW,
1339             &rack_rwnd_block_ends_measure, 0,
1340             "Does an rwnd just-return end the measurement window (app limited -- not persists)");
1341         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1342             SYSCTL_CHILDREN(rack_measure),
1343             OID_AUTO, "min_target", CTLFLAG_RW,
1344             &rack_def_data_window, 20,
1345             "What is the minimum target window (in mss) for a GP measurements");
1346         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1347             SYSCTL_CHILDREN(rack_measure),
1348             OID_AUTO, "goal_bdp", CTLFLAG_RW,
1349             &rack_goal_bdp, 2,
1350             "What is the goal BDP to measure");
1351         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1352             SYSCTL_CHILDREN(rack_measure),
1353             OID_AUTO, "min_srtts", CTLFLAG_RW,
1354             &rack_min_srtts, 1,
1355             "What is the goal BDP to measure");
1356         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1357             SYSCTL_CHILDREN(rack_measure),
1358             OID_AUTO, "min_measure_tim", CTLFLAG_RW,
1359             &rack_min_measure_usec, 0,
1360             "What is the Minimum time time for a measurement if 0, this is off");
1361         /* Misc rack controls */
1362         rack_misc = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1363             SYSCTL_CHILDREN(rack_sysctl_root),
1364             OID_AUTO,
1365             "misc",
1366             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1367             "Misc related controls");
1368 #ifdef TCP_ACCOUNTING
1369         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1370             SYSCTL_CHILDREN(rack_misc),
1371             OID_AUTO, "tcp_acct", CTLFLAG_RW,
1372             &rack_tcp_accounting, 0,
1373             "Should we turn on TCP accounting for all rack sessions?");
1374 #endif
1375         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1376             SYSCTL_CHILDREN(rack_misc),
1377             OID_AUTO, "prr_addback_max", CTLFLAG_RW,
1378             &rack_prr_addbackmax, 2,
1379             "What is the maximum number of MSS we allow to be added back if prr can't send all its data?");
1380         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1381             SYSCTL_CHILDREN(rack_misc),
1382             OID_AUTO, "stats_gets_ms", CTLFLAG_RW,
1383             &rack_stats_gets_ms_rtt, 1,
1384             "What do we feed the stats framework (1 = ms_rtt, 0 = us_rtt, 2 = ms_rtt from hdwr, > 2 usec rtt from hdwr)?");
1385         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1386             SYSCTL_CHILDREN(rack_misc),
1387             OID_AUTO, "clientlowbuf", CTLFLAG_RW,
1388             &rack_client_low_buf, 0,
1389             "Client low buffer level (below this we are more aggressive in DGP exiting recovery (0 = off)?");
1390         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1391             SYSCTL_CHILDREN(rack_misc),
1392             OID_AUTO, "defprofile", CTLFLAG_RW,
1393             &rack_def_profile, 0,
1394             "Should RACK use a default profile (0=no, num == profile num)?");
1395         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1396             SYSCTL_CHILDREN(rack_misc),
1397             OID_AUTO, "cmpack", CTLFLAG_RW,
1398             &rack_use_cmp_acks, 1,
1399             "Should RACK have LRO send compressed acks");
1400         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1401             SYSCTL_CHILDREN(rack_misc),
1402             OID_AUTO, "fsb", CTLFLAG_RW,
1403             &rack_use_fsb, 1,
1404             "Should RACK use the fast send block?");
1405         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1406             SYSCTL_CHILDREN(rack_misc),
1407             OID_AUTO, "rfo", CTLFLAG_RW,
1408             &rack_use_rfo, 1,
1409             "Should RACK use rack_fast_output()?");
1410         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1411             SYSCTL_CHILDREN(rack_misc),
1412             OID_AUTO, "rsmrfo", CTLFLAG_RW,
1413             &rack_use_rsm_rfo, 1,
1414             "Should RACK use rack_fast_rsm_output()?");
1415         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1416             SYSCTL_CHILDREN(rack_misc),
1417             OID_AUTO, "shared_cwnd", CTLFLAG_RW,
1418             &rack_enable_shared_cwnd, 1,
1419             "Should RACK try to use the shared cwnd on connections where allowed");
1420         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1421             SYSCTL_CHILDREN(rack_misc),
1422             OID_AUTO, "limits_on_scwnd", CTLFLAG_RW,
1423             &rack_limits_scwnd, 1,
1424             "Should RACK place low end time limits on the shared cwnd feature");
1425         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1426             SYSCTL_CHILDREN(rack_misc),
1427             OID_AUTO, "non_paced_lro_queue", CTLFLAG_RW,
1428             &rack_enable_mqueue_for_nonpaced, 0,
1429             "Should RACK use mbuf queuing for non-paced connections");
1430         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1431             SYSCTL_CHILDREN(rack_misc),
1432             OID_AUTO, "iMac_dack", CTLFLAG_RW,
1433             &rack_use_imac_dack, 0,
1434             "Should RACK try to emulate iMac delayed ack");
1435         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1436             SYSCTL_CHILDREN(rack_misc),
1437             OID_AUTO, "no_prr", CTLFLAG_RW,
1438             &rack_disable_prr, 0,
1439             "Should RACK not use prr and only pace (must have pacing on)");
1440         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1441             SYSCTL_CHILDREN(rack_misc),
1442             OID_AUTO, "bb_verbose", CTLFLAG_RW,
1443             &rack_verbose_logging, 0,
1444             "Should RACK black box logging be verbose");
1445         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1446             SYSCTL_CHILDREN(rack_misc),
1447             OID_AUTO, "data_after_close", CTLFLAG_RW,
1448             &rack_ignore_data_after_close, 1,
1449             "Do we hold off sending a RST until all pending data is ack'd");
1450         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1451             SYSCTL_CHILDREN(rack_misc),
1452             OID_AUTO, "no_sack_needed", CTLFLAG_RW,
1453             &rack_sack_not_required, 1,
1454             "Do we allow rack to run on connections not supporting SACK");
1455         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1456             SYSCTL_CHILDREN(rack_misc),
1457             OID_AUTO, "prr_sendalot", CTLFLAG_RW,
1458             &rack_send_a_lot_in_prr, 1,
1459             "Send a lot in prr");
1460         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1461             SYSCTL_CHILDREN(rack_misc),
1462             OID_AUTO, "autoscale", CTLFLAG_RW,
1463             &rack_autosndbuf_inc, 20,
1464             "What percentage should rack scale up its snd buffer by?");
1465         /* Sack Attacker detection stuff */
1466         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1467             SYSCTL_CHILDREN(rack_attack),
1468             OID_AUTO, "detect_highsackratio", CTLFLAG_RW,
1469             &rack_highest_sack_thresh_seen, 0,
1470             "Highest sack to ack ratio seen");
1471         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1472             SYSCTL_CHILDREN(rack_attack),
1473             OID_AUTO, "detect_highmoveratio", CTLFLAG_RW,
1474             &rack_highest_move_thresh_seen, 0,
1475             "Highest move to non-move ratio seen");
1476         rack_ack_total = counter_u64_alloc(M_WAITOK);
1477         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1478             SYSCTL_CHILDREN(rack_attack),
1479             OID_AUTO, "acktotal", CTLFLAG_RD,
1480             &rack_ack_total,
1481             "Total number of Ack's");
1482         rack_express_sack = counter_u64_alloc(M_WAITOK);
1483         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1484             SYSCTL_CHILDREN(rack_attack),
1485             OID_AUTO, "exp_sacktotal", CTLFLAG_RD,
1486             &rack_express_sack,
1487             "Total expresss number of Sack's");
1488         rack_sack_total = counter_u64_alloc(M_WAITOK);
1489         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1490             SYSCTL_CHILDREN(rack_attack),
1491             OID_AUTO, "sacktotal", CTLFLAG_RD,
1492             &rack_sack_total,
1493             "Total number of SACKs");
1494         rack_move_none = counter_u64_alloc(M_WAITOK);
1495         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1496             SYSCTL_CHILDREN(rack_attack),
1497             OID_AUTO, "move_none", CTLFLAG_RD,
1498             &rack_move_none,
1499             "Total number of SACK index reuse of postions under threshold");
1500         rack_move_some = counter_u64_alloc(M_WAITOK);
1501         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1502             SYSCTL_CHILDREN(rack_attack),
1503             OID_AUTO, "move_some", CTLFLAG_RD,
1504             &rack_move_some,
1505             "Total number of SACK index reuse of postions over threshold");
1506         rack_sack_attacks_detected = counter_u64_alloc(M_WAITOK);
1507         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1508             SYSCTL_CHILDREN(rack_attack),
1509             OID_AUTO, "attacks", CTLFLAG_RD,
1510             &rack_sack_attacks_detected,
1511             "Total number of SACK attackers that had sack disabled");
1512         rack_sack_attacks_reversed = counter_u64_alloc(M_WAITOK);
1513         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1514             SYSCTL_CHILDREN(rack_attack),
1515             OID_AUTO, "reversed", CTLFLAG_RD,
1516             &rack_sack_attacks_reversed,
1517             "Total number of SACK attackers that were later determined false positive");
1518         rack_sack_used_next_merge = counter_u64_alloc(M_WAITOK);
1519         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1520             SYSCTL_CHILDREN(rack_attack),
1521             OID_AUTO, "nextmerge", CTLFLAG_RD,
1522             &rack_sack_used_next_merge,
1523             "Total number of times we used the next merge");
1524         rack_sack_used_prev_merge = counter_u64_alloc(M_WAITOK);
1525         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1526             SYSCTL_CHILDREN(rack_attack),
1527             OID_AUTO, "prevmerge", CTLFLAG_RD,
1528             &rack_sack_used_prev_merge,
1529             "Total number of times we used the prev merge");
1530         /* Counters */
1531         rack_fto_send = counter_u64_alloc(M_WAITOK);
1532         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1533             SYSCTL_CHILDREN(rack_counters),
1534             OID_AUTO, "fto_send", CTLFLAG_RD,
1535             &rack_fto_send, "Total number of rack_fast_output sends");
1536         rack_fto_rsm_send = counter_u64_alloc(M_WAITOK);
1537         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1538             SYSCTL_CHILDREN(rack_counters),
1539             OID_AUTO, "fto_rsm_send", CTLFLAG_RD,
1540             &rack_fto_rsm_send, "Total number of rack_fast_rsm_output sends");
1541         rack_nfto_resend = counter_u64_alloc(M_WAITOK);
1542         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1543             SYSCTL_CHILDREN(rack_counters),
1544             OID_AUTO, "nfto_resend", CTLFLAG_RD,
1545             &rack_nfto_resend, "Total number of rack_output retransmissions");
1546         rack_non_fto_send = counter_u64_alloc(M_WAITOK);
1547         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1548             SYSCTL_CHILDREN(rack_counters),
1549             OID_AUTO, "nfto_send", CTLFLAG_RD,
1550             &rack_non_fto_send, "Total number of rack_output first sends");
1551         rack_extended_rfo = counter_u64_alloc(M_WAITOK);
1552         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1553             SYSCTL_CHILDREN(rack_counters),
1554             OID_AUTO, "rfo_extended", CTLFLAG_RD,
1555             &rack_extended_rfo, "Total number of times we extended rfo");
1556
1557         rack_hw_pace_init_fail = counter_u64_alloc(M_WAITOK);
1558         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1559             SYSCTL_CHILDREN(rack_counters),
1560             OID_AUTO, "hwpace_init_fail", CTLFLAG_RD,
1561             &rack_hw_pace_init_fail, "Total number of times we failed to initialize hw pacing");
1562         rack_hw_pace_lost = counter_u64_alloc(M_WAITOK);
1563
1564         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1565             SYSCTL_CHILDREN(rack_counters),
1566             OID_AUTO, "hwpace_lost", CTLFLAG_RD,
1567             &rack_hw_pace_lost, "Total number of times we failed to initialize hw pacing");
1568
1569
1570
1571         rack_badfr = counter_u64_alloc(M_WAITOK);
1572         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1573             SYSCTL_CHILDREN(rack_counters),
1574             OID_AUTO, "badfr", CTLFLAG_RD,
1575             &rack_badfr, "Total number of bad FRs");
1576         rack_badfr_bytes = counter_u64_alloc(M_WAITOK);
1577         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1578             SYSCTL_CHILDREN(rack_counters),
1579             OID_AUTO, "badfr_bytes", CTLFLAG_RD,
1580             &rack_badfr_bytes, "Total number of bad FRs");
1581         rack_rtm_prr_retran = counter_u64_alloc(M_WAITOK);
1582         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1583             SYSCTL_CHILDREN(rack_counters),
1584             OID_AUTO, "prrsndret", CTLFLAG_RD,
1585             &rack_rtm_prr_retran,
1586             "Total number of prr based retransmits");
1587         rack_rtm_prr_newdata = counter_u64_alloc(M_WAITOK);
1588         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1589             SYSCTL_CHILDREN(rack_counters),
1590             OID_AUTO, "prrsndnew", CTLFLAG_RD,
1591             &rack_rtm_prr_newdata,
1592             "Total number of prr based new transmits");
1593         rack_timestamp_mismatch = counter_u64_alloc(M_WAITOK);
1594         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1595             SYSCTL_CHILDREN(rack_counters),
1596             OID_AUTO, "tsnf", CTLFLAG_RD,
1597             &rack_timestamp_mismatch,
1598             "Total number of timestamps that we could not find the reported ts");
1599         rack_find_high = counter_u64_alloc(M_WAITOK);
1600         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1601             SYSCTL_CHILDREN(rack_counters),
1602             OID_AUTO, "findhigh", CTLFLAG_RD,
1603             &rack_find_high,
1604             "Total number of FIN causing find-high");
1605         rack_reorder_seen = counter_u64_alloc(M_WAITOK);
1606         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1607             SYSCTL_CHILDREN(rack_counters),
1608             OID_AUTO, "reordering", CTLFLAG_RD,
1609             &rack_reorder_seen,
1610             "Total number of times we added delay due to reordering");
1611         rack_tlp_tot = counter_u64_alloc(M_WAITOK);
1612         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1613             SYSCTL_CHILDREN(rack_counters),
1614             OID_AUTO, "tlp_to_total", CTLFLAG_RD,
1615             &rack_tlp_tot,
1616             "Total number of tail loss probe expirations");
1617         rack_tlp_newdata = counter_u64_alloc(M_WAITOK);
1618         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1619             SYSCTL_CHILDREN(rack_counters),
1620             OID_AUTO, "tlp_new", CTLFLAG_RD,
1621             &rack_tlp_newdata,
1622             "Total number of tail loss probe sending new data");
1623         rack_tlp_retran = counter_u64_alloc(M_WAITOK);
1624         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1625             SYSCTL_CHILDREN(rack_counters),
1626             OID_AUTO, "tlp_retran", CTLFLAG_RD,
1627             &rack_tlp_retran,
1628             "Total number of tail loss probe sending retransmitted data");
1629         rack_tlp_retran_bytes = counter_u64_alloc(M_WAITOK);
1630         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1631             SYSCTL_CHILDREN(rack_counters),
1632             OID_AUTO, "tlp_retran_bytes", CTLFLAG_RD,
1633             &rack_tlp_retran_bytes,
1634             "Total bytes of tail loss probe sending retransmitted data");
1635         rack_tlp_retran_fail = counter_u64_alloc(M_WAITOK);
1636         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1637             SYSCTL_CHILDREN(rack_counters),
1638             OID_AUTO, "tlp_retran_fail", CTLFLAG_RD,
1639             &rack_tlp_retran_fail,
1640             "Total number of tail loss probe sending retransmitted data that failed (wait for t3)");
1641         rack_to_tot = counter_u64_alloc(M_WAITOK);
1642         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1643             SYSCTL_CHILDREN(rack_counters),
1644             OID_AUTO, "rack_to_tot", CTLFLAG_RD,
1645             &rack_to_tot,
1646             "Total number of times the rack to expired");
1647         rack_to_arm_rack = counter_u64_alloc(M_WAITOK);
1648         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1649             SYSCTL_CHILDREN(rack_counters),
1650             OID_AUTO, "arm_rack", CTLFLAG_RD,
1651             &rack_to_arm_rack,
1652             "Total number of times the rack timer armed");
1653         rack_to_arm_tlp = counter_u64_alloc(M_WAITOK);
1654         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1655             SYSCTL_CHILDREN(rack_counters),
1656             OID_AUTO, "arm_tlp", CTLFLAG_RD,
1657             &rack_to_arm_tlp,
1658             "Total number of times the tlp timer armed");
1659         rack_calc_zero = counter_u64_alloc(M_WAITOK);
1660         rack_calc_nonzero = counter_u64_alloc(M_WAITOK);
1661         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1662             SYSCTL_CHILDREN(rack_counters),
1663             OID_AUTO, "calc_zero", CTLFLAG_RD,
1664             &rack_calc_zero,
1665             "Total number of times pacing time worked out to zero");
1666         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1667             SYSCTL_CHILDREN(rack_counters),
1668             OID_AUTO, "calc_nonzero", CTLFLAG_RD,
1669             &rack_calc_nonzero,
1670             "Total number of times pacing time worked out to non-zero");
1671         rack_paced_segments = counter_u64_alloc(M_WAITOK);
1672         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1673             SYSCTL_CHILDREN(rack_counters),
1674             OID_AUTO, "paced", CTLFLAG_RD,
1675             &rack_paced_segments,
1676             "Total number of times a segment send caused hptsi");
1677         rack_unpaced_segments = counter_u64_alloc(M_WAITOK);
1678         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1679             SYSCTL_CHILDREN(rack_counters),
1680             OID_AUTO, "unpaced", CTLFLAG_RD,
1681             &rack_unpaced_segments,
1682             "Total number of times a segment did not cause hptsi");
1683         rack_saw_enobuf = counter_u64_alloc(M_WAITOK);
1684         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1685             SYSCTL_CHILDREN(rack_counters),
1686             OID_AUTO, "saw_enobufs", CTLFLAG_RD,
1687             &rack_saw_enobuf,
1688             "Total number of times a sends returned enobuf for non-hdwr paced connections");
1689         rack_saw_enobuf_hw = counter_u64_alloc(M_WAITOK);
1690         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1691             SYSCTL_CHILDREN(rack_counters),
1692             OID_AUTO, "saw_enobufs_hw", CTLFLAG_RD,
1693             &rack_saw_enobuf_hw,
1694             "Total number of times a send returned enobuf for hdwr paced connections");
1695         rack_saw_enetunreach = counter_u64_alloc(M_WAITOK);
1696         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1697             SYSCTL_CHILDREN(rack_counters),
1698             OID_AUTO, "saw_enetunreach", CTLFLAG_RD,
1699             &rack_saw_enetunreach,
1700             "Total number of times a send received a enetunreachable");
1701         rack_hot_alloc = counter_u64_alloc(M_WAITOK);
1702         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1703             SYSCTL_CHILDREN(rack_counters),
1704             OID_AUTO, "alloc_hot", CTLFLAG_RD,
1705             &rack_hot_alloc,
1706             "Total allocations from the top of our list");
1707         rack_to_alloc = counter_u64_alloc(M_WAITOK);
1708         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1709             SYSCTL_CHILDREN(rack_counters),
1710             OID_AUTO, "allocs", CTLFLAG_RD,
1711             &rack_to_alloc,
1712             "Total allocations of tracking structures");
1713         rack_to_alloc_hard = counter_u64_alloc(M_WAITOK);
1714         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1715             SYSCTL_CHILDREN(rack_counters),
1716             OID_AUTO, "allochard", CTLFLAG_RD,
1717             &rack_to_alloc_hard,
1718             "Total allocations done with sleeping the hard way");
1719         rack_to_alloc_emerg = counter_u64_alloc(M_WAITOK);
1720         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1721             SYSCTL_CHILDREN(rack_counters),
1722             OID_AUTO, "allocemerg", CTLFLAG_RD,
1723             &rack_to_alloc_emerg,
1724             "Total allocations done from emergency cache");
1725         rack_to_alloc_limited = counter_u64_alloc(M_WAITOK);
1726         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1727             SYSCTL_CHILDREN(rack_counters),
1728             OID_AUTO, "alloc_limited", CTLFLAG_RD,
1729             &rack_to_alloc_limited,
1730             "Total allocations dropped due to limit");
1731         rack_alloc_limited_conns = counter_u64_alloc(M_WAITOK);
1732         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1733             SYSCTL_CHILDREN(rack_counters),
1734             OID_AUTO, "alloc_limited_conns", CTLFLAG_RD,
1735             &rack_alloc_limited_conns,
1736             "Connections with allocations dropped due to limit");
1737         rack_split_limited = counter_u64_alloc(M_WAITOK);
1738         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1739             SYSCTL_CHILDREN(rack_counters),
1740             OID_AUTO, "split_limited", CTLFLAG_RD,
1741             &rack_split_limited,
1742             "Split allocations dropped due to limit");
1743
1744         for (i = 0; i < MAX_NUM_OF_CNTS; i++) {
1745                 char name[32];
1746                 sprintf(name, "cmp_ack_cnt_%d", i);
1747                 rack_proc_comp_ack[i] = counter_u64_alloc(M_WAITOK);
1748                 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1749                                        SYSCTL_CHILDREN(rack_counters),
1750                                        OID_AUTO, name, CTLFLAG_RD,
1751                                        &rack_proc_comp_ack[i],
1752                                        "Number of compressed acks we processed");
1753         }
1754         rack_large_ackcmp = counter_u64_alloc(M_WAITOK);
1755         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1756             SYSCTL_CHILDREN(rack_counters),
1757             OID_AUTO, "cmp_large_mbufs", CTLFLAG_RD,
1758             &rack_large_ackcmp,
1759             "Number of TCP connections with large mbuf's for compressed acks");
1760         rack_small_ackcmp = counter_u64_alloc(M_WAITOK);
1761         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1762             SYSCTL_CHILDREN(rack_counters),
1763             OID_AUTO, "cmp_small_mbufs", CTLFLAG_RD,
1764             &rack_small_ackcmp,
1765             "Number of TCP connections with small mbuf's for compressed acks");
1766 #ifdef INVARIANTS
1767         rack_adjust_map_bw = counter_u64_alloc(M_WAITOK);
1768         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1769             SYSCTL_CHILDREN(rack_counters),
1770             OID_AUTO, "map_adjust_req", CTLFLAG_RD,
1771             &rack_adjust_map_bw,
1772             "Number of times we hit the case where the sb went up and down on a sendmap entry");
1773 #endif
1774         rack_multi_single_eq = counter_u64_alloc(M_WAITOK);
1775         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1776             SYSCTL_CHILDREN(rack_counters),
1777             OID_AUTO, "cmp_ack_equiv", CTLFLAG_RD,
1778             &rack_multi_single_eq,
1779             "Number of compressed acks total represented");
1780         rack_proc_non_comp_ack = counter_u64_alloc(M_WAITOK);
1781         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1782             SYSCTL_CHILDREN(rack_counters),
1783             OID_AUTO, "cmp_ack_not", CTLFLAG_RD,
1784             &rack_proc_non_comp_ack,
1785             "Number of non compresseds acks that we processed");
1786
1787
1788         rack_sack_proc_all = counter_u64_alloc(M_WAITOK);
1789         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1790             SYSCTL_CHILDREN(rack_counters),
1791             OID_AUTO, "sack_long", CTLFLAG_RD,
1792             &rack_sack_proc_all,
1793             "Total times we had to walk whole list for sack processing");
1794         rack_sack_proc_restart = counter_u64_alloc(M_WAITOK);
1795         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1796             SYSCTL_CHILDREN(rack_counters),
1797             OID_AUTO, "sack_restart", CTLFLAG_RD,
1798             &rack_sack_proc_restart,
1799             "Total times we had to walk whole list due to a restart");
1800         rack_sack_proc_short = counter_u64_alloc(M_WAITOK);
1801         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1802             SYSCTL_CHILDREN(rack_counters),
1803             OID_AUTO, "sack_short", CTLFLAG_RD,
1804             &rack_sack_proc_short,
1805             "Total times we took shortcut for sack processing");
1806         rack_enter_tlp_calc = counter_u64_alloc(M_WAITOK);
1807         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1808             SYSCTL_CHILDREN(rack_counters),
1809             OID_AUTO, "tlp_calc_entered", CTLFLAG_RD,
1810             &rack_enter_tlp_calc,
1811             "Total times we called calc-tlp");
1812         rack_used_tlpmethod = counter_u64_alloc(M_WAITOK);
1813         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1814             SYSCTL_CHILDREN(rack_counters),
1815             OID_AUTO, "hit_tlp_method", CTLFLAG_RD,
1816             &rack_used_tlpmethod,
1817             "Total number of runt sacks");
1818         rack_used_tlpmethod2 = counter_u64_alloc(M_WAITOK);
1819         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1820             SYSCTL_CHILDREN(rack_counters),
1821             OID_AUTO, "hit_tlp_method2", CTLFLAG_RD,
1822             &rack_used_tlpmethod2,
1823             "Total number of times we hit TLP method 2");
1824         rack_sack_skipped_acked = counter_u64_alloc(M_WAITOK);
1825         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1826             SYSCTL_CHILDREN(rack_attack),
1827             OID_AUTO, "skipacked", CTLFLAG_RD,
1828             &rack_sack_skipped_acked,
1829             "Total number of times we skipped previously sacked");
1830         rack_sack_splits = counter_u64_alloc(M_WAITOK);
1831         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1832             SYSCTL_CHILDREN(rack_attack),
1833             OID_AUTO, "ofsplit", CTLFLAG_RD,
1834             &rack_sack_splits,
1835             "Total number of times we did the old fashion tree split");
1836         rack_progress_drops = counter_u64_alloc(M_WAITOK);
1837         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1838             SYSCTL_CHILDREN(rack_counters),
1839             OID_AUTO, "prog_drops", CTLFLAG_RD,
1840             &rack_progress_drops,
1841             "Total number of progress drops");
1842         rack_input_idle_reduces = counter_u64_alloc(M_WAITOK);
1843         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1844             SYSCTL_CHILDREN(rack_counters),
1845             OID_AUTO, "idle_reduce_oninput", CTLFLAG_RD,
1846             &rack_input_idle_reduces,
1847             "Total number of idle reductions on input");
1848         rack_collapsed_win = counter_u64_alloc(M_WAITOK);
1849         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1850             SYSCTL_CHILDREN(rack_counters),
1851             OID_AUTO, "collapsed_win", CTLFLAG_RD,
1852             &rack_collapsed_win,
1853             "Total number of collapsed windows");
1854         rack_tlp_does_nada = counter_u64_alloc(M_WAITOK);
1855         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1856             SYSCTL_CHILDREN(rack_counters),
1857             OID_AUTO, "tlp_nada", CTLFLAG_RD,
1858             &rack_tlp_does_nada,
1859             "Total number of nada tlp calls");
1860         rack_try_scwnd = counter_u64_alloc(M_WAITOK);
1861         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1862             SYSCTL_CHILDREN(rack_counters),
1863             OID_AUTO, "tried_scwnd", CTLFLAG_RD,
1864             &rack_try_scwnd,
1865             "Total number of scwnd attempts");
1866
1867         rack_per_timer_hole = counter_u64_alloc(M_WAITOK);
1868         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1869             SYSCTL_CHILDREN(rack_counters),
1870             OID_AUTO, "timer_hole", CTLFLAG_RD,
1871             &rack_per_timer_hole,
1872             "Total persists start in timer hole");
1873
1874         rack_sbsndptr_wrong = counter_u64_alloc(M_WAITOK);
1875         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1876             SYSCTL_CHILDREN(rack_counters),
1877             OID_AUTO, "sndptr_wrong", CTLFLAG_RD,
1878             &rack_sbsndptr_wrong, "Total number of times the saved sbsndptr was incorret");
1879         rack_sbsndptr_right = counter_u64_alloc(M_WAITOK);
1880         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1881             SYSCTL_CHILDREN(rack_counters),
1882             OID_AUTO, "sndptr_right", CTLFLAG_RD,
1883             &rack_sbsndptr_right, "Total number of times the saved sbsndptr was corret");
1884
1885         COUNTER_ARRAY_ALLOC(rack_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK);
1886         SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1887             OID_AUTO, "outsize", CTLFLAG_RD,
1888             rack_out_size, TCP_MSS_ACCT_SIZE, "MSS send sizes");
1889         COUNTER_ARRAY_ALLOC(rack_opts_arry, RACK_OPTS_SIZE, M_WAITOK);
1890         SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1891             OID_AUTO, "opts", CTLFLAG_RD,
1892             rack_opts_arry, RACK_OPTS_SIZE, "RACK Option Stats");
1893         SYSCTL_ADD_PROC(&rack_sysctl_ctx,
1894             SYSCTL_CHILDREN(rack_sysctl_root),
1895             OID_AUTO, "clear", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1896             &rack_clear_counter, 0, sysctl_rack_clear, "IU", "Clear counters");
1897 }
1898
1899 static __inline int
1900 rb_map_cmp(struct rack_sendmap *b, struct rack_sendmap *a)
1901 {
1902         if (SEQ_GEQ(b->r_start, a->r_start) &&
1903             SEQ_LT(b->r_start, a->r_end)) {
1904                 /*
1905                  * The entry b is within the
1906                  * block a. i.e.:
1907                  * a --   |-------------|
1908                  * b --   |----|
1909                  * <or>
1910                  * b --       |------|
1911                  * <or>
1912                  * b --       |-----------|
1913                  */
1914                 return (0);
1915         } else if (SEQ_GEQ(b->r_start, a->r_end)) {
1916                 /*
1917                  * b falls as either the next
1918                  * sequence block after a so a
1919                  * is said to be smaller than b.
1920                  * i.e:
1921                  * a --   |------|
1922                  * b --          |--------|
1923                  * or
1924                  * b --              |-----|
1925                  */
1926                 return (1);
1927         }
1928         /*
1929          * Whats left is where a is
1930          * larger than b. i.e:
1931          * a --         |-------|
1932          * b --  |---|
1933          * or even possibly
1934          * b --   |--------------|
1935          */
1936         return (-1);
1937 }
1938
1939 RB_PROTOTYPE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp);
1940 RB_GENERATE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp);
1941
1942 static uint32_t
1943 rc_init_window(struct tcp_rack *rack)
1944 {
1945         uint32_t win;
1946
1947         if (rack->rc_init_win == 0) {
1948                 /*
1949                  * Nothing set by the user, use the system stack
1950                  * default.
1951                  */
1952                 return (tcp_compute_initwnd(tcp_maxseg(rack->rc_tp)));
1953         }
1954         win = ctf_fixed_maxseg(rack->rc_tp) * rack->rc_init_win;
1955         return (win);
1956 }
1957
1958 static uint64_t
1959 rack_get_fixed_pacing_bw(struct tcp_rack *rack)
1960 {
1961         if (IN_FASTRECOVERY(rack->rc_tp->t_flags))
1962                 return (rack->r_ctl.rc_fixed_pacing_rate_rec);
1963         else if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
1964                 return (rack->r_ctl.rc_fixed_pacing_rate_ss);
1965         else
1966                 return (rack->r_ctl.rc_fixed_pacing_rate_ca);
1967 }
1968
1969 static uint64_t
1970 rack_get_bw(struct tcp_rack *rack)
1971 {
1972         if (rack->use_fixed_rate) {
1973                 /* Return the fixed pacing rate */
1974                 return (rack_get_fixed_pacing_bw(rack));
1975         }
1976         if (rack->r_ctl.gp_bw == 0) {
1977                 /*
1978                  * We have yet no b/w measurement,
1979                  * if we have a user set initial bw
1980                  * return it. If we don't have that and
1981                  * we have an srtt, use the tcp IW (10) to
1982                  * calculate a fictional b/w over the SRTT
1983                  * which is more or less a guess. Note
1984                  * we don't use our IW from rack on purpose
1985                  * so if we have like IW=30, we are not
1986                  * calculating a "huge" b/w.
1987                  */
1988                 uint64_t bw, srtt;
1989                 if (rack->r_ctl.init_rate)
1990                         return (rack->r_ctl.init_rate);
1991
1992                 /* Has the user set a max peak rate? */
1993 #ifdef NETFLIX_PEAKRATE
1994                 if (rack->rc_tp->t_maxpeakrate)
1995                         return (rack->rc_tp->t_maxpeakrate);
1996 #endif
1997                 /* Ok lets come up with the IW guess, if we have a srtt */
1998                 if (rack->rc_tp->t_srtt == 0) {
1999                         /*
2000                          * Go with old pacing method
2001                          * i.e. burst mitigation only.
2002                          */
2003                         return (0);
2004                 }
2005                 /* Ok lets get the initial TCP win (not racks) */
2006                 bw = tcp_compute_initwnd(tcp_maxseg(rack->rc_tp));
2007                 srtt = (uint64_t)rack->rc_tp->t_srtt;
2008                 bw *= (uint64_t)USECS_IN_SECOND;
2009                 bw /= srtt;
2010                 if (rack->r_ctl.bw_rate_cap && (bw > rack->r_ctl.bw_rate_cap))
2011                         bw = rack->r_ctl.bw_rate_cap;
2012                 return (bw);
2013         } else {
2014                 uint64_t bw;
2015
2016                 if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
2017                         /* Averaging is done, we can return the value */
2018                         bw = rack->r_ctl.gp_bw;
2019                 } else {
2020                         /* Still doing initial average must calculate */
2021                         bw = rack->r_ctl.gp_bw / rack->r_ctl.num_measurements;
2022                 }
2023 #ifdef NETFLIX_PEAKRATE
2024                 if ((rack->rc_tp->t_maxpeakrate) &&
2025                     (bw > rack->rc_tp->t_maxpeakrate)) {
2026                         /* The user has set a peak rate to pace at
2027                          * don't allow us to pace faster than that.
2028                          */
2029                         return (rack->rc_tp->t_maxpeakrate);
2030                 }
2031 #endif
2032                 if (rack->r_ctl.bw_rate_cap && (bw > rack->r_ctl.bw_rate_cap))
2033                         bw = rack->r_ctl.bw_rate_cap;
2034                 return (bw);
2035         }
2036 }
2037
2038 static uint16_t
2039 rack_get_output_gain(struct tcp_rack *rack, struct rack_sendmap *rsm)
2040 {
2041         if (rack->use_fixed_rate) {
2042                 return (100);
2043         } else if (rack->in_probe_rtt && (rsm == NULL))
2044                 return (rack->r_ctl.rack_per_of_gp_probertt);
2045         else if ((IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
2046                   rack->r_ctl.rack_per_of_gp_rec)) {
2047                 if (rsm) {
2048                         /* a retransmission always use the recovery rate */
2049                         return (rack->r_ctl.rack_per_of_gp_rec);
2050                 } else if (rack->rack_rec_nonrxt_use_cr) {
2051                         /* Directed to use the configured rate */
2052                         goto configured_rate;
2053                 } else if (rack->rack_no_prr &&
2054                            (rack->r_ctl.rack_per_of_gp_rec > 100)) {
2055                         /* No PRR, lets just use the b/w estimate only */
2056                         return (100);
2057                 } else {
2058                         /*
2059                          * Here we may have a non-retransmit but we
2060                          * have no overrides, so just use the recovery
2061                          * rate (prr is in effect).
2062                          */
2063                         return (rack->r_ctl.rack_per_of_gp_rec);
2064                 }
2065         }
2066 configured_rate:
2067         /* For the configured rate we look at our cwnd vs the ssthresh */
2068         if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
2069                 return (rack->r_ctl.rack_per_of_gp_ss);
2070         else
2071                 return (rack->r_ctl.rack_per_of_gp_ca);
2072 }
2073
2074 static void
2075 rack_log_hdwr_pacing(struct tcp_rack *rack,
2076                      uint64_t rate, uint64_t hw_rate, int line,
2077                      int error, uint16_t mod)
2078 {
2079         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2080                 union tcp_log_stackspecific log;
2081                 struct timeval tv;
2082                 const struct ifnet *ifp;
2083
2084                 memset(&log, 0, sizeof(log));
2085                 log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff);
2086                 log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff);
2087                 if (rack->r_ctl.crte) {
2088                         ifp = rack->r_ctl.crte->ptbl->rs_ifp;
2089                 } else if (rack->rc_inp->inp_route.ro_nh &&
2090                            rack->rc_inp->inp_route.ro_nh->nh_ifp) {
2091                         ifp = rack->rc_inp->inp_route.ro_nh->nh_ifp;
2092                 } else
2093                         ifp = NULL;
2094                 if (ifp) {
2095                         log.u_bbr.flex3 = (((uint64_t)ifp  >> 32) & 0x00000000ffffffff);
2096                         log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff);
2097                 }
2098                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2099                 log.u_bbr.bw_inuse = rate;
2100                 log.u_bbr.flex5 = line;
2101                 log.u_bbr.flex6 = error;
2102                 log.u_bbr.flex7 = mod;
2103                 log.u_bbr.applimited = rack->r_ctl.rc_pace_max_segs;
2104                 log.u_bbr.flex8 = rack->use_fixed_rate;
2105                 log.u_bbr.flex8 <<= 1;
2106                 log.u_bbr.flex8 |= rack->rack_hdrw_pacing;
2107                 log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
2108                 log.u_bbr.delRate = rack->r_ctl.crte_prev_rate;
2109                 if (rack->r_ctl.crte)
2110                         log.u_bbr.cur_del_rate = rack->r_ctl.crte->rate;
2111                 else
2112                         log.u_bbr.cur_del_rate = 0;
2113                 log.u_bbr.rttProp = rack->r_ctl.last_hw_bw_req;
2114                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2115                     &rack->rc_inp->inp_socket->so_rcv,
2116                     &rack->rc_inp->inp_socket->so_snd,
2117                     BBR_LOG_HDWR_PACE, 0,
2118                     0, &log, false, &tv);
2119         }
2120 }
2121
2122 static uint64_t
2123 rack_get_output_bw(struct tcp_rack *rack, uint64_t bw, struct rack_sendmap *rsm, int *capped)
2124 {
2125         /*
2126          * We allow rack_per_of_gp_xx to dictate our bw rate we want.
2127          */
2128         uint64_t bw_est, high_rate;
2129         uint64_t gain;
2130
2131         gain = (uint64_t)rack_get_output_gain(rack, rsm);
2132         bw_est = bw * gain;
2133         bw_est /= (uint64_t)100;
2134         /* Never fall below the minimum (def 64kbps) */
2135         if (bw_est < RACK_MIN_BW)
2136                 bw_est = RACK_MIN_BW;
2137         if (rack->r_rack_hw_rate_caps) {
2138                 /* Rate caps are in place */
2139                 if (rack->r_ctl.crte != NULL) {
2140                         /* We have a hdwr rate already */
2141                         high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
2142                         if (bw_est >= high_rate) {
2143                                 /* We are capping bw at the highest rate table entry */
2144                                 rack_log_hdwr_pacing(rack,
2145                                                      bw_est, high_rate, __LINE__,
2146                                                      0, 3);
2147                                 bw_est = high_rate;
2148                                 if (capped)
2149                                         *capped = 1;
2150                         }
2151                 } else if ((rack->rack_hdrw_pacing == 0) &&
2152                            (rack->rack_hdw_pace_ena) &&
2153                            (rack->rack_attempt_hdwr_pace == 0) &&
2154                            (rack->rc_inp->inp_route.ro_nh != NULL) &&
2155                            (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
2156                         /*
2157                          * Special case, we have not yet attempted hardware
2158                          * pacing, and yet we may, when we do, find out if we are
2159                          * above the highest rate. We need to know the maxbw for the interface
2160                          * in question (if it supports ratelimiting). We get back
2161                          * a 0, if the interface is not found in the RL lists.
2162                          */
2163                         high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
2164                         if (high_rate) {
2165                                 /* Yep, we have a rate is it above this rate? */
2166                                 if (bw_est > high_rate) {
2167                                         bw_est = high_rate;
2168                                         if (capped)
2169                                                 *capped = 1;
2170                                 }
2171                         }
2172                 }
2173         }
2174         return (bw_est);
2175 }
2176
2177 static void
2178 rack_log_retran_reason(struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t tsused, uint32_t thresh, int mod)
2179 {
2180         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2181                 union tcp_log_stackspecific log;
2182                 struct timeval tv;
2183
2184                 if ((mod != 1) && (rack_verbose_logging == 0)) {
2185                         /*
2186                          * We get 3 values currently for mod
2187                          * 1 - We are retransmitting and this tells the reason.
2188                          * 2 - We are clearing a dup-ack count.
2189                          * 3 - We are incrementing a dup-ack count.
2190                          *
2191                          * The clear/increment are only logged
2192                          * if you have BBverbose on.
2193                          */
2194                         return;
2195                 }
2196                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2197                 log.u_bbr.flex1 = tsused;
2198                 log.u_bbr.flex2 = thresh;
2199                 log.u_bbr.flex3 = rsm->r_flags;
2200                 log.u_bbr.flex4 = rsm->r_dupack;
2201                 log.u_bbr.flex5 = rsm->r_start;
2202                 log.u_bbr.flex6 = rsm->r_end;
2203                 log.u_bbr.flex8 = mod;
2204                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2205                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2206                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2207                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2208                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2209                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2210                 log.u_bbr.pacing_gain = rack->r_must_retran;
2211                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2212                     &rack->rc_inp->inp_socket->so_rcv,
2213                     &rack->rc_inp->inp_socket->so_snd,
2214                     BBR_LOG_SETTINGS_CHG, 0,
2215                     0, &log, false, &tv);
2216         }
2217 }
2218
2219 static void
2220 rack_log_to_start(struct tcp_rack *rack, uint32_t cts, uint32_t to, int32_t slot, uint8_t which)
2221 {
2222         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2223                 union tcp_log_stackspecific log;
2224                 struct timeval tv;
2225
2226                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2227                 log.u_bbr.flex1 = rack->rc_tp->t_srtt;
2228                 log.u_bbr.flex2 = to;
2229                 log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags;
2230                 log.u_bbr.flex4 = slot;
2231                 log.u_bbr.flex5 = rack->rc_inp->inp_hptsslot;
2232                 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2233                 log.u_bbr.flex7 = rack->rc_in_persist;
2234                 log.u_bbr.flex8 = which;
2235                 if (rack->rack_no_prr)
2236                         log.u_bbr.pkts_out = 0;
2237                 else
2238                         log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
2239                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2240                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2241                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2242                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2243                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2244                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2245                 log.u_bbr.pacing_gain = rack->r_must_retran;
2246                 log.u_bbr.lt_epoch = rack->rc_tp->t_rxtshift;
2247                 log.u_bbr.lost = rack_rto_min;
2248                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2249                     &rack->rc_inp->inp_socket->so_rcv,
2250                     &rack->rc_inp->inp_socket->so_snd,
2251                     BBR_LOG_TIMERSTAR, 0,
2252                     0, &log, false, &tv);
2253         }
2254 }
2255
2256 static void
2257 rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm)
2258 {
2259         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2260                 union tcp_log_stackspecific log;
2261                 struct timeval tv;
2262
2263                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2264                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2265                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2266                 log.u_bbr.flex8 = to_num;
2267                 log.u_bbr.flex1 = rack->r_ctl.rc_rack_min_rtt;
2268                 log.u_bbr.flex2 = rack->rc_rack_rtt;
2269                 if (rsm == NULL)
2270                         log.u_bbr.flex3 = 0;
2271                 else
2272                         log.u_bbr.flex3 = rsm->r_end - rsm->r_start;
2273                 if (rack->rack_no_prr)
2274                         log.u_bbr.flex5 = 0;
2275                 else
2276                         log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2277                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2278                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2279                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2280                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2281                 log.u_bbr.pacing_gain = rack->r_must_retran;
2282                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2283                     &rack->rc_inp->inp_socket->so_rcv,
2284                     &rack->rc_inp->inp_socket->so_snd,
2285                     BBR_LOG_RTO, 0,
2286                     0, &log, false, &tv);
2287         }
2288 }
2289
2290 static void
2291 rack_log_map_chg(struct tcpcb *tp, struct tcp_rack *rack,
2292                  struct rack_sendmap *prev,
2293                  struct rack_sendmap *rsm,
2294                  struct rack_sendmap *next,
2295                  int flag, uint32_t th_ack, int line)
2296 {
2297         if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) {
2298                 union tcp_log_stackspecific log;
2299                 struct timeval tv;
2300
2301                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2302                 log.u_bbr.flex8 = flag;
2303                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2304                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2305                 log.u_bbr.cur_del_rate = (uint64_t)prev;
2306                 log.u_bbr.delRate = (uint64_t)rsm;
2307                 log.u_bbr.rttProp = (uint64_t)next;
2308                 log.u_bbr.flex7 = 0;
2309                 if (prev) {
2310                         log.u_bbr.flex1 = prev->r_start;
2311                         log.u_bbr.flex2 = prev->r_end;
2312                         log.u_bbr.flex7 |= 0x4;
2313                 }
2314                 if (rsm) {
2315                         log.u_bbr.flex3 = rsm->r_start;
2316                         log.u_bbr.flex4 = rsm->r_end;
2317                         log.u_bbr.flex7 |= 0x2;
2318                 }
2319                 if (next) {
2320                         log.u_bbr.flex5 = next->r_start;
2321                         log.u_bbr.flex6 = next->r_end;
2322                         log.u_bbr.flex7 |= 0x1;
2323                 }
2324                 log.u_bbr.applimited = line;
2325                 log.u_bbr.pkts_out = th_ack;
2326                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2327                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2328                 if (rack->rack_no_prr)
2329                         log.u_bbr.lost = 0;
2330                 else
2331                         log.u_bbr.lost = rack->r_ctl.rc_prr_sndcnt;
2332                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2333                     &rack->rc_inp->inp_socket->so_rcv,
2334                     &rack->rc_inp->inp_socket->so_snd,
2335                     TCP_LOG_MAPCHG, 0,
2336                     0, &log, false, &tv);
2337         }
2338 }
2339
2340 static void
2341 rack_log_rtt_upd(struct tcpcb *tp, struct tcp_rack *rack, uint32_t t, uint32_t len,
2342                  struct rack_sendmap *rsm, int conf)
2343 {
2344         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
2345                 union tcp_log_stackspecific log;
2346                 struct timeval tv;
2347                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2348                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2349                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2350                 log.u_bbr.flex1 = t;
2351                 log.u_bbr.flex2 = len;
2352                 log.u_bbr.flex3 = rack->r_ctl.rc_rack_min_rtt;
2353                 log.u_bbr.flex4 = rack->r_ctl.rack_rs.rs_rtt_lowest;
2354                 log.u_bbr.flex5 = rack->r_ctl.rack_rs.rs_rtt_highest;
2355                 log.u_bbr.flex6 = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2356                 log.u_bbr.flex7 = conf;
2357                 log.u_bbr.rttProp = (uint64_t)rack->r_ctl.rack_rs.rs_rtt_tot;
2358                 log.u_bbr.flex8 = rack->r_ctl.rc_rate_sample_method;
2359                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2360                 log.u_bbr.delivered = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2361                 log.u_bbr.pkts_out = rack->r_ctl.rack_rs.rs_flags;
2362                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2363                 if (rsm) {
2364                         log.u_bbr.pkt_epoch = rsm->r_start;
2365                         log.u_bbr.lost = rsm->r_end;
2366                         log.u_bbr.cwnd_gain = rsm->r_rtr_cnt;
2367                         log.u_bbr.pacing_gain = rsm->r_flags;
2368                 } else {
2369                         /* Its a SYN */
2370                         log.u_bbr.pkt_epoch = rack->rc_tp->iss;
2371                         log.u_bbr.lost = 0;
2372                         log.u_bbr.cwnd_gain = 0;
2373                         log.u_bbr.pacing_gain = 0;
2374                 }
2375                 /* Write out general bits of interest rrs here */
2376                 log.u_bbr.use_lt_bw = rack->rc_highly_buffered;
2377                 log.u_bbr.use_lt_bw <<= 1;
2378                 log.u_bbr.use_lt_bw |= rack->forced_ack;
2379                 log.u_bbr.use_lt_bw <<= 1;
2380                 log.u_bbr.use_lt_bw |= rack->rc_gp_dyn_mul;
2381                 log.u_bbr.use_lt_bw <<= 1;
2382                 log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
2383                 log.u_bbr.use_lt_bw <<= 1;
2384                 log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
2385                 log.u_bbr.use_lt_bw <<= 1;
2386                 log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
2387                 log.u_bbr.use_lt_bw <<= 1;
2388                 log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
2389                 log.u_bbr.use_lt_bw <<= 1;
2390                 log.u_bbr.use_lt_bw |= rack->rc_dragged_bottom;
2391                 log.u_bbr.applimited = rack->r_ctl.rc_target_probertt_flight;
2392                 log.u_bbr.epoch = rack->r_ctl.rc_time_probertt_starts;
2393                 log.u_bbr.lt_epoch = rack->r_ctl.rc_time_probertt_entered;
2394                 log.u_bbr.cur_del_rate = rack->r_ctl.rc_lower_rtt_us_cts;
2395                 log.u_bbr.delRate = rack->r_ctl.rc_gp_srtt;
2396                 log.u_bbr.bw_inuse = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
2397                 log.u_bbr.bw_inuse <<= 32;
2398                 if (rsm)
2399                         log.u_bbr.bw_inuse |= ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]);
2400                 TCP_LOG_EVENTP(tp, NULL,
2401                     &rack->rc_inp->inp_socket->so_rcv,
2402                     &rack->rc_inp->inp_socket->so_snd,
2403                     BBR_LOG_BBRRTT, 0,
2404                     0, &log, false, &tv);
2405
2406
2407         }
2408 }
2409
2410 static void
2411 rack_log_rtt_sample(struct tcp_rack *rack, uint32_t rtt)
2412 {
2413         /*
2414          * Log the rtt sample we are
2415          * applying to the srtt algorithm in
2416          * useconds.
2417          */
2418         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2419                 union tcp_log_stackspecific log;
2420                 struct timeval tv;
2421
2422                 /* Convert our ms to a microsecond */
2423                 memset(&log, 0, sizeof(log));
2424                 log.u_bbr.flex1 = rtt;
2425                 log.u_bbr.flex2 = rack->r_ctl.ack_count;
2426                 log.u_bbr.flex3 = rack->r_ctl.sack_count;
2427                 log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
2428                 log.u_bbr.flex5 = rack->r_ctl.sack_moved_extra;
2429                 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2430                 log.u_bbr.flex7 = 1;
2431                 log.u_bbr.flex8 = rack->sack_attack_disable;
2432                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2433                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2434                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2435                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2436                 log.u_bbr.pacing_gain = rack->r_must_retran;
2437                 /*
2438                  * We capture in delRate the upper 32 bits as
2439                  * the confidence level we had declared, and the
2440                  * lower 32 bits as the actual RTT using the arrival
2441                  * timestamp.
2442                  */
2443                 log.u_bbr.delRate = rack->r_ctl.rack_rs.confidence;
2444                 log.u_bbr.delRate <<= 32;
2445                 log.u_bbr.delRate |= rack->r_ctl.rack_rs.rs_us_rtt;
2446                 /* Lets capture all the things that make up t_rtxcur */
2447                 log.u_bbr.applimited = rack_rto_min;
2448                 log.u_bbr.epoch = rack_rto_max;
2449                 log.u_bbr.lt_epoch = rack->r_ctl.timer_slop;
2450                 log.u_bbr.lost = rack_rto_min;
2451                 log.u_bbr.pkt_epoch = TICKS_2_USEC(tcp_rexmit_slop);
2452                 log.u_bbr.rttProp = RACK_REXMTVAL(rack->rc_tp);
2453                 log.u_bbr.bw_inuse = rack->r_ctl.act_rcv_time.tv_sec;
2454                 log.u_bbr.bw_inuse *= HPTS_USEC_IN_SEC;
2455                 log.u_bbr.bw_inuse += rack->r_ctl.act_rcv_time.tv_usec;
2456                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2457                     &rack->rc_inp->inp_socket->so_rcv,
2458                     &rack->rc_inp->inp_socket->so_snd,
2459                     TCP_LOG_RTT, 0,
2460                     0, &log, false, &tv);
2461         }
2462 }
2463
2464 static void
2465 rack_log_rtt_sample_calc(struct tcp_rack *rack, uint32_t rtt, uint32_t send_time, uint32_t ack_time, int where)
2466 {
2467         if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2468                 union tcp_log_stackspecific log;
2469                 struct timeval tv;
2470
2471                 /* Convert our ms to a microsecond */
2472                 memset(&log, 0, sizeof(log));
2473                 log.u_bbr.flex1 = rtt;
2474                 log.u_bbr.flex2 = send_time;
2475                 log.u_bbr.flex3 = ack_time;
2476                 log.u_bbr.flex4 = where;
2477                 log.u_bbr.flex7 = 2;
2478                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2479                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2480                     &rack->rc_inp->inp_socket->so_rcv,
2481                     &rack->rc_inp->inp_socket->so_snd,
2482                     TCP_LOG_RTT, 0,
2483                     0, &log, false, &tv);
2484         }
2485 }
2486
2487
2488
2489 static inline void
2490 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line)
2491 {
2492         if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) {
2493                 union tcp_log_stackspecific log;
2494                 struct timeval tv;
2495
2496                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2497                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2498                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2499                 log.u_bbr.flex1 = line;
2500                 log.u_bbr.flex2 = tick;
2501                 log.u_bbr.flex3 = tp->t_maxunacktime;
2502                 log.u_bbr.flex4 = tp->t_acktime;
2503                 log.u_bbr.flex8 = event;
2504                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2505                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2506                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2507                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2508                 log.u_bbr.pacing_gain = rack->r_must_retran;
2509                 TCP_LOG_EVENTP(tp, NULL,
2510                     &rack->rc_inp->inp_socket->so_rcv,
2511                     &rack->rc_inp->inp_socket->so_snd,
2512                     BBR_LOG_PROGRESS, 0,
2513                     0, &log, false, &tv);
2514         }
2515 }
2516
2517 static void
2518 rack_log_type_bbrsnd(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint32_t cts, struct timeval *tv)
2519 {
2520         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2521                 union tcp_log_stackspecific log;
2522
2523                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2524                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2525                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2526                 log.u_bbr.flex1 = slot;
2527                 if (rack->rack_no_prr)
2528                         log.u_bbr.flex2 = 0;
2529                 else
2530                         log.u_bbr.flex2 = rack->r_ctl.rc_prr_sndcnt;
2531                 log.u_bbr.flex7 = (0x0000ffff & rack->r_ctl.rc_hpts_flags);
2532                 log.u_bbr.flex8 = rack->rc_in_persist;
2533                 log.u_bbr.timeStamp = cts;
2534                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2535                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2536                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2537                 log.u_bbr.pacing_gain = rack->r_must_retran;
2538                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2539                     &rack->rc_inp->inp_socket->so_rcv,
2540                     &rack->rc_inp->inp_socket->so_snd,
2541                     BBR_LOG_BBRSND, 0,
2542                     0, &log, false, tv);
2543         }
2544 }
2545
2546 static void
2547 rack_log_doseg_done(struct tcp_rack *rack, uint32_t cts, int32_t nxt_pkt, int32_t did_out, int way_out, int nsegs)
2548 {
2549         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2550                 union tcp_log_stackspecific log;
2551                 struct timeval tv;
2552
2553                 memset(&log, 0, sizeof(log));
2554                 log.u_bbr.flex1 = did_out;
2555                 log.u_bbr.flex2 = nxt_pkt;
2556                 log.u_bbr.flex3 = way_out;
2557                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
2558                 if (rack->rack_no_prr)
2559                         log.u_bbr.flex5 = 0;
2560                 else
2561                         log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2562                 log.u_bbr.flex6 = nsegs;
2563                 log.u_bbr.applimited = rack->r_ctl.rc_pace_min_segs;
2564                 log.u_bbr.flex7 = rack->rc_ack_can_sendout_data;        /* Do we have ack-can-send set */
2565                 log.u_bbr.flex7 <<= 1;
2566                 log.u_bbr.flex7 |= rack->r_fast_output; /* is fast output primed */
2567                 log.u_bbr.flex7 <<= 1;
2568                 log.u_bbr.flex7 |= rack->r_wanted_output;       /* Do we want output */
2569                 log.u_bbr.flex8 = rack->rc_in_persist;
2570                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2571                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2572                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2573                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
2574                 log.u_bbr.use_lt_bw <<= 1;
2575                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
2576                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2577                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2578                 log.u_bbr.pacing_gain = rack->r_must_retran;
2579                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2580                     &rack->rc_inp->inp_socket->so_rcv,
2581                     &rack->rc_inp->inp_socket->so_snd,
2582                     BBR_LOG_DOSEG_DONE, 0,
2583                     0, &log, false, &tv);
2584         }
2585 }
2586
2587 static void
2588 rack_log_type_pacing_sizes(struct tcpcb *tp, struct tcp_rack *rack, uint32_t arg1, uint32_t arg2, uint32_t arg3, uint8_t frm)
2589 {
2590         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
2591                 union tcp_log_stackspecific log;
2592                 struct timeval tv;
2593                 uint32_t cts;
2594
2595                 memset(&log, 0, sizeof(log));
2596                 cts = tcp_get_usecs(&tv);
2597                 log.u_bbr.flex1 = rack->r_ctl.rc_pace_min_segs;
2598                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
2599                 log.u_bbr.flex4 = arg1;
2600                 log.u_bbr.flex5 = arg2;
2601                 log.u_bbr.flex6 = arg3;
2602                 log.u_bbr.flex8 = frm;
2603                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2604                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2605                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2606                 log.u_bbr.applimited = rack->r_ctl.rc_sacked;
2607                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2608                 log.u_bbr.pacing_gain = rack->r_must_retran;
2609                 TCP_LOG_EVENTP(tp, NULL,
2610                     &tp->t_inpcb->inp_socket->so_rcv,
2611                     &tp->t_inpcb->inp_socket->so_snd,
2612                     TCP_HDWR_PACE_SIZE, 0,
2613                     0, &log, false, &tv);
2614         }
2615 }
2616
2617 static void
2618 rack_log_type_just_return(struct tcp_rack *rack, uint32_t cts, uint32_t tlen, uint32_t slot,
2619                           uint8_t hpts_calling, int reason, uint32_t cwnd_to_use)
2620 {
2621         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2622                 union tcp_log_stackspecific log;
2623                 struct timeval tv;
2624
2625                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2626                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2627                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2628                 log.u_bbr.flex1 = slot;
2629                 log.u_bbr.flex2 = rack->r_ctl.rc_hpts_flags;
2630                 log.u_bbr.flex4 = reason;
2631                 if (rack->rack_no_prr)
2632                         log.u_bbr.flex5 = 0;
2633                 else
2634                         log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2635                 log.u_bbr.flex7 = hpts_calling;
2636                 log.u_bbr.flex8 = rack->rc_in_persist;
2637                 log.u_bbr.lt_epoch = cwnd_to_use;
2638                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2639                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2640                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2641                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2642                 log.u_bbr.pacing_gain = rack->r_must_retran;
2643                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2644                     &rack->rc_inp->inp_socket->so_rcv,
2645                     &rack->rc_inp->inp_socket->so_snd,
2646                     BBR_LOG_JUSTRET, 0,
2647                     tlen, &log, false, &tv);
2648         }
2649 }
2650
2651 static void
2652 rack_log_to_cancel(struct tcp_rack *rack, int32_t hpts_removed, int line, uint32_t us_cts,
2653                    struct timeval *tv, uint32_t flags_on_entry)
2654 {
2655         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2656                 union tcp_log_stackspecific log;
2657
2658                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2659                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2660                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2661                 log.u_bbr.flex1 = line;
2662                 log.u_bbr.flex2 = rack->r_ctl.rc_last_output_to;
2663                 log.u_bbr.flex3 = flags_on_entry;
2664                 log.u_bbr.flex4 = us_cts;
2665                 if (rack->rack_no_prr)
2666                         log.u_bbr.flex5 = 0;
2667                 else
2668                         log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2669                 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2670                 log.u_bbr.flex7 = hpts_removed;
2671                 log.u_bbr.flex8 = 1;
2672                 log.u_bbr.applimited = rack->r_ctl.rc_hpts_flags;
2673                 log.u_bbr.timeStamp = us_cts;
2674                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2675                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2676                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2677                 log.u_bbr.pacing_gain = rack->r_must_retran;
2678                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2679                     &rack->rc_inp->inp_socket->so_rcv,
2680                     &rack->rc_inp->inp_socket->so_snd,
2681                     BBR_LOG_TIMERCANC, 0,
2682                     0, &log, false, tv);
2683         }
2684 }
2685
2686 static void
2687 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
2688                           uint32_t flex1, uint32_t flex2,
2689                           uint32_t flex3, uint32_t flex4,
2690                           uint32_t flex5, uint32_t flex6,
2691                           uint16_t flex7, uint8_t mod)
2692 {
2693         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2694                 union tcp_log_stackspecific log;
2695                 struct timeval tv;
2696
2697                 if (mod == 1) {
2698                         /* No you can't use 1, its for the real to cancel */
2699                         return;
2700                 }
2701                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2702                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2703                 log.u_bbr.flex1 = flex1;
2704                 log.u_bbr.flex2 = flex2;
2705                 log.u_bbr.flex3 = flex3;
2706                 log.u_bbr.flex4 = flex4;
2707                 log.u_bbr.flex5 = flex5;
2708                 log.u_bbr.flex6 = flex6;
2709                 log.u_bbr.flex7 = flex7;
2710                 log.u_bbr.flex8 = mod;
2711                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2712                     &rack->rc_inp->inp_socket->so_rcv,
2713                     &rack->rc_inp->inp_socket->so_snd,
2714                     BBR_LOG_TIMERCANC, 0,
2715                     0, &log, false, &tv);
2716         }
2717 }
2718
2719 static void
2720 rack_log_to_processing(struct tcp_rack *rack, uint32_t cts, int32_t ret, int32_t timers)
2721 {
2722         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2723                 union tcp_log_stackspecific log;
2724                 struct timeval tv;
2725
2726                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2727                 log.u_bbr.flex1 = timers;
2728                 log.u_bbr.flex2 = ret;
2729                 log.u_bbr.flex3 = rack->r_ctl.rc_timer_exp;
2730                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
2731                 log.u_bbr.flex5 = cts;
2732                 if (rack->rack_no_prr)
2733                         log.u_bbr.flex6 = 0;
2734                 else
2735                         log.u_bbr.flex6 = rack->r_ctl.rc_prr_sndcnt;
2736                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2737                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2738                 log.u_bbr.pacing_gain = rack->r_must_retran;
2739                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2740                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2741                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2742                     &rack->rc_inp->inp_socket->so_rcv,
2743                     &rack->rc_inp->inp_socket->so_snd,
2744                     BBR_LOG_TO_PROCESS, 0,
2745                     0, &log, false, &tv);
2746         }
2747 }
2748
2749 static void
2750 rack_log_to_prr(struct tcp_rack *rack, int frm, int orig_cwnd)
2751 {
2752         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2753                 union tcp_log_stackspecific log;
2754                 struct timeval tv;
2755
2756                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2757                 log.u_bbr.flex1 = rack->r_ctl.rc_prr_out;
2758                 log.u_bbr.flex2 = rack->r_ctl.rc_prr_recovery_fs;
2759                 if (rack->rack_no_prr)
2760                         log.u_bbr.flex3 = 0;
2761                 else
2762                         log.u_bbr.flex3 = rack->r_ctl.rc_prr_sndcnt;
2763                 log.u_bbr.flex4 = rack->r_ctl.rc_prr_delivered;
2764                 log.u_bbr.flex5 = rack->r_ctl.rc_sacked;
2765                 log.u_bbr.flex6 = rack->r_ctl.rc_holes_rxt;
2766                 log.u_bbr.flex8 = frm;
2767                 log.u_bbr.pkts_out = orig_cwnd;
2768                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2769                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2770                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
2771                 log.u_bbr.use_lt_bw <<= 1;
2772                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
2773                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2774                     &rack->rc_inp->inp_socket->so_rcv,
2775                     &rack->rc_inp->inp_socket->so_snd,
2776                     BBR_LOG_BBRUPD, 0,
2777                     0, &log, false, &tv);
2778         }
2779 }
2780
2781 #ifdef NETFLIX_EXP_DETECTION
2782 static void
2783 rack_log_sad(struct tcp_rack *rack, int event)
2784 {
2785         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2786                 union tcp_log_stackspecific log;
2787                 struct timeval tv;
2788
2789                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2790                 log.u_bbr.flex1 = rack->r_ctl.sack_count;
2791                 log.u_bbr.flex2 = rack->r_ctl.ack_count;
2792                 log.u_bbr.flex3 = rack->r_ctl.sack_moved_extra;
2793                 log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
2794                 log.u_bbr.flex5 = rack->r_ctl.rc_num_maps_alloced;
2795                 log.u_bbr.flex6 = tcp_sack_to_ack_thresh;
2796                 log.u_bbr.pkts_out = tcp_sack_to_move_thresh;
2797                 log.u_bbr.lt_epoch = (tcp_force_detection << 8);
2798                 log.u_bbr.lt_epoch |= rack->do_detection;
2799                 log.u_bbr.applimited = tcp_map_minimum;
2800                 log.u_bbr.flex7 = rack->sack_attack_disable;
2801                 log.u_bbr.flex8 = event;
2802                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2803                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2804                 log.u_bbr.delivered = tcp_sad_decay_val;
2805                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2806                     &rack->rc_inp->inp_socket->so_rcv,
2807                     &rack->rc_inp->inp_socket->so_snd,
2808                     TCP_SAD_DETECTION, 0,
2809                     0, &log, false, &tv);
2810         }
2811 }
2812 #endif
2813
2814 static void
2815 rack_counter_destroy(void)
2816 {
2817         int i;
2818
2819         counter_u64_free(rack_fto_send);
2820         counter_u64_free(rack_fto_rsm_send);
2821         counter_u64_free(rack_nfto_resend);
2822         counter_u64_free(rack_hw_pace_init_fail);
2823         counter_u64_free(rack_hw_pace_lost);
2824         counter_u64_free(rack_non_fto_send);
2825         counter_u64_free(rack_extended_rfo);
2826         counter_u64_free(rack_ack_total);
2827         counter_u64_free(rack_express_sack);
2828         counter_u64_free(rack_sack_total);
2829         counter_u64_free(rack_move_none);
2830         counter_u64_free(rack_move_some);
2831         counter_u64_free(rack_sack_attacks_detected);
2832         counter_u64_free(rack_sack_attacks_reversed);
2833         counter_u64_free(rack_sack_used_next_merge);
2834         counter_u64_free(rack_sack_used_prev_merge);
2835         counter_u64_free(rack_badfr);
2836         counter_u64_free(rack_badfr_bytes);
2837         counter_u64_free(rack_rtm_prr_retran);
2838         counter_u64_free(rack_rtm_prr_newdata);
2839         counter_u64_free(rack_timestamp_mismatch);
2840         counter_u64_free(rack_find_high);
2841         counter_u64_free(rack_reorder_seen);
2842         counter_u64_free(rack_tlp_tot);
2843         counter_u64_free(rack_tlp_newdata);
2844         counter_u64_free(rack_tlp_retran);
2845         counter_u64_free(rack_tlp_retran_bytes);
2846         counter_u64_free(rack_tlp_retran_fail);
2847         counter_u64_free(rack_to_tot);
2848         counter_u64_free(rack_to_arm_rack);
2849         counter_u64_free(rack_to_arm_tlp);
2850         counter_u64_free(rack_calc_zero);
2851         counter_u64_free(rack_calc_nonzero);
2852         counter_u64_free(rack_paced_segments);
2853         counter_u64_free(rack_unpaced_segments);
2854         counter_u64_free(rack_saw_enobuf);
2855         counter_u64_free(rack_saw_enobuf_hw);
2856         counter_u64_free(rack_saw_enetunreach);
2857         counter_u64_free(rack_hot_alloc);
2858         counter_u64_free(rack_to_alloc);
2859         counter_u64_free(rack_to_alloc_hard);
2860         counter_u64_free(rack_to_alloc_emerg);
2861         counter_u64_free(rack_to_alloc_limited);
2862         counter_u64_free(rack_alloc_limited_conns);
2863         counter_u64_free(rack_split_limited);
2864         for (i = 0; i < MAX_NUM_OF_CNTS; i++) {
2865                 counter_u64_free(rack_proc_comp_ack[i]);
2866         }
2867         counter_u64_free(rack_multi_single_eq);
2868         counter_u64_free(rack_proc_non_comp_ack);
2869         counter_u64_free(rack_sack_proc_all);
2870         counter_u64_free(rack_sack_proc_restart);
2871         counter_u64_free(rack_sack_proc_short);
2872         counter_u64_free(rack_enter_tlp_calc);
2873         counter_u64_free(rack_used_tlpmethod);
2874         counter_u64_free(rack_used_tlpmethod2);
2875         counter_u64_free(rack_sack_skipped_acked);
2876         counter_u64_free(rack_sack_splits);
2877         counter_u64_free(rack_progress_drops);
2878         counter_u64_free(rack_input_idle_reduces);
2879         counter_u64_free(rack_collapsed_win);
2880         counter_u64_free(rack_tlp_does_nada);
2881         counter_u64_free(rack_try_scwnd);
2882         counter_u64_free(rack_per_timer_hole);
2883         counter_u64_free(rack_large_ackcmp);
2884         counter_u64_free(rack_small_ackcmp);
2885 #ifdef INVARIANTS
2886         counter_u64_free(rack_adjust_map_bw);
2887 #endif
2888         COUNTER_ARRAY_FREE(rack_out_size, TCP_MSS_ACCT_SIZE);
2889         COUNTER_ARRAY_FREE(rack_opts_arry, RACK_OPTS_SIZE);
2890 }
2891
2892 static struct rack_sendmap *
2893 rack_alloc(struct tcp_rack *rack)
2894 {
2895         struct rack_sendmap *rsm;
2896
2897         /*
2898          * First get the top of the list it in
2899          * theory is the "hottest" rsm we have,
2900          * possibly just freed by ack processing.
2901          */
2902         if (rack->rc_free_cnt > rack_free_cache) {
2903                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
2904                 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2905                 counter_u64_add(rack_hot_alloc, 1);
2906                 rack->rc_free_cnt--;
2907                 return (rsm);
2908         }
2909         /*
2910          * Once we get under our free cache we probably
2911          * no longer have a "hot" one available. Lets
2912          * get one from UMA.
2913          */
2914         rsm = uma_zalloc(rack_zone, M_NOWAIT);
2915         if (rsm) {
2916                 rack->r_ctl.rc_num_maps_alloced++;
2917                 counter_u64_add(rack_to_alloc, 1);
2918                 return (rsm);
2919         }
2920         /*
2921          * Dig in to our aux rsm's (the last two) since
2922          * UMA failed to get us one.
2923          */
2924         if (rack->rc_free_cnt) {
2925                 counter_u64_add(rack_to_alloc_emerg, 1);
2926                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
2927                 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2928                 rack->rc_free_cnt--;
2929                 return (rsm);
2930         }
2931         return (NULL);
2932 }
2933
2934 static struct rack_sendmap *
2935 rack_alloc_full_limit(struct tcp_rack *rack)
2936 {
2937         if ((V_tcp_map_entries_limit > 0) &&
2938             (rack->do_detection == 0) &&
2939             (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
2940                 counter_u64_add(rack_to_alloc_limited, 1);
2941                 if (!rack->alloc_limit_reported) {
2942                         rack->alloc_limit_reported = 1;
2943                         counter_u64_add(rack_alloc_limited_conns, 1);
2944                 }
2945                 return (NULL);
2946         }
2947         return (rack_alloc(rack));
2948 }
2949
2950 /* wrapper to allocate a sendmap entry, subject to a specific limit */
2951 static struct rack_sendmap *
2952 rack_alloc_limit(struct tcp_rack *rack, uint8_t limit_type)
2953 {
2954         struct rack_sendmap *rsm;
2955
2956         if (limit_type) {
2957                 /* currently there is only one limit type */
2958                 if (V_tcp_map_split_limit > 0 &&
2959                     (rack->do_detection == 0) &&
2960                     rack->r_ctl.rc_num_split_allocs >= V_tcp_map_split_limit) {
2961                         counter_u64_add(rack_split_limited, 1);
2962                         if (!rack->alloc_limit_reported) {
2963                                 rack->alloc_limit_reported = 1;
2964                                 counter_u64_add(rack_alloc_limited_conns, 1);
2965                         }
2966                         return (NULL);
2967                 }
2968         }
2969
2970         /* allocate and mark in the limit type, if set */
2971         rsm = rack_alloc(rack);
2972         if (rsm != NULL && limit_type) {
2973                 rsm->r_limit_type = limit_type;
2974                 rack->r_ctl.rc_num_split_allocs++;
2975         }
2976         return (rsm);
2977 }
2978
2979 static void
2980 rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm)
2981 {
2982         if (rsm->r_flags & RACK_APP_LIMITED) {
2983                 if (rack->r_ctl.rc_app_limited_cnt > 0) {
2984                         rack->r_ctl.rc_app_limited_cnt--;
2985                 }
2986         }
2987         if (rsm->r_limit_type) {
2988                 /* currently there is only one limit type */
2989                 rack->r_ctl.rc_num_split_allocs--;
2990         }
2991         if (rsm == rack->r_ctl.rc_first_appl) {
2992                 if (rack->r_ctl.rc_app_limited_cnt == 0)
2993                         rack->r_ctl.rc_first_appl = NULL;
2994                 else {
2995                         /* Follow the next one out */
2996                         struct rack_sendmap fe;
2997
2998                         fe.r_start = rsm->r_nseq_appl;
2999                         rack->r_ctl.rc_first_appl = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
3000                 }
3001         }
3002         if (rsm == rack->r_ctl.rc_resend)
3003                 rack->r_ctl.rc_resend = NULL;
3004         if (rsm == rack->r_ctl.rc_rsm_at_retran)
3005                 rack->r_ctl.rc_rsm_at_retran = NULL;
3006         if (rsm == rack->r_ctl.rc_end_appl)
3007                 rack->r_ctl.rc_end_appl = NULL;
3008         if (rack->r_ctl.rc_tlpsend == rsm)
3009                 rack->r_ctl.rc_tlpsend = NULL;
3010         if (rack->r_ctl.rc_sacklast == rsm)
3011                 rack->r_ctl.rc_sacklast = NULL;
3012         memset(rsm, 0, sizeof(struct rack_sendmap));
3013         TAILQ_INSERT_HEAD(&rack->r_ctl.rc_free, rsm, r_tnext);
3014         rack->rc_free_cnt++;
3015 }
3016
3017 static void
3018 rack_free_trim(struct tcp_rack *rack)
3019 {
3020         struct rack_sendmap *rsm;
3021
3022         /*
3023          * Free up all the tail entries until
3024          * we get our list down to the limit.
3025          */
3026         while (rack->rc_free_cnt > rack_free_cache) {
3027                 rsm = TAILQ_LAST(&rack->r_ctl.rc_free, rack_head);
3028                 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3029                 rack->rc_free_cnt--;
3030                 uma_zfree(rack_zone, rsm);
3031         }
3032 }
3033
3034
3035 static uint32_t
3036 rack_get_measure_window(struct tcpcb *tp, struct tcp_rack *rack)
3037 {
3038         uint64_t srtt, bw, len, tim;
3039         uint32_t segsiz, def_len, minl;
3040
3041         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3042         def_len = rack_def_data_window * segsiz;
3043         if (rack->rc_gp_filled == 0) {
3044                 /*
3045                  * We have no measurement (IW is in flight?) so
3046                  * we can only guess using our data_window sysctl
3047                  * value (usually 100MSS).
3048                  */
3049                 return (def_len);
3050         }
3051         /*
3052          * Now we have a number of factors to consider.
3053          *
3054          * 1) We have a desired BDP which is usually
3055          *    at least 2.
3056          * 2) We have a minimum number of rtt's usually 1 SRTT
3057          *    but we allow it too to be more.
3058          * 3) We want to make sure a measurement last N useconds (if
3059          *    we have set rack_min_measure_usec.
3060          *
3061          * We handle the first concern here by trying to create a data
3062          * window of max(rack_def_data_window, DesiredBDP). The
3063          * second concern we handle in not letting the measurement
3064          * window end normally until at least the required SRTT's
3065          * have gone by which is done further below in
3066          * rack_enough_for_measurement(). Finally the third concern
3067          * we also handle here by calculating how long that time
3068          * would take at the current BW and then return the
3069          * max of our first calculation and that length. Note
3070          * that if rack_min_measure_usec is 0, we don't deal
3071          * with concern 3. Also for both Concern 1 and 3 an
3072          * application limited period could end the measurement
3073          * earlier.
3074          *
3075          * So lets calculate the BDP with the "known" b/w using
3076          * the SRTT has our rtt and then multiply it by the
3077          * goal.
3078          */
3079         bw = rack_get_bw(rack);
3080         srtt = (uint64_t)tp->t_srtt;
3081         len = bw * srtt;
3082         len /= (uint64_t)HPTS_USEC_IN_SEC;
3083         len *= max(1, rack_goal_bdp);
3084         /* Now we need to round up to the nearest MSS */
3085         len = roundup(len, segsiz);
3086         if (rack_min_measure_usec) {
3087                 /* Now calculate our min length for this b/w */
3088                 tim = rack_min_measure_usec;
3089                 minl = (tim * bw) / (uint64_t)HPTS_USEC_IN_SEC;
3090                 if (minl == 0)
3091                         minl = 1;
3092                 minl = roundup(minl, segsiz);
3093                 if (len < minl)
3094                         len = minl;
3095         }
3096         /*
3097          * Now if we have a very small window we want
3098          * to attempt to get the window that is
3099          * as small as possible. This happens on
3100          * low b/w connections and we don't want to
3101          * span huge numbers of rtt's between measurements.
3102          *
3103          * We basically include 2 over our "MIN window" so
3104          * that the measurement can be shortened (possibly) by
3105          * an ack'ed packet.
3106          */
3107         if (len < def_len)
3108                 return (max((uint32_t)len, ((MIN_GP_WIN+2) * segsiz)));
3109         else
3110                 return (max((uint32_t)len, def_len));
3111
3112 }
3113
3114 static int
3115 rack_enough_for_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq th_ack)
3116 {
3117         uint32_t tim, srtts, segsiz;
3118
3119         /*
3120          * Has enough time passed for the GP measurement to be valid?
3121          */
3122         if ((tp->snd_max == tp->snd_una) ||
3123             (th_ack == tp->snd_max)){
3124                 /* All is acked */
3125                 return (1);
3126         }
3127         if (SEQ_LT(th_ack, tp->gput_seq)) {
3128                 /* Not enough bytes yet */
3129                 return (0);
3130         }
3131         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3132         if (SEQ_LT(th_ack, tp->gput_ack) &&
3133             ((th_ack - tp->gput_seq) < max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
3134                 /* Not enough bytes yet */
3135                 return (0);
3136         }
3137         if (rack->r_ctl.rc_first_appl &&
3138             (rack->r_ctl.rc_first_appl->r_start == th_ack)) {
3139                 /*
3140                  * We are up to the app limited point
3141                  * we have to measure irrespective of the time..
3142                  */
3143                 return (1);
3144         }
3145         /* Now what about time? */
3146         srtts = (rack->r_ctl.rc_gp_srtt * rack_min_srtts);
3147         tim = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - tp->gput_ts;
3148         if (tim >= srtts) {
3149                 return (1);
3150         }
3151         /* Nope not even a full SRTT has passed */
3152         return (0);
3153 }
3154
3155 static void
3156 rack_log_timely(struct tcp_rack *rack,
3157                 uint32_t logged, uint64_t cur_bw, uint64_t low_bnd,
3158                 uint64_t up_bnd, int line, uint8_t method)
3159 {
3160         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
3161                 union tcp_log_stackspecific log;
3162                 struct timeval tv;
3163
3164                 memset(&log, 0, sizeof(log));
3165                 log.u_bbr.flex1 = logged;
3166                 log.u_bbr.flex2 = rack->rc_gp_timely_inc_cnt;
3167                 log.u_bbr.flex2 <<= 4;
3168                 log.u_bbr.flex2 |= rack->rc_gp_timely_dec_cnt;
3169                 log.u_bbr.flex2 <<= 4;
3170                 log.u_bbr.flex2 |= rack->rc_gp_incr;
3171                 log.u_bbr.flex2 <<= 4;
3172                 log.u_bbr.flex2 |= rack->rc_gp_bwred;
3173                 log.u_bbr.flex3 = rack->rc_gp_incr;
3174                 log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3175                 log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ca;
3176                 log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_rec;
3177                 log.u_bbr.flex7 = rack->rc_gp_bwred;
3178                 log.u_bbr.flex8 = method;
3179                 log.u_bbr.cur_del_rate = cur_bw;
3180                 log.u_bbr.delRate = low_bnd;
3181                 log.u_bbr.bw_inuse = up_bnd;
3182                 log.u_bbr.rttProp = rack_get_bw(rack);
3183                 log.u_bbr.pkt_epoch = line;
3184                 log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3185                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3186                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3187                 log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3188                 log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3189                 log.u_bbr.cwnd_gain = rack->rc_dragged_bottom;
3190                 log.u_bbr.cwnd_gain <<= 1;
3191                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_rec;
3192                 log.u_bbr.cwnd_gain <<= 1;
3193                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
3194                 log.u_bbr.cwnd_gain <<= 1;
3195                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
3196                 log.u_bbr.lost = rack->r_ctl.rc_loss_count;
3197                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3198                     &rack->rc_inp->inp_socket->so_rcv,
3199                     &rack->rc_inp->inp_socket->so_snd,
3200                     TCP_TIMELY_WORK, 0,
3201                     0, &log, false, &tv);
3202         }
3203 }
3204
3205 static int
3206 rack_bw_can_be_raised(struct tcp_rack *rack, uint64_t cur_bw, uint64_t last_bw_est, uint16_t mult)
3207 {
3208         /*
3209          * Before we increase we need to know if
3210          * the estimate just made was less than
3211          * our pacing goal (i.e. (cur_bw * mult) > last_bw_est)
3212          *
3213          * If we already are pacing at a fast enough
3214          * rate to push us faster there is no sense of
3215          * increasing.
3216          *
3217          * We first caculate our actual pacing rate (ss or ca multipler
3218          * times our cur_bw).
3219          *
3220          * Then we take the last measured rate and multipy by our
3221          * maximum pacing overage to give us a max allowable rate.
3222          *
3223          * If our act_rate is smaller than our max_allowable rate
3224          * then we should increase. Else we should hold steady.
3225          *
3226          */
3227         uint64_t act_rate, max_allow_rate;
3228
3229         if (rack_timely_no_stopping)
3230                 return (1);
3231
3232         if ((cur_bw == 0) || (last_bw_est == 0)) {
3233                 /*
3234                  * Initial startup case or
3235                  * everything is acked case.
3236                  */
3237                 rack_log_timely(rack,  mult, cur_bw, 0, 0,
3238                                 __LINE__, 9);
3239                 return (1);
3240         }
3241         if (mult <= 100) {
3242                 /*
3243                  * We can always pace at or slightly above our rate.
3244                  */
3245                 rack_log_timely(rack,  mult, cur_bw, 0, 0,
3246                                 __LINE__, 9);
3247                 return (1);
3248         }
3249         act_rate = cur_bw * (uint64_t)mult;
3250         act_rate /= 100;
3251         max_allow_rate = last_bw_est * ((uint64_t)rack_max_per_above + (uint64_t)100);
3252         max_allow_rate /= 100;
3253         if (act_rate < max_allow_rate) {
3254                 /*
3255                  * Here the rate we are actually pacing at
3256                  * is smaller than 10% above our last measurement.
3257                  * This means we are pacing below what we would
3258                  * like to try to achieve (plus some wiggle room).
3259                  */
3260                 rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3261                                 __LINE__, 9);
3262                 return (1);
3263         } else {
3264                 /*
3265                  * Here we are already pacing at least rack_max_per_above(10%)
3266                  * what we are getting back. This indicates most likely
3267                  * that we are being limited (cwnd/rwnd/app) and can't
3268                  * get any more b/w. There is no sense of trying to
3269                  * raise up the pacing rate its not speeding us up
3270                  * and we already are pacing faster than we are getting.
3271                  */
3272                 rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3273                                 __LINE__, 8);
3274                 return (0);
3275         }
3276 }
3277
3278 static void
3279 rack_validate_multipliers_at_or_above100(struct tcp_rack *rack)
3280 {
3281         /*
3282          * When we drag bottom, we want to assure
3283          * that no multiplier is below 1.0, if so
3284          * we want to restore it to at least that.
3285          */
3286         if (rack->r_ctl.rack_per_of_gp_rec  < 100) {
3287                 /* This is unlikely we usually do not touch recovery */
3288                 rack->r_ctl.rack_per_of_gp_rec = 100;
3289         }
3290         if (rack->r_ctl.rack_per_of_gp_ca < 100) {
3291                 rack->r_ctl.rack_per_of_gp_ca = 100;
3292         }
3293         if (rack->r_ctl.rack_per_of_gp_ss < 100) {
3294                 rack->r_ctl.rack_per_of_gp_ss = 100;
3295         }
3296 }
3297
3298 static void
3299 rack_validate_multipliers_at_or_below_100(struct tcp_rack *rack)
3300 {
3301         if (rack->r_ctl.rack_per_of_gp_ca > 100) {
3302                 rack->r_ctl.rack_per_of_gp_ca = 100;
3303         }
3304         if (rack->r_ctl.rack_per_of_gp_ss > 100) {
3305                 rack->r_ctl.rack_per_of_gp_ss = 100;
3306         }
3307 }
3308
3309 static void
3310 rack_increase_bw_mul(struct tcp_rack *rack, int timely_says, uint64_t cur_bw, uint64_t last_bw_est, int override)
3311 {
3312         int32_t  calc, logged, plus;
3313
3314         logged = 0;
3315
3316         if (override) {
3317                 /*
3318                  * override is passed when we are
3319                  * loosing b/w and making one last
3320                  * gasp at trying to not loose out
3321                  * to a new-reno flow.
3322                  */
3323                 goto extra_boost;
3324         }
3325         /* In classic timely we boost by 5x if we have 5 increases in a row, lets not */
3326         if (rack->rc_gp_incr &&
3327             ((rack->rc_gp_timely_inc_cnt + 1) >= RACK_TIMELY_CNT_BOOST)) {
3328                 /*
3329                  * Reset and get 5 strokes more before the boost. Note
3330                  * that the count is 0 based so we have to add one.
3331                  */
3332 extra_boost:
3333                 plus = (uint32_t)rack_gp_increase_per * RACK_TIMELY_CNT_BOOST;
3334                 rack->rc_gp_timely_inc_cnt = 0;
3335         } else
3336                 plus = (uint32_t)rack_gp_increase_per;
3337         /* Must be at least 1% increase for true timely increases */
3338         if ((plus < 1) &&
3339             ((rack->r_ctl.rc_rtt_diff <= 0) || (timely_says <= 0)))
3340                 plus = 1;
3341         if (rack->rc_gp_saw_rec &&
3342             (rack->rc_gp_no_rec_chg == 0) &&
3343             rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3344                                   rack->r_ctl.rack_per_of_gp_rec)) {
3345                 /* We have been in recovery ding it too */
3346                 calc = rack->r_ctl.rack_per_of_gp_rec + plus;
3347                 if (calc > 0xffff)
3348                         calc = 0xffff;
3349                 logged |= 1;
3350                 rack->r_ctl.rack_per_of_gp_rec = (uint16_t)calc;
3351                 if (rack_per_upper_bound_ss &&
3352                     (rack->rc_dragged_bottom == 0) &&
3353                     (rack->r_ctl.rack_per_of_gp_rec > rack_per_upper_bound_ss))
3354                         rack->r_ctl.rack_per_of_gp_rec = rack_per_upper_bound_ss;
3355         }
3356         if (rack->rc_gp_saw_ca &&
3357             (rack->rc_gp_saw_ss == 0) &&
3358             rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3359                                   rack->r_ctl.rack_per_of_gp_ca)) {
3360                 /* In CA */
3361                 calc = rack->r_ctl.rack_per_of_gp_ca + plus;
3362                 if (calc > 0xffff)
3363                         calc = 0xffff;
3364                 logged |= 2;
3365                 rack->r_ctl.rack_per_of_gp_ca = (uint16_t)calc;
3366                 if (rack_per_upper_bound_ca &&
3367                     (rack->rc_dragged_bottom == 0) &&
3368                     (rack->r_ctl.rack_per_of_gp_ca > rack_per_upper_bound_ca))
3369                         rack->r_ctl.rack_per_of_gp_ca = rack_per_upper_bound_ca;
3370         }
3371         if (rack->rc_gp_saw_ss &&
3372             rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3373                                   rack->r_ctl.rack_per_of_gp_ss)) {
3374                 /* In SS */
3375                 calc = rack->r_ctl.rack_per_of_gp_ss + plus;
3376                 if (calc > 0xffff)
3377                         calc = 0xffff;
3378                 rack->r_ctl.rack_per_of_gp_ss = (uint16_t)calc;
3379                 if (rack_per_upper_bound_ss &&
3380                     (rack->rc_dragged_bottom == 0) &&
3381                     (rack->r_ctl.rack_per_of_gp_ss > rack_per_upper_bound_ss))
3382                         rack->r_ctl.rack_per_of_gp_ss = rack_per_upper_bound_ss;
3383                 logged |= 4;
3384         }
3385         if (logged &&
3386             (rack->rc_gp_incr == 0)){
3387                 /* Go into increment mode */
3388                 rack->rc_gp_incr = 1;
3389                 rack->rc_gp_timely_inc_cnt = 0;
3390         }
3391         if (rack->rc_gp_incr &&
3392             logged &&
3393             (rack->rc_gp_timely_inc_cnt < RACK_TIMELY_CNT_BOOST)) {
3394                 rack->rc_gp_timely_inc_cnt++;
3395         }
3396         rack_log_timely(rack,  logged, plus, 0, 0,
3397                         __LINE__, 1);
3398 }
3399
3400 static uint32_t
3401 rack_get_decrease(struct tcp_rack *rack, uint32_t curper, int32_t rtt_diff)
3402 {
3403         /*
3404          * norm_grad = rtt_diff / minrtt;
3405          * new_per = curper * (1 - B * norm_grad)
3406          *
3407          * B = rack_gp_decrease_per (default 10%)
3408          * rtt_dif = input var current rtt-diff
3409          * curper = input var current percentage
3410          * minrtt = from rack filter
3411          *
3412          */
3413         uint64_t perf;
3414
3415         perf = (((uint64_t)curper * ((uint64_t)1000000 -
3416                     ((uint64_t)rack_gp_decrease_per * (uint64_t)10000 *
3417                      (((uint64_t)rtt_diff * (uint64_t)1000000)/
3418                       (uint64_t)get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt)))/
3419                      (uint64_t)1000000)) /
3420                 (uint64_t)1000000);
3421         if (perf > curper) {
3422                 /* TSNH */
3423                 perf = curper - 1;
3424         }
3425         return ((uint32_t)perf);
3426 }
3427
3428 static uint32_t
3429 rack_decrease_highrtt(struct tcp_rack *rack, uint32_t curper, uint32_t rtt)
3430 {
3431         /*
3432          *                                   highrttthresh
3433          * result = curper * (1 - (B * ( 1 -  ------          ))
3434          *                                     gp_srtt
3435          *
3436          * B = rack_gp_decrease_per (default 10%)
3437          * highrttthresh = filter_min * rack_gp_rtt_maxmul
3438          */
3439         uint64_t perf;
3440         uint32_t highrttthresh;
3441
3442         highrttthresh = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
3443
3444         perf = (((uint64_t)curper * ((uint64_t)1000000 -
3445                                      ((uint64_t)rack_gp_decrease_per * ((uint64_t)1000000 -
3446                                         ((uint64_t)highrttthresh * (uint64_t)1000000) /
3447                                                     (uint64_t)rtt)) / 100)) /(uint64_t)1000000);
3448         return (perf);
3449 }
3450
3451 static void
3452 rack_decrease_bw_mul(struct tcp_rack *rack, int timely_says, uint32_t rtt, int32_t rtt_diff)
3453 {
3454         uint64_t logvar, logvar2, logvar3;
3455         uint32_t logged, new_per, ss_red, ca_red, rec_red, alt, val;
3456
3457         if (rack->rc_gp_incr) {
3458                 /* Turn off increment counting */
3459                 rack->rc_gp_incr = 0;
3460                 rack->rc_gp_timely_inc_cnt = 0;
3461         }
3462         ss_red = ca_red = rec_red = 0;
3463         logged = 0;
3464         /* Calculate the reduction value */
3465         if (rtt_diff < 0) {
3466                 rtt_diff *= -1;
3467         }
3468         /* Must be at least 1% reduction */
3469         if (rack->rc_gp_saw_rec && (rack->rc_gp_no_rec_chg == 0)) {
3470                 /* We have been in recovery ding it too */
3471                 if (timely_says == 2) {
3472                         new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_rec, rtt);
3473                         alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3474                         if (alt < new_per)
3475                                 val = alt;
3476                         else
3477                                 val = new_per;
3478                 } else
3479                          val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3480                 if (rack->r_ctl.rack_per_of_gp_rec > val) {
3481                         rec_red = (rack->r_ctl.rack_per_of_gp_rec - val);
3482                         rack->r_ctl.rack_per_of_gp_rec = (uint16_t)val;
3483                 } else {
3484                         rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
3485                         rec_red = 0;
3486                 }
3487                 if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_rec)
3488                         rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
3489                 logged |= 1;
3490         }
3491         if (rack->rc_gp_saw_ss) {
3492                 /* Sent in SS */
3493                 if (timely_says == 2) {
3494                         new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ss, rtt);
3495                         alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3496                         if (alt < new_per)
3497                                 val = alt;
3498                         else
3499                                 val = new_per;
3500                 } else
3501                         val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff);
3502                 if (rack->r_ctl.rack_per_of_gp_ss > new_per) {
3503                         ss_red = rack->r_ctl.rack_per_of_gp_ss - val;
3504                         rack->r_ctl.rack_per_of_gp_ss = (uint16_t)val;
3505                 } else {
3506                         ss_red = new_per;
3507                         rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
3508                         logvar = new_per;
3509                         logvar <<= 32;
3510                         logvar |= alt;
3511                         logvar2 = (uint32_t)rtt;
3512                         logvar2 <<= 32;
3513                         logvar2 |= (uint32_t)rtt_diff;
3514                         logvar3 = rack_gp_rtt_maxmul;
3515                         logvar3 <<= 32;
3516                         logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3517                         rack_log_timely(rack, timely_says,
3518                                         logvar2, logvar3,
3519                                         logvar, __LINE__, 10);
3520                 }
3521                 if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ss)
3522                         rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
3523                 logged |= 4;
3524         } else if (rack->rc_gp_saw_ca) {
3525                 /* Sent in CA */
3526                 if (timely_says == 2) {
3527                         new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ca, rtt);
3528                         alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3529                         if (alt < new_per)
3530                                 val = alt;
3531                         else
3532                                 val = new_per;
3533                 } else
3534                         val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff);
3535                 if (rack->r_ctl.rack_per_of_gp_ca > val) {
3536                         ca_red = rack->r_ctl.rack_per_of_gp_ca - val;
3537                         rack->r_ctl.rack_per_of_gp_ca = (uint16_t)val;
3538                 } else {
3539                         rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
3540                         ca_red = 0;
3541                         logvar = new_per;
3542                         logvar <<= 32;
3543                         logvar |= alt;
3544                         logvar2 = (uint32_t)rtt;
3545                         logvar2 <<= 32;
3546                         logvar2 |= (uint32_t)rtt_diff;
3547                         logvar3 = rack_gp_rtt_maxmul;
3548                         logvar3 <<= 32;
3549                         logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3550                         rack_log_timely(rack, timely_says,
3551                                         logvar2, logvar3,
3552                                         logvar, __LINE__, 10);
3553                 }
3554                 if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ca)
3555                         rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
3556                 logged |= 2;
3557         }
3558         if (rack->rc_gp_timely_dec_cnt < 0x7) {
3559                 rack->rc_gp_timely_dec_cnt++;
3560                 if (rack_timely_dec_clear &&
3561                     (rack->rc_gp_timely_dec_cnt == rack_timely_dec_clear))
3562                         rack->rc_gp_timely_dec_cnt = 0;
3563         }
3564         logvar = ss_red;
3565         logvar <<= 32;
3566         logvar |= ca_red;
3567         rack_log_timely(rack,  logged, rec_red, rack_per_lower_bound, logvar,
3568                         __LINE__, 2);
3569 }
3570
3571 static void
3572 rack_log_rtt_shrinks(struct tcp_rack *rack, uint32_t us_cts,
3573                      uint32_t rtt, uint32_t line, uint8_t reas)
3574 {
3575         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
3576                 union tcp_log_stackspecific log;
3577                 struct timeval tv;
3578
3579                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3580                 log.u_bbr.flex1 = line;
3581                 log.u_bbr.flex2 = rack->r_ctl.rc_time_probertt_starts;
3582                 log.u_bbr.flex3 = rack->r_ctl.rc_lower_rtt_us_cts;
3583                 log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3584                 log.u_bbr.flex5 = rtt;
3585                 log.u_bbr.flex6 = rack->rc_highly_buffered;
3586                 log.u_bbr.flex6 <<= 1;
3587                 log.u_bbr.flex6 |= rack->forced_ack;
3588                 log.u_bbr.flex6 <<= 1;
3589                 log.u_bbr.flex6 |= rack->rc_gp_dyn_mul;
3590                 log.u_bbr.flex6 <<= 1;
3591                 log.u_bbr.flex6 |= rack->in_probe_rtt;
3592                 log.u_bbr.flex6 <<= 1;
3593                 log.u_bbr.flex6 |= rack->measure_saw_probe_rtt;
3594                 log.u_bbr.flex7 = rack->r_ctl.rack_per_of_gp_probertt;
3595                 log.u_bbr.pacing_gain = rack->r_ctl.rack_per_of_gp_ca;
3596                 log.u_bbr.cwnd_gain = rack->r_ctl.rack_per_of_gp_rec;
3597                 log.u_bbr.flex8 = reas;
3598                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3599                 log.u_bbr.delRate = rack_get_bw(rack);
3600                 log.u_bbr.cur_del_rate = rack->r_ctl.rc_highest_us_rtt;
3601                 log.u_bbr.cur_del_rate <<= 32;
3602                 log.u_bbr.cur_del_rate |= rack->r_ctl.rc_lowest_us_rtt;
3603                 log.u_bbr.applimited = rack->r_ctl.rc_time_probertt_entered;
3604                 log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3605                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3606                 log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3607                 log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3608                 log.u_bbr.pkt_epoch = rack->r_ctl.rc_lower_rtt_us_cts;
3609                 log.u_bbr.delivered = rack->r_ctl.rc_target_probertt_flight;
3610                 log.u_bbr.lost = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3611                 log.u_bbr.rttProp = us_cts;
3612                 log.u_bbr.rttProp <<= 32;
3613                 log.u_bbr.rttProp |= rack->r_ctl.rc_entry_gp_rtt;
3614                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3615                     &rack->rc_inp->inp_socket->so_rcv,
3616                     &rack->rc_inp->inp_socket->so_snd,
3617                     BBR_LOG_RTT_SHRINKS, 0,
3618                     0, &log, false, &rack->r_ctl.act_rcv_time);
3619         }
3620 }
3621
3622 static void
3623 rack_set_prtt_target(struct tcp_rack *rack, uint32_t segsiz, uint32_t rtt)
3624 {
3625         uint64_t bwdp;
3626
3627         bwdp = rack_get_bw(rack);
3628         bwdp *= (uint64_t)rtt;
3629         bwdp /= (uint64_t)HPTS_USEC_IN_SEC;
3630         rack->r_ctl.rc_target_probertt_flight = roundup((uint32_t)bwdp, segsiz);
3631         if (rack->r_ctl.rc_target_probertt_flight < (segsiz * rack_timely_min_segs)) {
3632                 /*
3633                  * A window protocol must be able to have 4 packets
3634                  * outstanding as the floor in order to function
3635                  * (especially considering delayed ack :D).
3636                  */
3637                 rack->r_ctl.rc_target_probertt_flight = (segsiz * rack_timely_min_segs);
3638         }
3639 }
3640
3641 static void
3642 rack_enter_probertt(struct tcp_rack *rack, uint32_t us_cts)
3643 {
3644         /**
3645          * ProbeRTT is a bit different in rack_pacing than in
3646          * BBR. It is like BBR in that it uses the lowering of
3647          * the RTT as a signal that we saw something new and
3648          * counts from there for how long between. But it is
3649          * different in that its quite simple. It does not
3650          * play with the cwnd and wait until we get down
3651          * to N segments outstanding and hold that for
3652          * 200ms. Instead it just sets the pacing reduction
3653          * rate to a set percentage (70 by default) and hold
3654          * that for a number of recent GP Srtt's.
3655          */
3656         uint32_t segsiz;
3657
3658         if (rack->rc_gp_dyn_mul == 0)
3659                 return;
3660
3661         if (rack->rc_tp->snd_max == rack->rc_tp->snd_una) {
3662                 /* We are idle */
3663                 return;
3664         }
3665         if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
3666             SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
3667                 /*
3668                  * Stop the goodput now, the idea here is
3669                  * that future measurements with in_probe_rtt
3670                  * won't register if they are not greater so
3671                  * we want to get what info (if any) is available
3672                  * now.
3673                  */
3674                 rack_do_goodput_measurement(rack->rc_tp, rack,
3675                                             rack->rc_tp->snd_una, __LINE__);
3676         }
3677         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
3678         rack->r_ctl.rc_time_probertt_entered = us_cts;
3679         segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
3680                      rack->r_ctl.rc_pace_min_segs);
3681         rack->in_probe_rtt = 1;
3682         rack->measure_saw_probe_rtt = 1;
3683         rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
3684         rack->r_ctl.rc_time_probertt_starts = 0;
3685         rack->r_ctl.rc_entry_gp_rtt = rack->r_ctl.rc_gp_srtt;
3686         if (rack_probertt_use_min_rtt_entry)
3687                 rack_set_prtt_target(rack, segsiz, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
3688         else
3689                 rack_set_prtt_target(rack, segsiz, rack->r_ctl.rc_gp_srtt);
3690         rack_log_rtt_shrinks(rack,  us_cts,  get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3691                              __LINE__, RACK_RTTS_ENTERPROBE);
3692 }
3693
3694 static void
3695 rack_exit_probertt(struct tcp_rack *rack, uint32_t us_cts)
3696 {
3697         struct rack_sendmap *rsm;
3698         uint32_t segsiz;
3699
3700         segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
3701                      rack->r_ctl.rc_pace_min_segs);
3702         rack->in_probe_rtt = 0;
3703         if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
3704             SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
3705                 /*
3706                  * Stop the goodput now, the idea here is
3707                  * that future measurements with in_probe_rtt
3708                  * won't register if they are not greater so
3709                  * we want to get what info (if any) is available
3710                  * now.
3711                  */
3712                 rack_do_goodput_measurement(rack->rc_tp, rack,
3713                                             rack->rc_tp->snd_una, __LINE__);
3714         } else if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
3715                 /*
3716                  * We don't have enough data to make a measurement.
3717                  * So lets just stop and start here after exiting
3718                  * probe-rtt. We probably are not interested in
3719                  * the results anyway.
3720                  */
3721                 rack->rc_tp->t_flags &= ~TF_GPUTINPROG;
3722         }
3723         /*
3724          * Measurements through the current snd_max are going
3725          * to be limited by the slower pacing rate.
3726          *
3727          * We need to mark these as app-limited so we
3728          * don't collapse the b/w.
3729          */
3730         rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
3731         if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
3732                 if (rack->r_ctl.rc_app_limited_cnt == 0)
3733                         rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
3734                 else {
3735                         /*
3736                          * Go out to the end app limited and mark
3737                          * this new one as next and move the end_appl up
3738                          * to this guy.
3739                          */
3740                         if (rack->r_ctl.rc_end_appl)
3741                                 rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
3742                         rack->r_ctl.rc_end_appl = rsm;
3743                 }
3744                 rsm->r_flags |= RACK_APP_LIMITED;
3745                 rack->r_ctl.rc_app_limited_cnt++;
3746         }
3747         /*
3748          * Now, we need to examine our pacing rate multipliers.
3749          * If its under 100%, we need to kick it back up to
3750          * 100%. We also don't let it be over our "max" above
3751          * the actual rate i.e. 100% + rack_clamp_atexit_prtt.
3752          * Note setting clamp_atexit_prtt to 0 has the effect
3753          * of setting CA/SS to 100% always at exit (which is
3754          * the default behavior).
3755          */
3756         if (rack_probertt_clear_is) {
3757                 rack->rc_gp_incr = 0;
3758                 rack->rc_gp_bwred = 0;
3759                 rack->rc_gp_timely_inc_cnt = 0;
3760                 rack->rc_gp_timely_dec_cnt = 0;
3761         }
3762         /* Do we do any clamping at exit? */
3763         if (rack->rc_highly_buffered && rack_atexit_prtt_hbp) {
3764                 rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt_hbp;
3765                 rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt_hbp;
3766         }
3767         if ((rack->rc_highly_buffered == 0) && rack_atexit_prtt) {
3768                 rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt;
3769                 rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt;
3770         }
3771         /*
3772          * Lets set rtt_diff to 0, so that we will get a "boost"
3773          * after exiting.
3774          */
3775         rack->r_ctl.rc_rtt_diff = 0;
3776
3777         /* Clear all flags so we start fresh */
3778         rack->rc_tp->t_bytes_acked = 0;
3779         rack->rc_tp->ccv->flags &= ~CCF_ABC_SENTAWND;
3780         /*
3781          * If configured to, set the cwnd and ssthresh to
3782          * our targets.
3783          */
3784         if (rack_probe_rtt_sets_cwnd) {
3785                 uint64_t ebdp;
3786                 uint32_t setto;
3787
3788                 /* Set ssthresh so we get into CA once we hit our target */
3789                 if (rack_probertt_use_min_rtt_exit == 1) {
3790                         /* Set to min rtt */
3791                         rack_set_prtt_target(rack, segsiz,
3792                                              get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
3793                 } else if (rack_probertt_use_min_rtt_exit == 2) {
3794                         /* Set to current gp rtt */
3795                         rack_set_prtt_target(rack, segsiz,
3796                                              rack->r_ctl.rc_gp_srtt);
3797                 } else if (rack_probertt_use_min_rtt_exit == 3) {
3798                         /* Set to entry gp rtt */
3799                         rack_set_prtt_target(rack, segsiz,
3800                                              rack->r_ctl.rc_entry_gp_rtt);
3801                 } else {
3802                         uint64_t sum;
3803                         uint32_t setval;
3804
3805                         sum = rack->r_ctl.rc_entry_gp_rtt;
3806                         sum *= 10;
3807                         sum /= (uint64_t)(max(1, rack->r_ctl.rc_gp_srtt));
3808                         if (sum >= 20) {
3809                                 /*
3810                                  * A highly buffered path needs
3811                                  * cwnd space for timely to work.
3812                                  * Lets set things up as if
3813                                  * we are heading back here again.
3814                                  */
3815                                 setval = rack->r_ctl.rc_entry_gp_rtt;
3816                         } else if (sum >= 15) {
3817                                 /*
3818                                  * Lets take the smaller of the
3819                                  * two since we are just somewhat
3820                                  * buffered.
3821                                  */
3822                                 setval = rack->r_ctl.rc_gp_srtt;
3823                                 if (setval > rack->r_ctl.rc_entry_gp_rtt)
3824                                         setval = rack->r_ctl.rc_entry_gp_rtt;
3825                         } else {
3826                                 /*
3827                                  * Here we are not highly buffered
3828                                  * and should pick the min we can to
3829                                  * keep from causing loss.
3830                                  */
3831                                 setval = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3832                         }
3833                         rack_set_prtt_target(rack, segsiz,
3834                                              setval);
3835                 }
3836                 if (rack_probe_rtt_sets_cwnd > 1) {
3837                         /* There is a percentage here to boost */
3838                         ebdp = rack->r_ctl.rc_target_probertt_flight;
3839                         ebdp *= rack_probe_rtt_sets_cwnd;
3840                         ebdp /= 100;
3841                         setto = rack->r_ctl.rc_target_probertt_flight + ebdp;
3842                 } else
3843                         setto = rack->r_ctl.rc_target_probertt_flight;
3844                 rack->rc_tp->snd_cwnd = roundup(setto, segsiz);
3845                 if (rack->rc_tp->snd_cwnd < (segsiz * rack_timely_min_segs)) {
3846                         /* Enforce a min */
3847                         rack->rc_tp->snd_cwnd = segsiz * rack_timely_min_segs;
3848                 }
3849                 /* If we set in the cwnd also set the ssthresh point so we are in CA */
3850                 rack->rc_tp->snd_ssthresh = (rack->rc_tp->snd_cwnd - 1);
3851         }
3852         rack_log_rtt_shrinks(rack,  us_cts,
3853                              get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3854                              __LINE__, RACK_RTTS_EXITPROBE);
3855         /* Clear times last so log has all the info */
3856         rack->r_ctl.rc_probertt_sndmax_atexit = rack->rc_tp->snd_max;
3857         rack->r_ctl.rc_time_probertt_entered = us_cts;
3858         rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
3859         rack->r_ctl.rc_time_of_last_probertt = us_cts;
3860 }
3861
3862 static void
3863 rack_check_probe_rtt(struct tcp_rack *rack, uint32_t us_cts)
3864 {
3865         /* Check in on probe-rtt */
3866         if (rack->rc_gp_filled == 0) {
3867                 /* We do not do p-rtt unless we have gp measurements */
3868                 return;
3869         }
3870         if (rack->in_probe_rtt) {
3871                 uint64_t no_overflow;
3872                 uint32_t endtime, must_stay;
3873
3874                 if (rack->r_ctl.rc_went_idle_time &&
3875                     ((us_cts - rack->r_ctl.rc_went_idle_time) > rack_min_probertt_hold)) {
3876                         /*
3877                          * We went idle during prtt, just exit now.
3878                          */
3879                         rack_exit_probertt(rack, us_cts);
3880                 } else if (rack_probe_rtt_safety_val &&
3881                     TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered) &&
3882                     ((us_cts - rack->r_ctl.rc_time_probertt_entered) > rack_probe_rtt_safety_val)) {
3883                         /*
3884                          * Probe RTT safety value triggered!
3885                          */
3886                         rack_log_rtt_shrinks(rack,  us_cts,
3887                                              get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3888                                              __LINE__, RACK_RTTS_SAFETY);
3889                         rack_exit_probertt(rack, us_cts);
3890                 }
3891                 /* Calculate the max we will wait */
3892                 endtime = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_max_drain_wait);
3893                 if (rack->rc_highly_buffered)
3894                         endtime += (rack->r_ctl.rc_gp_srtt * rack_max_drain_hbp);
3895                 /* Calculate the min we must wait */
3896                 must_stay = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_must_drain);
3897                 if ((ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.rc_target_probertt_flight) &&
3898                     TSTMP_LT(us_cts, endtime)) {
3899                         uint32_t calc;
3900                         /* Do we lower more? */
3901 no_exit:
3902                         if (TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered))
3903                                 calc = us_cts - rack->r_ctl.rc_time_probertt_entered;
3904                         else
3905                                 calc = 0;
3906                         calc /= max(rack->r_ctl.rc_gp_srtt, 1);
3907                         if (calc) {
3908                                 /* Maybe */
3909                                 calc *= rack_per_of_gp_probertt_reduce;
3910                                 rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt - calc;
3911                                 /* Limit it too */
3912                                 if (rack->r_ctl.rack_per_of_gp_probertt < rack_per_of_gp_lowthresh)
3913                                         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_lowthresh;
3914                         }
3915                         /* We must reach target or the time set */
3916                         return;
3917                 }
3918                 if (rack->r_ctl.rc_time_probertt_starts == 0) {
3919                         if ((TSTMP_LT(us_cts, must_stay) &&
3920                              rack->rc_highly_buffered) ||
3921                              (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) >
3922                               rack->r_ctl.rc_target_probertt_flight)) {
3923                                 /* We are not past the must_stay time */
3924                                 goto no_exit;
3925                         }
3926                         rack_log_rtt_shrinks(rack,  us_cts,
3927                                              get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3928                                              __LINE__, RACK_RTTS_REACHTARGET);
3929                         rack->r_ctl.rc_time_probertt_starts = us_cts;
3930                         if (rack->r_ctl.rc_time_probertt_starts == 0)
3931                                 rack->r_ctl.rc_time_probertt_starts = 1;
3932                         /* Restore back to our rate we want to pace at in prtt */
3933                         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
3934                 }
3935                 /*
3936                  * Setup our end time, some number of gp_srtts plus 200ms.
3937                  */
3938                 no_overflow = ((uint64_t)rack->r_ctl.rc_gp_srtt *
3939                                (uint64_t)rack_probertt_gpsrtt_cnt_mul);
3940                 if (rack_probertt_gpsrtt_cnt_div)
3941                         endtime = (uint32_t)(no_overflow / (uint64_t)rack_probertt_gpsrtt_cnt_div);
3942                 else
3943                         endtime = 0;
3944                 endtime += rack_min_probertt_hold;
3945                 endtime += rack->r_ctl.rc_time_probertt_starts;
3946                 if (TSTMP_GEQ(us_cts,  endtime)) {
3947                         /* yes, exit probertt */
3948                         rack_exit_probertt(rack, us_cts);
3949                 }
3950
3951         } else if ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= rack_time_between_probertt) {
3952                 /* Go into probertt, its been too long since we went lower */
3953                 rack_enter_probertt(rack, us_cts);
3954         }
3955 }
3956
3957 static void
3958 rack_update_multiplier(struct tcp_rack *rack, int32_t timely_says, uint64_t last_bw_est,
3959                        uint32_t rtt, int32_t rtt_diff)
3960 {
3961         uint64_t cur_bw, up_bnd, low_bnd, subfr;
3962         uint32_t losses;
3963
3964         if ((rack->rc_gp_dyn_mul == 0) ||
3965             (rack->use_fixed_rate) ||
3966             (rack->in_probe_rtt) ||
3967             (rack->rc_always_pace == 0)) {
3968                 /* No dynamic GP multipler in play */
3969                 return;
3970         }
3971         losses = rack->r_ctl.rc_loss_count - rack->r_ctl.rc_loss_at_start;
3972         cur_bw = rack_get_bw(rack);
3973         /* Calculate our up and down range */
3974         up_bnd = rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_up;
3975         up_bnd /= 100;
3976         up_bnd += rack->r_ctl.last_gp_comp_bw;
3977
3978         subfr = (uint64_t)rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_down;
3979         subfr /= 100;
3980         low_bnd = rack->r_ctl.last_gp_comp_bw - subfr;
3981         if ((timely_says == 2) && (rack->r_ctl.rc_no_push_at_mrtt)) {
3982                 /*
3983                  * This is the case where our RTT is above
3984                  * the max target and we have been configured
3985                  * to just do timely no bonus up stuff in that case.
3986                  *
3987                  * There are two configurations, set to 1, and we
3988                  * just do timely if we are over our max. If its
3989                  * set above 1 then we slam the multipliers down
3990                  * to 100 and then decrement per timely.
3991                  */
3992                 rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3993                                 __LINE__, 3);
3994                 if (rack->r_ctl.rc_no_push_at_mrtt > 1)
3995                         rack_validate_multipliers_at_or_below_100(rack);
3996                 rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
3997         } else if ((last_bw_est < low_bnd) && !losses) {
3998                 /*
3999                  * We are decreasing this is a bit complicated this
4000                  * means we are loosing ground. This could be
4001                  * because another flow entered and we are competing
4002                  * for b/w with it. This will push the RTT up which
4003                  * makes timely unusable unless we want to get shoved
4004                  * into a corner and just be backed off (the age
4005                  * old problem with delay based CC).
4006                  *
4007                  * On the other hand if it was a route change we
4008                  * would like to stay somewhat contained and not
4009                  * blow out the buffers.
4010                  */
4011                 rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4012                                 __LINE__, 3);
4013                 rack->r_ctl.last_gp_comp_bw = cur_bw;
4014                 if (rack->rc_gp_bwred == 0) {
4015                         /* Go into reduction counting */
4016                         rack->rc_gp_bwred = 1;
4017                         rack->rc_gp_timely_dec_cnt = 0;
4018                 }
4019                 if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) ||
4020                     (timely_says == 0)) {
4021                         /*
4022                          * Push another time with a faster pacing
4023                          * to try to gain back (we include override to
4024                          * get a full raise factor).
4025                          */
4026                         if ((rack->rc_gp_saw_ca && rack->r_ctl.rack_per_of_gp_ca <= rack_down_raise_thresh) ||
4027                             (rack->rc_gp_saw_ss && rack->r_ctl.rack_per_of_gp_ss <= rack_down_raise_thresh) ||
4028                             (timely_says == 0) ||
4029                             (rack_down_raise_thresh == 0)) {
4030                                 /*
4031                                  * Do an override up in b/w if we were
4032                                  * below the threshold or if the threshold
4033                                  * is zero we always do the raise.
4034                                  */
4035                                 rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 1);
4036                         } else {
4037                                 /* Log it stays the same */
4038                                 rack_log_timely(rack,  0, last_bw_est, low_bnd, 0,
4039                                                 __LINE__, 11);
4040                         }
4041                         rack->rc_gp_timely_dec_cnt++;
4042                         /* We are not incrementing really no-count */
4043                         rack->rc_gp_incr = 0;
4044                         rack->rc_gp_timely_inc_cnt = 0;
4045                 } else {
4046                         /*
4047                          * Lets just use the RTT
4048                          * information and give up
4049                          * pushing.
4050                          */
4051                         goto use_timely;
4052                 }
4053         } else if ((timely_says != 2) &&
4054                     !losses &&
4055                     (last_bw_est > up_bnd)) {
4056                 /*
4057                  * We are increasing b/w lets keep going, updating
4058                  * our b/w and ignoring any timely input, unless
4059                  * of course we are at our max raise (if there is one).
4060                  */
4061
4062                 rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4063                                 __LINE__, 3);
4064                 rack->r_ctl.last_gp_comp_bw = cur_bw;
4065                 if (rack->rc_gp_saw_ss &&
4066                     rack_per_upper_bound_ss &&
4067                      (rack->r_ctl.rack_per_of_gp_ss == rack_per_upper_bound_ss)) {
4068                             /*
4069                              * In cases where we can't go higher
4070                              * we should just use timely.
4071                              */
4072                             goto use_timely;
4073                 }
4074                 if (rack->rc_gp_saw_ca &&
4075                     rack_per_upper_bound_ca &&
4076                     (rack->r_ctl.rack_per_of_gp_ca == rack_per_upper_bound_ca)) {
4077                             /*
4078                              * In cases where we can't go higher
4079                              * we should just use timely.
4080                              */
4081                             goto use_timely;
4082                 }
4083                 rack->rc_gp_bwred = 0;
4084                 rack->rc_gp_timely_dec_cnt = 0;
4085                 /* You get a set number of pushes if timely is trying to reduce */
4086                 if ((rack->rc_gp_incr < rack_timely_max_push_rise) || (timely_says == 0)) {
4087                         rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4088                 } else {
4089                         /* Log it stays the same */
4090                         rack_log_timely(rack,  0, last_bw_est, up_bnd, 0,
4091                             __LINE__, 12);
4092                 }
4093                 return;
4094         } else {
4095                 /*
4096                  * We are staying between the lower and upper range bounds
4097                  * so use timely to decide.
4098                  */
4099                 rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4100                                 __LINE__, 3);
4101 use_timely:
4102                 if (timely_says) {
4103                         rack->rc_gp_incr = 0;
4104                         rack->rc_gp_timely_inc_cnt = 0;
4105                         if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) &&
4106                             !losses &&
4107                             (last_bw_est < low_bnd)) {
4108                                 /* We are loosing ground */
4109                                 rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4110                                 rack->rc_gp_timely_dec_cnt++;
4111                                 /* We are not incrementing really no-count */
4112                                 rack->rc_gp_incr = 0;
4113                                 rack->rc_gp_timely_inc_cnt = 0;
4114                         } else
4115                                 rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
4116                 } else {
4117                         rack->rc_gp_bwred = 0;
4118                         rack->rc_gp_timely_dec_cnt = 0;
4119                         rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4120                 }
4121         }
4122 }
4123
4124 static int32_t
4125 rack_make_timely_judgement(struct tcp_rack *rack, uint32_t rtt, int32_t rtt_diff, uint32_t prev_rtt)
4126 {
4127         int32_t timely_says;
4128         uint64_t log_mult, log_rtt_a_diff;
4129
4130         log_rtt_a_diff = rtt;
4131         log_rtt_a_diff <<= 32;
4132         log_rtt_a_diff |= (uint32_t)rtt_diff;
4133         if (rtt >= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) *
4134                     rack_gp_rtt_maxmul)) {
4135                 /* Reduce the b/w multipler */
4136                 timely_says = 2;
4137                 log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
4138                 log_mult <<= 32;
4139                 log_mult |= prev_rtt;
4140                 rack_log_timely(rack,  timely_says, log_mult,
4141                                 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4142                                 log_rtt_a_diff, __LINE__, 4);
4143         } else if (rtt <= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4144                            ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4145                             max(rack_gp_rtt_mindiv , 1)))) {
4146                 /* Increase the b/w multipler */
4147                 log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4148                         ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4149                          max(rack_gp_rtt_mindiv , 1));
4150                 log_mult <<= 32;
4151                 log_mult |= prev_rtt;
4152                 timely_says = 0;
4153                 rack_log_timely(rack,  timely_says, log_mult ,
4154                                 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4155                                 log_rtt_a_diff, __LINE__, 5);
4156         } else {
4157                 /*
4158                  * Use a gradient to find it the timely gradient
4159                  * is:
4160                  * grad = rc_rtt_diff / min_rtt;
4161                  *
4162                  * anything below or equal to 0 will be
4163                  * a increase indication. Anything above
4164                  * zero is a decrease. Note we take care
4165                  * of the actual gradient calculation
4166                  * in the reduction (its not needed for
4167                  * increase).
4168                  */
4169                 log_mult = prev_rtt;
4170                 if (rtt_diff <= 0) {
4171                         /*
4172                          * Rttdiff is less than zero, increase the
4173                          * b/w multipler (its 0 or negative)
4174                          */
4175                         timely_says = 0;
4176                         rack_log_timely(rack,  timely_says, log_mult,
4177                                         get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 6);
4178                 } else {
4179                         /* Reduce the b/w multipler */
4180                         timely_says = 1;
4181                         rack_log_timely(rack,  timely_says, log_mult,
4182                                         get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 7);
4183                 }
4184         }
4185         return (timely_says);
4186 }
4187
4188 static void
4189 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
4190                             tcp_seq th_ack, int line)
4191 {
4192         uint64_t tim, bytes_ps, ltim, stim, utim;
4193         uint32_t segsiz, bytes, reqbytes, us_cts;
4194         int32_t gput, new_rtt_diff, timely_says;
4195         uint64_t  resid_bw, subpart = 0, addpart = 0, srtt;
4196         int did_add = 0;
4197
4198         us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4199         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
4200         if (TSTMP_GEQ(us_cts, tp->gput_ts))
4201                 tim = us_cts - tp->gput_ts;
4202         else
4203                 tim = 0;
4204
4205         if (rack->r_ctl.rc_gp_cumack_ts > rack->r_ctl.rc_gp_output_ts)
4206                 stim = rack->r_ctl.rc_gp_cumack_ts - rack->r_ctl.rc_gp_output_ts;
4207         else
4208                 stim = 0;
4209         /*
4210          * Use the larger of the send time or ack time. This prevents us
4211          * from being influenced by ack artifacts to come up with too
4212          * high of measurement. Note that since we are spanning over many more
4213          * bytes in most of our measurements hopefully that is less likely to
4214          * occur.
4215          */
4216         if (tim > stim)
4217                 utim = max(tim, 1);
4218         else
4219                 utim = max(stim, 1);
4220         /* Lets get a msec time ltim too for the old stuff */
4221         ltim = max(1, (utim / HPTS_USEC_IN_MSEC));
4222         gput = (((uint64_t) (th_ack - tp->gput_seq)) << 3) / ltim;
4223         reqbytes = min(rc_init_window(rack), (MIN_GP_WIN * segsiz));
4224         if ((tim == 0) && (stim == 0)) {
4225                 /*
4226                  * Invalid measurement time, maybe
4227                  * all on one ack/one send?
4228                  */
4229                 bytes = 0;
4230                 bytes_ps = 0;
4231                 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4232                                            0, 0, 0, 10, __LINE__, NULL);
4233                 goto skip_measurement;
4234         }
4235         if (rack->r_ctl.rc_gp_lowrtt == 0xffffffff) {
4236                 /* We never made a us_rtt measurement? */
4237                 bytes = 0;
4238                 bytes_ps = 0;
4239                 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4240                                            0, 0, 0, 10, __LINE__, NULL);
4241                 goto skip_measurement;
4242         }
4243         /*
4244          * Calculate the maximum possible b/w this connection
4245          * could have. We base our calculation on the lowest
4246          * rtt we have seen during the measurement and the
4247          * largest rwnd the client has given us in that time. This
4248          * forms a BDP that is the maximum that we could ever
4249          * get to the client. Anything larger is not valid.
4250          *
4251          * I originally had code here that rejected measurements
4252          * where the time was less than 1/2 the latest us_rtt.
4253          * But after thinking on that I realized its wrong since
4254          * say you had a 150Mbps or even 1Gbps link, and you
4255          * were a long way away.. example I am in Europe (100ms rtt)
4256          * talking to my 1Gbps link in S.C. Now measuring say 150,000
4257          * bytes my time would be 1.2ms, and yet my rtt would say
4258          * the measurement was invalid the time was < 50ms. The
4259          * same thing is true for 150Mb (8ms of time).
4260          *
4261          * A better way I realized is to look at what the maximum
4262          * the connection could possibly do. This is gated on
4263          * the lowest RTT we have seen and the highest rwnd.
4264          * We should in theory never exceed that, if we are
4265          * then something on the path is storing up packets
4266          * and then feeding them all at once to our endpoint
4267          * messing up our measurement.
4268          */
4269         rack->r_ctl.last_max_bw = rack->r_ctl.rc_gp_high_rwnd;
4270         rack->r_ctl.last_max_bw *= HPTS_USEC_IN_SEC;
4271         rack->r_ctl.last_max_bw /= rack->r_ctl.rc_gp_lowrtt;
4272         if (SEQ_LT(th_ack, tp->gput_seq)) {
4273                 /* No measurement can be made */
4274                 bytes = 0;
4275                 bytes_ps = 0;
4276                 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4277                                            0, 0, 0, 10, __LINE__, NULL);
4278                 goto skip_measurement;
4279         } else
4280                 bytes = (th_ack - tp->gput_seq);
4281         bytes_ps = (uint64_t)bytes;
4282         /*
4283          * Don't measure a b/w for pacing unless we have gotten at least
4284          * an initial windows worth of data in this measurement interval.
4285          *
4286          * Small numbers of bytes get badly influenced by delayed ack and
4287          * other artifacts. Note we take the initial window or our
4288          * defined minimum GP (defaulting to 10 which hopefully is the
4289          * IW).
4290          */
4291         if (rack->rc_gp_filled == 0) {
4292                 /*
4293                  * The initial estimate is special. We
4294                  * have blasted out an IW worth of packets
4295                  * without a real valid ack ts results. We
4296                  * then setup the app_limited_needs_set flag,
4297                  * this should get the first ack in (probably 2
4298                  * MSS worth) to be recorded as the timestamp.
4299                  * We thus allow a smaller number of bytes i.e.
4300                  * IW - 2MSS.
4301                  */
4302                 reqbytes -= (2 * segsiz);
4303                 /* Also lets fill previous for our first measurement to be neutral */
4304                 rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
4305         }
4306         if ((bytes_ps < reqbytes) || rack->app_limited_needs_set) {
4307                 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4308                                            rack->r_ctl.rc_app_limited_cnt,
4309                                            0, 0, 10, __LINE__, NULL);
4310                 goto skip_measurement;
4311         }
4312         /*
4313          * We now need to calculate the Timely like status so
4314          * we can update (possibly) the b/w multipliers.
4315          */
4316         new_rtt_diff = (int32_t)rack->r_ctl.rc_gp_srtt - (int32_t)rack->r_ctl.rc_prev_gp_srtt;
4317         if (rack->rc_gp_filled == 0) {
4318                 /* No previous reading */
4319                 rack->r_ctl.rc_rtt_diff = new_rtt_diff;
4320         } else {
4321                 if (rack->measure_saw_probe_rtt == 0) {
4322                         /*
4323                          * We don't want a probertt to be counted
4324                          * since it will be negative incorrectly. We
4325                          * expect to be reducing the RTT when we
4326                          * pace at a slower rate.
4327                          */
4328                         rack->r_ctl.rc_rtt_diff -= (rack->r_ctl.rc_rtt_diff / 8);
4329                         rack->r_ctl.rc_rtt_diff += (new_rtt_diff / 8);
4330                 }
4331         }
4332         timely_says = rack_make_timely_judgement(rack,
4333                 rack->r_ctl.rc_gp_srtt,
4334                 rack->r_ctl.rc_rtt_diff,
4335                 rack->r_ctl.rc_prev_gp_srtt
4336                 );
4337         bytes_ps *= HPTS_USEC_IN_SEC;
4338         bytes_ps /= utim;
4339         if (bytes_ps > rack->r_ctl.last_max_bw) {
4340                 /*
4341                  * Something is on path playing
4342                  * since this b/w is not possible based
4343                  * on our BDP (highest rwnd and lowest rtt
4344                  * we saw in the measurement window).
4345                  *
4346                  * Another option here would be to
4347                  * instead skip the measurement.
4348                  */
4349                 rack_log_pacing_delay_calc(rack, bytes, reqbytes,
4350                                            bytes_ps, rack->r_ctl.last_max_bw, 0,
4351                                            11, __LINE__, NULL);
4352                 bytes_ps = rack->r_ctl.last_max_bw;
4353         }
4354         /* We store gp for b/w in bytes per second */
4355         if (rack->rc_gp_filled == 0) {
4356                 /* Initial measurment */
4357                 if (bytes_ps) {
4358                         rack->r_ctl.gp_bw = bytes_ps;
4359                         rack->rc_gp_filled = 1;
4360                         rack->r_ctl.num_measurements = 1;
4361                         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
4362                 } else {
4363                         rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4364                                                    rack->r_ctl.rc_app_limited_cnt,
4365                                                    0, 0, 10, __LINE__, NULL);
4366                 }
4367                 if (rack->rc_inp->inp_in_hpts &&
4368                     (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
4369                         /*
4370                          * Ok we can't trust the pacer in this case
4371                          * where we transition from un-paced to paced.
4372                          * Or for that matter when the burst mitigation
4373                          * was making a wild guess and got it wrong.
4374                          * Stop the pacer and clear up all the aggregate
4375                          * delays etc.
4376                          */
4377                         tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
4378                         rack->r_ctl.rc_hpts_flags = 0;
4379                         rack->r_ctl.rc_last_output_to = 0;
4380                 }
4381                 did_add = 2;
4382         } else if (rack->r_ctl.num_measurements < RACK_REQ_AVG) {
4383                 /* Still a small number run an average */
4384                 rack->r_ctl.gp_bw += bytes_ps;
4385                 addpart = rack->r_ctl.num_measurements;
4386                 rack->r_ctl.num_measurements++;
4387                 if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
4388                         /* We have collected enought to move forward */
4389                         rack->r_ctl.gp_bw /= (uint64_t)rack->r_ctl.num_measurements;
4390                 }
4391                 did_add = 3;
4392         } else {
4393                 /*
4394                  * We want to take 1/wma of the goodput and add in to 7/8th
4395                  * of the old value weighted by the srtt. So if your measurement
4396                  * period is say 2 SRTT's long you would get 1/4 as the
4397                  * value, if it was like 1/2 SRTT then you would get 1/16th.
4398                  *
4399                  * But we must be careful not to take too much i.e. if the
4400                  * srtt is say 20ms and the measurement is taken over
4401                  * 400ms our weight would be 400/20 i.e. 20. On the
4402                  * other hand if we get a measurement over 1ms with a
4403                  * 10ms rtt we only want to take a much smaller portion.
4404                  */
4405                 if (rack->r_ctl.num_measurements < 0xff) {
4406                         rack->r_ctl.num_measurements++;
4407                 }
4408                 srtt = (uint64_t)tp->t_srtt;
4409                 if (srtt == 0) {
4410                         /*
4411                          * Strange why did t_srtt go back to zero?
4412                          */
4413                         if (rack->r_ctl.rc_rack_min_rtt)
4414                                 srtt = rack->r_ctl.rc_rack_min_rtt;
4415                         else
4416                                 srtt = HPTS_USEC_IN_MSEC;
4417                 }
4418                 /*
4419                  * XXXrrs: Note for reviewers, in playing with
4420                  * dynamic pacing I discovered this GP calculation
4421                  * as done originally leads to some undesired results.
4422                  * Basically you can get longer measurements contributing
4423                  * too much to the WMA. Thus I changed it if you are doing
4424                  * dynamic adjustments to only do the aportioned adjustment
4425                  * if we have a very small (time wise) measurement. Longer
4426                  * measurements just get there weight (defaulting to 1/8)
4427                  * add to the WMA. We may want to think about changing
4428                  * this to always do that for both sides i.e. dynamic
4429                  * and non-dynamic... but considering lots of folks
4430                  * were playing with this I did not want to change the
4431                  * calculation per.se. without your thoughts.. Lawerence?
4432                  * Peter??
4433                  */
4434                 if (rack->rc_gp_dyn_mul == 0) {
4435                         subpart = rack->r_ctl.gp_bw * utim;
4436                         subpart /= (srtt * 8);
4437                         if (subpart < (rack->r_ctl.gp_bw / 2)) {
4438                                 /*
4439                                  * The b/w update takes no more
4440                                  * away then 1/2 our running total
4441                                  * so factor it in.
4442                                  */
4443                                 addpart = bytes_ps * utim;
4444                                 addpart /= (srtt * 8);
4445                         } else {
4446                                 /*
4447                                  * Don't allow a single measurement
4448                                  * to account for more than 1/2 of the
4449                                  * WMA. This could happen on a retransmission
4450                                  * where utim becomes huge compared to
4451                                  * srtt (multiple retransmissions when using
4452                                  * the sending rate which factors in all the
4453                                  * transmissions from the first one).
4454                                  */
4455                                 subpart = rack->r_ctl.gp_bw / 2;
4456                                 addpart = bytes_ps / 2;
4457                         }
4458                         resid_bw = rack->r_ctl.gp_bw - subpart;
4459                         rack->r_ctl.gp_bw = resid_bw + addpart;
4460                         did_add = 1;
4461                 } else {
4462                         if ((utim / srtt) <= 1) {
4463                                 /*
4464                                  * The b/w update was over a small period
4465                                  * of time. The idea here is to prevent a small
4466                                  * measurement time period from counting
4467                                  * too much. So we scale it based on the
4468                                  * time so it attributes less than 1/rack_wma_divisor
4469                                  * of its measurement.
4470                                  */
4471                                 subpart = rack->r_ctl.gp_bw * utim;
4472                                 subpart /= (srtt * rack_wma_divisor);
4473                                 addpart = bytes_ps * utim;
4474                                 addpart /= (srtt * rack_wma_divisor);
4475                         } else {
4476                                 /*
4477                                  * The scaled measurement was long
4478                                  * enough so lets just add in the
4479                                  * portion of the measurment i.e. 1/rack_wma_divisor
4480                                  */
4481                                 subpart = rack->r_ctl.gp_bw / rack_wma_divisor;
4482                                 addpart = bytes_ps / rack_wma_divisor;
4483                         }
4484                         if ((rack->measure_saw_probe_rtt == 0) ||
4485                             (bytes_ps > rack->r_ctl.gp_bw)) {
4486                                 /*
4487                                  * For probe-rtt we only add it in
4488                                  * if its larger, all others we just
4489                                  * add in.
4490                                  */
4491                                 did_add = 1;
4492                                 resid_bw = rack->r_ctl.gp_bw - subpart;
4493                                 rack->r_ctl.gp_bw = resid_bw + addpart;
4494                         }
4495                 }
4496         }
4497         if ((rack->gp_ready == 0) &&
4498             (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
4499                 /* We have enough measurements now */
4500                 rack->gp_ready = 1;
4501                 rack_set_cc_pacing(rack);
4502                 if (rack->defer_options)
4503                         rack_apply_deferred_options(rack);
4504         }
4505         rack_log_pacing_delay_calc(rack, subpart, addpart, bytes_ps, stim,
4506                                    rack_get_bw(rack), 22, did_add, NULL);
4507         /* We do not update any multipliers if we are in or have seen a probe-rtt */
4508         if ((rack->measure_saw_probe_rtt == 0) && rack->rc_gp_rtt_set)
4509                 rack_update_multiplier(rack, timely_says, bytes_ps,
4510                                        rack->r_ctl.rc_gp_srtt,
4511                                        rack->r_ctl.rc_rtt_diff);
4512         rack_log_pacing_delay_calc(rack, bytes, tim, bytes_ps, stim,
4513                                    rack_get_bw(rack), 3, line, NULL);
4514         /* reset the gp srtt and setup the new prev */
4515         rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
4516         /* Record the lost count for the next measurement */
4517         rack->r_ctl.rc_loss_at_start = rack->r_ctl.rc_loss_count;
4518         /*
4519          * We restart our diffs based on the gpsrtt in the
4520          * measurement window.
4521          */
4522         rack->rc_gp_rtt_set = 0;
4523         rack->rc_gp_saw_rec = 0;
4524         rack->rc_gp_saw_ca = 0;
4525         rack->rc_gp_saw_ss = 0;
4526         rack->rc_dragged_bottom = 0;
4527 skip_measurement:
4528
4529 #ifdef STATS
4530         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT,
4531                                  gput);
4532         /*
4533          * XXXLAS: This is a temporary hack, and should be
4534          * chained off VOI_TCP_GPUT when stats(9) grows an
4535          * API to deal with chained VOIs.
4536          */
4537         if (tp->t_stats_gput_prev > 0)
4538                 stats_voi_update_abs_s32(tp->t_stats,
4539                                          VOI_TCP_GPUT_ND,
4540                                          ((gput - tp->t_stats_gput_prev) * 100) /
4541                                          tp->t_stats_gput_prev);
4542 #endif
4543         tp->t_flags &= ~TF_GPUTINPROG;
4544         tp->t_stats_gput_prev = gput;
4545         /*
4546          * Now are we app limited now and there is space from where we
4547          * were to where we want to go?
4548          *
4549          * We don't do the other case i.e. non-applimited here since
4550          * the next send will trigger us picking up the missing data.
4551          */
4552         if (rack->r_ctl.rc_first_appl &&
4553             TCPS_HAVEESTABLISHED(tp->t_state) &&
4554             rack->r_ctl.rc_app_limited_cnt &&
4555             (SEQ_GT(rack->r_ctl.rc_first_appl->r_start, th_ack)) &&
4556             ((rack->r_ctl.rc_first_appl->r_start - th_ack) >
4557              max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
4558                 /*
4559                  * Yep there is enough outstanding to make a measurement here.
4560                  */
4561                 struct rack_sendmap *rsm, fe;
4562
4563                 tp->t_flags |= TF_GPUTINPROG;
4564                 rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
4565                 rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
4566                 tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4567                 rack->app_limited_needs_set = 0;
4568                 tp->gput_seq = th_ack;
4569                 if (rack->in_probe_rtt)
4570                         rack->measure_saw_probe_rtt = 1;
4571                 else if ((rack->measure_saw_probe_rtt) &&
4572                          (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
4573                         rack->measure_saw_probe_rtt = 0;
4574                 if ((rack->r_ctl.rc_first_appl->r_start - th_ack) >= rack_get_measure_window(tp, rack)) {
4575                         /* There is a full window to gain info from */
4576                         tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
4577                 } else {
4578                         /* We can only measure up to the applimited point */
4579                         tp->gput_ack = tp->gput_seq + (rack->r_ctl.rc_first_appl->r_start - th_ack);
4580                 }
4581                 /*
4582                  * Now we need to find the timestamp of the send at tp->gput_seq
4583                  * for the send based measurement.
4584                  */
4585                 fe.r_start = tp->gput_seq;
4586                 rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
4587                 if (rsm) {
4588                         /* Ok send-based limit is set */
4589                         if (SEQ_LT(rsm->r_start, tp->gput_seq)) {
4590                                 /*
4591                                  * Move back to include the earlier part
4592                                  * so our ack time lines up right (this may
4593                                  * make an overlapping measurement but thats
4594                                  * ok).
4595                                  */
4596                                 tp->gput_seq = rsm->r_start;
4597                         }
4598                         if (rsm->r_flags & RACK_ACKED)
4599                                 tp->gput_ts = (uint32_t)rsm->r_ack_arrival;
4600                         else
4601                                 rack->app_limited_needs_set = 1;
4602                         rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
4603                 } else {
4604                         /*
4605                          * If we don't find the rsm due to some
4606                          * send-limit set the current time, which
4607                          * basically disables the send-limit.
4608                          */
4609                         struct timeval tv;
4610
4611                         microuptime(&tv);
4612                         rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
4613                 }
4614                 rack_log_pacing_delay_calc(rack,
4615                                            tp->gput_seq,
4616                                            tp->gput_ack,
4617                                            (uint64_t)rsm,
4618                                            tp->gput_ts,
4619                                            rack->r_ctl.rc_app_limited_cnt,
4620                                            9,
4621                                            __LINE__, NULL);
4622         }
4623 }
4624
4625 /*
4626  * CC wrapper hook functions
4627  */
4628 static void
4629 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, uint32_t th_ack, uint16_t nsegs,
4630     uint16_t type, int32_t recovery)
4631 {
4632         uint32_t prior_cwnd, acked;
4633         struct tcp_log_buffer *lgb = NULL;
4634         uint8_t labc_to_use;
4635
4636         INP_WLOCK_ASSERT(tp->t_inpcb);
4637         tp->ccv->nsegs = nsegs;
4638         acked = tp->ccv->bytes_this_ack = (th_ack - tp->snd_una);
4639         if ((recovery) && (rack->r_ctl.rc_early_recovery_segs)) {
4640                 uint32_t max;
4641
4642                 max = rack->r_ctl.rc_early_recovery_segs * ctf_fixed_maxseg(tp);
4643                 if (tp->ccv->bytes_this_ack > max) {
4644                         tp->ccv->bytes_this_ack = max;
4645                 }
4646         }
4647 #ifdef STATS
4648         stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF,
4649             ((int32_t)rack->r_ctl.cwnd_to_use) - tp->snd_wnd);
4650 #endif
4651         if ((tp->t_flags & TF_GPUTINPROG) &&
4652             rack_enough_for_measurement(tp, rack, th_ack)) {
4653                 /* Measure the Goodput */
4654                 rack_do_goodput_measurement(tp, rack, th_ack, __LINE__);
4655 #ifdef NETFLIX_PEAKRATE
4656                 if ((type == CC_ACK) &&
4657                     (tp->t_maxpeakrate)) {
4658                         /*
4659                          * We update t_peakrate_thr. This gives us roughly
4660                          * one update per round trip time. Note
4661                          * it will only be used if pace_always is off i.e
4662                          * we don't do this for paced flows.
4663                          */
4664                         rack_update_peakrate_thr(tp);
4665                 }
4666 #endif
4667         }
4668         /* Which way our we limited, if not cwnd limited no advance in CA */
4669         if (tp->snd_cwnd <= tp->snd_wnd)
4670                 tp->ccv->flags |= CCF_CWND_LIMITED;
4671         else
4672                 tp->ccv->flags &= ~CCF_CWND_LIMITED;
4673         if (tp->snd_cwnd > tp->snd_ssthresh) {
4674                 tp->t_bytes_acked += min(tp->ccv->bytes_this_ack,
4675                          nsegs * V_tcp_abc_l_var * ctf_fixed_maxseg(tp));
4676                 /* For the setting of a window past use the actual scwnd we are using */
4677                 if (tp->t_bytes_acked >= rack->r_ctl.cwnd_to_use) {
4678                         tp->t_bytes_acked -= rack->r_ctl.cwnd_to_use;
4679                         tp->ccv->flags |= CCF_ABC_SENTAWND;
4680                 }
4681         } else {
4682                 tp->ccv->flags &= ~CCF_ABC_SENTAWND;
4683                 tp->t_bytes_acked = 0;
4684         }
4685         prior_cwnd = tp->snd_cwnd;
4686         if ((recovery == 0) || (rack_max_abc_post_recovery == 0) || rack->r_use_labc_for_rec ||
4687             (rack_client_low_buf && (rack->client_bufferlvl < rack_client_low_buf)))
4688                 labc_to_use = rack->rc_labc;
4689         else
4690                 labc_to_use = rack_max_abc_post_recovery;
4691         if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
4692                 union tcp_log_stackspecific log;
4693                 struct timeval tv;
4694
4695                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4696                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4697                 log.u_bbr.flex1 = th_ack;
4698                 log.u_bbr.flex2 = tp->ccv->flags;
4699                 log.u_bbr.flex3 = tp->ccv->bytes_this_ack;
4700                 log.u_bbr.flex4 = tp->ccv->nsegs;
4701                 log.u_bbr.flex5 = labc_to_use;
4702                 log.u_bbr.flex6 = prior_cwnd;
4703                 log.u_bbr.flex7 = V_tcp_do_newsack;
4704                 log.u_bbr.flex8 = 1;
4705                 lgb = tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
4706                                      0, &log, false, NULL, NULL, 0, &tv);
4707         }
4708         if (CC_ALGO(tp)->ack_received != NULL) {
4709                 /* XXXLAS: Find a way to live without this */
4710                 tp->ccv->curack = th_ack;
4711                 tp->ccv->labc = labc_to_use;
4712                 tp->ccv->flags |= CCF_USE_LOCAL_ABC;
4713                 CC_ALGO(tp)->ack_received(tp->ccv, type);
4714         }
4715         if (lgb) {
4716                 lgb->tlb_stackinfo.u_bbr.flex6 = tp->snd_cwnd;
4717         }
4718         if (rack->r_must_retran) {
4719                 if (SEQ_GEQ(th_ack, rack->r_ctl.rc_snd_max_at_rto)) {
4720                         /*
4721                          * We now are beyond the rxt point so lets disable
4722                          * the flag.
4723                          */
4724                         rack->r_ctl.rc_out_at_rto = 0;
4725                         rack->r_must_retran = 0;
4726                 } else if ((prior_cwnd + ctf_fixed_maxseg(tp)) <= tp->snd_cwnd) {
4727                         /*
4728                          * Only decrement the rc_out_at_rto if the cwnd advances
4729                          * at least a whole segment. Otherwise next time the peer
4730                          * acks, we won't be able to send this generaly happens
4731                          * when we are in Congestion Avoidance.
4732                          */
4733                         if (acked <= rack->r_ctl.rc_out_at_rto){
4734                                 rack->r_ctl.rc_out_at_rto -= acked;
4735                         } else {
4736                                 rack->r_ctl.rc_out_at_rto = 0;
4737                         }
4738                 }
4739         }
4740 #ifdef STATS
4741         stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, rack->r_ctl.cwnd_to_use);
4742 #endif
4743         if (rack->r_ctl.rc_rack_largest_cwnd < rack->r_ctl.cwnd_to_use) {
4744                 rack->r_ctl.rc_rack_largest_cwnd = rack->r_ctl.cwnd_to_use;
4745         }
4746 #ifdef NETFLIX_PEAKRATE
4747         /* we enforce max peak rate if it is set and we are not pacing */
4748         if ((rack->rc_always_pace == 0) &&
4749             tp->t_peakrate_thr &&
4750             (tp->snd_cwnd > tp->t_peakrate_thr)) {
4751                 tp->snd_cwnd = tp->t_peakrate_thr;
4752         }
4753 #endif
4754 }
4755
4756 static void
4757 tcp_rack_partialack(struct tcpcb *tp)
4758 {
4759         struct tcp_rack *rack;
4760
4761         rack = (struct tcp_rack *)tp->t_fb_ptr;
4762         INP_WLOCK_ASSERT(tp->t_inpcb);
4763         /*
4764          * If we are doing PRR and have enough
4765          * room to send <or> we are pacing and prr
4766          * is disabled we will want to see if we
4767          * can send data (by setting r_wanted_output to
4768          * true).
4769          */
4770         if ((rack->r_ctl.rc_prr_sndcnt > 0) ||
4771             rack->rack_no_prr)
4772                 rack->r_wanted_output = 1;
4773 }
4774
4775 static void
4776 rack_post_recovery(struct tcpcb *tp, uint32_t th_ack)
4777 {
4778         struct tcp_rack *rack;
4779         uint32_t orig_cwnd;
4780
4781         orig_cwnd = tp->snd_cwnd;
4782         INP_WLOCK_ASSERT(tp->t_inpcb);
4783         rack = (struct tcp_rack *)tp->t_fb_ptr;
4784         /* only alert CC if we alerted when we entered */
4785         if (CC_ALGO(tp)->post_recovery != NULL) {
4786                 tp->ccv->curack = th_ack;
4787                 CC_ALGO(tp)->post_recovery(tp->ccv);
4788                 if (tp->snd_cwnd < tp->snd_ssthresh) {
4789                         /*
4790                          * Rack has burst control and pacing
4791                          * so lets not set this any lower than
4792                          * snd_ssthresh per RFC-6582 (option 2).
4793                          */
4794                         tp->snd_cwnd = tp->snd_ssthresh;
4795                 }
4796         }
4797         if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
4798                 union tcp_log_stackspecific log;
4799                 struct timeval tv;
4800
4801                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4802                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4803                 log.u_bbr.flex1 = th_ack;
4804                 log.u_bbr.flex2 = tp->ccv->flags;
4805                 log.u_bbr.flex3 = tp->ccv->bytes_this_ack;
4806                 log.u_bbr.flex4 = tp->ccv->nsegs;
4807                 log.u_bbr.flex5 = V_tcp_abc_l_var;
4808                 log.u_bbr.flex6 = orig_cwnd;
4809                 log.u_bbr.flex7 = V_tcp_do_newsack;
4810                 log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
4811                 log.u_bbr.flex8 = 2;
4812                 tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
4813                                0, &log, false, NULL, NULL, 0, &tv);
4814         }
4815         if ((rack->rack_no_prr == 0) &&
4816             (rack->no_prr_addback == 0) &&
4817             (rack->r_ctl.rc_prr_sndcnt > 0)) {
4818                 /*
4819                  * Suck the next prr cnt back into cwnd, but
4820                  * only do that if we are not application limited.
4821                  */
4822                 if (ctf_outstanding(tp) <= sbavail(&(tp->t_inpcb->inp_socket->so_snd))) {
4823                         /*
4824                          * We are allowed to add back to the cwnd the amount we did
4825                          * not get out if:
4826                          * a) no_prr_addback is off.
4827                          * b) we are not app limited
4828                          * c) we are doing prr
4829                          * <and>
4830                          * d) it is bounded by rack_prr_addbackmax (if addback is 0, then none).
4831                          */
4832                         tp->snd_cwnd += min((ctf_fixed_maxseg(tp) * rack_prr_addbackmax),
4833                                             rack->r_ctl.rc_prr_sndcnt);
4834                 }
4835                 rack->r_ctl.rc_prr_sndcnt = 0;
4836                 rack_log_to_prr(rack, 1, 0);
4837         }
4838         rack_log_to_prr(rack, 14, orig_cwnd);
4839         tp->snd_recover = tp->snd_una;
4840         EXIT_RECOVERY(tp->t_flags);
4841 }
4842
4843 static void
4844 rack_cong_signal(struct tcpcb *tp, uint32_t type, uint32_t ack)
4845 {
4846         struct tcp_rack *rack;
4847         uint32_t ssthresh_enter, cwnd_enter, in_rec_at_entry, orig_cwnd;
4848
4849         INP_WLOCK_ASSERT(tp->t_inpcb);
4850 #ifdef STATS
4851         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type);
4852 #endif
4853         if (IN_RECOVERY(tp->t_flags) == 0) {
4854                 in_rec_at_entry = 0;
4855                 ssthresh_enter = tp->snd_ssthresh;
4856                 cwnd_enter = tp->snd_cwnd;
4857         } else
4858                 in_rec_at_entry = 1;
4859         rack = (struct tcp_rack *)tp->t_fb_ptr;
4860         switch (type) {
4861         case CC_NDUPACK:
4862                 tp->t_flags &= ~TF_WASFRECOVERY;
4863                 tp->t_flags &= ~TF_WASCRECOVERY;
4864                 if (!IN_FASTRECOVERY(tp->t_flags)) {
4865                         rack->r_ctl.rc_prr_delivered = 0;
4866                         rack->r_ctl.rc_prr_out = 0;
4867                         if (rack->rack_no_prr == 0) {
4868                                 rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
4869                                 rack_log_to_prr(rack, 2, in_rec_at_entry);
4870                         }
4871                         rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una;
4872                         tp->snd_recover = tp->snd_max;
4873                         if (tp->t_flags2 & TF2_ECN_PERMIT)
4874                                 tp->t_flags2 |= TF2_ECN_SND_CWR;
4875                 }
4876                 break;
4877         case CC_ECN:
4878                 if (!IN_CONGRECOVERY(tp->t_flags) ||
4879                     /*
4880                      * Allow ECN reaction on ACK to CWR, if
4881                      * that data segment was also CE marked.
4882                      */
4883                     SEQ_GEQ(ack, tp->snd_recover)) {
4884                         EXIT_CONGRECOVERY(tp->t_flags);
4885                         KMOD_TCPSTAT_INC(tcps_ecn_rcwnd);
4886                         tp->snd_recover = tp->snd_max + 1;
4887                         if (tp->t_flags2 & TF2_ECN_PERMIT)
4888                                 tp->t_flags2 |= TF2_ECN_SND_CWR;
4889                 }
4890                 break;
4891         case CC_RTO:
4892                 tp->t_dupacks = 0;
4893                 tp->t_bytes_acked = 0;
4894                 EXIT_RECOVERY(tp->t_flags);
4895                 tp->snd_ssthresh = max(2, min(tp->snd_wnd, rack->r_ctl.cwnd_to_use) / 2 /
4896                     ctf_fixed_maxseg(tp)) * ctf_fixed_maxseg(tp);
4897                 orig_cwnd = tp->snd_cwnd;
4898                 tp->snd_cwnd = ctf_fixed_maxseg(tp);
4899                 rack_log_to_prr(rack, 16, orig_cwnd);
4900                 if (tp->t_flags2 & TF2_ECN_PERMIT)
4901                         tp->t_flags2 |= TF2_ECN_SND_CWR;
4902                 break;
4903         case CC_RTO_ERR:
4904                 KMOD_TCPSTAT_INC(tcps_sndrexmitbad);
4905                 /* RTO was unnecessary, so reset everything. */
4906                 tp->snd_cwnd = tp->snd_cwnd_prev;
4907                 tp->snd_ssthresh = tp->snd_ssthresh_prev;
4908                 tp->snd_recover = tp->snd_recover_prev;
4909                 if (tp->t_flags & TF_WASFRECOVERY) {
4910                         ENTER_FASTRECOVERY(tp->t_flags);
4911                         tp->t_flags &= ~TF_WASFRECOVERY;
4912                 }
4913                 if (tp->t_flags & TF_WASCRECOVERY) {
4914                         ENTER_CONGRECOVERY(tp->t_flags);
4915                         tp->t_flags &= ~TF_WASCRECOVERY;
4916                 }
4917                 tp->snd_nxt = tp->snd_max;
4918                 tp->t_badrxtwin = 0;
4919                 break;
4920         }
4921         if ((CC_ALGO(tp)->cong_signal != NULL)  &&
4922             (type != CC_RTO)){
4923                 tp->ccv->curack = ack;
4924                 CC_ALGO(tp)->cong_signal(tp->ccv, type);
4925         }
4926         if ((in_rec_at_entry == 0) && IN_RECOVERY(tp->t_flags)) {
4927                 rack_log_to_prr(rack, 15, cwnd_enter);
4928                 rack->r_ctl.dsack_byte_cnt = 0;
4929                 rack->r_ctl.retran_during_recovery = 0;
4930                 rack->r_ctl.rc_cwnd_at_erec = cwnd_enter;
4931                 rack->r_ctl.rc_ssthresh_at_erec = ssthresh_enter;
4932                 rack->r_ent_rec_ns = 1;
4933         }
4934 }
4935
4936 static inline void
4937 rack_cc_after_idle(struct tcp_rack *rack, struct tcpcb *tp)
4938 {
4939         uint32_t i_cwnd;
4940
4941         INP_WLOCK_ASSERT(tp->t_inpcb);
4942
4943 #ifdef NETFLIX_STATS
4944         KMOD_TCPSTAT_INC(tcps_idle_restarts);
4945         if (tp->t_state == TCPS_ESTABLISHED)
4946                 KMOD_TCPSTAT_INC(tcps_idle_estrestarts);
4947 #endif
4948         if (CC_ALGO(tp)->after_idle != NULL)
4949                 CC_ALGO(tp)->after_idle(tp->ccv);
4950
4951         if (tp->snd_cwnd == 1)
4952                 i_cwnd = tp->t_maxseg;          /* SYN(-ACK) lost */
4953         else
4954                 i_cwnd = rc_init_window(rack);
4955
4956         /*
4957          * Being idle is no differnt than the initial window. If the cc
4958          * clamps it down below the initial window raise it to the initial
4959          * window.
4960          */
4961         if (tp->snd_cwnd < i_cwnd) {
4962                 tp->snd_cwnd = i_cwnd;
4963         }
4964 }
4965
4966 /*
4967  * Indicate whether this ack should be delayed.  We can delay the ack if
4968  * following conditions are met:
4969  *      - There is no delayed ack timer in progress.
4970  *      - Our last ack wasn't a 0-sized window. We never want to delay
4971  *        the ack that opens up a 0-sized window.
4972  *      - LRO wasn't used for this segment. We make sure by checking that the
4973  *        segment size is not larger than the MSS.
4974  *      - Delayed acks are enabled or this is a half-synchronized T/TCP
4975  *        connection.
4976  */
4977 #define DELAY_ACK(tp, tlen)                      \
4978         (((tp->t_flags & TF_RXWIN0SENT) == 0) && \
4979         ((tp->t_flags & TF_DELACK) == 0) &&      \
4980         (tlen <= tp->t_maxseg) &&                \
4981         (tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN)))
4982
4983 static struct rack_sendmap *
4984 rack_find_lowest_rsm(struct tcp_rack *rack)
4985 {
4986         struct rack_sendmap *rsm;
4987
4988         /*
4989          * Walk the time-order transmitted list looking for an rsm that is
4990          * not acked. This will be the one that was sent the longest time
4991          * ago that is still outstanding.
4992          */
4993         TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
4994                 if (rsm->r_flags & RACK_ACKED) {
4995                         continue;
4996                 }
4997                 goto finish;
4998         }
4999 finish:
5000         return (rsm);
5001 }
5002
5003 static struct rack_sendmap *
5004 rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm)
5005 {
5006         struct rack_sendmap *prsm;
5007
5008         /*
5009          * Walk the sequence order list backward until we hit and arrive at
5010          * the highest seq not acked. In theory when this is called it
5011          * should be the last segment (which it was not).
5012          */
5013         counter_u64_add(rack_find_high, 1);
5014         prsm = rsm;
5015         RB_FOREACH_REVERSE_FROM(prsm, rack_rb_tree_head, rsm) {
5016                 if (prsm->r_flags & (RACK_ACKED | RACK_HAS_FIN)) {
5017                         continue;
5018                 }
5019                 return (prsm);
5020         }
5021         return (NULL);
5022 }
5023
5024 static uint32_t
5025 rack_calc_thresh_rack(struct tcp_rack *rack, uint32_t srtt, uint32_t cts)
5026 {
5027         int32_t lro;
5028         uint32_t thresh;
5029
5030         /*
5031          * lro is the flag we use to determine if we have seen reordering.
5032          * If it gets set we have seen reordering. The reorder logic either
5033          * works in one of two ways:
5034          *
5035          * If reorder-fade is configured, then we track the last time we saw
5036          * re-ordering occur. If we reach the point where enough time as
5037          * passed we no longer consider reordering has occuring.
5038          *
5039          * Or if reorder-face is 0, then once we see reordering we consider
5040          * the connection to alway be subject to reordering and just set lro
5041          * to 1.
5042          *
5043          * In the end if lro is non-zero we add the extra time for
5044          * reordering in.
5045          */
5046         if (srtt == 0)
5047                 srtt = 1;
5048         if (rack->r_ctl.rc_reorder_ts) {
5049                 if (rack->r_ctl.rc_reorder_fade) {
5050                         if (SEQ_GEQ(cts, rack->r_ctl.rc_reorder_ts)) {
5051                                 lro = cts - rack->r_ctl.rc_reorder_ts;
5052                                 if (lro == 0) {
5053                                         /*
5054                                          * No time as passed since the last
5055                                          * reorder, mark it as reordering.
5056                                          */
5057                                         lro = 1;
5058                                 }
5059                         } else {
5060                                 /* Negative time? */
5061                                 lro = 0;
5062                         }
5063                         if (lro > rack->r_ctl.rc_reorder_fade) {
5064                                 /* Turn off reordering seen too */
5065                                 rack->r_ctl.rc_reorder_ts = 0;
5066                                 lro = 0;
5067                         }
5068                 } else {
5069                         /* Reodering does not fade */
5070                         lro = 1;
5071                 }
5072         } else {
5073                 lro = 0;
5074         }
5075         thresh = srtt + rack->r_ctl.rc_pkt_delay;
5076         if (lro) {
5077                 /* It must be set, if not you get 1/4 rtt */
5078                 if (rack->r_ctl.rc_reorder_shift)
5079                         thresh += (srtt >> rack->r_ctl.rc_reorder_shift);
5080                 else
5081                         thresh += (srtt >> 2);
5082         } else {
5083                 thresh += 1;
5084         }
5085         /* We don't let the rack timeout be above a RTO */
5086         if (thresh > rack->rc_tp->t_rxtcur) {
5087                 thresh = rack->rc_tp->t_rxtcur;
5088         }
5089         /* And we don't want it above the RTO max either */
5090         if (thresh > rack_rto_max) {
5091                 thresh = rack_rto_max;
5092         }
5093         return (thresh);
5094 }
5095
5096 static uint32_t
5097 rack_calc_thresh_tlp(struct tcpcb *tp, struct tcp_rack *rack,
5098                      struct rack_sendmap *rsm, uint32_t srtt)
5099 {
5100         struct rack_sendmap *prsm;
5101         uint32_t thresh, len;
5102         int segsiz;
5103
5104         if (srtt == 0)
5105                 srtt = 1;
5106         if (rack->r_ctl.rc_tlp_threshold)
5107                 thresh = srtt + (srtt / rack->r_ctl.rc_tlp_threshold);
5108         else
5109                 thresh = (srtt * 2);
5110
5111         /* Get the previous sent packet, if any */
5112         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
5113         counter_u64_add(rack_enter_tlp_calc, 1);
5114         len = rsm->r_end - rsm->r_start;
5115         if (rack->rack_tlp_threshold_use == TLP_USE_ID) {
5116                 /* Exactly like the ID */
5117                 if (((tp->snd_max - tp->snd_una) - rack->r_ctl.rc_sacked + rack->r_ctl.rc_holes_rxt) <= segsiz) {
5118                         uint32_t alt_thresh;
5119                         /*
5120                          * Compensate for delayed-ack with the d-ack time.
5121                          */
5122                         counter_u64_add(rack_used_tlpmethod, 1);
5123                         alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5124                         if (alt_thresh > thresh)
5125                                 thresh = alt_thresh;
5126                 }
5127         } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_ONE) {
5128                 /* 2.1 behavior */
5129                 prsm = TAILQ_PREV(rsm, rack_head, r_tnext);
5130                 if (prsm && (len <= segsiz)) {
5131                         /*
5132                          * Two packets outstanding, thresh should be (2*srtt) +
5133                          * possible inter-packet delay (if any).
5134                          */
5135                         uint32_t inter_gap = 0;
5136                         int idx, nidx;
5137
5138                         counter_u64_add(rack_used_tlpmethod, 1);
5139                         idx = rsm->r_rtr_cnt - 1;
5140                         nidx = prsm->r_rtr_cnt - 1;
5141                         if (rsm->r_tim_lastsent[nidx] >= prsm->r_tim_lastsent[idx]) {
5142                                 /* Yes it was sent later (or at the same time) */
5143                                 inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx];
5144                         }
5145                         thresh += inter_gap;
5146                 } else if (len <= segsiz) {
5147                         /*
5148                          * Possibly compensate for delayed-ack.
5149                          */
5150                         uint32_t alt_thresh;
5151
5152                         counter_u64_add(rack_used_tlpmethod2, 1);
5153                         alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5154                         if (alt_thresh > thresh)
5155                                 thresh = alt_thresh;
5156                 }
5157         } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_TWO) {
5158                 /* 2.2 behavior */
5159                 if (len <= segsiz) {
5160                         uint32_t alt_thresh;
5161                         /*
5162                          * Compensate for delayed-ack with the d-ack time.
5163                          */
5164                         counter_u64_add(rack_used_tlpmethod, 1);
5165                         alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5166                         if (alt_thresh > thresh)
5167                                 thresh = alt_thresh;
5168                 }
5169         }
5170         /* Not above an RTO */
5171         if (thresh > tp->t_rxtcur) {
5172                 thresh = tp->t_rxtcur;
5173         }
5174         /* Not above a RTO max */
5175         if (thresh > rack_rto_max) {
5176                 thresh = rack_rto_max;
5177         }
5178         /* Apply user supplied min TLP */
5179         if (thresh < rack_tlp_min) {
5180                 thresh = rack_tlp_min;
5181         }
5182         return (thresh);
5183 }
5184
5185 static uint32_t
5186 rack_grab_rtt(struct tcpcb *tp, struct tcp_rack *rack)
5187 {
5188         /*
5189          * We want the rack_rtt which is the
5190          * last rtt we measured. However if that
5191          * does not exist we fallback to the srtt (which
5192          * we probably will never do) and then as a last
5193          * resort we use RACK_INITIAL_RTO if no srtt is
5194          * yet set.
5195          */
5196         if (rack->rc_rack_rtt)
5197                 return (rack->rc_rack_rtt);
5198         else if (tp->t_srtt == 0)
5199                 return (RACK_INITIAL_RTO);
5200         return (tp->t_srtt);
5201 }
5202
5203 static struct rack_sendmap *
5204 rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused)
5205 {
5206         /*
5207          * Check to see that we don't need to fall into recovery. We will
5208          * need to do so if our oldest transmit is past the time we should
5209          * have had an ack.
5210          */
5211         struct tcp_rack *rack;
5212         struct rack_sendmap *rsm;
5213         int32_t idx;
5214         uint32_t srtt, thresh;
5215
5216         rack = (struct tcp_rack *)tp->t_fb_ptr;
5217         if (RB_EMPTY(&rack->r_ctl.rc_mtree)) {
5218                 return (NULL);
5219         }
5220         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5221         if (rsm == NULL)
5222                 return (NULL);
5223
5224         if (rsm->r_flags & RACK_ACKED) {
5225                 rsm = rack_find_lowest_rsm(rack);
5226                 if (rsm == NULL)
5227                         return (NULL);
5228         }
5229         idx = rsm->r_rtr_cnt - 1;
5230         srtt = rack_grab_rtt(tp, rack);
5231         thresh = rack_calc_thresh_rack(rack, srtt, tsused);
5232         if (TSTMP_LT(tsused, ((uint32_t)rsm->r_tim_lastsent[idx]))) {
5233                 return (NULL);
5234         }
5235         if ((tsused - ((uint32_t)rsm->r_tim_lastsent[idx])) < thresh) {
5236                 return (NULL);
5237         }
5238         /* Ok if we reach here we are over-due and this guy can be sent */
5239         if (IN_RECOVERY(tp->t_flags) == 0) {
5240                 /*
5241                  * For the one that enters us into recovery record undo
5242                  * info.
5243                  */
5244                 rack->r_ctl.rc_rsm_start = rsm->r_start;
5245                 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
5246                 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
5247         }
5248         rack_cong_signal(tp, CC_NDUPACK, tp->snd_una);
5249         return (rsm);
5250 }
5251
5252 static uint32_t
5253 rack_get_persists_timer_val(struct tcpcb *tp, struct tcp_rack *rack)
5254 {
5255         int32_t t;
5256         int32_t tt;
5257         uint32_t ret_val;
5258
5259         t = (tp->t_srtt + (tp->t_rttvar << 2));
5260         RACK_TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
5261             rack_persist_min, rack_persist_max, rack->r_ctl.timer_slop);
5262         if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
5263                 tp->t_rxtshift++;
5264         rack->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT;
5265         ret_val = (uint32_t)tt;
5266         return (ret_val);
5267 }
5268
5269 static uint32_t
5270 rack_timer_start(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int sup_rack)
5271 {
5272         /*
5273          * Start the FR timer, we do this based on getting the first one in
5274          * the rc_tmap. Note that if its NULL we must stop the timer. in all
5275          * events we need to stop the running timer (if its running) before
5276          * starting the new one.
5277          */
5278         uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse;
5279         uint32_t srtt_cur;
5280         int32_t idx;
5281         int32_t is_tlp_timer = 0;
5282         struct rack_sendmap *rsm;
5283
5284         if (rack->t_timers_stopped) {
5285                 /* All timers have been stopped none are to run */
5286                 return (0);
5287         }
5288         if (rack->rc_in_persist) {
5289                 /* We can't start any timer in persists */
5290                 return (rack_get_persists_timer_val(tp, rack));
5291         }
5292         rack->rc_on_min_to = 0;
5293         if ((tp->t_state < TCPS_ESTABLISHED) ||
5294             ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
5295                 goto activate_rxt;
5296         }
5297         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5298         if ((rsm == NULL) || sup_rack) {
5299                 /* Nothing on the send map or no rack */
5300 activate_rxt:
5301                 time_since_sent = 0;
5302                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5303                 if (rsm) {
5304                         /*
5305                          * Should we discount the RTX timer any?
5306                          *
5307                          * We want to discount it the smallest amount.
5308                          * If a timer (Rack/TLP or RXT) has gone off more
5309                          * recently thats the discount we want to use (now - timer time).
5310                          * If the retransmit of the oldest packet was more recent then
5311                          * we want to use that (now - oldest-packet-last_transmit_time).
5312                          *
5313                          */
5314                         idx = rsm->r_rtr_cnt - 1;
5315                         if (TSTMP_GEQ(rack->r_ctl.rc_tlp_rxt_last_time, ((uint32_t)rsm->r_tim_lastsent[idx])))
5316                                 tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
5317                         else
5318                                 tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
5319                         if (TSTMP_GT(cts, tstmp_touse))
5320                             time_since_sent = cts - tstmp_touse;
5321                 }
5322                 if (SEQ_LT(tp->snd_una, tp->snd_max) || sbavail(&(tp->t_inpcb->inp_socket->so_snd))) {
5323                         rack->r_ctl.rc_hpts_flags |= PACE_TMR_RXT;
5324                         to = tp->t_rxtcur;
5325                         if (to > time_since_sent)
5326                                 to -= time_since_sent;
5327                         else
5328                                 to = rack->r_ctl.rc_min_to;
5329                         if (to == 0)
5330                                 to = 1;
5331                         /* Special case for KEEPINIT */
5332                         if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
5333                             (TP_KEEPINIT(tp) != 0) &&
5334                             rsm) {
5335                                 /*
5336                                  * We have to put a ceiling on the rxt timer
5337                                  * of the keep-init timeout.
5338                                  */
5339                                 uint32_t max_time, red;
5340
5341                                 max_time = TICKS_2_USEC(TP_KEEPINIT(tp));
5342                                 if (TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) {
5343                                         red = (cts - (uint32_t)rsm->r_tim_lastsent[0]);
5344                                         if (red < max_time)
5345                                                 max_time -= red;
5346                                         else
5347                                                 max_time = 1;
5348                                 }
5349                                 /* Reduce timeout to the keep value if needed */
5350                                 if (max_time < to)
5351                                         to = max_time;
5352                         }
5353                         return (to);
5354                 }
5355                 return (0);
5356         }
5357         if (rsm->r_flags & RACK_ACKED) {
5358                 rsm = rack_find_lowest_rsm(rack);
5359                 if (rsm == NULL) {
5360                         /* No lowest? */
5361                         goto activate_rxt;
5362                 }
5363         }
5364         if (rack->sack_attack_disable) {
5365                 /*
5366                  * We don't want to do
5367                  * any TLP's if you are an attacker.
5368                  * Though if you are doing what
5369                  * is expected you may still have
5370                  * SACK-PASSED marks.
5371                  */
5372                 goto activate_rxt;
5373         }
5374         /* Convert from ms to usecs */
5375         if ((rsm->r_flags & RACK_SACK_PASSED) || (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
5376                 if ((tp->t_flags & TF_SENTFIN) &&
5377                     ((tp->snd_max - tp->snd_una) == 1) &&
5378                     (rsm->r_flags & RACK_HAS_FIN)) {
5379                         /*
5380                          * We don't start a rack timer if all we have is a
5381                          * FIN outstanding.
5382                          */
5383                         goto activate_rxt;
5384                 }
5385                 if ((rack->use_rack_rr == 0) &&
5386                     (IN_FASTRECOVERY(tp->t_flags)) &&
5387                     (rack->rack_no_prr == 0) &&
5388                      (rack->r_ctl.rc_prr_sndcnt  < ctf_fixed_maxseg(tp))) {
5389                         /*
5390                          * We are not cheating, in recovery  and
5391                          * not enough ack's to yet get our next
5392                          * retransmission out.
5393                          *
5394                          * Note that classified attackers do not
5395                          * get to use the rack-cheat.
5396                          */
5397                         goto activate_tlp;
5398                 }
5399                 srtt = rack_grab_rtt(tp, rack);
5400                 thresh = rack_calc_thresh_rack(rack, srtt, cts);
5401                 idx = rsm->r_rtr_cnt - 1;
5402                 exp = ((uint32_t)rsm->r_tim_lastsent[idx]) + thresh;
5403                 if (SEQ_GEQ(exp, cts)) {
5404                         to = exp - cts;
5405                         if (to < rack->r_ctl.rc_min_to) {
5406                                 to = rack->r_ctl.rc_min_to;
5407                                 if (rack->r_rr_config == 3)
5408                                         rack->rc_on_min_to = 1;
5409                         }
5410                 } else {
5411                         to = rack->r_ctl.rc_min_to;
5412                         if (rack->r_rr_config == 3)
5413                                 rack->rc_on_min_to = 1;
5414                 }
5415         } else {
5416                 /* Ok we need to do a TLP not RACK */
5417 activate_tlp:
5418                 if ((rack->rc_tlp_in_progress != 0) &&
5419                     (rack->r_ctl.rc_tlp_cnt_out >= rack_tlp_limit)) {
5420                         /*
5421                          * The previous send was a TLP and we have sent
5422                          * N TLP's without sending new data.
5423                          */
5424                         goto activate_rxt;
5425                 }
5426                 rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
5427                 if (rsm == NULL) {
5428                         /* We found no rsm to TLP with. */
5429                         goto activate_rxt;
5430                 }
5431                 if (rsm->r_flags & RACK_HAS_FIN) {
5432                         /* If its a FIN we dont do TLP */
5433                         rsm = NULL;
5434                         goto activate_rxt;
5435                 }
5436                 idx = rsm->r_rtr_cnt - 1;
5437                 time_since_sent = 0;
5438                 if (TSTMP_GEQ(((uint32_t)rsm->r_tim_lastsent[idx]), rack->r_ctl.rc_tlp_rxt_last_time))
5439                         tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
5440                 else
5441                         tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
5442                 if (TSTMP_GT(cts, tstmp_touse))
5443                     time_since_sent = cts - tstmp_touse;
5444                 is_tlp_timer = 1;
5445                 if (tp->t_srtt) {
5446                         if ((rack->rc_srtt_measure_made == 0) &&
5447                             (tp->t_srtt == 1)) {
5448                                 /*
5449                                  * If another stack as run and set srtt to 1,
5450                                  * then the srtt was 0, so lets use the initial.
5451                                  */
5452                                 srtt = RACK_INITIAL_RTO;
5453                         } else {
5454                                 srtt_cur = tp->t_srtt;
5455                                 srtt = srtt_cur;
5456                         }
5457                 } else
5458                         srtt = RACK_INITIAL_RTO;
5459                 /*
5460                  * If the SRTT is not keeping up and the
5461                  * rack RTT has spiked we want to use
5462                  * the last RTT not the smoothed one.
5463                  */
5464                 if (rack_tlp_use_greater &&
5465                     tp->t_srtt &&
5466                     (srtt < rack_grab_rtt(tp, rack))) {
5467                         srtt = rack_grab_rtt(tp, rack);
5468                 }
5469                 thresh = rack_calc_thresh_tlp(tp, rack, rsm, srtt);
5470                 if (thresh > time_since_sent) {
5471                         to = thresh - time_since_sent;
5472                 } else {
5473                         to = rack->r_ctl.rc_min_to;
5474                         rack_log_alt_to_to_cancel(rack,
5475                                                   thresh,               /* flex1 */
5476                                                   time_since_sent,      /* flex2 */
5477                                                   tstmp_touse,          /* flex3 */
5478                                                   rack->r_ctl.rc_tlp_rxt_last_time, /* flex4 */
5479                                                   (uint32_t)rsm->r_tim_lastsent[idx],
5480                                                   srtt,
5481                                                   idx, 99);
5482                 }
5483                 if (to < rack_tlp_min) {
5484                         to = rack_tlp_min;
5485                 }
5486                 if (to > TICKS_2_USEC(TCPTV_REXMTMAX)) {
5487                         /*
5488                          * If the TLP time works out to larger than the max
5489                          * RTO lets not do TLP.. just RTO.
5490                          */
5491                         goto activate_rxt;
5492                 }
5493         }
5494         if (is_tlp_timer == 0) {
5495                 rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK;
5496         } else {
5497                 rack->r_ctl.rc_hpts_flags |= PACE_TMR_TLP;
5498         }
5499         if (to == 0)
5500                 to = 1;
5501         return (to);
5502 }
5503
5504 static void
5505 rack_enter_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5506 {
5507         if (rack->rc_in_persist == 0) {
5508                 if (tp->t_flags & TF_GPUTINPROG) {
5509                         /*
5510                          * Stop the goodput now, the calling of the
5511                          * measurement function clears the flag.
5512                          */
5513                         rack_do_goodput_measurement(tp, rack, tp->snd_una, __LINE__);
5514                 }
5515 #ifdef NETFLIX_SHARED_CWND
5516                 if (rack->r_ctl.rc_scw) {
5517                         tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
5518                         rack->rack_scwnd_is_idle = 1;
5519                 }
5520 #endif
5521                 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
5522                 if (rack->r_ctl.rc_went_idle_time == 0)
5523                         rack->r_ctl.rc_went_idle_time = 1;
5524                 rack_timer_cancel(tp, rack, cts, __LINE__);
5525                 tp->t_rxtshift = 0;
5526                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
5527                               rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
5528                 rack->rc_in_persist = 1;
5529         }
5530 }
5531
5532 static void
5533 rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5534 {
5535         if (rack->rc_inp->inp_in_hpts) {
5536                 tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
5537                 rack->r_ctl.rc_hpts_flags = 0;
5538         }
5539 #ifdef NETFLIX_SHARED_CWND
5540         if (rack->r_ctl.rc_scw) {
5541                 tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
5542                 rack->rack_scwnd_is_idle = 0;
5543         }
5544 #endif
5545         if (rack->rc_gp_dyn_mul &&
5546             (rack->use_fixed_rate == 0) &&
5547             (rack->rc_always_pace)) {
5548                 /*
5549                  * Do we count this as if a probe-rtt just
5550                  * finished?
5551                  */
5552                 uint32_t time_idle, idle_min;
5553
5554                 time_idle = tcp_get_usecs(NULL) - rack->r_ctl.rc_went_idle_time;
5555                 idle_min = rack_min_probertt_hold;
5556                 if (rack_probertt_gpsrtt_cnt_div) {
5557                         uint64_t extra;
5558                         extra = (uint64_t)rack->r_ctl.rc_gp_srtt *
5559                                 (uint64_t)rack_probertt_gpsrtt_cnt_mul;
5560                         extra /= (uint64_t)rack_probertt_gpsrtt_cnt_div;
5561                         idle_min += (uint32_t)extra;
5562                 }
5563                 if (time_idle >= idle_min) {
5564                         /* Yes, we count it as a probe-rtt. */
5565                         uint32_t us_cts;
5566
5567                         us_cts = tcp_get_usecs(NULL);
5568                         if (rack->in_probe_rtt == 0) {
5569                                 rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
5570                                 rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
5571                                 rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
5572                                 rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
5573                         } else {
5574                                 rack_exit_probertt(rack, us_cts);
5575                         }
5576                 }
5577         }
5578         rack->rc_in_persist = 0;
5579         rack->r_ctl.rc_went_idle_time = 0;
5580         tp->t_rxtshift = 0;
5581         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
5582            rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
5583         rack->r_ctl.rc_agg_delayed = 0;
5584         rack->r_early = 0;
5585         rack->r_late = 0;
5586         rack->r_ctl.rc_agg_early = 0;
5587 }
5588
5589 static void
5590 rack_log_hpts_diag(struct tcp_rack *rack, uint32_t cts,
5591                    struct hpts_diag *diag, struct timeval *tv)
5592 {
5593         if (rack_verbose_logging && rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
5594                 union tcp_log_stackspecific log;
5595
5596                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5597                 log.u_bbr.flex1 = diag->p_nxt_slot;
5598                 log.u_bbr.flex2 = diag->p_cur_slot;
5599                 log.u_bbr.flex3 = diag->slot_req;
5600                 log.u_bbr.flex4 = diag->inp_hptsslot;
5601                 log.u_bbr.flex5 = diag->slot_remaining;
5602                 log.u_bbr.flex6 = diag->need_new_to;
5603                 log.u_bbr.flex7 = diag->p_hpts_active;
5604                 log.u_bbr.flex8 = diag->p_on_min_sleep;
5605                 /* Hijack other fields as needed */
5606                 log.u_bbr.epoch = diag->have_slept;
5607                 log.u_bbr.lt_epoch = diag->yet_to_sleep;
5608                 log.u_bbr.pkts_out = diag->co_ret;
5609                 log.u_bbr.applimited = diag->hpts_sleep_time;
5610                 log.u_bbr.delivered = diag->p_prev_slot;
5611                 log.u_bbr.inflight = diag->p_runningtick;
5612                 log.u_bbr.bw_inuse = diag->wheel_tick;
5613                 log.u_bbr.rttProp = diag->wheel_cts;
5614                 log.u_bbr.timeStamp = cts;
5615                 log.u_bbr.delRate = diag->maxticks;
5616                 log.u_bbr.cur_del_rate = diag->p_curtick;
5617                 log.u_bbr.cur_del_rate <<= 32;
5618                 log.u_bbr.cur_del_rate |= diag->p_lasttick;
5619                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
5620                     &rack->rc_inp->inp_socket->so_rcv,
5621                     &rack->rc_inp->inp_socket->so_snd,
5622                     BBR_LOG_HPTSDIAG, 0,
5623                     0, &log, false, tv);
5624         }
5625
5626 }
5627
5628 static void
5629 rack_log_wakeup(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb, uint32_t len, int type)
5630 {
5631         if (rack_verbose_logging && rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
5632                 union tcp_log_stackspecific log;
5633                 struct timeval tv;
5634
5635                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5636                 log.u_bbr.flex1 = sb->sb_flags;
5637                 log.u_bbr.flex2 = len;
5638                 log.u_bbr.flex3 = sb->sb_state;
5639                 log.u_bbr.flex8 = type;
5640                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5641                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
5642                     &rack->rc_inp->inp_socket->so_rcv,
5643                     &rack->rc_inp->inp_socket->so_snd,
5644                     TCP_LOG_SB_WAKE, 0,
5645                     len, &log, false, &tv);
5646         }
5647 }
5648
5649 static void
5650 rack_start_hpts_timer(struct tcp_rack *rack, struct tcpcb *tp, uint32_t cts,
5651       int32_t slot, uint32_t tot_len_this_send, int sup_rack)
5652 {
5653         struct hpts_diag diag;
5654         struct inpcb *inp;
5655         struct timeval tv;
5656         uint32_t delayed_ack = 0;
5657         uint32_t hpts_timeout;
5658         uint32_t entry_slot = slot;
5659         uint8_t stopped;
5660         uint32_t left = 0;
5661         uint32_t us_cts;
5662
5663         inp = tp->t_inpcb;
5664         if ((tp->t_state == TCPS_CLOSED) ||
5665             (tp->t_state == TCPS_LISTEN)) {
5666                 return;
5667         }
5668         if (inp->inp_in_hpts) {
5669                 /* Already on the pacer */
5670                 return;
5671         }
5672         stopped = rack->rc_tmr_stopped;
5673         if (stopped && TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
5674                 left = rack->r_ctl.rc_timer_exp - cts;
5675         }
5676         rack->r_ctl.rc_timer_exp = 0;
5677         rack->r_ctl.rc_hpts_flags = 0;
5678         us_cts = tcp_get_usecs(&tv);
5679         /* Now early/late accounting */
5680         rack_log_pacing_delay_calc(rack, entry_slot, slot, 0, 0, 0, 26, __LINE__, NULL);
5681         if (rack->r_early && (rack->rc_ack_can_sendout_data == 0)) {
5682                 /*
5683                  * We have a early carry over set,
5684                  * we can always add more time so we
5685                  * can always make this compensation.
5686                  *
5687                  * Note if ack's are allowed to wake us do not
5688                  * penalize the next timer for being awoke
5689                  * by an ack aka the rc_agg_early (non-paced mode).
5690                  */
5691                 slot += rack->r_ctl.rc_agg_early;
5692                 rack->r_early = 0;
5693                 rack->r_ctl.rc_agg_early = 0;
5694         }
5695         if (rack->r_late) {
5696                 /*
5697                  * This is harder, we can
5698                  * compensate some but it
5699                  * really depends on what
5700                  * the current pacing time is.
5701                  */
5702                 if (rack->r_ctl.rc_agg_delayed >= slot) {
5703                         /*
5704                          * We can't compensate for it all.
5705                          * And we have to have some time
5706                          * on the clock. We always have a min
5707                          * 10 slots (10 x 10 i.e. 100 usecs).
5708                          */
5709                         if (slot <= HPTS_TICKS_PER_USEC) {
5710                                 /* We gain delay */
5711                                 rack->r_ctl.rc_agg_delayed += (HPTS_TICKS_PER_USEC - slot);
5712                                 slot = HPTS_TICKS_PER_USEC;
5713                         } else {
5714                                 /* We take off some */
5715                                 rack->r_ctl.rc_agg_delayed -= (slot - HPTS_TICKS_PER_USEC);
5716                                 slot = HPTS_TICKS_PER_USEC;
5717                         }
5718                 } else {
5719                         slot -= rack->r_ctl.rc_agg_delayed;
5720                         rack->r_ctl.rc_agg_delayed = 0;
5721                         /* Make sure we have 100 useconds at minimum */
5722                         if (slot < HPTS_TICKS_PER_USEC) {
5723                                 rack->r_ctl.rc_agg_delayed = HPTS_TICKS_PER_USEC - slot;
5724                                 slot = HPTS_TICKS_PER_USEC;
5725                         }
5726                         if (rack->r_ctl.rc_agg_delayed == 0)
5727                                 rack->r_late = 0;
5728                 }
5729         }
5730         if (slot) {
5731                 /* We are pacing too */
5732                 rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT;
5733         }
5734         hpts_timeout = rack_timer_start(tp, rack, cts, sup_rack);
5735 #ifdef NETFLIX_EXP_DETECTION
5736         if (rack->sack_attack_disable &&
5737             (slot < tcp_sad_pacing_interval)) {
5738                 /*
5739                  * We have a potential attacker on
5740                  * the line. We have possibly some
5741                  * (or now) pacing time set. We want to
5742                  * slow down the processing of sacks by some
5743                  * amount (if it is an attacker). Set the default
5744                  * slot for attackers in place (unless the orginal
5745                  * interval is longer). Its stored in
5746                  * micro-seconds, so lets convert to msecs.
5747                  */
5748                 slot = tcp_sad_pacing_interval;
5749         }
5750 #endif
5751         if (tp->t_flags & TF_DELACK) {
5752                 delayed_ack = TICKS_2_USEC(tcp_delacktime);
5753                 rack->r_ctl.rc_hpts_flags |= PACE_TMR_DELACK;
5754         }
5755         if (delayed_ack && ((hpts_timeout == 0) ||
5756                             (delayed_ack < hpts_timeout)))
5757                 hpts_timeout = delayed_ack;
5758         else
5759                 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
5760         /*
5761          * If no timers are going to run and we will fall off the hptsi
5762          * wheel, we resort to a keep-alive timer if its configured.
5763          */
5764         if ((hpts_timeout == 0) &&
5765             (slot == 0)) {
5766                 if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
5767                     (tp->t_state <= TCPS_CLOSING)) {
5768                         /*
5769                          * Ok we have no timer (persists, rack, tlp, rxt  or
5770                          * del-ack), we don't have segments being paced. So
5771                          * all that is left is the keepalive timer.
5772                          */
5773                         if (TCPS_HAVEESTABLISHED(tp->t_state)) {
5774                                 /* Get the established keep-alive time */
5775                                 hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp));
5776                         } else {
5777                                 /*
5778                                  * Get the initial setup keep-alive time,
5779                                  * note that this is probably not going to
5780                                  * happen, since rack will be running a rxt timer
5781                                  * if a SYN of some sort is outstanding. It is
5782                                  * actually handled in rack_timeout_rxt().
5783                                  */
5784                                 hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp));
5785                         }
5786                         rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP;
5787                         if (rack->in_probe_rtt) {
5788                                 /*
5789                                  * We want to instead not wake up a long time from
5790                                  * now but to wake up about the time we would
5791                                  * exit probe-rtt and initiate a keep-alive ack.
5792                                  * This will get us out of probe-rtt and update
5793                                  * our min-rtt.
5794                                  */
5795                                 hpts_timeout = rack_min_probertt_hold;
5796                         }
5797                 }
5798         }
5799         if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) ==
5800             (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) {
5801                 /*
5802                  * RACK, TLP, persists and RXT timers all are restartable
5803                  * based on actions input .. i.e we received a packet (ack
5804                  * or sack) and that changes things (rw, or snd_una etc).
5805                  * Thus we can restart them with a new value. For
5806                  * keep-alive, delayed_ack we keep track of what was left
5807                  * and restart the timer with a smaller value.
5808                  */
5809                 if (left < hpts_timeout)
5810                         hpts_timeout = left;
5811         }
5812         if (hpts_timeout) {
5813                 /*
5814                  * Hack alert for now we can't time-out over 2,147,483
5815                  * seconds (a bit more than 596 hours), which is probably ok
5816                  * :).
5817                  */
5818                 if (hpts_timeout > 0x7ffffffe)
5819                         hpts_timeout = 0x7ffffffe;
5820                 rack->r_ctl.rc_timer_exp = cts + hpts_timeout;
5821         }
5822         rack_log_pacing_delay_calc(rack, entry_slot, slot, hpts_timeout, 0, 0, 27, __LINE__, NULL);
5823         if ((rack->gp_ready == 0) &&
5824             (rack->use_fixed_rate == 0) &&
5825             (hpts_timeout < slot) &&
5826             (rack->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) {
5827                 /*
5828                  * We have no good estimate yet for the
5829                  * old clunky burst mitigation or the
5830                  * real pacing. And the tlp or rxt is smaller
5831                  * than the pacing calculation. Lets not
5832                  * pace that long since we know the calculation
5833                  * so far is not accurate.
5834                  */
5835                 slot = hpts_timeout;
5836         }
5837         rack->r_ctl.last_pacing_time = slot;
5838         /**
5839          * Turn off all the flags for queuing by default. The
5840          * flags have important meanings to what happens when
5841          * LRO interacts with the transport. Most likely (by default now)
5842          * mbuf_queueing and ack compression are on. So the transport
5843          * has a couple of flags that control what happens (if those
5844          * are not on then these flags won't have any effect since it
5845          * won't go through the queuing LRO path).
5846          *
5847          * INP_MBUF_QUEUE_READY - This flags says that I am busy
5848          *                        pacing output, so don't disturb. But
5849          *                        it also means LRO can wake me if there
5850          *                        is a SACK arrival.
5851          *
5852          * INP_DONT_SACK_QUEUE - This flag is used in conjunction
5853          *                       with the above flag (QUEUE_READY) and
5854          *                       when present it says don't even wake me
5855          *                       if a SACK arrives.
5856          *
5857          * The idea behind these flags is that if we are pacing we
5858          * set the MBUF_QUEUE_READY and only get woken up if
5859          * a SACK arrives (which could change things) or if
5860          * our pacing timer expires. If, however, we have a rack
5861          * timer running, then we don't even want a sack to wake
5862          * us since the rack timer has to expire before we can send.
5863          *
5864          * Other cases should usually have none of the flags set
5865          * so LRO can call into us.
5866          */
5867         inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY);
5868         if (slot) {
5869                 rack->r_ctl.rc_last_output_to = us_cts + slot;
5870                 /*
5871                  * A pacing timer (slot) is being set, in
5872                  * such a case we cannot send (we are blocked by
5873                  * the timer). So lets tell LRO that it should not
5874                  * wake us unless there is a SACK. Note this only
5875                  * will be effective if mbuf queueing is on or
5876                  * compressed acks are being processed.
5877                  */
5878                 inp->inp_flags2 |= INP_MBUF_QUEUE_READY;
5879                 /*
5880                  * But wait if we have a Rack timer running
5881                  * even a SACK should not disturb us (with
5882                  * the exception of r_rr_config 3).
5883                  */
5884                 if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK) &&
5885                     (rack->r_rr_config != 3))
5886                         inp->inp_flags2 |= INP_DONT_SACK_QUEUE;
5887                 if (rack->rc_ack_can_sendout_data) {
5888                         /*
5889                          * Ahh but wait, this is that special case
5890                          * where the pacing timer can be disturbed
5891                          * backout the changes (used for non-paced
5892                          * burst limiting).
5893                          */
5894                         inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY);
5895                 }
5896                 if ((rack->use_rack_rr) &&
5897                     (rack->r_rr_config < 2) &&
5898                     ((hpts_timeout) && (hpts_timeout < slot))) {
5899                         /*
5900                          * Arrange for the hpts to kick back in after the
5901                          * t-o if the t-o does not cause a send.
5902                          */
5903                         (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(hpts_timeout),
5904                                                    __LINE__, &diag);
5905                         rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5906                         rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
5907                 } else {
5908                         (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(slot),
5909                                                    __LINE__, &diag);
5910                         rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5911                         rack_log_to_start(rack, cts, hpts_timeout, slot, 1);
5912                 }
5913         } else if (hpts_timeout) {
5914                 /*
5915                  * With respect to inp_flags2 here, lets let any new acks wake
5916                  * us up here. Since we are not pacing (no pacing timer), output
5917                  * can happen so we should let it. If its a Rack timer, then any inbound
5918                  * packet probably won't change the sending (we will be blocked)
5919                  * but it may change the prr stats so letting it in (the set defaults
5920                  * at the start of this block) are good enough.
5921                  */
5922                 (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(hpts_timeout),
5923                                            __LINE__, &diag);
5924                 rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5925                 rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
5926         } else {
5927                 /* No timer starting */
5928 #ifdef INVARIANTS
5929                 if (SEQ_GT(tp->snd_max, tp->snd_una)) {
5930                         panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?",
5931                             tp, rack, tot_len_this_send, cts, slot, hpts_timeout);
5932                 }
5933 #endif
5934         }
5935         rack->rc_tmr_stopped = 0;
5936         if (slot)
5937                 rack_log_type_bbrsnd(rack, tot_len_this_send, slot, us_cts, &tv);
5938 }
5939
5940 /*
5941  * RACK Timer, here we simply do logging and house keeping.
5942  * the normal rack_output() function will call the
5943  * appropriate thing to check if we need to do a RACK retransmit.
5944  * We return 1, saying don't proceed with rack_output only
5945  * when all timers have been stopped (destroyed PCB?).
5946  */
5947 static int
5948 rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5949 {
5950         /*
5951          * This timer simply provides an internal trigger to send out data.
5952          * The check_recovery_mode call will see if there are needed
5953          * retransmissions, if so we will enter fast-recovery. The output
5954          * call may or may not do the same thing depending on sysctl
5955          * settings.
5956          */
5957         struct rack_sendmap *rsm;
5958
5959         if (tp->t_timers->tt_flags & TT_STOPPED) {
5960                 return (1);
5961         }
5962         counter_u64_add(rack_to_tot, 1);
5963         if (rack->r_state && (rack->r_state != tp->t_state))
5964                 rack_set_state(tp, rack);
5965         rack->rc_on_min_to = 0;
5966         rsm = rack_check_recovery_mode(tp, cts);
5967         rack_log_to_event(rack, RACK_TO_FRM_RACK, rsm);
5968         if (rsm) {
5969                 rack->r_ctl.rc_resend = rsm;
5970                 rack->r_timer_override = 1;
5971                 if (rack->use_rack_rr) {
5972                         /*
5973                          * Don't accumulate extra pacing delay
5974                          * we are allowing the rack timer to
5975                          * over-ride pacing i.e. rrr takes precedence
5976                          * if the pacing interval is longer than the rrr
5977                          * time (in other words we get the min pacing
5978                          * time versus rrr pacing time).
5979                          */
5980                         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
5981                 }
5982         }
5983         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK;
5984         if (rsm == NULL) {
5985                 /* restart a timer and return 1 */
5986                 rack_start_hpts_timer(rack, tp, cts,
5987                                       0, 0, 0);
5988                 return (1);
5989         }
5990         return (0);
5991 }
5992
5993 static void
5994 rack_adjust_orig_mlen(struct rack_sendmap *rsm)
5995 {
5996         if (rsm->m->m_len > rsm->orig_m_len) {
5997                 /*
5998                  * Mbuf grew, caused by sbcompress, our offset does
5999                  * not change.
6000                  */
6001                 rsm->orig_m_len = rsm->m->m_len;
6002         } else if (rsm->m->m_len < rsm->orig_m_len) {
6003                 /*
6004                  * Mbuf shrank, trimmed off the top by an ack, our
6005                  * offset changes.
6006                  */
6007                 rsm->soff -= (rsm->orig_m_len - rsm->m->m_len);
6008                 rsm->orig_m_len = rsm->m->m_len;
6009         }
6010 }
6011
6012 static void
6013 rack_setup_offset_for_rsm(struct rack_sendmap *src_rsm, struct rack_sendmap *rsm)
6014 {
6015         struct mbuf *m;
6016         uint32_t soff;
6017
6018         if (src_rsm->m && (src_rsm->orig_m_len != src_rsm->m->m_len)) {
6019                 /* Fix up the orig_m_len and possibly the mbuf offset */
6020                 rack_adjust_orig_mlen(src_rsm);
6021         }
6022         m = src_rsm->m;
6023         soff = src_rsm->soff + (src_rsm->r_end - src_rsm->r_start);
6024         while (soff >= m->m_len) {
6025                 /* Move out past this mbuf */
6026                 soff -= m->m_len;
6027                 m = m->m_next;
6028                 KASSERT((m != NULL),
6029                         ("rsm:%p nrsm:%p hit at soff:%u null m",
6030                          src_rsm, rsm, soff));
6031         }
6032         rsm->m = m;
6033         rsm->soff = soff;
6034         rsm->orig_m_len = m->m_len;
6035 }
6036
6037 static __inline void
6038 rack_clone_rsm(struct tcp_rack *rack, struct rack_sendmap *nrsm,
6039                struct rack_sendmap *rsm, uint32_t start)
6040 {
6041         int idx;
6042
6043         nrsm->r_start = start;
6044         nrsm->r_end = rsm->r_end;
6045         nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
6046         nrsm->r_flags = rsm->r_flags;
6047         nrsm->r_dupack = rsm->r_dupack;
6048         nrsm->r_no_rtt_allowed = rsm->r_no_rtt_allowed;
6049         nrsm->r_rtr_bytes = 0;
6050         rsm->r_end = nrsm->r_start;
6051         nrsm->r_just_ret = rsm->r_just_ret;
6052         for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
6053                 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
6054         }
6055         /* Now if we have SYN flag we keep it on the left edge */
6056         if (nrsm->r_flags & RACK_HAS_SYN)
6057                 nrsm->r_flags &= ~RACK_HAS_SYN;
6058         /* Now if we have a FIN flag we keep it on the right edge */
6059         if (rsm->r_flags & RACK_HAS_FIN)
6060                 rsm->r_flags &= ~RACK_HAS_FIN;
6061         /* Push bit must go to the right edge as well */
6062         if (rsm->r_flags & RACK_HAD_PUSH)
6063                 rsm->r_flags &= ~RACK_HAD_PUSH;
6064
6065         /*
6066          * Now we need to find nrsm's new location in the mbuf chain
6067          * we basically calculate a new offset, which is soff +
6068          * how much is left in original rsm. Then we walk out the mbuf
6069          * chain to find the righ postion, it may be the same mbuf
6070          * or maybe not.
6071          */
6072         KASSERT(((rsm->m != NULL) ||
6073                  (rsm->r_flags & (RACK_HAS_SYN|RACK_HAS_FIN))),
6074                 ("rsm:%p nrsm:%p rack:%p -- rsm->m is NULL?", rsm, nrsm, rack));
6075         if (rsm->m)
6076                 rack_setup_offset_for_rsm(rsm, nrsm);
6077 }
6078
6079 static struct rack_sendmap *
6080 rack_merge_rsm(struct tcp_rack *rack,
6081                struct rack_sendmap *l_rsm,
6082                struct rack_sendmap *r_rsm)
6083 {
6084         /*
6085          * We are merging two ack'd RSM's,
6086          * the l_rsm is on the left (lower seq
6087          * values) and the r_rsm is on the right
6088          * (higher seq value). The simplest way
6089          * to merge these is to move the right
6090          * one into the left. I don't think there
6091          * is any reason we need to try to find
6092          * the oldest (or last oldest retransmitted).
6093          */
6094         struct rack_sendmap *rm;
6095
6096         rack_log_map_chg(rack->rc_tp, rack, NULL,
6097                          l_rsm, r_rsm, MAP_MERGE, r_rsm->r_end, __LINE__);
6098         l_rsm->r_end = r_rsm->r_end;
6099         if (l_rsm->r_dupack < r_rsm->r_dupack)
6100                 l_rsm->r_dupack = r_rsm->r_dupack;
6101         if (r_rsm->r_rtr_bytes)
6102                 l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes;
6103         if (r_rsm->r_in_tmap) {
6104                 /* This really should not happen */
6105                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, r_rsm, r_tnext);
6106                 r_rsm->r_in_tmap = 0;
6107         }
6108
6109         /* Now the flags */
6110         if (r_rsm->r_flags & RACK_HAS_FIN)
6111                 l_rsm->r_flags |= RACK_HAS_FIN;
6112         if (r_rsm->r_flags & RACK_TLP)
6113                 l_rsm->r_flags |= RACK_TLP;
6114         if (r_rsm->r_flags & RACK_RWND_COLLAPSED)
6115                 l_rsm->r_flags |= RACK_RWND_COLLAPSED;
6116         if ((r_rsm->r_flags & RACK_APP_LIMITED)  &&
6117             ((l_rsm->r_flags & RACK_APP_LIMITED) == 0)) {
6118                 /*
6119                  * If both are app-limited then let the
6120                  * free lower the count. If right is app
6121                  * limited and left is not, transfer.
6122                  */
6123                 l_rsm->r_flags |= RACK_APP_LIMITED;
6124                 r_rsm->r_flags &= ~RACK_APP_LIMITED;
6125                 if (r_rsm == rack->r_ctl.rc_first_appl)
6126                         rack->r_ctl.rc_first_appl = l_rsm;
6127         }
6128         rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, r_rsm);
6129 #ifdef INVARIANTS
6130         if (rm != r_rsm) {
6131                 panic("removing head in rack:%p rsm:%p rm:%p",
6132                       rack, r_rsm, rm);
6133         }
6134 #endif
6135         if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) {
6136                 /* Transfer the split limit to the map we free */
6137                 r_rsm->r_limit_type = l_rsm->r_limit_type;
6138                 l_rsm->r_limit_type = 0;
6139         }
6140         rack_free(rack, r_rsm);
6141         return (l_rsm);
6142 }
6143
6144 /*
6145  * TLP Timer, here we simply setup what segment we want to
6146  * have the TLP expire on, the normal rack_output() will then
6147  * send it out.
6148  *
6149  * We return 1, saying don't proceed with rack_output only
6150  * when all timers have been stopped (destroyed PCB?).
6151  */
6152 static int
6153 rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6154 {
6155         /*
6156          * Tail Loss Probe.
6157          */
6158         struct rack_sendmap *rsm = NULL;
6159         struct rack_sendmap *insret;
6160         struct socket *so;
6161         uint32_t amm;
6162         uint32_t out, avail;
6163         int collapsed_win = 0;
6164
6165         if (tp->t_timers->tt_flags & TT_STOPPED) {
6166                 return (1);
6167         }
6168         if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
6169                 /* Its not time yet */
6170                 return (0);
6171         }
6172         if (ctf_progress_timeout_check(tp, true)) {
6173                 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6174                 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
6175                 return (1);
6176         }
6177         /*
6178          * A TLP timer has expired. We have been idle for 2 rtts. So we now
6179          * need to figure out how to force a full MSS segment out.
6180          */
6181         rack_log_to_event(rack, RACK_TO_FRM_TLP, NULL);
6182         rack->r_ctl.retran_during_recovery = 0;
6183         rack->r_ctl.dsack_byte_cnt = 0;
6184         counter_u64_add(rack_tlp_tot, 1);
6185         if (rack->r_state && (rack->r_state != tp->t_state))
6186                 rack_set_state(tp, rack);
6187         so = tp->t_inpcb->inp_socket;
6188         avail = sbavail(&so->so_snd);
6189         out = tp->snd_max - tp->snd_una;
6190         if (out > tp->snd_wnd) {
6191                 /* special case, we need a retransmission */
6192                 collapsed_win = 1;
6193                 goto need_retran;
6194         }
6195         /*
6196          * Check our send oldest always settings, and if
6197          * there is an oldest to send jump to the need_retran.
6198          */
6199         if (rack_always_send_oldest && (TAILQ_EMPTY(&rack->r_ctl.rc_tmap) == 0))
6200                 goto need_retran;
6201
6202         if (avail > out) {
6203                 /* New data is available */
6204                 amm = avail - out;
6205                 if (amm > ctf_fixed_maxseg(tp)) {
6206                         amm = ctf_fixed_maxseg(tp);
6207                         if ((amm + out) > tp->snd_wnd) {
6208                                 /* We are rwnd limited */
6209                                 goto need_retran;
6210                         }
6211                 } else if (amm < ctf_fixed_maxseg(tp)) {
6212                         /* not enough to fill a MTU */
6213                         goto need_retran;
6214                 }
6215                 if (IN_FASTRECOVERY(tp->t_flags)) {
6216                         /* Unlikely */
6217                         if (rack->rack_no_prr == 0) {
6218                                 if (out + amm <= tp->snd_wnd) {
6219                                         rack->r_ctl.rc_prr_sndcnt = amm;
6220                                         rack_log_to_prr(rack, 4, 0);
6221                                 }
6222                         } else
6223                                 goto need_retran;
6224                 } else {
6225                         /* Set the send-new override */
6226                         if (out + amm <= tp->snd_wnd)
6227                                 rack->r_ctl.rc_tlp_new_data = amm;
6228                         else
6229                                 goto need_retran;
6230                 }
6231                 rack->r_ctl.rc_tlpsend = NULL;
6232                 counter_u64_add(rack_tlp_newdata, 1);
6233                 goto send;
6234         }
6235 need_retran:
6236         /*
6237          * Ok we need to arrange the last un-acked segment to be re-sent, or
6238          * optionally the first un-acked segment.
6239          */
6240         if (collapsed_win == 0) {
6241                 if (rack_always_send_oldest)
6242                         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6243                 else {
6244                         rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6245                         if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) {
6246                                 rsm = rack_find_high_nonack(rack, rsm);
6247                         }
6248                 }
6249                 if (rsm == NULL) {
6250                         counter_u64_add(rack_tlp_does_nada, 1);
6251 #ifdef TCP_BLACKBOX
6252                         tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
6253 #endif
6254                         goto out;
6255                 }
6256         } else {
6257                 /*
6258                  * We must find the last segment
6259                  * that was acceptable by the client.
6260                  */
6261                 RB_FOREACH_REVERSE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
6262                         if ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0) {
6263                                 /* Found one */
6264                                 break;
6265                         }
6266                 }
6267                 if (rsm == NULL) {
6268                         /* None? if so send the first */
6269                         rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6270                         if (rsm == NULL) {
6271                                 counter_u64_add(rack_tlp_does_nada, 1);
6272 #ifdef TCP_BLACKBOX
6273                                 tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
6274 #endif
6275                                 goto out;
6276                         }
6277                 }
6278         }
6279         if ((rsm->r_end - rsm->r_start) > ctf_fixed_maxseg(tp)) {
6280                 /*
6281                  * We need to split this the last segment in two.
6282                  */
6283                 struct rack_sendmap *nrsm;
6284
6285                 nrsm = rack_alloc_full_limit(rack);
6286                 if (nrsm == NULL) {
6287                         /*
6288                          * No memory to split, we will just exit and punt
6289                          * off to the RXT timer.
6290                          */
6291                         counter_u64_add(rack_tlp_does_nada, 1);
6292                         goto out;
6293                 }
6294                 rack_clone_rsm(rack, nrsm, rsm,
6295                                (rsm->r_end - ctf_fixed_maxseg(tp)));
6296                 rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
6297                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
6298 #ifdef INVARIANTS
6299                 if (insret != NULL) {
6300                         panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
6301                               nrsm, insret, rack, rsm);
6302                 }
6303 #endif
6304                 if (rsm->r_in_tmap) {
6305                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
6306                         nrsm->r_in_tmap = 1;
6307                 }
6308                 rsm->r_flags &= (~RACK_HAS_FIN);
6309                 rsm = nrsm;
6310         }
6311         rack->r_ctl.rc_tlpsend = rsm;
6312 send:
6313         rack->r_timer_override = 1;
6314         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
6315         return (0);
6316 out:
6317         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
6318         return (0);
6319 }
6320
6321 /*
6322  * Delayed ack Timer, here we simply need to setup the
6323  * ACK_NOW flag and remove the DELACK flag. From there
6324  * the output routine will send the ack out.
6325  *
6326  * We only return 1, saying don't proceed, if all timers
6327  * are stopped (destroyed PCB?).
6328  */
6329 static int
6330 rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6331 {
6332         if (tp->t_timers->tt_flags & TT_STOPPED) {
6333                 return (1);
6334         }
6335         rack_log_to_event(rack, RACK_TO_FRM_DELACK, NULL);
6336         tp->t_flags &= ~TF_DELACK;
6337         tp->t_flags |= TF_ACKNOW;
6338         KMOD_TCPSTAT_INC(tcps_delack);
6339         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
6340         return (0);
6341 }
6342
6343 /*
6344  * Persists timer, here we simply send the
6345  * same thing as a keepalive will.
6346  * the one byte send.
6347  *
6348  * We only return 1, saying don't proceed, if all timers
6349  * are stopped (destroyed PCB?).
6350  */
6351 static int
6352 rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6353 {
6354         struct tcptemp *t_template;
6355         struct inpcb *inp;
6356         int32_t retval = 1;
6357
6358         inp = tp->t_inpcb;
6359
6360         if (tp->t_timers->tt_flags & TT_STOPPED) {
6361                 return (1);
6362         }
6363         if (rack->rc_in_persist == 0)
6364                 return (0);
6365         if (ctf_progress_timeout_check(tp, false)) {
6366                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6367                 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6368                 tcp_set_inp_to_drop(inp, ETIMEDOUT);
6369                 return (1);
6370         }
6371         KASSERT(inp != NULL, ("%s: tp %p tp->t_inpcb == NULL", __func__, tp));
6372         /*
6373          * Persistence timer into zero window. Force a byte to be output, if
6374          * possible.
6375          */
6376         KMOD_TCPSTAT_INC(tcps_persisttimeo);
6377         /*
6378          * Hack: if the peer is dead/unreachable, we do not time out if the
6379          * window is closed.  After a full backoff, drop the connection if
6380          * the idle time (no responses to probes) reaches the maximum
6381          * backoff that we would use if retransmitting.
6382          */
6383         if (tp->t_rxtshift == TCP_MAXRXTSHIFT &&
6384             (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
6385              TICKS_2_USEC(ticks - tp->t_rcvtime) >= RACK_REXMTVAL(tp) * tcp_totbackoff)) {
6386                 KMOD_TCPSTAT_INC(tcps_persistdrop);
6387                 retval = 1;
6388                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6389                 tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT);
6390                 goto out;
6391         }
6392         if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) &&
6393             tp->snd_una == tp->snd_max)
6394                 rack_exit_persist(tp, rack, cts);
6395         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT;
6396         /*
6397          * If the user has closed the socket then drop a persisting
6398          * connection after a much reduced timeout.
6399          */
6400         if (tp->t_state > TCPS_CLOSE_WAIT &&
6401             (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
6402                 retval = 1;
6403                 KMOD_TCPSTAT_INC(tcps_persistdrop);
6404                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6405                 tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT);
6406                 goto out;
6407         }
6408         t_template = tcpip_maketemplate(rack->rc_inp);
6409         if (t_template) {
6410                 /* only set it if we were answered */
6411                 if (rack->forced_ack == 0) {
6412                         rack->forced_ack = 1;
6413                         rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
6414                 }
6415                 tcp_respond(tp, t_template->tt_ipgen,
6416                             &t_template->tt_t, (struct mbuf *)NULL,
6417                             tp->rcv_nxt, tp->snd_una - 1, 0);
6418                 /* This sends an ack */
6419                 if (tp->t_flags & TF_DELACK)
6420                         tp->t_flags &= ~TF_DELACK;
6421                 free(t_template, M_TEMP);
6422         }
6423         if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
6424                 tp->t_rxtshift++;
6425 out:
6426         rack_log_to_event(rack, RACK_TO_FRM_PERSIST, NULL);
6427         rack_start_hpts_timer(rack, tp, cts,
6428                               0, 0, 0);
6429         return (retval);
6430 }
6431
6432 /*
6433  * If a keepalive goes off, we had no other timers
6434  * happening. We always return 1 here since this
6435  * routine either drops the connection or sends
6436  * out a segment with respond.
6437  */
6438 static int
6439 rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6440 {
6441         struct tcptemp *t_template;
6442         struct inpcb *inp;
6443
6444         if (tp->t_timers->tt_flags & TT_STOPPED) {
6445                 return (1);
6446         }
6447         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP;
6448         inp = tp->t_inpcb;
6449         rack_log_to_event(rack, RACK_TO_FRM_KEEP, NULL);
6450         /*
6451          * Keep-alive timer went off; send something or drop connection if
6452          * idle for too long.
6453          */
6454         KMOD_TCPSTAT_INC(tcps_keeptimeo);
6455         if (tp->t_state < TCPS_ESTABLISHED)
6456                 goto dropit;
6457         if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
6458             tp->t_state <= TCPS_CLOSING) {
6459                 if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
6460                         goto dropit;
6461                 /*
6462                  * Send a packet designed to force a response if the peer is
6463                  * up and reachable: either an ACK if the connection is
6464                  * still alive, or an RST if the peer has closed the
6465                  * connection due to timeout or reboot. Using sequence
6466                  * number tp->snd_una-1 causes the transmitted zero-length
6467                  * segment to lie outside the receive window; by the
6468                  * protocol spec, this requires the correspondent TCP to
6469                  * respond.
6470                  */
6471                 KMOD_TCPSTAT_INC(tcps_keepprobe);
6472                 t_template = tcpip_maketemplate(inp);
6473                 if (t_template) {
6474                         if (rack->forced_ack == 0) {
6475                                 rack->forced_ack = 1;
6476                                 rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
6477                         }
6478                         tcp_respond(tp, t_template->tt_ipgen,
6479                             &t_template->tt_t, (struct mbuf *)NULL,
6480                             tp->rcv_nxt, tp->snd_una - 1, 0);
6481                         free(t_template, M_TEMP);
6482                 }
6483         }
6484         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
6485         return (1);
6486 dropit:
6487         KMOD_TCPSTAT_INC(tcps_keepdrops);
6488         tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
6489         tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT);
6490         return (1);
6491 }
6492
6493 /*
6494  * Retransmit helper function, clear up all the ack
6495  * flags and take care of important book keeping.
6496  */
6497 static void
6498 rack_remxt_tmr(struct tcpcb *tp)
6499 {
6500         /*
6501          * The retransmit timer went off, all sack'd blocks must be
6502          * un-acked.
6503          */
6504         struct rack_sendmap *rsm, *trsm = NULL;
6505         struct tcp_rack *rack;
6506
6507         rack = (struct tcp_rack *)tp->t_fb_ptr;
6508         rack_timer_cancel(tp, rack, tcp_get_usecs(NULL), __LINE__);
6509         rack_log_to_event(rack, RACK_TO_FRM_TMR, NULL);
6510         if (rack->r_state && (rack->r_state != tp->t_state))
6511                 rack_set_state(tp, rack);
6512         /*
6513          * Ideally we would like to be able to
6514          * mark SACK-PASS on anything not acked here.
6515          *
6516          * However, if we do that we would burst out
6517          * all that data 1ms apart. This would be unwise,
6518          * so for now we will just let the normal rxt timer
6519          * and tlp timer take care of it.
6520          *
6521          * Also we really need to stick them back in sequence
6522          * order. This way we send in the proper order and any
6523          * sacks that come floating in will "re-ack" the data.
6524          * To do this we zap the tmap with an INIT and then
6525          * walk through and place every rsm in the RB tree
6526          * back in its seq ordered place.
6527          */
6528         TAILQ_INIT(&rack->r_ctl.rc_tmap);
6529         RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
6530                 rsm->r_dupack = 0;
6531                 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
6532                 /* We must re-add it back to the tlist */
6533                 if (trsm == NULL) {
6534                         TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
6535                 } else {
6536                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext);
6537                 }
6538                 rsm->r_in_tmap = 1;
6539                 trsm = rsm;
6540                 if (rsm->r_flags & RACK_ACKED)
6541                         rsm->r_flags |= RACK_WAS_ACKED;
6542                 rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS);
6543         }
6544         /* Clear the count (we just un-acked them) */
6545         rack->r_ctl.rc_last_timeout_snduna = tp->snd_una;
6546         rack->r_ctl.rc_sacked = 0;
6547         rack->r_ctl.rc_sacklast = NULL;
6548         rack->r_ctl.rc_agg_delayed = 0;
6549         rack->r_early = 0;
6550         rack->r_ctl.rc_agg_early = 0;
6551         rack->r_late = 0;
6552         /* Clear the tlp rtx mark */
6553         rack->r_ctl.rc_resend = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6554         if (rack->r_ctl.rc_resend != NULL)
6555                 rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
6556         rack->r_ctl.rc_prr_sndcnt = 0;
6557         rack_log_to_prr(rack, 6, 0);
6558         rack->r_timer_override = 1;
6559         if ((((tp->t_flags & TF_SACK_PERMIT) == 0)
6560 #ifdef NETFLIX_EXP_DETECTION
6561             || (rack->sack_attack_disable != 0)
6562 #endif
6563                     ) && ((tp->t_flags & TF_SENTFIN) == 0)) {
6564                 /*
6565                  * For non-sack customers new data
6566                  * needs to go out as retransmits until
6567                  * we retransmit up to snd_max.
6568                  */
6569                 rack->r_must_retran = 1;
6570                 rack->r_ctl.rc_out_at_rto = ctf_flight_size(rack->rc_tp,
6571                                                 rack->r_ctl.rc_sacked);
6572         }
6573         rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
6574 }
6575
6576 static void
6577 rack_convert_rtts(struct tcpcb *tp)
6578 {
6579         if (tp->t_srtt > 1) {
6580                 uint32_t val, frac;
6581
6582                 val = tp->t_srtt >> TCP_RTT_SHIFT;
6583                 frac = tp->t_srtt & 0x1f;
6584                 tp->t_srtt = TICKS_2_USEC(val);
6585                 /*
6586                  * frac is the fractional part of the srtt (if any)
6587                  * but its in ticks and every bit represents
6588                  * 1/32nd of a hz.
6589                  */
6590                 if (frac) {
6591                         if (hz == 1000) {
6592                                 frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_MSEC) / (uint64_t)TCP_RTT_SCALE);
6593                         } else {
6594                                 frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_SEC) / ((uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE));
6595                         }
6596                         tp->t_srtt += frac;
6597                 }
6598         }
6599         if (tp->t_rttvar) {
6600                 uint32_t val, frac;
6601
6602                 val = tp->t_rttvar >> TCP_RTTVAR_SHIFT;
6603                 frac = tp->t_rttvar & 0x1f;
6604                 tp->t_rttvar = TICKS_2_USEC(val);
6605                 /*
6606                  * frac is the fractional part of the srtt (if any)
6607                  * but its in ticks and every bit represents
6608                  * 1/32nd of a hz.
6609                  */
6610                 if (frac) {
6611                         if (hz == 1000) {
6612                                 frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_MSEC) / (uint64_t)TCP_RTT_SCALE);
6613                         } else {
6614                                 frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_SEC) / ((uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE));
6615                         }
6616                         tp->t_rttvar += frac;
6617                 }
6618         }
6619         tp->t_rxtcur = RACK_REXMTVAL(tp);
6620         if (TCPS_HAVEESTABLISHED(tp->t_state)) {
6621                 tp->t_rxtcur += TICKS_2_USEC(tcp_rexmit_slop);
6622         }
6623         if (tp->t_rxtcur > rack_rto_max) {
6624                 tp->t_rxtcur = rack_rto_max;
6625         }
6626 }
6627
6628 static void
6629 rack_cc_conn_init(struct tcpcb *tp)
6630 {
6631         struct tcp_rack *rack;
6632         uint32_t srtt;
6633
6634         rack = (struct tcp_rack *)tp->t_fb_ptr;
6635         srtt = tp->t_srtt;
6636         cc_conn_init(tp);
6637         /*
6638          * Now convert to rack's internal format,
6639          * if required.
6640          */
6641         if ((srtt == 0) && (tp->t_srtt != 0))
6642                 rack_convert_rtts(tp);
6643         /*
6644          * We want a chance to stay in slowstart as
6645          * we create a connection. TCP spec says that
6646          * initially ssthresh is infinite. For our
6647          * purposes that is the snd_wnd.
6648          */
6649         if (tp->snd_ssthresh < tp->snd_wnd) {
6650                 tp->snd_ssthresh = tp->snd_wnd;
6651         }
6652         /*
6653          * We also want to assure a IW worth of
6654          * data can get inflight.
6655          */
6656         if (rc_init_window(rack) < tp->snd_cwnd)
6657                 tp->snd_cwnd = rc_init_window(rack);
6658 }
6659
6660 /*
6661  * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
6662  * we will setup to retransmit the lowest seq number outstanding.
6663  */
6664 static int
6665 rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6666 {
6667         int32_t rexmt;
6668         struct inpcb *inp;
6669         int32_t retval = 0;
6670         bool isipv6;
6671
6672         inp = tp->t_inpcb;
6673         if (tp->t_timers->tt_flags & TT_STOPPED) {
6674                 return (1);
6675         }
6676         if (ctf_progress_timeout_check(tp, false)) {
6677                 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
6678                 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6679                 tcp_set_inp_to_drop(inp, ETIMEDOUT);
6680                 return (1);
6681         }
6682         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
6683         rack->r_ctl.retran_during_recovery = 0;
6684         rack->r_ctl.dsack_byte_cnt = 0;
6685         if (IN_FASTRECOVERY(tp->t_flags))
6686                 tp->t_flags |= TF_WASFRECOVERY;
6687         else
6688                 tp->t_flags &= ~TF_WASFRECOVERY;
6689         if (IN_CONGRECOVERY(tp->t_flags))
6690                 tp->t_flags |= TF_WASCRECOVERY;
6691         else
6692                 tp->t_flags &= ~TF_WASCRECOVERY;
6693         if (TCPS_HAVEESTABLISHED(tp->t_state) &&
6694             (tp->snd_una == tp->snd_max)) {
6695                 /* Nothing outstanding .. nothing to do */
6696                 return (0);
6697         }
6698         /*
6699          * Rack can only run one timer  at a time, so we cannot
6700          * run a KEEPINIT (gating SYN sending) and a retransmit
6701          * timer for the SYN. So if we are in a front state and
6702          * have a KEEPINIT timer we need to check the first transmit
6703          * against now to see if we have exceeded the KEEPINIT time
6704          * (if one is set).
6705          */
6706         if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
6707             (TP_KEEPINIT(tp) != 0)) {
6708                 struct rack_sendmap *rsm;
6709
6710                 rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6711                 if (rsm) {
6712                         /* Ok we have something outstanding to test keepinit with */
6713                         if ((TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) &&
6714                             ((cts - (uint32_t)rsm->r_tim_lastsent[0]) >= TICKS_2_USEC(TP_KEEPINIT(tp)))) {
6715                                 /* We have exceeded the KEEPINIT time */
6716                                 tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
6717                                 goto drop_it;
6718                         }
6719                 }
6720         }
6721         /*
6722          * Retransmission timer went off.  Message has not been acked within
6723          * retransmit interval.  Back off to a longer retransmit interval
6724          * and retransmit one segment.
6725          */
6726         rack_remxt_tmr(tp);
6727         if ((rack->r_ctl.rc_resend == NULL) ||
6728             ((rack->r_ctl.rc_resend->r_flags & RACK_RWND_COLLAPSED) == 0)) {
6729                 /*
6730                  * If the rwnd collapsed on
6731                  * the one we are retransmitting
6732                  * it does not count against the
6733                  * rxt count.
6734                  */
6735                 tp->t_rxtshift++;
6736         }
6737         if (tp->t_rxtshift > TCP_MAXRXTSHIFT) {
6738                 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
6739 drop_it:
6740                 tp->t_rxtshift = TCP_MAXRXTSHIFT;
6741                 KMOD_TCPSTAT_INC(tcps_timeoutdrop);
6742                 retval = 1;
6743                 tcp_set_inp_to_drop(rack->rc_inp,
6744                     (tp->t_softerror ? (uint16_t) tp->t_softerror : ETIMEDOUT));
6745                 goto out;
6746         }
6747         if (tp->t_state == TCPS_SYN_SENT) {
6748                 /*
6749                  * If the SYN was retransmitted, indicate CWND to be limited
6750                  * to 1 segment in cc_conn_init().
6751                  */
6752                 tp->snd_cwnd = 1;
6753         } else if (tp->t_rxtshift == 1) {
6754                 /*
6755                  * first retransmit; record ssthresh and cwnd so they can be
6756                  * recovered if this turns out to be a "bad" retransmit. A
6757                  * retransmit is considered "bad" if an ACK for this segment
6758                  * is received within RTT/2 interval; the assumption here is
6759                  * that the ACK was already in flight.  See "On Estimating
6760                  * End-to-End Network Path Properties" by Allman and Paxson
6761                  * for more details.
6762                  */
6763                 tp->snd_cwnd_prev = tp->snd_cwnd;
6764                 tp->snd_ssthresh_prev = tp->snd_ssthresh;
6765                 tp->snd_recover_prev = tp->snd_recover;
6766                 tp->t_badrxtwin = ticks + (USEC_2_TICKS(tp->t_srtt)/2);
6767                 tp->t_flags |= TF_PREVVALID;
6768         } else if ((tp->t_flags & TF_RCVD_TSTMP) == 0)
6769                 tp->t_flags &= ~TF_PREVVALID;
6770         KMOD_TCPSTAT_INC(tcps_rexmttimeo);
6771         if ((tp->t_state == TCPS_SYN_SENT) ||
6772             (tp->t_state == TCPS_SYN_RECEIVED))
6773                 rexmt = RACK_INITIAL_RTO * tcp_backoff[tp->t_rxtshift];
6774         else
6775                 rexmt = max(rack_rto_min, (tp->t_srtt + (tp->t_rttvar << 2))) * tcp_backoff[tp->t_rxtshift];
6776
6777         RACK_TCPT_RANGESET(tp->t_rxtcur, rexmt,
6778            max(rack_rto_min, rexmt), rack_rto_max, rack->r_ctl.timer_slop);
6779         /*
6780          * We enter the path for PLMTUD if connection is established or, if
6781          * connection is FIN_WAIT_1 status, reason for the last is that if
6782          * amount of data we send is very small, we could send it in couple
6783          * of packets and process straight to FIN. In that case we won't
6784          * catch ESTABLISHED state.
6785          */
6786 #ifdef INET6
6787         isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) ? true : false;
6788 #else
6789         isipv6 = false;
6790 #endif
6791         if (((V_tcp_pmtud_blackhole_detect == 1) ||
6792             (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
6793             (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
6794             ((tp->t_state == TCPS_ESTABLISHED) ||
6795             (tp->t_state == TCPS_FIN_WAIT_1))) {
6796                 /*
6797                  * Idea here is that at each stage of mtu probe (usually,
6798                  * 1448 -> 1188 -> 524) should be given 2 chances to recover
6799                  * before further clamping down. 'tp->t_rxtshift % 2 == 0'
6800                  * should take care of that.
6801                  */
6802                 if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
6803                     (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
6804                     (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
6805                     tp->t_rxtshift % 2 == 0)) {
6806                         /*
6807                          * Enter Path MTU Black-hole Detection mechanism: -
6808                          * Disable Path MTU Discovery (IP "DF" bit). -
6809                          * Reduce MTU to lower value than what we negotiated
6810                          * with peer.
6811                          */
6812                         if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
6813                                 /* Record that we may have found a black hole. */
6814                                 tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
6815                                 /* Keep track of previous MSS. */
6816                                 tp->t_pmtud_saved_maxseg = tp->t_maxseg;
6817                         }
6818
6819                         /*
6820                          * Reduce the MSS to blackhole value or to the
6821                          * default in an attempt to retransmit.
6822                          */
6823 #ifdef INET6
6824                         if (isipv6 &&
6825                             tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
6826                                 /* Use the sysctl tuneable blackhole MSS. */
6827                                 tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
6828                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
6829                         } else if (isipv6) {
6830                                 /* Use the default MSS. */
6831                                 tp->t_maxseg = V_tcp_v6mssdflt;
6832                                 /*
6833                                  * Disable Path MTU Discovery when we switch
6834                                  * to minmss.
6835                                  */
6836                                 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
6837                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
6838                         }
6839 #endif
6840 #if defined(INET6) && defined(INET)
6841                         else
6842 #endif
6843 #ifdef INET
6844                         if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
6845                                 /* Use the sysctl tuneable blackhole MSS. */
6846                                 tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
6847                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
6848                         } else {
6849                                 /* Use the default MSS. */
6850                                 tp->t_maxseg = V_tcp_mssdflt;
6851                                 /*
6852                                  * Disable Path MTU Discovery when we switch
6853                                  * to minmss.
6854                                  */
6855                                 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
6856                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
6857                         }
6858 #endif
6859                 } else {
6860                         /*
6861                          * If further retransmissions are still unsuccessful
6862                          * with a lowered MTU, maybe this isn't a blackhole
6863                          * and we restore the previous MSS and blackhole
6864                          * detection flags. The limit '6' is determined by
6865                          * giving each probe stage (1448, 1188, 524) 2
6866                          * chances to recover.
6867                          */
6868                         if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
6869                             (tp->t_rxtshift >= 6)) {
6870                                 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
6871                                 tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
6872                                 tp->t_maxseg = tp->t_pmtud_saved_maxseg;
6873                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed);
6874                         }
6875                 }
6876         }
6877         /*
6878          * Disable RFC1323 and SACK if we haven't got any response to
6879          * our third SYN to work-around some broken terminal servers
6880          * (most of which have hopefully been retired) that have bad VJ
6881          * header compression code which trashes TCP segments containing
6882          * unknown-to-them TCP options.
6883          */
6884         if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
6885             (tp->t_rxtshift == 3))
6886                 tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP|TF_SACK_PERMIT);
6887         /*
6888          * If we backed off this far, our srtt estimate is probably bogus.
6889          * Clobber it so we'll take the next rtt measurement as our srtt;
6890          * move the current srtt into rttvar to keep the current retransmit
6891          * times until then.
6892          */
6893         if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
6894 #ifdef INET6
6895                 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0)
6896                         in6_losing(tp->t_inpcb);
6897                 else
6898 #endif
6899                         in_losing(tp->t_inpcb);
6900                 tp->t_rttvar += tp->t_srtt;
6901                 tp->t_srtt = 0;
6902         }
6903         sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
6904         tp->snd_recover = tp->snd_max;
6905         tp->t_flags |= TF_ACKNOW;
6906         tp->t_rtttime = 0;
6907         rack_cong_signal(tp, CC_RTO, tp->snd_una);
6908 out:
6909         return (retval);
6910 }
6911
6912 static int
6913 rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling)
6914 {
6915         int32_t ret = 0;
6916         int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
6917
6918         if (timers == 0) {
6919                 return (0);
6920         }
6921         if (tp->t_state == TCPS_LISTEN) {
6922                 /* no timers on listen sockets */
6923                 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
6924                         return (0);
6925                 return (1);
6926         }
6927         if ((timers & PACE_TMR_RACK) &&
6928             rack->rc_on_min_to) {
6929                 /*
6930                  * For the rack timer when we
6931                  * are on a min-timeout (which means rrr_conf = 3)
6932                  * we don't want to check the timer. It may
6933                  * be going off for a pace and thats ok we
6934                  * want to send the retransmit (if its ready).
6935                  *
6936                  * If its on a normal rack timer (non-min) then
6937                  * we will check if its expired.
6938                  */
6939                 goto skip_time_check;
6940         }
6941         if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
6942                 uint32_t left;
6943
6944                 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
6945                         ret = -1;
6946                         rack_log_to_processing(rack, cts, ret, 0);
6947                         return (0);
6948                 }
6949                 if (hpts_calling == 0) {
6950                         /*
6951                          * A user send or queued mbuf (sack) has called us? We
6952                          * return 0 and let the pacing guards
6953                          * deal with it if they should or
6954                          * should not cause a send.
6955                          */
6956                         ret = -2;
6957                         rack_log_to_processing(rack, cts, ret, 0);
6958                         return (0);
6959                 }
6960                 /*
6961                  * Ok our timer went off early and we are not paced false
6962                  * alarm, go back to sleep.
6963                  */
6964                 ret = -3;
6965                 left = rack->r_ctl.rc_timer_exp - cts;
6966                 tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(left));
6967                 rack_log_to_processing(rack, cts, ret, left);
6968                 return (1);
6969         }
6970 skip_time_check:
6971         rack->rc_tmr_stopped = 0;
6972         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
6973         if (timers & PACE_TMR_DELACK) {
6974                 ret = rack_timeout_delack(tp, rack, cts);
6975         } else if (timers & PACE_TMR_RACK) {
6976                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
6977                 rack->r_fast_output = 0;
6978                 ret = rack_timeout_rack(tp, rack, cts);
6979         } else if (timers & PACE_TMR_TLP) {
6980                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
6981                 ret = rack_timeout_tlp(tp, rack, cts);
6982         } else if (timers & PACE_TMR_RXT) {
6983                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
6984                 rack->r_fast_output = 0;
6985                 ret = rack_timeout_rxt(tp, rack, cts);
6986         } else if (timers & PACE_TMR_PERSIT) {
6987                 ret = rack_timeout_persist(tp, rack, cts);
6988         } else if (timers & PACE_TMR_KEEP) {
6989                 ret = rack_timeout_keepalive(tp, rack, cts);
6990         }
6991         rack_log_to_processing(rack, cts, ret, timers);
6992         return (ret);
6993 }
6994
6995 static void
6996 rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line)
6997 {
6998         struct timeval tv;
6999         uint32_t us_cts, flags_on_entry;
7000         uint8_t hpts_removed = 0;
7001
7002         flags_on_entry = rack->r_ctl.rc_hpts_flags;
7003         us_cts = tcp_get_usecs(&tv);
7004         if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
7005             ((TSTMP_GEQ(us_cts, rack->r_ctl.rc_last_output_to)) ||
7006              ((tp->snd_max - tp->snd_una) == 0))) {
7007                 tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
7008                 hpts_removed = 1;
7009                 /* If we were not delayed cancel out the flag. */
7010                 if ((tp->snd_max - tp->snd_una) == 0)
7011                         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
7012                 rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
7013         }
7014         if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
7015                 rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
7016                 if (rack->rc_inp->inp_in_hpts &&
7017                     ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) {
7018                         /*
7019                          * Canceling timer's when we have no output being
7020                          * paced. We also must remove ourselves from the
7021                          * hpts.
7022                          */
7023                         tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
7024                         hpts_removed = 1;
7025                 }
7026                 rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
7027         }
7028         if (hpts_removed == 0)
7029                 rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
7030 }
7031
7032 static void
7033 rack_timer_stop(struct tcpcb *tp, uint32_t timer_type)
7034 {
7035         return;
7036 }
7037
7038 static int
7039 rack_stopall(struct tcpcb *tp)
7040 {
7041         struct tcp_rack *rack;
7042         rack = (struct tcp_rack *)tp->t_fb_ptr;
7043         rack->t_timers_stopped = 1;
7044         return (0);
7045 }
7046
7047 static void
7048 rack_timer_activate(struct tcpcb *tp, uint32_t timer_type, uint32_t delta)
7049 {
7050         return;
7051 }
7052
7053 static int
7054 rack_timer_active(struct tcpcb *tp, uint32_t timer_type)
7055 {
7056         return (0);
7057 }
7058
7059 static void
7060 rack_stop_all_timers(struct tcpcb *tp)
7061 {
7062         struct tcp_rack *rack;
7063
7064         /*
7065          * Assure no timers are running.
7066          */
7067         if (tcp_timer_active(tp, TT_PERSIST)) {
7068                 /* We enter in persists, set the flag appropriately */
7069                 rack = (struct tcp_rack *)tp->t_fb_ptr;
7070                 rack->rc_in_persist = 1;
7071         }
7072         tcp_timer_suspend(tp, TT_PERSIST);
7073         tcp_timer_suspend(tp, TT_REXMT);
7074         tcp_timer_suspend(tp, TT_KEEP);
7075         tcp_timer_suspend(tp, TT_DELACK);
7076 }
7077
7078 static void
7079 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
7080     struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag)
7081 {
7082         int32_t idx;
7083         uint16_t stripped_flags;
7084
7085         rsm->r_rtr_cnt++;
7086         rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7087         rsm->r_dupack = 0;
7088         if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) {
7089                 rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS;
7090                 rsm->r_flags |= RACK_OVERMAX;
7091         }
7092         if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & RACK_TLP) == 0)) {
7093                 rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
7094                 rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
7095         }
7096         idx = rsm->r_rtr_cnt - 1;
7097         rsm->r_tim_lastsent[idx] = ts;
7098         stripped_flags = rsm->r_flags & ~(RACK_SENT_SP|RACK_SENT_FP);
7099         if (rsm->r_flags & RACK_ACKED) {
7100                 /* Problably MTU discovery messing with us */
7101                 rsm->r_flags &= ~RACK_ACKED;
7102                 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
7103         }
7104         if (rsm->r_in_tmap) {
7105                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7106                 rsm->r_in_tmap = 0;
7107         }
7108         TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7109         rsm->r_in_tmap = 1;
7110         if (rsm->r_flags & RACK_SACK_PASSED) {
7111                 /* We have retransmitted due to the SACK pass */
7112                 rsm->r_flags &= ~RACK_SACK_PASSED;
7113                 rsm->r_flags |= RACK_WAS_SACKPASS;
7114         }
7115 }
7116
7117 static uint32_t
7118 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
7119     struct rack_sendmap *rsm, uint64_t ts, int32_t *lenp, uint16_t add_flag)
7120 {
7121         /*
7122          * We (re-)transmitted starting at rsm->r_start for some length
7123          * (possibly less than r_end.
7124          */
7125         struct rack_sendmap *nrsm, *insret;
7126         uint32_t c_end;
7127         int32_t len;
7128
7129         len = *lenp;
7130         c_end = rsm->r_start + len;
7131         if (SEQ_GEQ(c_end, rsm->r_end)) {
7132                 /*
7133                  * We retransmitted the whole piece or more than the whole
7134                  * slopping into the next rsm.
7135                  */
7136                 rack_update_rsm(tp, rack, rsm, ts, add_flag);
7137                 if (c_end == rsm->r_end) {
7138                         *lenp = 0;
7139                         return (0);
7140                 } else {
7141                         int32_t act_len;
7142
7143                         /* Hangs over the end return whats left */
7144                         act_len = rsm->r_end - rsm->r_start;
7145                         *lenp = (len - act_len);
7146                         return (rsm->r_end);
7147                 }
7148                 /* We don't get out of this block. */
7149         }
7150         /*
7151          * Here we retransmitted less than the whole thing which means we
7152          * have to split this into what was transmitted and what was not.
7153          */
7154         nrsm = rack_alloc_full_limit(rack);
7155         if (nrsm == NULL) {
7156                 /*
7157                  * We can't get memory, so lets not proceed.
7158                  */
7159                 *lenp = 0;
7160                 return (0);
7161         }
7162         /*
7163          * So here we are going to take the original rsm and make it what we
7164          * retransmitted. nrsm will be the tail portion we did not
7165          * retransmit. For example say the chunk was 1, 11 (10 bytes). And
7166          * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
7167          * 1, 6 and the new piece will be 6, 11.
7168          */
7169         rack_clone_rsm(rack, nrsm, rsm, c_end);
7170         nrsm->r_dupack = 0;
7171         rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
7172         insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7173 #ifdef INVARIANTS
7174         if (insret != NULL) {
7175                 panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7176                       nrsm, insret, rack, rsm);
7177         }
7178 #endif
7179         if (rsm->r_in_tmap) {
7180                 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7181                 nrsm->r_in_tmap = 1;
7182         }
7183         rsm->r_flags &= (~RACK_HAS_FIN);
7184         rack_update_rsm(tp, rack, rsm, ts, add_flag);
7185         /* Log a split of rsm into rsm and nrsm */
7186         rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7187         *lenp = 0;
7188         return (0);
7189 }
7190
7191 static void
7192 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
7193                 uint32_t seq_out, uint8_t th_flags, int32_t err, uint64_t cts,
7194                 struct rack_sendmap *hintrsm, uint16_t add_flag, struct mbuf *s_mb, uint32_t s_moff)
7195 {
7196         struct tcp_rack *rack;
7197         struct rack_sendmap *rsm, *nrsm, *insret, fe;
7198         register uint32_t snd_max, snd_una;
7199
7200         /*
7201          * Add to the RACK log of packets in flight or retransmitted. If
7202          * there is a TS option we will use the TS echoed, if not we will
7203          * grab a TS.
7204          *
7205          * Retransmissions will increment the count and move the ts to its
7206          * proper place. Note that if options do not include TS's then we
7207          * won't be able to effectively use the ACK for an RTT on a retran.
7208          *
7209          * Notes about r_start and r_end. Lets consider a send starting at
7210          * sequence 1 for 10 bytes. In such an example the r_start would be
7211          * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
7212          * This means that r_end is actually the first sequence for the next
7213          * slot (11).
7214          *
7215          */
7216         /*
7217          * If err is set what do we do XXXrrs? should we not add the thing?
7218          * -- i.e. return if err != 0 or should we pretend we sent it? --
7219          * i.e. proceed with add ** do this for now.
7220          */
7221         INP_WLOCK_ASSERT(tp->t_inpcb);
7222         if (err)
7223                 /*
7224                  * We don't log errors -- we could but snd_max does not
7225                  * advance in this case either.
7226                  */
7227                 return;
7228
7229         if (th_flags & TH_RST) {
7230                 /*
7231                  * We don't log resets and we return immediately from
7232                  * sending
7233                  */
7234                 return;
7235         }
7236         rack = (struct tcp_rack *)tp->t_fb_ptr;
7237         snd_una = tp->snd_una;
7238         snd_max = tp->snd_max;
7239         if (th_flags & (TH_SYN | TH_FIN)) {
7240                 /*
7241                  * The call to rack_log_output is made before bumping
7242                  * snd_max. This means we can record one extra byte on a SYN
7243                  * or FIN if seq_out is adding more on and a FIN is present
7244                  * (and we are not resending).
7245                  */
7246                 if ((th_flags & TH_SYN) && (seq_out == tp->iss))
7247                         len++;
7248                 if (th_flags & TH_FIN)
7249                         len++;
7250                 if (SEQ_LT(snd_max, tp->snd_nxt)) {
7251                         /*
7252                          * The add/update as not been done for the FIN/SYN
7253                          * yet.
7254                          */
7255                         snd_max = tp->snd_nxt;
7256                 }
7257         }
7258         if (SEQ_LEQ((seq_out + len), snd_una)) {
7259                 /* Are sending an old segment to induce an ack (keep-alive)? */
7260                 return;
7261         }
7262         if (SEQ_LT(seq_out, snd_una)) {
7263                 /* huh? should we panic? */
7264                 uint32_t end;
7265
7266                 end = seq_out + len;
7267                 seq_out = snd_una;
7268                 if (SEQ_GEQ(end, seq_out))
7269                         len = end - seq_out;
7270                 else
7271                         len = 0;
7272         }
7273         if (len == 0) {
7274                 /* We don't log zero window probes */
7275                 return;
7276         }
7277         rack->r_ctl.rc_time_last_sent = cts;
7278         if (IN_FASTRECOVERY(tp->t_flags)) {
7279                 rack->r_ctl.rc_prr_out += len;
7280         }
7281         /* First question is it a retransmission or new? */
7282         if (seq_out == snd_max) {
7283                 /* Its new */
7284 again:
7285                 rsm = rack_alloc(rack);
7286                 if (rsm == NULL) {
7287                         /*
7288                          * Hmm out of memory and the tcb got destroyed while
7289                          * we tried to wait.
7290                          */
7291                         return;
7292                 }
7293                 if (th_flags & TH_FIN) {
7294                         rsm->r_flags = RACK_HAS_FIN|add_flag;
7295                 } else {
7296                         rsm->r_flags = add_flag;
7297                 }
7298                 rsm->r_tim_lastsent[0] = cts;
7299                 rsm->r_rtr_cnt = 1;
7300                 rsm->r_rtr_bytes = 0;
7301                 if (th_flags & TH_SYN) {
7302                         /* The data space is one beyond snd_una */
7303                         rsm->r_flags |= RACK_HAS_SYN;
7304                 }
7305                 rsm->r_start = seq_out;
7306                 rsm->r_end = rsm->r_start + len;
7307                 rsm->r_dupack = 0;
7308                 /*
7309                  * save off the mbuf location that
7310                  * sndmbuf_noadv returned (which is
7311                  * where we started copying from)..
7312                  */
7313                 rsm->m = s_mb;
7314                 rsm->soff = s_moff;
7315                 /* rsm->m will be NULL if RACK_HAS_SYN or RACK_HAS_FIN is set */
7316                 if (rsm->m) {
7317                         if (rsm->m->m_len <= rsm->soff) {
7318                                 /*
7319                                  * XXXrrs Question, will this happen?
7320                                  *
7321                                  * If sbsndptr is set at the correct place
7322                                  * then s_moff should always be somewhere
7323                                  * within rsm->m. But if the sbsndptr was
7324                                  * off then that won't be true. If it occurs
7325                                  * we need to walkout to the correct location.
7326                                  */
7327                                 struct mbuf *lm;
7328
7329                                 lm = rsm->m;
7330                                 while (lm->m_len <= rsm->soff) {
7331                                         rsm->soff -= lm->m_len;
7332                                         lm = lm->m_next;
7333                                         KASSERT(lm != NULL, ("%s rack:%p lm goes null orig_off:%u origmb:%p rsm->soff:%u",
7334                                                              __func__, rack, s_moff, s_mb, rsm->soff));
7335                                 }
7336                                 rsm->m = lm;
7337                                 counter_u64_add(rack_sbsndptr_wrong, 1);
7338                         } else
7339                                 counter_u64_add(rack_sbsndptr_right, 1);
7340                         rsm->orig_m_len = rsm->m->m_len;
7341                 } else
7342                         rsm->orig_m_len = 0;
7343                 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7344                 /* Log a new rsm */
7345                 rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_NEW, 0, __LINE__);
7346                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7347 #ifdef INVARIANTS
7348                 if (insret != NULL) {
7349                         panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7350                               nrsm, insret, rack, rsm);
7351                 }
7352 #endif
7353                 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7354                 rsm->r_in_tmap = 1;
7355                 /*
7356                  * Special case detection, is there just a single
7357                  * packet outstanding when we are not in recovery?
7358                  *
7359                  * If this is true mark it so.
7360                  */
7361                 if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
7362                     (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) == ctf_fixed_maxseg(tp))) {
7363                         struct rack_sendmap *prsm;
7364
7365                         prsm = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7366                         if (prsm)
7367                                 prsm->r_one_out_nr = 1;
7368                 }
7369                 return;
7370         }
7371         /*
7372          * If we reach here its a retransmission and we need to find it.
7373          */
7374         memset(&fe, 0, sizeof(fe));
7375 more:
7376         if (hintrsm && (hintrsm->r_start == seq_out)) {
7377                 rsm = hintrsm;
7378                 hintrsm = NULL;
7379         } else {
7380                 /* No hints sorry */
7381                 rsm = NULL;
7382         }
7383         if ((rsm) && (rsm->r_start == seq_out)) {
7384                 seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag);
7385                 if (len == 0) {
7386                         return;
7387                 } else {
7388                         goto more;
7389                 }
7390         }
7391         /* Ok it was not the last pointer go through it the hard way. */
7392 refind:
7393         fe.r_start = seq_out;
7394         rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
7395         if (rsm) {
7396                 if (rsm->r_start == seq_out) {
7397                         seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag);
7398                         if (len == 0) {
7399                                 return;
7400                         } else {
7401                                 goto refind;
7402                         }
7403                 }
7404                 if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
7405                         /* Transmitted within this piece */
7406                         /*
7407                          * Ok we must split off the front and then let the
7408                          * update do the rest
7409                          */
7410                         nrsm = rack_alloc_full_limit(rack);
7411                         if (nrsm == NULL) {
7412                                 rack_update_rsm(tp, rack, rsm, cts, add_flag);
7413                                 return;
7414                         }
7415                         /*
7416                          * copy rsm to nrsm and then trim the front of rsm
7417                          * to not include this part.
7418                          */
7419                         rack_clone_rsm(rack, nrsm, rsm, seq_out);
7420                         insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7421                         rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7422 #ifdef INVARIANTS
7423                         if (insret != NULL) {
7424                                 panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7425                                       nrsm, insret, rack, rsm);
7426                         }
7427 #endif
7428                         if (rsm->r_in_tmap) {
7429                                 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7430                                 nrsm->r_in_tmap = 1;
7431                         }
7432                         rsm->r_flags &= (~RACK_HAS_FIN);
7433                         seq_out = rack_update_entry(tp, rack, nrsm, cts, &len, add_flag);
7434                         if (len == 0) {
7435                                 return;
7436                         } else if (len > 0)
7437                                 goto refind;
7438                 }
7439         }
7440         /*
7441          * Hmm not found in map did they retransmit both old and on into the
7442          * new?
7443          */
7444         if (seq_out == tp->snd_max) {
7445                 goto again;
7446         } else if (SEQ_LT(seq_out, tp->snd_max)) {
7447 #ifdef INVARIANTS
7448                 printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
7449                        seq_out, len, tp->snd_una, tp->snd_max);
7450                 printf("Starting Dump of all rack entries\n");
7451                 RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
7452                         printf("rsm:%p start:%u end:%u\n",
7453                                rsm, rsm->r_start, rsm->r_end);
7454                 }
7455                 printf("Dump complete\n");
7456                 panic("seq_out not found rack:%p tp:%p",
7457                       rack, tp);
7458 #endif
7459         } else {
7460 #ifdef INVARIANTS
7461                 /*
7462                  * Hmm beyond sndmax? (only if we are using the new rtt-pack
7463                  * flag)
7464                  */
7465                 panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
7466                       seq_out, len, tp->snd_max, tp);
7467 #endif
7468         }
7469 }
7470
7471 /*
7472  * Record one of the RTT updates from an ack into
7473  * our sample structure.
7474  */
7475
7476 static void
7477 tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt, uint32_t len, uint32_t us_rtt,
7478                     int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt)
7479 {
7480         if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7481             (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) {
7482                 rack->r_ctl.rack_rs.rs_rtt_lowest = rtt;
7483         }
7484         if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7485             (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) {
7486                 rack->r_ctl.rack_rs.rs_rtt_highest = rtt;
7487         }
7488         if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
7489             if (us_rtt < rack->r_ctl.rc_gp_lowrtt)
7490                 rack->r_ctl.rc_gp_lowrtt = us_rtt;
7491             if (rack->rc_tp->snd_wnd > rack->r_ctl.rc_gp_high_rwnd)
7492                     rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
7493         }
7494         if ((confidence == 1) &&
7495             ((rsm == NULL) ||
7496              (rsm->r_just_ret) ||
7497              (rsm->r_one_out_nr &&
7498               len < (ctf_fixed_maxseg(rack->rc_tp) * 2)))) {
7499                 /*
7500                  * If the rsm had a just return
7501                  * hit it then we can't trust the
7502                  * rtt measurement for buffer deterimination
7503                  * Note that a confidence of 2, indicates
7504                  * SACK'd which overrides the r_just_ret or
7505                  * the r_one_out_nr. If it was a CUM-ACK and
7506                  * we had only two outstanding, but get an
7507                  * ack for only 1. Then that also lowers our
7508                  * confidence.
7509                  */
7510                 confidence = 0;
7511         }
7512         if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7513             (rack->r_ctl.rack_rs.rs_us_rtt > us_rtt)) {
7514                 if (rack->r_ctl.rack_rs.confidence == 0) {
7515                         /*
7516                          * We take anything with no current confidence
7517                          * saved.
7518                          */
7519                         rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
7520                         rack->r_ctl.rack_rs.confidence = confidence;
7521                         rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
7522                 } else if (confidence || rack->r_ctl.rack_rs.confidence) {
7523                         /*
7524                          * Once we have a confident number,
7525                          * we can update it with a smaller
7526                          * value since this confident number
7527                          * may include the DSACK time until
7528                          * the next segment (the second one) arrived.
7529                          */
7530                         rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
7531                         rack->r_ctl.rack_rs.confidence = confidence;
7532                         rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
7533                 }
7534         }
7535         rack_log_rtt_upd(rack->rc_tp, rack, us_rtt, len, rsm, confidence);
7536         rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID;
7537         rack->r_ctl.rack_rs.rs_rtt_tot += rtt;
7538         rack->r_ctl.rack_rs.rs_rtt_cnt++;
7539 }
7540
7541 /*
7542  * Collect new round-trip time estimate
7543  * and update averages and current timeout.
7544  */
7545 static void
7546 tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp)
7547 {
7548         int32_t delta;
7549         uint32_t o_srtt, o_var;
7550         int32_t hrtt_up = 0;
7551         int32_t rtt;
7552
7553         if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY)
7554                 /* No valid sample */
7555                 return;
7556         if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) {
7557                 /* We are to use the lowest RTT seen in a single ack */
7558                 rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
7559         } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) {
7560                 /* We are to use the highest RTT seen in a single ack */
7561                 rtt = rack->r_ctl.rack_rs.rs_rtt_highest;
7562         } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) {
7563                 /* We are to use the average RTT seen in a single ack */
7564                 rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot /
7565                                 (uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt);
7566         } else {
7567 #ifdef INVARIANTS
7568                 panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method);
7569 #endif
7570                 return;
7571         }
7572         if (rtt == 0)
7573                 rtt = 1;
7574         if (rack->rc_gp_rtt_set == 0) {
7575                 /*
7576                  * With no RTT we have to accept
7577                  * even one we are not confident of.
7578                  */
7579                 rack->r_ctl.rc_gp_srtt = rack->r_ctl.rack_rs.rs_us_rtt;
7580                 rack->rc_gp_rtt_set = 1;
7581         } else if (rack->r_ctl.rack_rs.confidence) {
7582                 /* update the running gp srtt */
7583                 rack->r_ctl.rc_gp_srtt -= (rack->r_ctl.rc_gp_srtt/8);
7584                 rack->r_ctl.rc_gp_srtt += rack->r_ctl.rack_rs.rs_us_rtt / 8;
7585         }
7586         if (rack->r_ctl.rack_rs.confidence) {
7587                 /*
7588                  * record the low and high for highly buffered path computation,
7589                  * we only do this if we are confident (not a retransmission).
7590                  */
7591                 if (rack->r_ctl.rc_highest_us_rtt < rack->r_ctl.rack_rs.rs_us_rtt) {
7592                         rack->r_ctl.rc_highest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7593                         hrtt_up = 1;
7594                 }
7595                 if (rack->rc_highly_buffered == 0) {
7596                         /*
7597                          * Currently once we declare a path has
7598                          * highly buffered there is no going
7599                          * back, which may be a problem...
7600                          */
7601                         if ((rack->r_ctl.rc_highest_us_rtt / rack->r_ctl.rc_lowest_us_rtt) > rack_hbp_thresh) {
7602                                 rack_log_rtt_shrinks(rack, rack->r_ctl.rack_rs.rs_us_rtt,
7603                                                      rack->r_ctl.rc_highest_us_rtt,
7604                                                      rack->r_ctl.rc_lowest_us_rtt,
7605                                                      RACK_RTTS_SEEHBP);
7606                                 rack->rc_highly_buffered = 1;
7607                         }
7608                 }
7609         }
7610         if ((rack->r_ctl.rack_rs.confidence) ||
7611             (rack->r_ctl.rack_rs.rs_us_rtrcnt == 1)) {
7612                 /*
7613                  * If we are highly confident of it <or> it was
7614                  * never retransmitted we accept it as the last us_rtt.
7615                  */
7616                 rack->r_ctl.rc_last_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7617                 /* The lowest rtt can be set if its was not retransmited */
7618                 if (rack->r_ctl.rc_lowest_us_rtt > rack->r_ctl.rack_rs.rs_us_rtt) {
7619                         rack->r_ctl.rc_lowest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7620                         if (rack->r_ctl.rc_lowest_us_rtt == 0)
7621                                 rack->r_ctl.rc_lowest_us_rtt = 1;
7622                 }
7623         }
7624         o_srtt = tp->t_srtt;
7625         o_var = tp->t_rttvar;
7626         rack = (struct tcp_rack *)tp->t_fb_ptr;
7627         if (tp->t_srtt != 0) {
7628                 /*
7629                  * We keep a simple srtt in microseconds, like our rtt
7630                  * measurement. We don't need to do any tricks with shifting
7631                  * etc. Instead we just add in 1/8th of the new measurement
7632                  * and subtract out 1/8 of the old srtt. We do the same with
7633                  * the variance after finding the absolute value of the
7634                  * difference between this sample and the current srtt.
7635                  */
7636                 delta = tp->t_srtt - rtt;
7637                 /* Take off 1/8th of the current sRTT */
7638                 tp->t_srtt -= (tp->t_srtt >> 3);
7639                 /* Add in 1/8th of the new RTT just measured */
7640                 tp->t_srtt += (rtt >> 3);
7641                 if (tp->t_srtt <= 0)
7642                         tp->t_srtt = 1;
7643                 /* Now lets make the absolute value of the variance */
7644                 if (delta < 0)
7645                         delta = -delta;
7646                 /* Subtract out 1/8th */
7647                 tp->t_rttvar -= (tp->t_rttvar >> 3);
7648                 /* Add in 1/8th of the new variance we just saw */
7649                 tp->t_rttvar += (delta >> 3);
7650                 if (tp->t_rttvar <= 0)
7651                         tp->t_rttvar = 1;
7652                 if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar)
7653                         tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
7654         } else {
7655                 /*
7656                  * No rtt measurement yet - use the unsmoothed rtt. Set the
7657                  * variance to half the rtt (so our first retransmit happens
7658                  * at 3*rtt).
7659                  */
7660                 tp->t_srtt = rtt;
7661                 tp->t_rttvar = rtt >> 1;
7662                 tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
7663         }
7664         rack->rc_srtt_measure_made = 1;
7665         KMOD_TCPSTAT_INC(tcps_rttupdated);
7666         tp->t_rttupdated++;
7667 #ifdef STATS
7668         if (rack_stats_gets_ms_rtt == 0) {
7669                 /* Send in the microsecond rtt used for rxt timeout purposes */
7670                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt));
7671         } else if (rack_stats_gets_ms_rtt == 1) {
7672                 /* Send in the millisecond rtt used for rxt timeout purposes */
7673                 int32_t ms_rtt;
7674
7675                 /* Round up */
7676                 ms_rtt = (rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
7677                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
7678         } else if (rack_stats_gets_ms_rtt == 2) {
7679                 /* Send in the millisecond rtt has close to the path RTT as we can get  */
7680                 int32_t ms_rtt;
7681
7682                 /* Round up */
7683                 ms_rtt = (rack->r_ctl.rack_rs.rs_us_rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
7684                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
7685         }  else {
7686                 /* Send in the microsecond rtt has close to the path RTT as we can get  */
7687                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
7688         }
7689
7690 #endif
7691         /*
7692          * the retransmit should happen at rtt + 4 * rttvar. Because of the
7693          * way we do the smoothing, srtt and rttvar will each average +1/2
7694          * tick of bias.  When we compute the retransmit timer, we want 1/2
7695          * tick of rounding and 1 extra tick because of +-1/2 tick
7696          * uncertainty in the firing of the timer.  The bias will give us
7697          * exactly the 1.5 tick we need.  But, because the bias is
7698          * statistical, we have to test that we don't drop below the minimum
7699          * feasible timer (which is 2 ticks).
7700          */
7701         tp->t_rxtshift = 0;
7702         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
7703                       max(rack_rto_min, rtt + 2), rack_rto_max, rack->r_ctl.timer_slop);
7704         rack_log_rtt_sample(rack, rtt);
7705         tp->t_softerror = 0;
7706 }
7707
7708
7709 static void
7710 rack_apply_updated_usrtt(struct tcp_rack *rack, uint32_t us_rtt, uint32_t us_cts)
7711 {
7712         /*
7713          * Apply to filter the inbound us-rtt at us_cts.
7714          */
7715         uint32_t old_rtt;
7716
7717         old_rtt = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
7718         apply_filter_min_small(&rack->r_ctl.rc_gp_min_rtt,
7719                                us_rtt, us_cts);
7720         if (rack->r_ctl.last_pacing_time &&
7721             rack->rc_gp_dyn_mul &&
7722             (rack->r_ctl.last_pacing_time > us_rtt))
7723                 rack->pacing_longer_than_rtt = 1;
7724         else
7725                 rack->pacing_longer_than_rtt = 0;
7726         if (old_rtt > us_rtt) {
7727                 /* We just hit a new lower rtt time */
7728                 rack_log_rtt_shrinks(rack,  us_cts,  old_rtt,
7729                                      __LINE__, RACK_RTTS_NEWRTT);
7730                 /*
7731                  * Only count it if its lower than what we saw within our
7732                  * calculated range.
7733                  */
7734                 if ((old_rtt - us_rtt) > rack_min_rtt_movement) {
7735                         if (rack_probertt_lower_within &&
7736                             rack->rc_gp_dyn_mul &&
7737                             (rack->use_fixed_rate == 0) &&
7738                             (rack->rc_always_pace)) {
7739                                 /*
7740                                  * We are seeing a new lower rtt very close
7741                                  * to the time that we would have entered probe-rtt.
7742                                  * This is probably due to the fact that a peer flow
7743                                  * has entered probe-rtt. Lets go in now too.
7744                                  */
7745                                 uint32_t val;
7746
7747                                 val = rack_probertt_lower_within * rack_time_between_probertt;
7748                                 val /= 100;
7749                                 if ((rack->in_probe_rtt == 0)  &&
7750                                     ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= (rack_time_between_probertt - val))) {
7751                                         rack_enter_probertt(rack, us_cts);
7752                                 }
7753                         }
7754                         rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
7755                 }
7756         }
7757 }
7758
7759 static int
7760 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
7761     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack)
7762 {
7763         int32_t i, all;
7764         uint32_t t, len_acked;
7765
7766         if ((rsm->r_flags & RACK_ACKED) ||
7767             (rsm->r_flags & RACK_WAS_ACKED))
7768                 /* Already done */
7769                 return (0);
7770         if (rsm->r_no_rtt_allowed) {
7771                 /* Not allowed */
7772                 return (0);
7773         }
7774         if (ack_type == CUM_ACKED) {
7775                 if (SEQ_GT(th_ack, rsm->r_end)) {
7776                         len_acked = rsm->r_end - rsm->r_start;
7777                         all = 1;
7778                 } else {
7779                         len_acked = th_ack - rsm->r_start;
7780                         all = 0;
7781                 }
7782         } else {
7783                 len_acked = rsm->r_end - rsm->r_start;
7784                 all = 0;
7785         }
7786         if (rsm->r_rtr_cnt == 1) {
7787                 uint32_t us_rtt;
7788
7789                 t = cts - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7790                 if ((int)t <= 0)
7791                         t = 1;
7792                 if (!tp->t_rttlow || tp->t_rttlow > t)
7793                         tp->t_rttlow = t;
7794                 if (!rack->r_ctl.rc_rack_min_rtt ||
7795                     SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7796                         rack->r_ctl.rc_rack_min_rtt = t;
7797                         if (rack->r_ctl.rc_rack_min_rtt == 0) {
7798                                 rack->r_ctl.rc_rack_min_rtt = 1;
7799                         }
7800                 }
7801                 if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]))
7802                         us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
7803                 else
7804                         us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
7805                 if (us_rtt == 0)
7806                         us_rtt = 1;
7807                 rack_apply_updated_usrtt(rack, us_rtt, tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
7808                 if (ack_type == SACKED) {
7809                         rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 1);
7810                         tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt, 2 , rsm, rsm->r_rtr_cnt);
7811                 } else {
7812                         /*
7813                          * We need to setup what our confidence
7814                          * is in this ack.
7815                          *
7816                          * If the rsm was app limited and it is
7817                          * less than a mss in length (the end
7818                          * of the send) then we have a gap. If we
7819                          * were app limited but say we were sending
7820                          * multiple MSS's then we are more confident
7821                          * int it.
7822                          *
7823                          * When we are not app-limited then we see if
7824                          * the rsm is being included in the current
7825                          * measurement, we tell this by the app_limited_needs_set
7826                          * flag.
7827                          *
7828                          * Note that being cwnd blocked is not applimited
7829                          * as well as the pacing delay between packets which
7830                          * are sending only 1 or 2 MSS's also will show up
7831                          * in the RTT. We probably need to examine this algorithm
7832                          * a bit more and enhance it to account for the delay
7833                          * between rsm's. We could do that by saving off the
7834                          * pacing delay of each rsm (in an rsm) and then
7835                          * factoring that in somehow though for now I am
7836                          * not sure how :)
7837                          */
7838                         int calc_conf = 0;
7839
7840                         if (rsm->r_flags & RACK_APP_LIMITED) {
7841                                 if (all && (len_acked <= ctf_fixed_maxseg(tp)))
7842                                         calc_conf = 0;
7843                                 else
7844                                         calc_conf = 1;
7845                         } else if (rack->app_limited_needs_set == 0) {
7846                                 calc_conf = 1;
7847                         } else {
7848                                 calc_conf = 0;
7849                         }
7850                         rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 2);
7851                         tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt,
7852                                             calc_conf, rsm, rsm->r_rtr_cnt);
7853                 }
7854                 if ((rsm->r_flags & RACK_TLP) &&
7855                     (!IN_FASTRECOVERY(tp->t_flags))) {
7856                         /* Segment was a TLP and our retrans matched */
7857                         if (rack->r_ctl.rc_tlp_cwnd_reduce) {
7858                                 rack->r_ctl.rc_rsm_start = tp->snd_max;
7859                                 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
7860                                 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
7861                                 rack_cong_signal(tp, CC_NDUPACK, tp->snd_una);
7862                         }
7863                 }
7864                 if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
7865                         /* New more recent rack_tmit_time */
7866                         rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7867                         rack->rc_rack_rtt = t;
7868                 }
7869                 return (1);
7870         }
7871         /*
7872          * We clear the soft/rxtshift since we got an ack.
7873          * There is no assurance we will call the commit() function
7874          * so we need to clear these to avoid incorrect handling.
7875          */
7876         tp->t_rxtshift = 0;
7877         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
7878                       rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
7879         tp->t_softerror = 0;
7880         if (to && (to->to_flags & TOF_TS) &&
7881             (ack_type == CUM_ACKED) &&
7882             (to->to_tsecr) &&
7883             ((rsm->r_flags & RACK_OVERMAX) == 0)) {
7884                 /*
7885                  * Now which timestamp does it match? In this block the ACK
7886                  * must be coming from a previous transmission.
7887                  */
7888                 for (i = 0; i < rsm->r_rtr_cnt; i++) {
7889                         if (rack_ts_to_msec(rsm->r_tim_lastsent[i]) == to->to_tsecr) {
7890                                 t = cts - (uint32_t)rsm->r_tim_lastsent[i];
7891                                 if ((int)t <= 0)
7892                                         t = 1;
7893                                 if ((i + 1) < rsm->r_rtr_cnt) {
7894                                         /*
7895                                          * The peer ack'd from our previous
7896                                          * transmission. We have a spurious
7897                                          * retransmission and thus we dont
7898                                          * want to update our rack_rtt.
7899                                          */
7900                                         return (0);
7901                                 }
7902                                 if (!tp->t_rttlow || tp->t_rttlow > t)
7903                                         tp->t_rttlow = t;
7904                                 if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7905                                         rack->r_ctl.rc_rack_min_rtt = t;
7906                                         if (rack->r_ctl.rc_rack_min_rtt == 0) {
7907                                                 rack->r_ctl.rc_rack_min_rtt = 1;
7908                                         }
7909                                 }
7910                                 if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
7911                                            (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
7912                                         /* New more recent rack_tmit_time */
7913                                         rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7914                                         rack->rc_rack_rtt = t;
7915                                 }
7916                                 rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[i], cts, 3);
7917                                 tcp_rack_xmit_timer(rack, t + 1, len_acked, t, 0, rsm,
7918                                                     rsm->r_rtr_cnt);
7919                                 return (1);
7920                         }
7921                 }
7922                 goto ts_not_found;
7923         } else {
7924                 /*
7925                  * Ok its a SACK block that we retransmitted. or a windows
7926                  * machine without timestamps. We can tell nothing from the
7927                  * time-stamp since its not there or the time the peer last
7928                  * recieved a segment that moved forward its cum-ack point.
7929                  */
7930 ts_not_found:
7931                 i = rsm->r_rtr_cnt - 1;
7932                 t = cts - (uint32_t)rsm->r_tim_lastsent[i];
7933                 if ((int)t <= 0)
7934                         t = 1;
7935                 if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7936                         /*
7937                          * We retransmitted and the ack came back in less
7938                          * than the smallest rtt we have observed. We most
7939                          * likely did an improper retransmit as outlined in
7940                          * 6.2 Step 2 point 2 in the rack-draft so we
7941                          * don't want to update our rack_rtt. We in
7942                          * theory (in future) might want to think about reverting our
7943                          * cwnd state but we won't for now.
7944                          */
7945                         return (0);
7946                 } else if (rack->r_ctl.rc_rack_min_rtt) {
7947                         /*
7948                          * We retransmitted it and the retransmit did the
7949                          * job.
7950                          */
7951                         if (!rack->r_ctl.rc_rack_min_rtt ||
7952                             SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7953                                 rack->r_ctl.rc_rack_min_rtt = t;
7954                                 if (rack->r_ctl.rc_rack_min_rtt == 0) {
7955                                         rack->r_ctl.rc_rack_min_rtt = 1;
7956                                 }
7957                         }
7958                         if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[i])) {
7959                                 /* New more recent rack_tmit_time */
7960                                 rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[i];
7961                                 rack->rc_rack_rtt = t;
7962                         }
7963                         return (1);
7964                 }
7965         }
7966         return (0);
7967 }
7968
7969 /*
7970  * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
7971  */
7972 static void
7973 rack_log_sack_passed(struct tcpcb *tp,
7974     struct tcp_rack *rack, struct rack_sendmap *rsm)
7975 {
7976         struct rack_sendmap *nrsm;
7977
7978         nrsm = rsm;
7979         TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap,
7980             rack_head, r_tnext) {
7981                 if (nrsm == rsm) {
7982                         /* Skip orginal segment he is acked */
7983                         continue;
7984                 }
7985                 if (nrsm->r_flags & RACK_ACKED) {
7986                         /*
7987                          * Skip ack'd segments, though we
7988                          * should not see these, since tmap
7989                          * should not have ack'd segments.
7990                          */
7991                         continue;
7992                 }
7993                 if (nrsm->r_flags & RACK_SACK_PASSED) {
7994                         /*
7995                          * We found one that is already marked
7996                          * passed, we have been here before and
7997                          * so all others below this are marked.
7998                          */
7999                         break;
8000                 }
8001                 nrsm->r_flags |= RACK_SACK_PASSED;
8002                 nrsm->r_flags &= ~RACK_WAS_SACKPASS;
8003         }
8004 }
8005
8006 static void
8007 rack_need_set_test(struct tcpcb *tp,
8008                    struct tcp_rack *rack,
8009                    struct rack_sendmap *rsm,
8010                    tcp_seq th_ack,
8011                    int line,
8012                    int use_which)
8013 {
8014
8015         if ((tp->t_flags & TF_GPUTINPROG) &&
8016             SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
8017                 /*
8018                  * We were app limited, and this ack
8019                  * butts up or goes beyond the point where we want
8020                  * to start our next measurement. We need
8021                  * to record the new gput_ts as here and
8022                  * possibly update the start sequence.
8023                  */
8024                 uint32_t seq, ts;
8025
8026                 if (rsm->r_rtr_cnt > 1) {
8027                         /*
8028                          * This is a retransmit, can we
8029                          * really make any assessment at this
8030                          * point?  We are not really sure of
8031                          * the timestamp, is it this or the
8032                          * previous transmission?
8033                          *
8034                          * Lets wait for something better that
8035                          * is not retransmitted.
8036                          */
8037                         return;
8038                 }
8039                 seq = tp->gput_seq;
8040                 ts = tp->gput_ts;
8041                 rack->app_limited_needs_set = 0;
8042                 tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
8043                 /* Do we start at a new end? */
8044                 if ((use_which == RACK_USE_BEG) &&
8045                     SEQ_GEQ(rsm->r_start, tp->gput_seq)) {
8046                         /*
8047                          * When we get an ACK that just eats
8048                          * up some of the rsm, we set RACK_USE_BEG
8049                          * since whats at r_start (i.e. th_ack)
8050                          * is left unacked and thats where the
8051                          * measurement not starts.
8052                          */
8053                         tp->gput_seq = rsm->r_start;
8054                         rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8055                 }
8056                 if ((use_which == RACK_USE_END) &&
8057                     SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
8058                             /*
8059                              * We use the end when the cumack
8060                              * is moving forward and completely
8061                              * deleting the rsm passed so basically
8062                              * r_end holds th_ack.
8063                              *
8064                              * For SACK's we also want to use the end
8065                              * since this piece just got sacked and
8066                              * we want to target anything after that
8067                              * in our measurement.
8068                              */
8069                             tp->gput_seq = rsm->r_end;
8070                             rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8071                 }
8072                 if (use_which == RACK_USE_END_OR_THACK) {
8073                         /*
8074                          * special case for ack moving forward,
8075                          * not a sack, we need to move all the
8076                          * way up to where this ack cum-ack moves
8077                          * to.
8078                          */
8079                         if (SEQ_GT(th_ack, rsm->r_end))
8080                                 tp->gput_seq = th_ack;
8081                         else
8082                                 tp->gput_seq = rsm->r_end;
8083                         rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8084                 }
8085                 if (SEQ_GT(tp->gput_seq, tp->gput_ack)) {
8086                         /*
8087                          * We moved beyond this guy's range, re-calculate
8088                          * the new end point.
8089                          */
8090                         if (rack->rc_gp_filled == 0) {
8091                                 tp->gput_ack = tp->gput_seq + max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
8092                         } else {
8093                                 tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
8094                         }
8095                 }
8096                 /*
8097                  * We are moving the goal post, we may be able to clear the
8098                  * measure_saw_probe_rtt flag.
8099                  */
8100                 if ((rack->in_probe_rtt == 0) &&
8101                     (rack->measure_saw_probe_rtt) &&
8102                     (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
8103                         rack->measure_saw_probe_rtt = 0;
8104                 rack_log_pacing_delay_calc(rack, ts, tp->gput_ts,
8105                                            seq, tp->gput_seq, 0, 5, line, NULL);
8106                 if (rack->rc_gp_filled &&
8107                     ((tp->gput_ack - tp->gput_seq) <
8108                      max(rc_init_window(rack), (MIN_GP_WIN *
8109                                                 ctf_fixed_maxseg(tp))))) {
8110                         uint32_t ideal_amount;
8111
8112                         ideal_amount = rack_get_measure_window(tp, rack);
8113                         if (ideal_amount > sbavail(&tp->t_inpcb->inp_socket->so_snd)) {
8114                                 /*
8115                                  * There is no sense of continuing this measurement
8116                                  * because its too small to gain us anything we
8117                                  * trust. Skip it and that way we can start a new
8118                                  * measurement quicker.
8119                                  */
8120                                 tp->t_flags &= ~TF_GPUTINPROG;
8121                                 rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
8122                                                            0, 0, 0, 6, __LINE__, NULL);
8123                         } else {
8124                                 /*
8125                                  * Reset the window further out.
8126                                  */
8127                                 tp->gput_ack = tp->gput_seq + ideal_amount;
8128                         }
8129                 }
8130         }
8131 }
8132
8133 static uint32_t
8134 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack,
8135                    struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts, int *moved_two)
8136 {
8137         uint32_t start, end, changed = 0;
8138         struct rack_sendmap stack_map;
8139         struct rack_sendmap *rsm, *nrsm, fe, *insret, *prev, *next;
8140         int32_t used_ref = 1;
8141         int moved = 0;
8142
8143         start = sack->start;
8144         end = sack->end;
8145         rsm = *prsm;
8146         memset(&fe, 0, sizeof(fe));
8147 do_rest_ofb:
8148         if ((rsm == NULL) ||
8149             (SEQ_LT(end, rsm->r_start)) ||
8150             (SEQ_GEQ(start, rsm->r_end)) ||
8151             (SEQ_LT(start, rsm->r_start))) {
8152                 /*
8153                  * We are not in the right spot,
8154                  * find the correct spot in the tree.
8155                  */
8156                 used_ref = 0;
8157                 fe.r_start = start;
8158                 rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
8159                 moved++;
8160         }
8161         if (rsm == NULL) {
8162                 /* TSNH */
8163                 goto out;
8164         }
8165         /* Ok we have an ACK for some piece of this rsm */
8166         if (rsm->r_start != start) {
8167                 if ((rsm->r_flags & RACK_ACKED) == 0) {
8168                         /**
8169                          * Need to split this in two pieces the before and after,
8170                          * the before remains in the map, the after must be
8171                          * added. In other words we have:
8172                          * rsm        |--------------|
8173                          * sackblk        |------->
8174                          * rsm will become
8175                          *     rsm    |---|
8176                          * and nrsm will be  the sacked piece
8177                          *     nrsm       |----------|
8178                          *
8179                          * But before we start down that path lets
8180                          * see if the sack spans over on top of
8181                          * the next guy and it is already sacked.
8182                          */
8183                         next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8184                         if (next && (next->r_flags & RACK_ACKED) &&
8185                             SEQ_GEQ(end, next->r_start)) {
8186                                 /**
8187                                  * So the next one is already acked, and
8188                                  * we can thus by hookery use our stack_map
8189                                  * to reflect the piece being sacked and
8190                                  * then adjust the two tree entries moving
8191                                  * the start and ends around. So we start like:
8192                                  *  rsm     |------------|             (not-acked)
8193                                  *  next                 |-----------| (acked)
8194                                  *  sackblk        |-------->
8195                                  *  We want to end like so:
8196                                  *  rsm     |------|                   (not-acked)
8197                                  *  next           |-----------------| (acked)
8198                                  *  nrsm           |-----|
8199                                  * Where nrsm is a temporary stack piece we
8200                                  * use to update all the gizmos.
8201                                  */
8202                                 /* Copy up our fudge block */
8203                                 nrsm = &stack_map;
8204                                 memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
8205                                 /* Now adjust our tree blocks */
8206                                 rsm->r_end = start;
8207                                 next->r_start = start;
8208                                 /* Now we must adjust back where next->m is */
8209                                 rack_setup_offset_for_rsm(rsm, next);
8210
8211                                 /* We don't need to adjust rsm, it did not change */
8212                                 /* Clear out the dup ack count of the remainder */
8213                                 rsm->r_dupack = 0;
8214                                 rsm->r_just_ret = 0;
8215                                 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8216                                 /* Now lets make sure our fudge block is right */
8217                                 nrsm->r_start = start;
8218                                 /* Now lets update all the stats and such */
8219                                 rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
8220                                 if (rack->app_limited_needs_set)
8221                                         rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
8222                                 changed += (nrsm->r_end - nrsm->r_start);
8223                                 rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
8224                                 if (nrsm->r_flags & RACK_SACK_PASSED) {
8225                                         counter_u64_add(rack_reorder_seen, 1);
8226                                         rack->r_ctl.rc_reorder_ts = cts;
8227                                 }
8228                                 /*
8229                                  * Now we want to go up from rsm (the
8230                                  * one left un-acked) to the next one
8231                                  * in the tmap. We do this so when
8232                                  * we walk backwards we include marking
8233                                  * sack-passed on rsm (The one passed in
8234                                  * is skipped since it is generally called
8235                                  * on something sacked before removing it
8236                                  * from the tmap).
8237                                  */
8238                                 if (rsm->r_in_tmap) {
8239                                         nrsm = TAILQ_NEXT(rsm, r_tnext);
8240                                         /*
8241                                          * Now that we have the next
8242                                          * one walk backwards from there.
8243                                          */
8244                                         if (nrsm && nrsm->r_in_tmap)
8245                                                 rack_log_sack_passed(tp, rack, nrsm);
8246                                 }
8247                                 /* Now are we done? */
8248                                 if (SEQ_LT(end, next->r_end) ||
8249                                     (end == next->r_end)) {
8250                                         /* Done with block */
8251                                         goto out;
8252                                 }
8253                                 rack_log_map_chg(tp, rack, &stack_map, rsm, next, MAP_SACK_M1, end, __LINE__);
8254                                 counter_u64_add(rack_sack_used_next_merge, 1);
8255                                 /* Postion for the next block */
8256                                 start = next->r_end;
8257                                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, next);
8258                                 if (rsm == NULL)
8259                                         goto out;
8260                         } else {
8261                                 /**
8262                                  * We can't use any hookery here, so we
8263                                  * need to split the map. We enter like
8264                                  * so:
8265                                  *  rsm      |--------|
8266                                  *  sackblk       |----->
8267                                  * We will add the new block nrsm and
8268                                  * that will be the new portion, and then
8269                                  * fall through after reseting rsm. So we
8270                                  * split and look like this:
8271                                  *  rsm      |----|
8272                                  *  sackblk       |----->
8273                                  *  nrsm          |---|
8274                                  * We then fall through reseting
8275                                  * rsm to nrsm, so the next block
8276                                  * picks it up.
8277                                  */
8278                                 nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
8279                                 if (nrsm == NULL) {
8280                                         /*
8281                                          * failed XXXrrs what can we do but loose the sack
8282                                          * info?
8283                                          */
8284                                         goto out;
8285                                 }
8286                                 counter_u64_add(rack_sack_splits, 1);
8287                                 rack_clone_rsm(rack, nrsm, rsm, start);
8288                                 rsm->r_just_ret = 0;
8289                                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8290 #ifdef INVARIANTS
8291                                 if (insret != NULL) {
8292                                         panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
8293                                               nrsm, insret, rack, rsm);
8294                                 }
8295 #endif
8296                                 if (rsm->r_in_tmap) {
8297                                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8298                                         nrsm->r_in_tmap = 1;
8299                                 }
8300                                 rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M2, end, __LINE__);
8301                                 rsm->r_flags &= (~RACK_HAS_FIN);
8302                                 /* Position us to point to the new nrsm that starts the sack blk */
8303                                 rsm = nrsm;
8304                         }
8305                 } else {
8306                         /* Already sacked this piece */
8307                         counter_u64_add(rack_sack_skipped_acked, 1);
8308                         moved++;
8309                         if (end == rsm->r_end) {
8310                                 /* Done with block */
8311                                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8312                                 goto out;
8313                         } else if (SEQ_LT(end, rsm->r_end)) {
8314                                 /* A partial sack to a already sacked block */
8315                                 moved++;
8316                                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8317                                 goto out;
8318                         } else {
8319                                 /*
8320                                  * The end goes beyond this guy
8321                                  * repostion the start to the
8322                                  * next block.
8323                                  */
8324                                 start = rsm->r_end;
8325                                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8326                                 if (rsm == NULL)
8327                                         goto out;
8328                         }
8329                 }
8330         }
8331         if (SEQ_GEQ(end, rsm->r_end)) {
8332                 /**
8333                  * The end of this block is either beyond this guy or right
8334                  * at this guy. I.e.:
8335                  *  rsm ---                 |-----|
8336                  *  end                     |-----|
8337                  *  <or>
8338                  *  end                     |---------|
8339                  */
8340                 if ((rsm->r_flags & RACK_ACKED) == 0) {
8341                         rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
8342                         changed += (rsm->r_end - rsm->r_start);
8343                         rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
8344                         if (rsm->r_in_tmap) /* should be true */
8345                                 rack_log_sack_passed(tp, rack, rsm);
8346                         /* Is Reordering occuring? */
8347                         if (rsm->r_flags & RACK_SACK_PASSED) {
8348                                 rsm->r_flags &= ~RACK_SACK_PASSED;
8349                                 counter_u64_add(rack_reorder_seen, 1);
8350                                 rack->r_ctl.rc_reorder_ts = cts;
8351                         }
8352                         if (rack->app_limited_needs_set)
8353                                 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
8354                         rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8355                         rsm->r_flags |= RACK_ACKED;
8356                         rsm->r_flags &= ~RACK_TLP;
8357                         if (rsm->r_in_tmap) {
8358                                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8359                                 rsm->r_in_tmap = 0;
8360                         }
8361                         rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_SACK_M3, end, __LINE__);
8362                 } else {
8363                         counter_u64_add(rack_sack_skipped_acked, 1);
8364                         moved++;
8365                 }
8366                 if (end == rsm->r_end) {
8367                         /* This block only - done, setup for next */
8368                         goto out;
8369                 }
8370                 /*
8371                  * There is more not coverend by this rsm move on
8372                  * to the next block in the RB tree.
8373                  */
8374                 nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8375                 start = rsm->r_end;
8376                 rsm = nrsm;
8377                 if (rsm == NULL)
8378                         goto out;
8379                 goto do_rest_ofb;
8380         }
8381         /**
8382          * The end of this sack block is smaller than
8383          * our rsm i.e.:
8384          *  rsm ---                 |-----|
8385          *  end                     |--|
8386          */
8387         if ((rsm->r_flags & RACK_ACKED) == 0) {
8388                 prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8389                 if (prev && (prev->r_flags & RACK_ACKED)) {
8390                         /**
8391                          * Goal, we want the right remainder of rsm to shrink
8392                          * in place and span from (rsm->r_start = end) to rsm->r_end.
8393                          * We want to expand prev to go all the way
8394                          * to prev->r_end <- end.
8395                          * so in the tree we have before:
8396                          *   prev     |--------|         (acked)
8397                          *   rsm               |-------| (non-acked)
8398                          *   sackblk           |-|
8399                          * We churn it so we end up with
8400                          *   prev     |----------|       (acked)
8401                          *   rsm                 |-----| (non-acked)
8402                          *   nrsm              |-| (temporary)
8403                          */
8404                         nrsm = &stack_map;
8405                         memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
8406                         prev->r_end = end;
8407                         rsm->r_start = end;
8408                         /* Now adjust nrsm (stack copy) to be
8409                          * the one that is the small
8410                          * piece that was "sacked".
8411                          */
8412                         nrsm->r_end = end;
8413                         rsm->r_dupack = 0;
8414                         rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8415                         /*
8416                          * Now that the rsm has had its start moved forward
8417                          * lets go ahead and get its new place in the world.
8418                          */
8419                         rack_setup_offset_for_rsm(prev, rsm);
8420                         /*
8421                          * Now nrsm is our new little piece
8422                          * that is acked (which was merged
8423                          * to prev). Update the rtt and changed
8424                          * based on that. Also check for reordering.
8425                          */
8426                         rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
8427                         if (rack->app_limited_needs_set)
8428                                 rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
8429                         changed += (nrsm->r_end - nrsm->r_start);
8430                         rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
8431                         if (nrsm->r_flags & RACK_SACK_PASSED) {
8432                                 counter_u64_add(rack_reorder_seen, 1);
8433                                 rack->r_ctl.rc_reorder_ts = cts;
8434                         }
8435                         rack_log_map_chg(tp, rack, prev, &stack_map, rsm, MAP_SACK_M4, end, __LINE__);
8436                         rsm = prev;
8437                         counter_u64_add(rack_sack_used_prev_merge, 1);
8438                 } else {
8439                         /**
8440                          * This is the case where our previous
8441                          * block is not acked either, so we must
8442                          * split the block in two.
8443                          */
8444                         nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
8445                         if (nrsm == NULL) {
8446                                 /* failed rrs what can we do but loose the sack info? */
8447                                 goto out;
8448                         }
8449                         /**
8450                          * In this case nrsm becomes
8451                          * nrsm->r_start = end;
8452                          * nrsm->r_end = rsm->r_end;
8453                          * which is un-acked.
8454                          * <and>
8455                          * rsm->r_end = nrsm->r_start;
8456                          * i.e. the remaining un-acked
8457                          * piece is left on the left
8458                          * hand side.
8459                          *
8460                          * So we start like this
8461                          * rsm      |----------| (not acked)
8462                          * sackblk  |---|
8463                          * build it so we have
8464                          * rsm      |---|         (acked)
8465                          * nrsm         |------|  (not acked)
8466                          */
8467                         counter_u64_add(rack_sack_splits, 1);
8468                         rack_clone_rsm(rack, nrsm, rsm, end);
8469                         rsm->r_flags &= (~RACK_HAS_FIN);
8470                         rsm->r_just_ret = 0;
8471                         insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8472 #ifdef INVARIANTS
8473                         if (insret != NULL) {
8474                                 panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
8475                                       nrsm, insret, rack, rsm);
8476                         }
8477 #endif
8478                         if (rsm->r_in_tmap) {
8479                                 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8480                                 nrsm->r_in_tmap = 1;
8481                         }
8482                         nrsm->r_dupack = 0;
8483                         rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
8484                         rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
8485                         changed += (rsm->r_end - rsm->r_start);
8486                         rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
8487                         if (rsm->r_in_tmap) /* should be true */
8488                                 rack_log_sack_passed(tp, rack, rsm);
8489                         /* Is Reordering occuring? */
8490                         if (rsm->r_flags & RACK_SACK_PASSED) {
8491                                 rsm->r_flags &= ~RACK_SACK_PASSED;
8492                                 counter_u64_add(rack_reorder_seen, 1);
8493                                 rack->r_ctl.rc_reorder_ts = cts;
8494                         }
8495                         if (rack->app_limited_needs_set)
8496                                 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
8497                         rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8498                         rsm->r_flags |= RACK_ACKED;
8499                         rsm->r_flags &= ~RACK_TLP;
8500                         rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M5, end, __LINE__);
8501                         if (rsm->r_in_tmap) {
8502                                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8503                                 rsm->r_in_tmap = 0;
8504                         }
8505                 }
8506         } else if (start != end){
8507                 /*
8508                  * The block was already acked.
8509                  */
8510                 counter_u64_add(rack_sack_skipped_acked, 1);
8511                 moved++;
8512         }
8513 out:
8514         if (rsm && (rsm->r_flags & RACK_ACKED)) {
8515                 /*
8516                  * Now can we merge where we worked
8517                  * with either the previous or
8518                  * next block?
8519                  */
8520                 next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8521                 while (next) {
8522                     if (next->r_flags & RACK_ACKED) {
8523                         /* yep this and next can be merged */
8524                         rsm = rack_merge_rsm(rack, rsm, next);
8525                         next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8526                     } else
8527                             break;
8528                 }
8529                 /* Now what about the previous? */
8530                 prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8531                 while (prev) {
8532                     if (prev->r_flags & RACK_ACKED) {
8533                         /* yep the previous and this can be merged */
8534                         rsm = rack_merge_rsm(rack, prev, rsm);
8535                         prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8536                     } else
8537                             break;
8538                 }
8539         }
8540         if (used_ref == 0) {
8541                 counter_u64_add(rack_sack_proc_all, 1);
8542         } else {
8543                 counter_u64_add(rack_sack_proc_short, 1);
8544         }
8545         /* Save off the next one for quick reference. */
8546         if (rsm)
8547                 nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8548         else
8549                 nrsm = NULL;
8550         *prsm = rack->r_ctl.rc_sacklast = nrsm;
8551         /* Pass back the moved. */
8552         *moved_two = moved;
8553         return (changed);
8554 }
8555
8556 static void inline
8557 rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack)
8558 {
8559         struct rack_sendmap *tmap;
8560
8561         tmap = NULL;
8562         while (rsm && (rsm->r_flags & RACK_ACKED)) {
8563                 /* Its no longer sacked, mark it so */
8564                 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
8565 #ifdef INVARIANTS
8566                 if (rsm->r_in_tmap) {
8567                         panic("rack:%p rsm:%p flags:0x%x in tmap?",
8568                               rack, rsm, rsm->r_flags);
8569                 }
8570 #endif
8571                 rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS);
8572                 /* Rebuild it into our tmap */
8573                 if (tmap == NULL) {
8574                         TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8575                         tmap = rsm;
8576                 } else {
8577                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext);
8578                         tmap = rsm;
8579                 }
8580                 tmap->r_in_tmap = 1;
8581                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8582         }
8583         /*
8584          * Now lets possibly clear the sack filter so we start
8585          * recognizing sacks that cover this area.
8586          */
8587         sack_filter_clear(&rack->r_ctl.rack_sf, th_ack);
8588
8589 }
8590
8591 static void
8592 rack_do_decay(struct tcp_rack *rack)
8593 {
8594         struct timeval res;
8595
8596 #define timersub(tvp, uvp, vvp)                                         \
8597         do {                                                            \
8598                 (vvp)->tv_sec = (tvp)->tv_sec - (uvp)->tv_sec;          \
8599                 (vvp)->tv_usec = (tvp)->tv_usec - (uvp)->tv_usec;       \
8600                 if ((vvp)->tv_usec < 0) {                               \
8601                         (vvp)->tv_sec--;                                \
8602                         (vvp)->tv_usec += 1000000;                      \
8603                 }                                                       \
8604         } while (0)
8605
8606         timersub(&rack->r_ctl.act_rcv_time, &rack->r_ctl.rc_last_time_decay, &res);
8607 #undef timersub
8608
8609         rack->r_ctl.input_pkt++;
8610         if ((rack->rc_in_persist) ||
8611             (res.tv_sec >= 1) ||
8612             (rack->rc_tp->snd_max == rack->rc_tp->snd_una)) {
8613                 /*
8614                  * Check for decay of non-SAD,
8615                  * we want all SAD detection metrics to
8616                  * decay 1/4 per second (or more) passed.
8617                  */
8618                 uint32_t pkt_delta;
8619
8620                 pkt_delta = rack->r_ctl.input_pkt - rack->r_ctl.saved_input_pkt;
8621                 /* Update our saved tracking values */
8622                 rack->r_ctl.saved_input_pkt = rack->r_ctl.input_pkt;
8623                 rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
8624                 /* Now do we escape without decay? */
8625 #ifdef NETFLIX_EXP_DETECTION
8626                 if (rack->rc_in_persist ||
8627                     (rack->rc_tp->snd_max == rack->rc_tp->snd_una) ||
8628                     (pkt_delta < tcp_sad_low_pps)){
8629                         /*
8630                          * We don't decay idle connections
8631                          * or ones that have a low input pps.
8632                          */
8633                         return;
8634                 }
8635                 /* Decay the counters */
8636                 rack->r_ctl.ack_count = ctf_decay_count(rack->r_ctl.ack_count,
8637                                                         tcp_sad_decay_val);
8638                 rack->r_ctl.sack_count = ctf_decay_count(rack->r_ctl.sack_count,
8639                                                          tcp_sad_decay_val);
8640                 rack->r_ctl.sack_moved_extra = ctf_decay_count(rack->r_ctl.sack_moved_extra,
8641                                                                tcp_sad_decay_val);
8642                 rack->r_ctl.sack_noextra_move = ctf_decay_count(rack->r_ctl.sack_noextra_move,
8643                                                                 tcp_sad_decay_val);
8644 #endif
8645         }
8646 }
8647
8648 static void
8649 rack_process_to_cumack(struct tcpcb *tp, struct tcp_rack *rack, register uint32_t th_ack, uint32_t cts, struct tcpopt *to)
8650 {
8651         struct rack_sendmap *rsm, *rm;
8652
8653         /*
8654          * The ACK point is advancing to th_ack, we must drop off
8655          * the packets in the rack log and calculate any eligble
8656          * RTT's.
8657          */
8658         rack->r_wanted_output = 1;
8659 more:
8660         rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
8661         if (rsm == NULL) {
8662                 if ((th_ack - 1) == tp->iss) {
8663                         /*
8664                          * For the SYN incoming case we will not
8665                          * have called tcp_output for the sending of
8666                          * the SYN, so there will be no map. All
8667                          * other cases should probably be a panic.
8668                          */
8669                         return;
8670                 }
8671                 if (tp->t_flags & TF_SENTFIN) {
8672                         /* if we sent a FIN we often will not have map */
8673                         return;
8674                 }
8675 #ifdef INVARIANTS
8676                 panic("No rack map tp:%p for state:%d ack:%u rack:%p snd_una:%u snd_max:%u snd_nxt:%u\n",
8677                       tp,
8678                       tp->t_state, th_ack, rack,
8679                       tp->snd_una, tp->snd_max, tp->snd_nxt);
8680 #endif
8681                 return;
8682         }
8683         if (SEQ_LT(th_ack, rsm->r_start)) {
8684                 /* Huh map is missing this */
8685 #ifdef INVARIANTS
8686                 printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n",
8687                        rsm->r_start,
8688                        th_ack, tp->t_state, rack->r_state);
8689 #endif
8690                 return;
8691         }
8692         rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED, th_ack);
8693         /* Now do we consume the whole thing? */
8694         if (SEQ_GEQ(th_ack, rsm->r_end)) {
8695                 /* Its all consumed. */
8696                 uint32_t left;
8697                 uint8_t newly_acked;
8698
8699                 rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_FREE, rsm->r_end, __LINE__);
8700                 rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
8701                 rsm->r_rtr_bytes = 0;
8702                 /* Record the time of highest cumack sent */
8703                 rack->r_ctl.rc_gp_cumack_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8704                 rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8705 #ifdef INVARIANTS
8706                 if (rm != rsm) {
8707                         panic("removing head in rack:%p rsm:%p rm:%p",
8708                               rack, rsm, rm);
8709                 }
8710 #endif
8711                 if (rsm->r_in_tmap) {
8712                         TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8713                         rsm->r_in_tmap = 0;
8714                 }
8715                 newly_acked = 1;
8716                 if (rsm->r_flags & RACK_ACKED) {
8717                         /*
8718                          * It was acked on the scoreboard -- remove
8719                          * it from total
8720                          */
8721                         rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
8722                         newly_acked = 0;
8723                 } else if (rsm->r_flags & RACK_SACK_PASSED) {
8724                         /*
8725                          * There are segments ACKED on the
8726                          * scoreboard further up. We are seeing
8727                          * reordering.
8728                          */
8729                         rsm->r_flags &= ~RACK_SACK_PASSED;
8730                         counter_u64_add(rack_reorder_seen, 1);
8731                         rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8732                         rsm->r_flags |= RACK_ACKED;
8733                         rack->r_ctl.rc_reorder_ts = cts;
8734                         if (rack->r_ent_rec_ns) {
8735                                 /*
8736                                  * We have sent no more, and we saw an sack
8737                                  * then ack arrive.
8738                                  */
8739                                 rack->r_might_revert = 1;
8740                         }
8741                 }
8742                 if ((rsm->r_flags & RACK_TO_REXT) &&
8743                     (tp->t_flags & TF_RCVD_TSTMP) &&
8744                     (to->to_flags & TOF_TS) &&
8745                     (tp->t_flags & TF_PREVVALID)) {
8746                         /*
8747                          * We can use the timestamp to see
8748                          * if this retransmission was from the
8749                          * first transmit. If so we made a mistake.
8750                          */
8751                         tp->t_flags &= ~TF_PREVVALID;
8752                         if (to->to_tsecr == rack_ts_to_msec(rsm->r_tim_lastsent[0])) {
8753                                 /* The first transmit is what this ack is for */
8754                                 rack_cong_signal(tp, CC_RTO_ERR, th_ack);
8755                         }
8756                 }
8757                 left = th_ack - rsm->r_end;
8758                 if (rack->app_limited_needs_set && newly_acked)
8759                         rack_need_set_test(tp, rack, rsm, th_ack, __LINE__, RACK_USE_END_OR_THACK);
8760                 /* Free back to zone */
8761                 rack_free(rack, rsm);
8762                 if (left) {
8763                         goto more;
8764                 }
8765                 /* Check for reneging */
8766                 rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
8767                 if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) {
8768                         /*
8769                          * The peer has moved snd_una up to
8770                          * the edge of this send, i.e. one
8771                          * that it had previously acked. The only
8772                          * way that can be true if the peer threw
8773                          * away data (space issues) that it had
8774                          * previously sacked (else it would have
8775                          * given us snd_una up to (rsm->r_end).
8776                          * We need to undo the acked markings here.
8777                          *
8778                          * Note we have to look to make sure th_ack is
8779                          * our rsm->r_start in case we get an old ack
8780                          * where th_ack is behind snd_una.
8781                          */
8782                         rack_peer_reneges(rack, rsm, th_ack);
8783                 }
8784                 return;
8785         }
8786         if (rsm->r_flags & RACK_ACKED) {
8787                 /*
8788                  * It was acked on the scoreboard -- remove it from
8789                  * total for the part being cum-acked.
8790                  */
8791                 rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
8792         }
8793         /*
8794          * Clear the dup ack count for
8795          * the piece that remains.
8796          */
8797         rsm->r_dupack = 0;
8798         rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8799         if (rsm->r_rtr_bytes) {
8800                 /*
8801                  * It was retransmitted adjust the
8802                  * sack holes for what was acked.
8803                  */
8804                 int ack_am;
8805
8806                 ack_am = (th_ack - rsm->r_start);
8807                 if (ack_am >= rsm->r_rtr_bytes) {
8808                         rack->r_ctl.rc_holes_rxt -= ack_am;
8809                         rsm->r_rtr_bytes -= ack_am;
8810                 }
8811         }
8812         /*
8813          * Update where the piece starts and record
8814          * the time of send of highest cumack sent.
8815          */
8816         rack->r_ctl.rc_gp_cumack_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8817         rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_TRIM_HEAD, th_ack, __LINE__);
8818         /* Now we need to move our offset forward too */
8819         if (rsm->m && (rsm->orig_m_len != rsm->m->m_len)) {
8820                 /* Fix up the orig_m_len and possibly the mbuf offset */
8821                 rack_adjust_orig_mlen(rsm);
8822         }
8823         rsm->soff += (th_ack - rsm->r_start);
8824         rsm->r_start = th_ack;
8825         /* Now do we need to move the mbuf fwd too? */
8826         if (rsm->m) {
8827                 while (rsm->soff >= rsm->m->m_len) {
8828                         rsm->soff -= rsm->m->m_len;
8829                         rsm->m = rsm->m->m_next;
8830                         KASSERT((rsm->m != NULL),
8831                                 (" nrsm:%p hit at soff:%u null m",
8832                                  rsm, rsm->soff));
8833                 }
8834                 rsm->orig_m_len = rsm->m->m_len;
8835         }
8836         if (rack->app_limited_needs_set)
8837                 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_BEG);
8838 }
8839
8840 static void
8841 rack_handle_might_revert(struct tcpcb *tp, struct tcp_rack *rack)
8842 {
8843         struct rack_sendmap *rsm;
8844         int sack_pass_fnd = 0;
8845
8846         if (rack->r_might_revert) {
8847                 /*
8848                  * Ok we have reordering, have not sent anything, we
8849                  * might want to revert the congestion state if nothing
8850                  * further has SACK_PASSED on it. Lets check.
8851                  *
8852                  * We also get here when we have DSACKs come in for
8853                  * all the data that we FR'd. Note that a rxt or tlp
8854                  * timer clears this from happening.
8855                  */
8856
8857                 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
8858                         if (rsm->r_flags & RACK_SACK_PASSED) {
8859                                 sack_pass_fnd = 1;
8860                                 break;
8861                         }
8862                 }
8863                 if (sack_pass_fnd == 0) {
8864                         /*
8865                          * We went into recovery
8866                          * incorrectly due to reordering!
8867                          */
8868                         int orig_cwnd;
8869
8870                         rack->r_ent_rec_ns = 0;
8871                         orig_cwnd = tp->snd_cwnd;
8872                         tp->snd_cwnd = rack->r_ctl.rc_cwnd_at_erec;
8873                         tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at_erec;
8874                         tp->snd_recover = tp->snd_una;
8875                         rack_log_to_prr(rack, 14, orig_cwnd);
8876                         EXIT_RECOVERY(tp->t_flags);
8877                 }
8878                 rack->r_might_revert = 0;
8879         }
8880 }
8881
8882 #ifdef NETFLIX_EXP_DETECTION
8883 static void
8884 rack_do_detection(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t bytes_this_ack, uint32_t segsiz)
8885 {
8886         if ((rack->do_detection || tcp_force_detection) &&
8887             tcp_sack_to_ack_thresh &&
8888             tcp_sack_to_move_thresh &&
8889             ((rack->r_ctl.rc_num_maps_alloced > tcp_map_minimum) || rack->sack_attack_disable)) {
8890                 /*
8891                  * We have thresholds set to find
8892                  * possible attackers and disable sack.
8893                  * Check them.
8894                  */
8895                 uint64_t ackratio, moveratio, movetotal;
8896
8897                 /* Log detecting */
8898                 rack_log_sad(rack, 1);
8899                 ackratio = (uint64_t)(rack->r_ctl.sack_count);
8900                 ackratio *= (uint64_t)(1000);
8901                 if (rack->r_ctl.ack_count)
8902                         ackratio /= (uint64_t)(rack->r_ctl.ack_count);
8903                 else {
8904                         /* We really should not hit here */
8905                         ackratio = 1000;
8906                 }
8907                 if ((rack->sack_attack_disable == 0) &&
8908                     (ackratio > rack_highest_sack_thresh_seen))
8909                         rack_highest_sack_thresh_seen = (uint32_t)ackratio;
8910                 movetotal = rack->r_ctl.sack_moved_extra;
8911                 movetotal += rack->r_ctl.sack_noextra_move;
8912                 moveratio = rack->r_ctl.sack_moved_extra;
8913                 moveratio *= (uint64_t)1000;
8914                 if (movetotal)
8915                         moveratio /= movetotal;
8916                 else {
8917                         /* No moves, thats pretty good */
8918                         moveratio = 0;
8919                 }
8920                 if ((rack->sack_attack_disable == 0) &&
8921                     (moveratio > rack_highest_move_thresh_seen))
8922                         rack_highest_move_thresh_seen = (uint32_t)moveratio;
8923                 if (rack->sack_attack_disable == 0) {
8924                         if ((ackratio > tcp_sack_to_ack_thresh) &&
8925                             (moveratio > tcp_sack_to_move_thresh)) {
8926                                 /* Disable sack processing */
8927                                 rack->sack_attack_disable = 1;
8928                                 if (rack->r_rep_attack == 0) {
8929                                         rack->r_rep_attack = 1;
8930                                         counter_u64_add(rack_sack_attacks_detected, 1);
8931                                 }
8932                                 if (tcp_attack_on_turns_on_logging) {
8933                                         /*
8934                                          * Turn on logging, used for debugging
8935                                          * false positives.
8936                                          */
8937                                         rack->rc_tp->t_logstate = tcp_attack_on_turns_on_logging;
8938                                 }
8939                                 /* Clamp the cwnd at flight size */
8940                                 rack->r_ctl.rc_saved_cwnd = rack->rc_tp->snd_cwnd;
8941                                 rack->rc_tp->snd_cwnd = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
8942                                 rack_log_sad(rack, 2);
8943                         }
8944                 } else {
8945                         /* We are sack-disabled check for false positives */
8946                         if ((ackratio <= tcp_restoral_thresh) ||
8947                             (rack->r_ctl.rc_num_maps_alloced  < tcp_map_minimum)) {
8948                                 rack->sack_attack_disable = 0;
8949                                 rack_log_sad(rack, 3);
8950                                 /* Restart counting */
8951                                 rack->r_ctl.sack_count = 0;
8952                                 rack->r_ctl.sack_moved_extra = 0;
8953                                 rack->r_ctl.sack_noextra_move = 1;
8954                                 rack->r_ctl.ack_count = max(1,
8955                                       (bytes_this_ack / segsiz));
8956
8957                                 if (rack->r_rep_reverse == 0) {
8958                                         rack->r_rep_reverse = 1;
8959                                         counter_u64_add(rack_sack_attacks_reversed, 1);
8960                                 }
8961                                 /* Restore the cwnd */
8962                                 if (rack->r_ctl.rc_saved_cwnd > rack->rc_tp->snd_cwnd)
8963                                         rack->rc_tp->snd_cwnd = rack->r_ctl.rc_saved_cwnd;
8964                         }
8965                 }
8966         }
8967 }
8968 #endif
8969
8970 static void
8971 rack_note_dsack(struct tcp_rack *rack, tcp_seq start, tcp_seq end)
8972 {
8973
8974         uint32_t am;
8975
8976         if (SEQ_GT(end, start))
8977                 am = end - start;
8978         else
8979                 am = 0;
8980         /*
8981          * We keep track of how many DSACK blocks we get
8982          * after a recovery incident.
8983          */
8984         rack->r_ctl.dsack_byte_cnt += am;
8985         if (!IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
8986             rack->r_ctl.retran_during_recovery &&
8987             (rack->r_ctl.dsack_byte_cnt >= rack->r_ctl.retran_during_recovery)) {
8988                 /*
8989                  * False recovery most likely culprit is reordering. If
8990                  * nothing else is missing we need to revert.
8991                  */
8992                 rack->r_might_revert = 1;
8993                 rack_handle_might_revert(rack->rc_tp, rack);
8994                 rack->r_might_revert = 0;
8995                 rack->r_ctl.retran_during_recovery = 0;
8996                 rack->r_ctl.dsack_byte_cnt = 0;
8997         }
8998 }
8999
9000 static void
9001 rack_update_prr(struct tcpcb *tp, struct tcp_rack *rack, uint32_t changed, tcp_seq th_ack)
9002 {
9003         /* Deal with changed and PRR here (in recovery only) */
9004         uint32_t pipe, snd_una;
9005
9006         rack->r_ctl.rc_prr_delivered += changed;
9007
9008         if (sbavail(&rack->rc_inp->inp_socket->so_snd) <= (tp->snd_max - tp->snd_una)) {
9009                 /*
9010                  * It is all outstanding, we are application limited
9011                  * and thus we don't need more room to send anything.
9012                  * Note we use tp->snd_una here and not th_ack because
9013                  * the data as yet not been cut from the sb.
9014                  */
9015                 rack->r_ctl.rc_prr_sndcnt = 0;
9016                 return;
9017         }
9018         /* Compute prr_sndcnt */
9019         if (SEQ_GT(tp->snd_una, th_ack)) {
9020                 snd_una = tp->snd_una;
9021         } else {
9022                 snd_una = th_ack;
9023         }
9024         pipe = ((tp->snd_max - snd_una) - rack->r_ctl.rc_sacked) + rack->r_ctl.rc_holes_rxt;
9025         if (pipe > tp->snd_ssthresh) {
9026                 long sndcnt;
9027
9028                 sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh;
9029                 if (rack->r_ctl.rc_prr_recovery_fs > 0)
9030                         sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs;
9031                 else {
9032                         rack->r_ctl.rc_prr_sndcnt = 0;
9033                         rack_log_to_prr(rack, 9, 0);
9034                         sndcnt = 0;
9035                 }
9036                 sndcnt++;
9037                 if (sndcnt > (long)rack->r_ctl.rc_prr_out)
9038                         sndcnt -= rack->r_ctl.rc_prr_out;
9039                 else
9040                         sndcnt = 0;
9041                 rack->r_ctl.rc_prr_sndcnt = sndcnt;
9042                 rack_log_to_prr(rack, 10, 0);
9043         } else {
9044                 uint32_t limit;
9045
9046                 if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out)
9047                         limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out);
9048                 else
9049                         limit = 0;
9050                 if (changed > limit)
9051                         limit = changed;
9052                 limit += ctf_fixed_maxseg(tp);
9053                 if (tp->snd_ssthresh > pipe) {
9054                         rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit);
9055                         rack_log_to_prr(rack, 11, 0);
9056                 } else {
9057                         rack->r_ctl.rc_prr_sndcnt = min(0, limit);
9058                         rack_log_to_prr(rack, 12, 0);
9059                 }
9060         }
9061 }
9062
9063 static void
9064 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, int entered_recovery, int dup_ack_struck)
9065 {
9066         uint32_t changed;
9067         struct tcp_rack *rack;
9068         struct rack_sendmap *rsm;
9069         struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
9070         register uint32_t th_ack;
9071         int32_t i, j, k, num_sack_blks = 0;
9072         uint32_t cts, acked, ack_point, sack_changed = 0;
9073         int loop_start = 0, moved_two = 0;
9074         uint32_t tsused;
9075
9076
9077         INP_WLOCK_ASSERT(tp->t_inpcb);
9078         if (th->th_flags & TH_RST) {
9079                 /* We don't log resets */
9080                 return;
9081         }
9082         rack = (struct tcp_rack *)tp->t_fb_ptr;
9083         cts = tcp_get_usecs(NULL);
9084         rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9085         changed = 0;
9086         th_ack = th->th_ack;
9087         if (rack->sack_attack_disable == 0)
9088                 rack_do_decay(rack);
9089         if (BYTES_THIS_ACK(tp, th) >= ctf_fixed_maxseg(rack->rc_tp)) {
9090                 /*
9091                  * You only get credit for
9092                  * MSS and greater (and you get extra
9093                  * credit for larger cum-ack moves).
9094                  */
9095                 int ac;
9096
9097                 ac = BYTES_THIS_ACK(tp, th) / ctf_fixed_maxseg(rack->rc_tp);
9098                 rack->r_ctl.ack_count += ac;
9099                 counter_u64_add(rack_ack_total, ac);
9100         }
9101         if (rack->r_ctl.ack_count > 0xfff00000) {
9102                 /*
9103                  * reduce the number to keep us under
9104                  * a uint32_t.
9105                  */
9106                 rack->r_ctl.ack_count /= 2;
9107                 rack->r_ctl.sack_count /= 2;
9108         }
9109         if (SEQ_GT(th_ack, tp->snd_una)) {
9110                 rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__);
9111                 tp->t_acktime = ticks;
9112         }
9113         if (rsm && SEQ_GT(th_ack, rsm->r_start))
9114                 changed = th_ack - rsm->r_start;
9115         if (changed) {
9116                 rack_process_to_cumack(tp, rack, th_ack, cts, to);
9117         }
9118         if ((to->to_flags & TOF_SACK) == 0) {
9119                 /* We are done nothing left and no sack. */
9120                 rack_handle_might_revert(tp, rack);
9121                 /*
9122                  * For cases where we struck a dup-ack
9123                  * with no SACK, add to the changes so
9124                  * PRR will work right.
9125                  */
9126                 if (dup_ack_struck && (changed == 0)) {
9127                         changed += ctf_fixed_maxseg(rack->rc_tp);
9128                 }
9129                 goto out;
9130         }
9131         /* Sack block processing */
9132         if (SEQ_GT(th_ack, tp->snd_una))
9133                 ack_point = th_ack;
9134         else
9135                 ack_point = tp->snd_una;
9136         for (i = 0; i < to->to_nsacks; i++) {
9137                 bcopy((to->to_sacks + i * TCPOLEN_SACK),
9138                       &sack, sizeof(sack));
9139                 sack.start = ntohl(sack.start);
9140                 sack.end = ntohl(sack.end);
9141                 if (SEQ_GT(sack.end, sack.start) &&
9142                     SEQ_GT(sack.start, ack_point) &&
9143                     SEQ_LT(sack.start, tp->snd_max) &&
9144                     SEQ_GT(sack.end, ack_point) &&
9145                     SEQ_LEQ(sack.end, tp->snd_max)) {
9146                         sack_blocks[num_sack_blks] = sack;
9147                         num_sack_blks++;
9148 #ifdef NETFLIX_STATS
9149                 } else if (SEQ_LEQ(sack.start, th_ack) &&
9150                            SEQ_LEQ(sack.end, th_ack)) {
9151                         /*
9152                          * Its a D-SACK block.
9153                          */
9154                         tcp_record_dsack(sack.start, sack.end);
9155 #endif
9156                         rack_note_dsack(rack, sack.start, sack.end);
9157                 }
9158         }
9159         /*
9160          * Sort the SACK blocks so we can update the rack scoreboard with
9161          * just one pass.
9162          */
9163         num_sack_blks = sack_filter_blks(&rack->r_ctl.rack_sf, sack_blocks,
9164                                          num_sack_blks, th->th_ack);
9165         ctf_log_sack_filter(rack->rc_tp, num_sack_blks, sack_blocks);
9166         if (num_sack_blks == 0) {
9167                 /* Nothing to sack (DSACKs?) */
9168                 goto out_with_totals;
9169         }
9170         if (num_sack_blks < 2) {
9171                 /* Only one, we don't need to sort */
9172                 goto do_sack_work;
9173         }
9174         /* Sort the sacks */
9175         for (i = 0; i < num_sack_blks; i++) {
9176                 for (j = i + 1; j < num_sack_blks; j++) {
9177                         if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
9178                                 sack = sack_blocks[i];
9179                                 sack_blocks[i] = sack_blocks[j];
9180                                 sack_blocks[j] = sack;
9181                         }
9182                 }
9183         }
9184         /*
9185          * Now are any of the sack block ends the same (yes some
9186          * implementations send these)?
9187          */
9188 again:
9189         if (num_sack_blks == 0)
9190                 goto out_with_totals;
9191         if (num_sack_blks > 1) {
9192                 for (i = 0; i < num_sack_blks; i++) {
9193                         for (j = i + 1; j < num_sack_blks; j++) {
9194                                 if (sack_blocks[i].end == sack_blocks[j].end) {
9195                                         /*
9196                                          * Ok these two have the same end we
9197                                          * want the smallest end and then
9198                                          * throw away the larger and start
9199                                          * again.
9200                                          */
9201                                         if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
9202                                                 /*
9203                                                  * The second block covers
9204                                                  * more area use that
9205                                                  */
9206                                                 sack_blocks[i].start = sack_blocks[j].start;
9207                                         }
9208                                         /*
9209                                          * Now collapse out the dup-sack and
9210                                          * lower the count
9211                                          */
9212                                         for (k = (j + 1); k < num_sack_blks; k++) {
9213                                                 sack_blocks[j].start = sack_blocks[k].start;
9214                                                 sack_blocks[j].end = sack_blocks[k].end;
9215                                                 j++;
9216                                         }
9217                                         num_sack_blks--;
9218                                         goto again;
9219                                 }
9220                         }
9221                 }
9222         }
9223 do_sack_work:
9224         /*
9225          * First lets look to see if
9226          * we have retransmitted and
9227          * can use the transmit next?
9228          */
9229         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
9230         if (rsm &&
9231             SEQ_GT(sack_blocks[0].end, rsm->r_start) &&
9232             SEQ_LT(sack_blocks[0].start, rsm->r_end)) {
9233                 /*
9234                  * We probably did the FR and the next
9235                  * SACK in continues as we would expect.
9236                  */
9237                 acked = rack_proc_sack_blk(tp, rack, &sack_blocks[0], to, &rsm, cts, &moved_two);
9238                 if (acked) {
9239                         rack->r_wanted_output = 1;
9240                         changed += acked;
9241                         sack_changed += acked;
9242                 }
9243                 if (num_sack_blks == 1) {
9244                         /*
9245                          * This is what we would expect from
9246                          * a normal implementation to happen
9247                          * after we have retransmitted the FR,
9248                          * i.e the sack-filter pushes down
9249                          * to 1 block and the next to be retransmitted
9250                          * is the sequence in the sack block (has more
9251                          * are acked). Count this as ACK'd data to boost
9252                          * up the chances of recovering any false positives.
9253                          */
9254                         rack->r_ctl.ack_count += (acked / ctf_fixed_maxseg(rack->rc_tp));
9255                         counter_u64_add(rack_ack_total, (acked / ctf_fixed_maxseg(rack->rc_tp)));
9256                         counter_u64_add(rack_express_sack, 1);
9257                         if (rack->r_ctl.ack_count > 0xfff00000) {
9258                                 /*
9259                                  * reduce the number to keep us under
9260                                  * a uint32_t.
9261                                  */
9262                                 rack->r_ctl.ack_count /= 2;
9263                                 rack->r_ctl.sack_count /= 2;
9264                         }
9265                         goto out_with_totals;
9266                 } else {
9267                         /*
9268                          * Start the loop through the
9269                          * rest of blocks, past the first block.
9270                          */
9271                         moved_two = 0;
9272                         loop_start = 1;
9273                 }
9274         }
9275         /* Its a sack of some sort */
9276         rack->r_ctl.sack_count++;
9277         if (rack->r_ctl.sack_count > 0xfff00000) {
9278                 /*
9279                  * reduce the number to keep us under
9280                  * a uint32_t.
9281                  */
9282                 rack->r_ctl.ack_count /= 2;
9283                 rack->r_ctl.sack_count /= 2;
9284         }
9285         counter_u64_add(rack_sack_total, 1);
9286         if (rack->sack_attack_disable) {
9287                 /* An attacker disablement is in place */
9288                 if (num_sack_blks > 1) {
9289                         rack->r_ctl.sack_count += (num_sack_blks - 1);
9290                         rack->r_ctl.sack_moved_extra++;
9291                         counter_u64_add(rack_move_some, 1);
9292                         if (rack->r_ctl.sack_moved_extra > 0xfff00000) {
9293                                 rack->r_ctl.sack_moved_extra /= 2;
9294                                 rack->r_ctl.sack_noextra_move /= 2;
9295                         }
9296                 }
9297                 goto out;
9298         }
9299         rsm = rack->r_ctl.rc_sacklast;
9300         for (i = loop_start; i < num_sack_blks; i++) {
9301                 acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts, &moved_two);
9302                 if (acked) {
9303                         rack->r_wanted_output = 1;
9304                         changed += acked;
9305                         sack_changed += acked;
9306                 }
9307                 if (moved_two) {
9308                         /*
9309                          * If we did not get a SACK for at least a MSS and
9310                          * had to move at all, or if we moved more than our
9311                          * threshold, it counts against the "extra" move.
9312                          */
9313                         rack->r_ctl.sack_moved_extra += moved_two;
9314                         counter_u64_add(rack_move_some, 1);
9315                 } else {
9316                         /*
9317                          * else we did not have to move
9318                          * any more than we would expect.
9319                          */
9320                         rack->r_ctl.sack_noextra_move++;
9321                         counter_u64_add(rack_move_none, 1);
9322                 }
9323                 if (moved_two && (acked < ctf_fixed_maxseg(rack->rc_tp))) {
9324                         /*
9325                          * If the SACK was not a full MSS then
9326                          * we add to sack_count the number of
9327                          * MSS's (or possibly more than
9328                          * a MSS if its a TSO send) we had to skip by.
9329                          */
9330                         rack->r_ctl.sack_count += moved_two;
9331                         counter_u64_add(rack_sack_total, moved_two);
9332                 }
9333                 /*
9334                  * Now we need to setup for the next
9335                  * round. First we make sure we won't
9336                  * exceed the size of our uint32_t on
9337                  * the various counts, and then clear out
9338                  * moved_two.
9339                  */
9340                 if ((rack->r_ctl.sack_moved_extra > 0xfff00000) ||
9341                     (rack->r_ctl.sack_noextra_move > 0xfff00000)) {
9342                         rack->r_ctl.sack_moved_extra /= 2;
9343                         rack->r_ctl.sack_noextra_move /= 2;
9344                 }
9345                 if (rack->r_ctl.sack_count > 0xfff00000) {
9346                         rack->r_ctl.ack_count /= 2;
9347                         rack->r_ctl.sack_count /= 2;
9348                 }
9349                 moved_two = 0;
9350         }
9351 out_with_totals:
9352         if (num_sack_blks > 1) {
9353                 /*
9354                  * You get an extra stroke if
9355                  * you have more than one sack-blk, this
9356                  * could be where we are skipping forward
9357                  * and the sack-filter is still working, or
9358                  * it could be an attacker constantly
9359                  * moving us.
9360                  */
9361                 rack->r_ctl.sack_moved_extra++;
9362                 counter_u64_add(rack_move_some, 1);
9363         }
9364 out:
9365 #ifdef NETFLIX_EXP_DETECTION
9366         rack_do_detection(tp, rack, BYTES_THIS_ACK(tp, th), ctf_fixed_maxseg(rack->rc_tp));
9367 #endif
9368         if (changed) {
9369                 /* Something changed cancel the rack timer */
9370                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
9371         }
9372         tsused = tcp_get_usecs(NULL);
9373         rsm = tcp_rack_output(tp, rack, tsused);
9374         if ((!IN_FASTRECOVERY(tp->t_flags)) &&
9375             rsm) {
9376                 /* Enter recovery */
9377                 rack->r_ctl.rc_rsm_start = rsm->r_start;
9378                 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
9379                 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
9380                 entered_recovery = 1;
9381                 rack_cong_signal(tp, CC_NDUPACK, tp->snd_una);
9382                 /*
9383                  * When we enter recovery we need to assure we send
9384                  * one packet.
9385                  */
9386                 if (rack->rack_no_prr == 0) {
9387                         rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
9388                         rack_log_to_prr(rack, 8, 0);
9389                 }
9390                 rack->r_timer_override = 1;
9391                 rack->r_early = 0;
9392                 rack->r_ctl.rc_agg_early = 0;
9393         } else if (IN_FASTRECOVERY(tp->t_flags) &&
9394                    rsm &&
9395                    (rack->r_rr_config == 3)) {
9396                 /*
9397                  * Assure we can output and we get no
9398                  * remembered pace time except the retransmit.
9399                  */
9400                 rack->r_timer_override = 1;
9401                 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
9402                 rack->r_ctl.rc_resend = rsm;
9403         }
9404         if (IN_FASTRECOVERY(tp->t_flags) &&
9405             (rack->rack_no_prr == 0) &&
9406             (entered_recovery == 0)) {
9407                 rack_update_prr(tp, rack, changed, th_ack);
9408                 if ((rsm && (rack->r_ctl.rc_prr_sndcnt >= ctf_fixed_maxseg(tp)) &&
9409                      ((rack->rc_inp->inp_in_hpts == 0) &&
9410                       ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)))) {
9411                         /*
9412                          * If you are pacing output you don't want
9413                          * to override.
9414                          */
9415                         rack->r_early = 0;
9416                         rack->r_ctl.rc_agg_early = 0;
9417                         rack->r_timer_override = 1;
9418                 }
9419         }
9420 }
9421
9422 static void
9423 rack_strike_dupack(struct tcp_rack *rack)
9424 {
9425         struct rack_sendmap *rsm;
9426
9427         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
9428         while (rsm && (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
9429                 rsm = TAILQ_NEXT(rsm, r_tnext);
9430         }
9431         if (rsm && (rsm->r_dupack < 0xff)) {
9432                 rsm->r_dupack++;
9433                 if (rsm->r_dupack >= DUP_ACK_THRESHOLD) {
9434                         struct timeval tv;
9435                         uint32_t cts;
9436                         /*
9437                          * Here we see if we need to retransmit. For
9438                          * a SACK type connection if enough time has passed
9439                          * we will get a return of the rsm. For a non-sack
9440                          * connection we will get the rsm returned if the
9441                          * dupack value is 3 or more.
9442                          */
9443                         cts = tcp_get_usecs(&tv);
9444                         rack->r_ctl.rc_resend = tcp_rack_output(rack->rc_tp, rack, cts);
9445                         if (rack->r_ctl.rc_resend != NULL) {
9446                                 if (!IN_FASTRECOVERY(rack->rc_tp->t_flags)) {
9447                                         rack_cong_signal(rack->rc_tp, CC_NDUPACK,
9448                                                          rack->rc_tp->snd_una);
9449                                 }
9450                                 rack->r_wanted_output = 1;
9451                                 rack->r_timer_override = 1;
9452                                 rack_log_retran_reason(rack, rsm, __LINE__, 1, 3);
9453                         }
9454                 } else {
9455                         rack_log_retran_reason(rack, rsm, __LINE__, 0, 3);
9456                 }
9457         }
9458 }
9459
9460 static void
9461 rack_check_bottom_drag(struct tcpcb *tp,
9462                        struct tcp_rack *rack,
9463                        struct socket *so, int32_t acked)
9464 {
9465         uint32_t segsiz, minseg;
9466
9467         segsiz = ctf_fixed_maxseg(tp);
9468         minseg = segsiz;
9469
9470         if (tp->snd_max == tp->snd_una) {
9471                 /*
9472                  * We are doing dynamic pacing and we are way
9473                  * under. Basically everything got acked while
9474                  * we were still waiting on the pacer to expire.
9475                  *
9476                  * This means we need to boost the b/w in
9477                  * addition to any earlier boosting of
9478                  * the multipler.
9479                  */
9480                 rack->rc_dragged_bottom = 1;
9481                 rack_validate_multipliers_at_or_above100(rack);
9482                 /*
9483                  * Lets use the segment bytes acked plus
9484                  * the lowest RTT seen as the basis to
9485                  * form a b/w estimate. This will be off
9486                  * due to the fact that the true estimate
9487                  * should be around 1/2 the time of the RTT
9488                  * but we can settle for that.
9489                  */
9490                 if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_VALID) &&
9491                     acked) {
9492                         uint64_t bw, calc_bw, rtt;
9493
9494                         rtt = rack->r_ctl.rack_rs.rs_us_rtt;
9495                         if (rtt == 0) {
9496                                 /* no us sample is there a ms one? */
9497                                 if (rack->r_ctl.rack_rs.rs_rtt_lowest) {
9498                                         rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
9499                                 } else {
9500                                         goto no_measurement;
9501                                 }
9502                         }
9503                         bw = acked;
9504                         calc_bw = bw * 1000000;
9505                         calc_bw /= rtt;
9506                         if (rack->r_ctl.last_max_bw &&
9507                             (rack->r_ctl.last_max_bw < calc_bw)) {
9508                                 /*
9509                                  * If we have a last calculated max bw
9510                                  * enforce it.
9511                                  */
9512                                 calc_bw = rack->r_ctl.last_max_bw;
9513                         }
9514                         /* now plop it in */
9515                         if (rack->rc_gp_filled == 0) {
9516                                 if (calc_bw > ONE_POINT_TWO_MEG) {
9517                                         /*
9518                                          * If we have no measurement
9519                                          * don't let us set in more than
9520                                          * 1.2Mbps. If we are still too
9521                                          * low after pacing with this we
9522                                          * will hopefully have a max b/w
9523                                          * available to sanity check things.
9524                                          */
9525                                         calc_bw = ONE_POINT_TWO_MEG;
9526                                 }
9527                                 rack->r_ctl.rc_rtt_diff = 0;
9528                                 rack->r_ctl.gp_bw = calc_bw;
9529                                 rack->rc_gp_filled = 1;
9530                                 if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
9531                                         rack->r_ctl.num_measurements = RACK_REQ_AVG;
9532                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
9533                         } else if (calc_bw > rack->r_ctl.gp_bw) {
9534                                 rack->r_ctl.rc_rtt_diff = 0;
9535                                 if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
9536                                         rack->r_ctl.num_measurements = RACK_REQ_AVG;
9537                                 rack->r_ctl.gp_bw = calc_bw;
9538                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
9539                         } else
9540                                 rack_increase_bw_mul(rack, -1, 0, 0, 1);
9541                         if ((rack->gp_ready == 0) &&
9542                             (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
9543                                 /* We have enough measurements now */
9544                                 rack->gp_ready = 1;
9545                                 rack_set_cc_pacing(rack);
9546                                 if (rack->defer_options)
9547                                         rack_apply_deferred_options(rack);
9548                         }
9549                         /*
9550                          * For acks over 1mss we do a extra boost to simulate
9551                          * where we would get 2 acks (we want 110 for the mul).
9552                          */
9553                         if (acked > segsiz)
9554                                 rack_increase_bw_mul(rack, -1, 0, 0, 1);
9555                 } else {
9556                         /*
9557                          * zero rtt possibly?, settle for just an old increase.
9558                          */
9559 no_measurement:
9560                         rack_increase_bw_mul(rack, -1, 0, 0, 1);
9561                 }
9562         } else if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
9563                    (sbavail(&so->so_snd) > max((segsiz * (4 + rack_req_segs)),
9564                                                minseg)) &&
9565                    (rack->r_ctl.cwnd_to_use > max((segsiz * (rack_req_segs + 2)), minseg)) &&
9566                    (tp->snd_wnd > max((segsiz * (rack_req_segs + 2)), minseg)) &&
9567                    (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) <=
9568                     (segsiz * rack_req_segs))) {
9569                 /*
9570                  * We are doing dynamic GP pacing and
9571                  * we have everything except 1MSS or less
9572                  * bytes left out. We are still pacing away.
9573                  * And there is data that could be sent, This
9574                  * means we are inserting delayed ack time in
9575                  * our measurements because we are pacing too slow.
9576                  */
9577                 rack_validate_multipliers_at_or_above100(rack);
9578                 rack->rc_dragged_bottom = 1;
9579                 rack_increase_bw_mul(rack, -1, 0, 0, 1);
9580         }
9581 }
9582
9583
9584
9585 static void
9586 rack_gain_for_fastoutput(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t acked_amount)
9587 {
9588         /*
9589          * The fast output path is enabled and we
9590          * have moved the cumack forward. Lets see if
9591          * we can expand forward the fast path length by
9592          * that amount. What we would ideally like to
9593          * do is increase the number of bytes in the
9594          * fast path block (left_to_send) by the
9595          * acked amount. However we have to gate that
9596          * by two factors:
9597          * 1) The amount outstanding and the rwnd of the peer
9598          *    (i.e. we don't want to exceed the rwnd of the peer).
9599          *    <and>
9600          * 2) The amount of data left in the socket buffer (i.e.
9601          *    we can't send beyond what is in the buffer).
9602          *
9603          * Note that this does not take into account any increase
9604          * in the cwnd. We will only extend the fast path by
9605          * what was acked.
9606          */
9607         uint32_t new_total, gating_val;
9608
9609         new_total = acked_amount + rack->r_ctl.fsb.left_to_send;
9610         gating_val = min((sbavail(&so->so_snd) - (tp->snd_max - tp->snd_una)),
9611                          (tp->snd_wnd - (tp->snd_max - tp->snd_una)));
9612         if (new_total <= gating_val) {
9613                 /* We can increase left_to_send by the acked amount */
9614                 counter_u64_add(rack_extended_rfo, 1);
9615                 rack->r_ctl.fsb.left_to_send = new_total;
9616                 KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(&rack->rc_inp->inp_socket->so_snd) - (tp->snd_max - tp->snd_una))),
9617                         ("rack:%p left_to_send:%u sbavail:%u out:%u",
9618                          rack, rack->r_ctl.fsb.left_to_send,
9619                          sbavail(&rack->rc_inp->inp_socket->so_snd),
9620                          (tp->snd_max - tp->snd_una)));
9621
9622         }
9623 }
9624
9625 static void
9626 rack_adjust_sendmap(struct tcp_rack *rack, struct sockbuf *sb, tcp_seq snd_una)
9627 {
9628         /*
9629          * Here any sendmap entry that points to the
9630          * beginning mbuf must be adjusted to the correct
9631          * offset. This must be called with:
9632          * 1) The socket buffer locked
9633          * 2) snd_una adjusted to its new postion.
9634          *
9635          * Note that (2) implies rack_ack_received has also
9636          * been called.
9637          *
9638          * We grab the first mbuf in the socket buffer and
9639          * then go through the front of the sendmap, recalculating
9640          * the stored offset for any sendmap entry that has
9641          * that mbuf. We must use the sb functions to do this
9642          * since its possible an add was done has well as
9643          * the subtraction we may have just completed. This should
9644          * not be a penalty though, since we just referenced the sb
9645          * to go in and trim off the mbufs that we freed (of course
9646          * there will be a penalty for the sendmap references though).
9647          */
9648         struct mbuf *m;
9649         struct rack_sendmap *rsm;
9650
9651         SOCKBUF_LOCK_ASSERT(sb);
9652         m = sb->sb_mb;
9653         rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9654         if ((rsm == NULL) || (m == NULL)) {
9655                 /* Nothing outstanding */
9656                 return;
9657         }
9658         while (rsm->m && (rsm->m == m)) {
9659                 /* one to adjust */
9660 #ifdef INVARIANTS
9661                 struct mbuf *tm;
9662                 uint32_t soff;
9663
9664                 tm = sbsndmbuf(sb, (rsm->r_start - snd_una), &soff);
9665                 if (rsm->orig_m_len != m->m_len) {
9666                         rack_adjust_orig_mlen(rsm);
9667                 }
9668                 if (rsm->soff != soff) {
9669                         /*
9670                          * This is not a fatal error, we anticipate it
9671                          * might happen (the else code), so we count it here
9672                          * so that under invariant we can see that it really
9673                          * does happen.
9674                          */
9675                         counter_u64_add(rack_adjust_map_bw, 1);
9676                 }
9677                 rsm->m = tm;
9678                 rsm->soff = soff;
9679                 if (tm)
9680                         rsm->orig_m_len = rsm->m->m_len;
9681                 else
9682                         rsm->orig_m_len = 0;
9683 #else
9684                 rsm->m = sbsndmbuf(sb, (rsm->r_start - snd_una), &rsm->soff);
9685                 if (rsm->m)
9686                         rsm->orig_m_len = rsm->m->m_len;
9687                 else
9688                         rsm->orig_m_len = 0;
9689 #endif
9690                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
9691                               rsm);
9692                 if (rsm == NULL)
9693                         break;
9694         }
9695 }
9696
9697 /*
9698  * Return value of 1, we do not need to call rack_process_data().
9699  * return value of 0, rack_process_data can be called.
9700  * For ret_val if its 0 the TCP is locked, if its non-zero
9701  * its unlocked and probably unsafe to touch the TCB.
9702  */
9703 static int
9704 rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
9705     struct tcpcb *tp, struct tcpopt *to,
9706     uint32_t tiwin, int32_t tlen,
9707     int32_t * ofia, int32_t thflags, int32_t *ret_val)
9708 {
9709         int32_t ourfinisacked = 0;
9710         int32_t nsegs, acked_amount;
9711         int32_t acked;
9712         struct mbuf *mfree;
9713         struct tcp_rack *rack;
9714         int32_t under_pacing = 0;
9715         int32_t recovery = 0;
9716
9717         rack = (struct tcp_rack *)tp->t_fb_ptr;
9718         if (SEQ_GT(th->th_ack, tp->snd_max)) {
9719                 __ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val,
9720                                       &rack->r_ctl.challenge_ack_ts,
9721                                       &rack->r_ctl.challenge_ack_cnt);
9722                 rack->r_wanted_output = 1;
9723                 return (1);
9724         }
9725         if (rack->gp_ready &&
9726             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
9727                 under_pacing = 1;
9728         }
9729         if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
9730                 int in_rec, dup_ack_struck = 0;
9731
9732                 in_rec = IN_FASTRECOVERY(tp->t_flags);
9733                 if (rack->rc_in_persist) {
9734                         tp->t_rxtshift = 0;
9735                         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
9736                                       rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
9737                 }
9738                 if ((th->th_ack == tp->snd_una) && (tiwin == tp->snd_wnd)) {
9739                         rack_strike_dupack(rack);
9740                         dup_ack_struck = 1;
9741                 }
9742                 rack_log_ack(tp, to, th, ((in_rec == 0) && IN_FASTRECOVERY(tp->t_flags)), dup_ack_struck);
9743         }
9744         if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
9745                 /*
9746                  * Old ack, behind (or duplicate to) the last one rcv'd
9747                  * Note: We mark reordering is occuring if its
9748                  * less than and we have not closed our window.
9749                  */
9750                 if (SEQ_LT(th->th_ack, tp->snd_una) && (sbspace(&so->so_rcv) > ctf_fixed_maxseg(tp))) {
9751                         counter_u64_add(rack_reorder_seen, 1);
9752                         rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
9753                 }
9754                 return (0);
9755         }
9756         /*
9757          * If we reach this point, ACK is not a duplicate, i.e., it ACKs
9758          * something we sent.
9759          */
9760         if (tp->t_flags & TF_NEEDSYN) {
9761                 /*
9762                  * T/TCP: Connection was half-synchronized, and our SYN has
9763                  * been ACK'd (so connection is now fully synchronized).  Go
9764                  * to non-starred state, increment snd_una for ACK of SYN,
9765                  * and check if we can do window scaling.
9766                  */
9767                 tp->t_flags &= ~TF_NEEDSYN;
9768                 tp->snd_una++;
9769                 /* Do window scaling? */
9770                 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
9771                     (TF_RCVD_SCALE | TF_REQ_SCALE)) {
9772                         tp->rcv_scale = tp->request_r_scale;
9773                         /* Send window already scaled. */
9774                 }
9775         }
9776         nsegs = max(1, m->m_pkthdr.lro_nsegs);
9777         INP_WLOCK_ASSERT(tp->t_inpcb);
9778
9779         acked = BYTES_THIS_ACK(tp, th);
9780         KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
9781         KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
9782         /*
9783          * If we just performed our first retransmit, and the ACK arrives
9784          * within our recovery window, then it was a mistake to do the
9785          * retransmit in the first place.  Recover our original cwnd and
9786          * ssthresh, and proceed to transmit where we left off.
9787          */
9788         if ((tp->t_flags & TF_PREVVALID) &&
9789             ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
9790                 tp->t_flags &= ~TF_PREVVALID;
9791                 if (tp->t_rxtshift == 1 &&
9792                     (int)(ticks - tp->t_badrxtwin) < 0)
9793                         rack_cong_signal(tp, CC_RTO_ERR, th->th_ack);
9794         }
9795         if (acked) {
9796                 /* assure we are not backed off */
9797                 tp->t_rxtshift = 0;
9798                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
9799                               rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
9800                 rack->rc_tlp_in_progress = 0;
9801                 rack->r_ctl.rc_tlp_cnt_out = 0;
9802                 /*
9803                  * If it is the RXT timer we want to
9804                  * stop it, so we can restart a TLP.
9805                  */
9806                 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
9807                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
9808 #ifdef NETFLIX_HTTP_LOGGING
9809                 tcp_http_check_for_comp(rack->rc_tp, th->th_ack);
9810 #endif
9811         }
9812         /*
9813          * If we have a timestamp reply, update smoothed round trip time. If
9814          * no timestamp is present but transmit timer is running and timed
9815          * sequence number was acked, update smoothed round trip time. Since
9816          * we now have an rtt measurement, cancel the timer backoff (cf.,
9817          * Phil Karn's retransmit alg.). Recompute the initial retransmit
9818          * timer.
9819          *
9820          * Some boxes send broken timestamp replies during the SYN+ACK
9821          * phase, ignore timestamps of 0 or we could calculate a huge RTT
9822          * and blow up the retransmit timer.
9823          */
9824         /*
9825          * If all outstanding data is acked, stop retransmit timer and
9826          * remember to restart (more output or persist). If there is more
9827          * data to be acked, restart retransmit timer, using current
9828          * (possibly backed-off) value.
9829          */
9830         if (acked == 0) {
9831                 if (ofia)
9832                         *ofia = ourfinisacked;
9833                 return (0);
9834         }
9835         if (IN_RECOVERY(tp->t_flags)) {
9836                 if (SEQ_LT(th->th_ack, tp->snd_recover) &&
9837                     (SEQ_LT(th->th_ack, tp->snd_max))) {
9838                         tcp_rack_partialack(tp);
9839                 } else {
9840                         rack_post_recovery(tp, th->th_ack);
9841                         recovery = 1;
9842                 }
9843         }
9844         /*
9845          * Let the congestion control algorithm update congestion control
9846          * related information. This typically means increasing the
9847          * congestion window.
9848          */
9849         rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, recovery);
9850         SOCKBUF_LOCK(&so->so_snd);
9851         acked_amount = min(acked, (int)sbavail(&so->so_snd));
9852         tp->snd_wnd -= acked_amount;
9853         mfree = sbcut_locked(&so->so_snd, acked_amount);
9854         if ((sbused(&so->so_snd) == 0) &&
9855             (acked > acked_amount) &&
9856             (tp->t_state >= TCPS_FIN_WAIT_1) &&
9857             (tp->t_flags & TF_SENTFIN)) {
9858                 /*
9859                  * We must be sure our fin
9860                  * was sent and acked (we can be
9861                  * in FIN_WAIT_1 without having
9862                  * sent the fin).
9863                  */
9864                 ourfinisacked = 1;
9865         }
9866         tp->snd_una = th->th_ack;
9867         if (acked_amount && sbavail(&so->so_snd))
9868                 rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
9869         rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
9870         /* NB: sowwakeup_locked() does an implicit unlock. */
9871         sowwakeup_locked(so);
9872         m_freem(mfree);
9873         if (SEQ_GT(tp->snd_una, tp->snd_recover))
9874                 tp->snd_recover = tp->snd_una;
9875
9876         if (SEQ_LT(tp->snd_nxt, tp->snd_una)) {
9877                 tp->snd_nxt = tp->snd_una;
9878         }
9879         if (under_pacing &&
9880             (rack->use_fixed_rate == 0) &&
9881             (rack->in_probe_rtt == 0) &&
9882             rack->rc_gp_dyn_mul &&
9883             rack->rc_always_pace) {
9884                 /* Check if we are dragging bottom */
9885                 rack_check_bottom_drag(tp, rack, so, acked);
9886         }
9887         if (tp->snd_una == tp->snd_max) {
9888                 /* Nothing left outstanding */
9889                 tp->t_flags &= ~TF_PREVVALID;
9890                 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
9891                 rack->r_ctl.retran_during_recovery = 0;
9892                 rack->r_ctl.dsack_byte_cnt = 0;
9893                 if (rack->r_ctl.rc_went_idle_time == 0)
9894                         rack->r_ctl.rc_went_idle_time = 1;
9895                 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
9896                 if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
9897                         tp->t_acktime = 0;
9898                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
9899                 /* Set need output so persist might get set */
9900                 rack->r_wanted_output = 1;
9901                 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
9902                 if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
9903                     (sbavail(&so->so_snd) == 0) &&
9904                     (tp->t_flags2 & TF2_DROP_AF_DATA)) {
9905                         /*
9906                          * The socket was gone and the
9907                          * peer sent data (now or in the past), time to
9908                          * reset him.
9909                          */
9910                         *ret_val = 1;
9911                         /* tcp_close will kill the inp pre-log the Reset */
9912                         tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
9913                         tp = tcp_close(tp);
9914                         ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen);
9915                         return (1);
9916                 }
9917         }
9918         if (ofia)
9919                 *ofia = ourfinisacked;
9920         return (0);
9921 }
9922
9923 static void
9924 rack_collapsed_window(struct tcp_rack *rack)
9925 {
9926         /*
9927          * Now we must walk the
9928          * send map and divide the
9929          * ones left stranded. These
9930          * guys can't cause us to abort
9931          * the connection and are really
9932          * "unsent". However if a buggy
9933          * client actually did keep some
9934          * of the data i.e. collapsed the win
9935          * and refused to ack and then opened
9936          * the win and acked that data. We would
9937          * get into an ack war, the simplier
9938          * method then of just pretending we
9939          * did not send those segments something
9940          * won't work.
9941          */
9942         struct rack_sendmap *rsm, *nrsm, fe, *insret;
9943         tcp_seq max_seq;
9944
9945         max_seq = rack->rc_tp->snd_una + rack->rc_tp->snd_wnd;
9946         memset(&fe, 0, sizeof(fe));
9947         fe.r_start = max_seq;
9948         /* Find the first seq past or at maxseq */
9949         rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
9950         if (rsm == NULL) {
9951                 /* Nothing to do strange */
9952                 rack->rc_has_collapsed = 0;
9953                 return;
9954         }
9955         /*
9956          * Now do we need to split at
9957          * the collapse point?
9958          */
9959         if (SEQ_GT(max_seq, rsm->r_start)) {
9960                 nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
9961                 if (nrsm == NULL) {
9962                         /* We can't get a rsm, mark all? */
9963                         nrsm = rsm;
9964                         goto no_split;
9965                 }
9966                 /* Clone it */
9967                 rack_clone_rsm(rack, nrsm, rsm, max_seq);
9968                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
9969 #ifdef INVARIANTS
9970                 if (insret != NULL) {
9971                         panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
9972                               nrsm, insret, rack, rsm);
9973                 }
9974 #endif
9975                 rack_log_map_chg(rack->rc_tp, rack, NULL, rsm, nrsm, MAP_SPLIT, max_seq, __LINE__);
9976                 if (rsm->r_in_tmap) {
9977                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
9978                         nrsm->r_in_tmap = 1;
9979                 }
9980                 /*
9981                  * Set in the new RSM as the
9982                  * collapsed starting point
9983                  */
9984                 rsm = nrsm;
9985         }
9986 no_split:
9987         counter_u64_add(rack_collapsed_win, 1);
9988         RB_FOREACH_FROM(nrsm, rack_rb_tree_head, rsm) {
9989                 nrsm->r_flags |= RACK_RWND_COLLAPSED;
9990         }
9991         rack->rc_has_collapsed = 1;
9992 }
9993
9994 static void
9995 rack_un_collapse_window(struct tcp_rack *rack)
9996 {
9997         struct rack_sendmap *rsm;
9998
9999         RB_FOREACH_REVERSE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
10000                 if (rsm->r_flags & RACK_RWND_COLLAPSED)
10001                         rsm->r_flags &= ~RACK_RWND_COLLAPSED;
10002                 else
10003                         break;
10004         }
10005         rack->rc_has_collapsed = 0;
10006 }
10007
10008 static void
10009 rack_handle_delayed_ack(struct tcpcb *tp, struct tcp_rack *rack,
10010                         int32_t tlen, int32_t tfo_syn)
10011 {
10012         if (DELAY_ACK(tp, tlen) || tfo_syn) {
10013                 if (rack->rc_dack_mode &&
10014                     (tlen > 500) &&
10015                     (rack->rc_dack_toggle == 1)) {
10016                         goto no_delayed_ack;
10017                 }
10018                 rack_timer_cancel(tp, rack,
10019                                   rack->r_ctl.rc_rcvtime, __LINE__);
10020                 tp->t_flags |= TF_DELACK;
10021         } else {
10022 no_delayed_ack:
10023                 rack->r_wanted_output = 1;
10024                 tp->t_flags |= TF_ACKNOW;
10025                 if (rack->rc_dack_mode) {
10026                         if (tp->t_flags & TF_DELACK)
10027                                 rack->rc_dack_toggle = 1;
10028                         else
10029                                 rack->rc_dack_toggle = 0;
10030                 }
10031         }
10032 }
10033
10034 static void
10035 rack_validate_fo_sendwin_up(struct tcpcb *tp, struct tcp_rack *rack)
10036 {
10037         /*
10038          * If fast output is in progress, lets validate that
10039          * the new window did not shrink on us and make it
10040          * so fast output should end.
10041          */
10042         if (rack->r_fast_output) {
10043                 uint32_t out;
10044
10045                 /*
10046                  * Calculate what we will send if left as is
10047                  * and compare that to our send window.
10048                  */
10049                 out = ctf_outstanding(tp);
10050                 if ((out + rack->r_ctl.fsb.left_to_send) > tp->snd_wnd) {
10051                         /* ok we have an issue */
10052                         if (out >= tp->snd_wnd) {
10053                                 /* Turn off fast output the window is met or collapsed */
10054                                 rack->r_fast_output = 0;
10055                         } else {
10056                                 /* we have some room left */
10057                                 rack->r_ctl.fsb.left_to_send = tp->snd_wnd - out;
10058                                 if (rack->r_ctl.fsb.left_to_send < ctf_fixed_maxseg(tp)) {
10059                                         /* If not at least 1 full segment never mind */
10060                                         rack->r_fast_output = 0;
10061                                 }
10062                         }
10063                 }
10064         }
10065 }
10066
10067
10068 /*
10069  * Return value of 1, the TCB is unlocked and most
10070  * likely gone, return value of 0, the TCP is still
10071  * locked.
10072  */
10073 static int
10074 rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
10075     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
10076     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
10077 {
10078         /*
10079          * Update window information. Don't look at window if no ACK: TAC's
10080          * send garbage on first SYN.
10081          */
10082         int32_t nsegs;
10083         int32_t tfo_syn;
10084         struct tcp_rack *rack;
10085
10086         rack = (struct tcp_rack *)tp->t_fb_ptr;
10087         INP_WLOCK_ASSERT(tp->t_inpcb);
10088         nsegs = max(1, m->m_pkthdr.lro_nsegs);
10089         if ((thflags & TH_ACK) &&
10090             (SEQ_LT(tp->snd_wl1, th->th_seq) ||
10091             (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
10092             (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
10093                 /* keep track of pure window updates */
10094                 if (tlen == 0 &&
10095                     tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
10096                         KMOD_TCPSTAT_INC(tcps_rcvwinupd);
10097                 tp->snd_wnd = tiwin;
10098                 rack_validate_fo_sendwin_up(tp, rack);
10099                 tp->snd_wl1 = th->th_seq;
10100                 tp->snd_wl2 = th->th_ack;
10101                 if (tp->snd_wnd > tp->max_sndwnd)
10102                         tp->max_sndwnd = tp->snd_wnd;
10103                 rack->r_wanted_output = 1;
10104         } else if (thflags & TH_ACK) {
10105                 if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
10106                         tp->snd_wnd = tiwin;
10107                         rack_validate_fo_sendwin_up(tp, rack);
10108                         tp->snd_wl1 = th->th_seq;
10109                         tp->snd_wl2 = th->th_ack;
10110                 }
10111         }
10112         if (tp->snd_wnd < ctf_outstanding(tp))
10113                 /* The peer collapsed the window */
10114                 rack_collapsed_window(rack);
10115         else if (rack->rc_has_collapsed)
10116                 rack_un_collapse_window(rack);
10117         /* Was persist timer active and now we have window space? */
10118         if ((rack->rc_in_persist != 0) &&
10119             (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
10120                                 rack->r_ctl.rc_pace_min_segs))) {
10121                 rack_exit_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10122                 tp->snd_nxt = tp->snd_max;
10123                 /* Make sure we output to start the timer */
10124                 rack->r_wanted_output = 1;
10125         }
10126         /* Do we enter persists? */
10127         if ((rack->rc_in_persist == 0) &&
10128             (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
10129             TCPS_HAVEESTABLISHED(tp->t_state) &&
10130             (tp->snd_max == tp->snd_una) &&
10131             sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
10132             (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
10133                 /*
10134                  * Here the rwnd is less than
10135                  * the pacing size, we are established,
10136                  * nothing is outstanding, and there is
10137                  * data to send. Enter persists.
10138                  */
10139                 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10140         }
10141         if (tp->t_flags2 & TF2_DROP_AF_DATA) {
10142                 m_freem(m);
10143                 return (0);
10144         }
10145         /*
10146          * don't process the URG bit, ignore them drag
10147          * along the up.
10148          */
10149         tp->rcv_up = tp->rcv_nxt;
10150         INP_WLOCK_ASSERT(tp->t_inpcb);
10151
10152         /*
10153          * Process the segment text, merging it into the TCP sequencing
10154          * queue, and arranging for acknowledgment of receipt if necessary.
10155          * This process logically involves adjusting tp->rcv_wnd as data is
10156          * presented to the user (this happens in tcp_usrreq.c, case
10157          * PRU_RCVD).  If a FIN has already been received on this connection
10158          * then we just ignore the text.
10159          */
10160         tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
10161                    IS_FASTOPEN(tp->t_flags));
10162         if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) &&
10163             TCPS_HAVERCVDFIN(tp->t_state) == 0) {
10164                 tcp_seq save_start = th->th_seq;
10165                 tcp_seq save_rnxt  = tp->rcv_nxt;
10166                 int     save_tlen  = tlen;
10167
10168                 m_adj(m, drop_hdrlen);  /* delayed header drop */
10169                 /*
10170                  * Insert segment which includes th into TCP reassembly
10171                  * queue with control block tp.  Set thflags to whether
10172                  * reassembly now includes a segment with FIN.  This handles
10173                  * the common case inline (segment is the next to be
10174                  * received on an established connection, and the queue is
10175                  * empty), avoiding linkage into and removal from the queue
10176                  * and repetition of various conversions. Set DELACK for
10177                  * segments received in order, but ack immediately when
10178                  * segments are out of order (so fast retransmit can work).
10179                  */
10180                 if (th->th_seq == tp->rcv_nxt &&
10181                     SEGQ_EMPTY(tp) &&
10182                     (TCPS_HAVEESTABLISHED(tp->t_state) ||
10183                     tfo_syn)) {
10184 #ifdef NETFLIX_SB_LIMITS
10185                         u_int mcnt, appended;
10186
10187                         if (so->so_rcv.sb_shlim) {
10188                                 mcnt = m_memcnt(m);
10189                                 appended = 0;
10190                                 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
10191                                     CFO_NOSLEEP, NULL) == false) {
10192                                         counter_u64_add(tcp_sb_shlim_fails, 1);
10193                                         m_freem(m);
10194                                         return (0);
10195                                 }
10196                         }
10197 #endif
10198                         rack_handle_delayed_ack(tp, rack, tlen, tfo_syn);
10199                         tp->rcv_nxt += tlen;
10200                         if (tlen &&
10201                             ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
10202                             (tp->t_fbyte_in == 0)) {
10203                                 tp->t_fbyte_in = ticks;
10204                                 if (tp->t_fbyte_in == 0)
10205                                         tp->t_fbyte_in = 1;
10206                                 if (tp->t_fbyte_out && tp->t_fbyte_in)
10207                                         tp->t_flags2 |= TF2_FBYTES_COMPLETE;
10208                         }
10209                         thflags = th->th_flags & TH_FIN;
10210                         KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
10211                         KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
10212                         SOCKBUF_LOCK(&so->so_rcv);
10213                         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10214                                 m_freem(m);
10215                         } else
10216 #ifdef NETFLIX_SB_LIMITS
10217                                 appended =
10218 #endif
10219                                         sbappendstream_locked(&so->so_rcv, m, 0);
10220
10221                         rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
10222                         /* NB: sorwakeup_locked() does an implicit unlock. */
10223                         sorwakeup_locked(so);
10224 #ifdef NETFLIX_SB_LIMITS
10225                         if (so->so_rcv.sb_shlim && appended != mcnt)
10226                                 counter_fo_release(so->so_rcv.sb_shlim,
10227                                     mcnt - appended);
10228 #endif
10229                 } else {
10230                         /*
10231                          * XXX: Due to the header drop above "th" is
10232                          * theoretically invalid by now.  Fortunately
10233                          * m_adj() doesn't actually frees any mbufs when
10234                          * trimming from the head.
10235                          */
10236                         tcp_seq temp = save_start;
10237
10238                         thflags = tcp_reass(tp, th, &temp, &tlen, m);
10239                         tp->t_flags |= TF_ACKNOW;
10240                         if (tp->t_flags & TF_WAKESOR) {
10241                                 tp->t_flags &= ~TF_WAKESOR;
10242                                 /* NB: sorwakeup_locked() does an implicit unlock. */
10243                                 sorwakeup_locked(so);
10244                         }
10245                 }
10246                 if ((tp->t_flags & TF_SACK_PERMIT) &&
10247                     (save_tlen > 0) &&
10248                     TCPS_HAVEESTABLISHED(tp->t_state)) {
10249                         if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) {
10250                                 /*
10251                                  * DSACK actually handled in the fastpath
10252                                  * above.
10253                                  */
10254                                 RACK_OPTS_INC(tcp_sack_path_1);
10255                                 tcp_update_sack_list(tp, save_start,
10256                                     save_start + save_tlen);
10257                         } else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) {
10258                                 if ((tp->rcv_numsacks >= 1) &&
10259                                     (tp->sackblks[0].end == save_start)) {
10260                                         /*
10261                                          * Partial overlap, recorded at todrop
10262                                          * above.
10263                                          */
10264                                         RACK_OPTS_INC(tcp_sack_path_2a);
10265                                         tcp_update_sack_list(tp,
10266                                             tp->sackblks[0].start,
10267                                             tp->sackblks[0].end);
10268                                 } else {
10269                                         RACK_OPTS_INC(tcp_sack_path_2b);
10270                                         tcp_update_dsack_list(tp, save_start,
10271                                             save_start + save_tlen);
10272                                 }
10273                         } else if (tlen >= save_tlen) {
10274                                 /* Update of sackblks. */
10275                                 RACK_OPTS_INC(tcp_sack_path_3);
10276                                 tcp_update_dsack_list(tp, save_start,
10277                                     save_start + save_tlen);
10278                         } else if (tlen > 0) {
10279                                 RACK_OPTS_INC(tcp_sack_path_4);
10280                                 tcp_update_dsack_list(tp, save_start,
10281                                     save_start + tlen);
10282                         }
10283                 }
10284         } else {
10285                 m_freem(m);
10286                 thflags &= ~TH_FIN;
10287         }
10288
10289         /*
10290          * If FIN is received ACK the FIN and let the user know that the
10291          * connection is closing.
10292          */
10293         if (thflags & TH_FIN) {
10294                 if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
10295                         /* The socket upcall is handled by socantrcvmore. */
10296                         socantrcvmore(so);
10297                         /*
10298                          * If connection is half-synchronized (ie NEEDSYN
10299                          * flag on) then delay ACK, so it may be piggybacked
10300                          * when SYN is sent. Otherwise, since we received a
10301                          * FIN then no more input can be expected, send ACK
10302                          * now.
10303                          */
10304                         if (tp->t_flags & TF_NEEDSYN) {
10305                                 rack_timer_cancel(tp, rack,
10306                                     rack->r_ctl.rc_rcvtime, __LINE__);
10307                                 tp->t_flags |= TF_DELACK;
10308                         } else {
10309                                 tp->t_flags |= TF_ACKNOW;
10310                         }
10311                         tp->rcv_nxt++;
10312                 }
10313                 switch (tp->t_state) {
10314                         /*
10315                          * In SYN_RECEIVED and ESTABLISHED STATES enter the
10316                          * CLOSE_WAIT state.
10317                          */
10318                 case TCPS_SYN_RECEIVED:
10319                         tp->t_starttime = ticks;
10320                         /* FALLTHROUGH */
10321                 case TCPS_ESTABLISHED:
10322                         rack_timer_cancel(tp, rack,
10323                             rack->r_ctl.rc_rcvtime, __LINE__);
10324                         tcp_state_change(tp, TCPS_CLOSE_WAIT);
10325                         break;
10326
10327                         /*
10328                          * If still in FIN_WAIT_1 STATE FIN has not been
10329                          * acked so enter the CLOSING state.
10330                          */
10331                 case TCPS_FIN_WAIT_1:
10332                         rack_timer_cancel(tp, rack,
10333                             rack->r_ctl.rc_rcvtime, __LINE__);
10334                         tcp_state_change(tp, TCPS_CLOSING);
10335                         break;
10336
10337                         /*
10338                          * In FIN_WAIT_2 state enter the TIME_WAIT state,
10339                          * starting the time-wait timer, turning off the
10340                          * other standard timers.
10341                          */
10342                 case TCPS_FIN_WAIT_2:
10343                         rack_timer_cancel(tp, rack,
10344                             rack->r_ctl.rc_rcvtime, __LINE__);
10345                         tcp_twstart(tp);
10346                         return (1);
10347                 }
10348         }
10349         /*
10350          * Return any desired output.
10351          */
10352         if ((tp->t_flags & TF_ACKNOW) ||
10353             (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) {
10354                 rack->r_wanted_output = 1;
10355         }
10356         INP_WLOCK_ASSERT(tp->t_inpcb);
10357         return (0);
10358 }
10359
10360 /*
10361  * Here nothing is really faster, its just that we
10362  * have broken out the fast-data path also just like
10363  * the fast-ack.
10364  */
10365 static int
10366 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
10367     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10368     uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos)
10369 {
10370         int32_t nsegs;
10371         int32_t newsize = 0;    /* automatic sockbuf scaling */
10372         struct tcp_rack *rack;
10373 #ifdef NETFLIX_SB_LIMITS
10374         u_int mcnt, appended;
10375 #endif
10376 #ifdef TCPDEBUG
10377         /*
10378          * The size of tcp_saveipgen must be the size of the max ip header,
10379          * now IPv6.
10380          */
10381         u_char tcp_saveipgen[IP6_HDR_LEN];
10382         struct tcphdr tcp_savetcp;
10383         short ostate = 0;
10384
10385 #endif
10386         /*
10387          * If last ACK falls within this segment's sequence numbers, record
10388          * the timestamp. NOTE that the test is modified according to the
10389          * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
10390          */
10391         if (__predict_false(th->th_seq != tp->rcv_nxt)) {
10392                 return (0);
10393         }
10394         if (__predict_false(tp->snd_nxt != tp->snd_max)) {
10395                 return (0);
10396         }
10397         if (tiwin && tiwin != tp->snd_wnd) {
10398                 return (0);
10399         }
10400         if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
10401                 return (0);
10402         }
10403         if (__predict_false((to->to_flags & TOF_TS) &&
10404             (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
10405                 return (0);
10406         }
10407         if (__predict_false((th->th_ack != tp->snd_una))) {
10408                 return (0);
10409         }
10410         if (__predict_false(tlen > sbspace(&so->so_rcv))) {
10411                 return (0);
10412         }
10413         if ((to->to_flags & TOF_TS) != 0 &&
10414             SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
10415                 tp->ts_recent_age = tcp_ts_getticks();
10416                 tp->ts_recent = to->to_tsval;
10417         }
10418         rack = (struct tcp_rack *)tp->t_fb_ptr;
10419         /*
10420          * This is a pure, in-sequence data packet with nothing on the
10421          * reassembly queue and we have enough buffer space to take it.
10422          */
10423         nsegs = max(1, m->m_pkthdr.lro_nsegs);
10424
10425 #ifdef NETFLIX_SB_LIMITS
10426         if (so->so_rcv.sb_shlim) {
10427                 mcnt = m_memcnt(m);
10428                 appended = 0;
10429                 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
10430                     CFO_NOSLEEP, NULL) == false) {
10431                         counter_u64_add(tcp_sb_shlim_fails, 1);
10432                         m_freem(m);
10433                         return (1);
10434                 }
10435         }
10436 #endif
10437         /* Clean receiver SACK report if present */
10438         if (tp->rcv_numsacks)
10439                 tcp_clean_sackreport(tp);
10440         KMOD_TCPSTAT_INC(tcps_preddat);
10441         tp->rcv_nxt += tlen;
10442         if (tlen &&
10443             ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
10444             (tp->t_fbyte_in == 0)) {
10445                 tp->t_fbyte_in = ticks;
10446                 if (tp->t_fbyte_in == 0)
10447                         tp->t_fbyte_in = 1;
10448                 if (tp->t_fbyte_out && tp->t_fbyte_in)
10449                         tp->t_flags2 |= TF2_FBYTES_COMPLETE;
10450         }
10451         /*
10452          * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
10453          */
10454         tp->snd_wl1 = th->th_seq;
10455         /*
10456          * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
10457          */
10458         tp->rcv_up = tp->rcv_nxt;
10459         KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
10460         KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
10461 #ifdef TCPDEBUG
10462         if (so->so_options & SO_DEBUG)
10463                 tcp_trace(TA_INPUT, ostate, tp,
10464                     (void *)tcp_saveipgen, &tcp_savetcp, 0);
10465 #endif
10466         newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
10467
10468         /* Add data to socket buffer. */
10469         SOCKBUF_LOCK(&so->so_rcv);
10470         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10471                 m_freem(m);
10472         } else {
10473                 /*
10474                  * Set new socket buffer size. Give up when limit is
10475                  * reached.
10476                  */
10477                 if (newsize)
10478                         if (!sbreserve_locked(&so->so_rcv,
10479                             newsize, so, NULL))
10480                                 so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
10481                 m_adj(m, drop_hdrlen);  /* delayed header drop */
10482 #ifdef NETFLIX_SB_LIMITS
10483                 appended =
10484 #endif
10485                         sbappendstream_locked(&so->so_rcv, m, 0);
10486                 ctf_calc_rwin(so, tp);
10487         }
10488         rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
10489         /* NB: sorwakeup_locked() does an implicit unlock. */
10490         sorwakeup_locked(so);
10491 #ifdef NETFLIX_SB_LIMITS
10492         if (so->so_rcv.sb_shlim && mcnt != appended)
10493                 counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended);
10494 #endif
10495         rack_handle_delayed_ack(tp, rack, tlen, 0);
10496         if (tp->snd_una == tp->snd_max)
10497                 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
10498         return (1);
10499 }
10500
10501 /*
10502  * This subfunction is used to try to highly optimize the
10503  * fast path. We again allow window updates that are
10504  * in sequence to remain in the fast-path. We also add
10505  * in the __predict's to attempt to help the compiler.
10506  * Note that if we return a 0, then we can *not* process
10507  * it and the caller should push the packet into the
10508  * slow-path.
10509  */
10510 static int
10511 rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
10512     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10513     uint32_t tiwin, int32_t nxt_pkt, uint32_t cts)
10514 {
10515         int32_t acked;
10516         int32_t nsegs;
10517 #ifdef TCPDEBUG
10518         /*
10519          * The size of tcp_saveipgen must be the size of the max ip header,
10520          * now IPv6.
10521          */
10522         u_char tcp_saveipgen[IP6_HDR_LEN];
10523         struct tcphdr tcp_savetcp;
10524         short ostate = 0;
10525 #endif
10526         int32_t under_pacing = 0;
10527         struct tcp_rack *rack;
10528
10529         if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
10530                 /* Old ack, behind (or duplicate to) the last one rcv'd */
10531                 return (0);
10532         }
10533         if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
10534                 /* Above what we have sent? */
10535                 return (0);
10536         }
10537         if (__predict_false(tp->snd_nxt != tp->snd_max)) {
10538                 /* We are retransmitting */
10539                 return (0);
10540         }
10541         if (__predict_false(tiwin == 0)) {
10542                 /* zero window */
10543                 return (0);
10544         }
10545         if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
10546                 /* We need a SYN or a FIN, unlikely.. */
10547                 return (0);
10548         }
10549         if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
10550                 /* Timestamp is behind .. old ack with seq wrap? */
10551                 return (0);
10552         }
10553         if (__predict_false(IN_RECOVERY(tp->t_flags))) {
10554                 /* Still recovering */
10555                 return (0);
10556         }
10557         rack = (struct tcp_rack *)tp->t_fb_ptr;
10558         if (rack->r_ctl.rc_sacked) {
10559                 /* We have sack holes on our scoreboard */
10560                 return (0);
10561         }
10562         /* Ok if we reach here, we can process a fast-ack */
10563         if (rack->gp_ready &&
10564             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
10565                 under_pacing = 1;
10566         }
10567         nsegs = max(1, m->m_pkthdr.lro_nsegs);
10568         rack_log_ack(tp, to, th, 0, 0);
10569         /* Did the window get updated? */
10570         if (tiwin != tp->snd_wnd) {
10571                 tp->snd_wnd = tiwin;
10572                 rack_validate_fo_sendwin_up(tp, rack);
10573                 tp->snd_wl1 = th->th_seq;
10574                 if (tp->snd_wnd > tp->max_sndwnd)
10575                         tp->max_sndwnd = tp->snd_wnd;
10576         }
10577         /* Do we exit persists? */
10578         if ((rack->rc_in_persist != 0) &&
10579             (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
10580                                rack->r_ctl.rc_pace_min_segs))) {
10581                 rack_exit_persist(tp, rack, cts);
10582         }
10583         /* Do we enter persists? */
10584         if ((rack->rc_in_persist == 0) &&
10585             (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
10586             TCPS_HAVEESTABLISHED(tp->t_state) &&
10587             (tp->snd_max == tp->snd_una) &&
10588             sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
10589             (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
10590                 /*
10591                  * Here the rwnd is less than
10592                  * the pacing size, we are established,
10593                  * nothing is outstanding, and there is
10594                  * data to send. Enter persists.
10595                  */
10596                 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10597         }
10598         /*
10599          * If last ACK falls within this segment's sequence numbers, record
10600          * the timestamp. NOTE that the test is modified according to the
10601          * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
10602          */
10603         if ((to->to_flags & TOF_TS) != 0 &&
10604             SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
10605                 tp->ts_recent_age = tcp_ts_getticks();
10606                 tp->ts_recent = to->to_tsval;
10607         }
10608         /*
10609          * This is a pure ack for outstanding data.
10610          */
10611         KMOD_TCPSTAT_INC(tcps_predack);
10612
10613         /*
10614          * "bad retransmit" recovery.
10615          */
10616         if ((tp->t_flags & TF_PREVVALID) &&
10617             ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
10618                 tp->t_flags &= ~TF_PREVVALID;
10619                 if (tp->t_rxtshift == 1 &&
10620                     (int)(ticks - tp->t_badrxtwin) < 0)
10621                         rack_cong_signal(tp, CC_RTO_ERR, th->th_ack);
10622         }
10623         /*
10624          * Recalculate the transmit timer / rtt.
10625          *
10626          * Some boxes send broken timestamp replies during the SYN+ACK
10627          * phase, ignore timestamps of 0 or we could calculate a huge RTT
10628          * and blow up the retransmit timer.
10629          */
10630         acked = BYTES_THIS_ACK(tp, th);
10631
10632 #ifdef TCP_HHOOK
10633         /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
10634         hhook_run_tcp_est_in(tp, th, to);
10635 #endif
10636         KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
10637         KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
10638         if (acked) {
10639                 struct mbuf *mfree;
10640
10641                 rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, 0);
10642                 SOCKBUF_LOCK(&so->so_snd);
10643                 mfree = sbcut_locked(&so->so_snd, acked);
10644                 tp->snd_una = th->th_ack;
10645                 /* Note we want to hold the sb lock through the sendmap adjust */
10646                 rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
10647                 /* Wake up the socket if we have room to write more */
10648                 rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
10649                 sowwakeup_locked(so);
10650                 m_freem(mfree);
10651                 tp->t_rxtshift = 0;
10652                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
10653                               rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
10654                 rack->rc_tlp_in_progress = 0;
10655                 rack->r_ctl.rc_tlp_cnt_out = 0;
10656                 /*
10657                  * If it is the RXT timer we want to
10658                  * stop it, so we can restart a TLP.
10659                  */
10660                 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
10661                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
10662 #ifdef NETFLIX_HTTP_LOGGING
10663                 tcp_http_check_for_comp(rack->rc_tp, th->th_ack);
10664 #endif
10665         }
10666         /*
10667          * Let the congestion control algorithm update congestion control
10668          * related information. This typically means increasing the
10669          * congestion window.
10670          */
10671         if (tp->snd_wnd < ctf_outstanding(tp)) {
10672                 /* The peer collapsed the window */
10673                 rack_collapsed_window(rack);
10674         } else if (rack->rc_has_collapsed)
10675                 rack_un_collapse_window(rack);
10676
10677         /*
10678          * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
10679          */
10680         tp->snd_wl2 = th->th_ack;
10681         tp->t_dupacks = 0;
10682         m_freem(m);
10683         /* ND6_HINT(tp);         *//* Some progress has been made. */
10684
10685         /*
10686          * If all outstanding data are acked, stop retransmit timer,
10687          * otherwise restart timer using current (possibly backed-off)
10688          * value. If process is waiting for space, wakeup/selwakeup/signal.
10689          * If data are ready to send, let tcp_output decide between more
10690          * output or persist.
10691          */
10692 #ifdef TCPDEBUG
10693         if (so->so_options & SO_DEBUG)
10694                 tcp_trace(TA_INPUT, ostate, tp,
10695                     (void *)tcp_saveipgen,
10696                     &tcp_savetcp, 0);
10697 #endif
10698         if (under_pacing &&
10699             (rack->use_fixed_rate == 0) &&
10700             (rack->in_probe_rtt == 0) &&
10701             rack->rc_gp_dyn_mul &&
10702             rack->rc_always_pace) {
10703                 /* Check if we are dragging bottom */
10704                 rack_check_bottom_drag(tp, rack, so, acked);
10705         }
10706         if (tp->snd_una == tp->snd_max) {
10707                 tp->t_flags &= ~TF_PREVVALID;
10708                 rack->r_ctl.retran_during_recovery = 0;
10709                 rack->r_ctl.dsack_byte_cnt = 0;
10710                 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
10711                 if (rack->r_ctl.rc_went_idle_time == 0)
10712                         rack->r_ctl.rc_went_idle_time = 1;
10713                 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
10714                 if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
10715                         tp->t_acktime = 0;
10716                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
10717         }
10718         if (acked && rack->r_fast_output)
10719                 rack_gain_for_fastoutput(rack, tp, so, (uint32_t)acked);
10720         if (sbavail(&so->so_snd)) {
10721                 rack->r_wanted_output = 1;
10722         }
10723         return (1);
10724 }
10725
10726 /*
10727  * Return value of 1, the TCB is unlocked and most
10728  * likely gone, return value of 0, the TCP is still
10729  * locked.
10730  */
10731 static int
10732 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
10733     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10734     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
10735 {
10736         int32_t ret_val = 0;
10737         int32_t todrop;
10738         int32_t ourfinisacked = 0;
10739         struct tcp_rack *rack;
10740
10741         ctf_calc_rwin(so, tp);
10742         /*
10743          * If the state is SYN_SENT: if seg contains an ACK, but not for our
10744          * SYN, drop the input. if seg contains a RST, then drop the
10745          * connection. if seg does not contain SYN, then drop it. Otherwise
10746          * this is an acceptable SYN segment initialize tp->rcv_nxt and
10747          * tp->irs if seg contains ack then advance tp->snd_una if seg
10748          * contains an ECE and ECN support is enabled, the stream is ECN
10749          * capable. if SYN has been acked change to ESTABLISHED else
10750          * SYN_RCVD state arrange for segment to be acked (eventually)
10751          * continue processing rest of data/controls.
10752          */
10753         if ((thflags & TH_ACK) &&
10754             (SEQ_LEQ(th->th_ack, tp->iss) ||
10755             SEQ_GT(th->th_ack, tp->snd_max))) {
10756                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
10757                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
10758                 return (1);
10759         }
10760         if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
10761                 TCP_PROBE5(connect__refused, NULL, tp,
10762                     mtod(m, const char *), tp, th);
10763                 tp = tcp_drop(tp, ECONNREFUSED);
10764                 ctf_do_drop(m, tp);
10765                 return (1);
10766         }
10767         if (thflags & TH_RST) {
10768                 ctf_do_drop(m, tp);
10769                 return (1);
10770         }
10771         if (!(thflags & TH_SYN)) {
10772                 ctf_do_drop(m, tp);
10773                 return (1);
10774         }
10775         tp->irs = th->th_seq;
10776         tcp_rcvseqinit(tp);
10777         rack = (struct tcp_rack *)tp->t_fb_ptr;
10778         if (thflags & TH_ACK) {
10779                 int tfo_partial = 0;
10780
10781                 KMOD_TCPSTAT_INC(tcps_connects);
10782                 soisconnected(so);
10783 #ifdef MAC
10784                 mac_socketpeer_set_from_mbuf(m, so);
10785 #endif
10786                 /* Do window scaling on this connection? */
10787                 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
10788                     (TF_RCVD_SCALE | TF_REQ_SCALE)) {
10789                         tp->rcv_scale = tp->request_r_scale;
10790                 }
10791                 tp->rcv_adv += min(tp->rcv_wnd,
10792                     TCP_MAXWIN << tp->rcv_scale);
10793                 /*
10794                  * If not all the data that was sent in the TFO SYN
10795                  * has been acked, resend the remainder right away.
10796                  */
10797                 if (IS_FASTOPEN(tp->t_flags) &&
10798                     (tp->snd_una != tp->snd_max)) {
10799                         tp->snd_nxt = th->th_ack;
10800                         tfo_partial = 1;
10801                 }
10802                 /*
10803                  * If there's data, delay ACK; if there's also a FIN ACKNOW
10804                  * will be turned on later.
10805                  */
10806                 if (DELAY_ACK(tp, tlen) && tlen != 0 && !tfo_partial) {
10807                         rack_timer_cancel(tp, rack,
10808                                           rack->r_ctl.rc_rcvtime, __LINE__);
10809                         tp->t_flags |= TF_DELACK;
10810                 } else {
10811                         rack->r_wanted_output = 1;
10812                         tp->t_flags |= TF_ACKNOW;
10813                         rack->rc_dack_toggle = 0;
10814                 }
10815                 if (((thflags & (TH_CWR | TH_ECE)) == TH_ECE) &&
10816                     (V_tcp_do_ecn == 1)) {
10817                         tp->t_flags2 |= TF2_ECN_PERMIT;
10818                         KMOD_TCPSTAT_INC(tcps_ecn_shs);
10819                 }
10820                 if (SEQ_GT(th->th_ack, tp->snd_una)) {
10821                         /*
10822                          * We advance snd_una for the
10823                          * fast open case. If th_ack is
10824                          * acknowledging data beyond
10825                          * snd_una we can't just call
10826                          * ack-processing since the
10827                          * data stream in our send-map
10828                          * will start at snd_una + 1 (one
10829                          * beyond the SYN). If its just
10830                          * equal we don't need to do that
10831                          * and there is no send_map.
10832                          */
10833                         tp->snd_una++;
10834                 }
10835                 /*
10836                  * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
10837                  * SYN_SENT  --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
10838                  */
10839                 tp->t_starttime = ticks;
10840                 if (tp->t_flags & TF_NEEDFIN) {
10841                         tcp_state_change(tp, TCPS_FIN_WAIT_1);
10842                         tp->t_flags &= ~TF_NEEDFIN;
10843                         thflags &= ~TH_SYN;
10844                 } else {
10845                         tcp_state_change(tp, TCPS_ESTABLISHED);
10846                         TCP_PROBE5(connect__established, NULL, tp,
10847                             mtod(m, const char *), tp, th);
10848                         rack_cc_conn_init(tp);
10849                 }
10850         } else {
10851                 /*
10852                  * Received initial SYN in SYN-SENT[*] state => simultaneous
10853                  * open.  If segment contains CC option and there is a
10854                  * cached CC, apply TAO test. If it succeeds, connection is *
10855                  * half-synchronized. Otherwise, do 3-way handshake:
10856                  * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
10857                  * there was no CC option, clear cached CC value.
10858                  */
10859                 tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN);
10860                 tcp_state_change(tp, TCPS_SYN_RECEIVED);
10861         }
10862         INP_WLOCK_ASSERT(tp->t_inpcb);
10863         /*
10864          * Advance th->th_seq to correspond to first data byte. If data,
10865          * trim to stay within window, dropping FIN if necessary.
10866          */
10867         th->th_seq++;
10868         if (tlen > tp->rcv_wnd) {
10869                 todrop = tlen - tp->rcv_wnd;
10870                 m_adj(m, -todrop);
10871                 tlen = tp->rcv_wnd;
10872                 thflags &= ~TH_FIN;
10873                 KMOD_TCPSTAT_INC(tcps_rcvpackafterwin);
10874                 KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
10875         }
10876         tp->snd_wl1 = th->th_seq - 1;
10877         tp->rcv_up = th->th_seq;
10878         /*
10879          * Client side of transaction: already sent SYN and data. If the
10880          * remote host used T/TCP to validate the SYN, our data will be
10881          * ACK'd; if so, enter normal data segment processing in the middle
10882          * of step 5, ack processing. Otherwise, goto step 6.
10883          */
10884         if (thflags & TH_ACK) {
10885                 /* For syn-sent we need to possibly update the rtt */
10886                 if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
10887                         uint32_t t, mcts;
10888
10889                         mcts = tcp_ts_getticks();
10890                         t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
10891                         if (!tp->t_rttlow || tp->t_rttlow > t)
10892                                 tp->t_rttlow = t;
10893                         rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 4);
10894                         tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
10895                         tcp_rack_xmit_timer_commit(rack, tp);
10896                 }
10897                 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val))
10898                         return (ret_val);
10899                 /* We may have changed to FIN_WAIT_1 above */
10900                 if (tp->t_state == TCPS_FIN_WAIT_1) {
10901                         /*
10902                          * In FIN_WAIT_1 STATE in addition to the processing
10903                          * for the ESTABLISHED state if our FIN is now
10904                          * acknowledged then enter FIN_WAIT_2.
10905                          */
10906                         if (ourfinisacked) {
10907                                 /*
10908                                  * If we can't receive any more data, then
10909                                  * closing user can proceed. Starting the
10910                                  * timer is contrary to the specification,
10911                                  * but if we don't get a FIN we'll hang
10912                                  * forever.
10913                                  *
10914                                  * XXXjl: we should release the tp also, and
10915                                  * use a compressed state.
10916                                  */
10917                                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10918                                         soisdisconnected(so);
10919                                         tcp_timer_activate(tp, TT_2MSL,
10920                                             (tcp_fast_finwait2_recycle ?
10921                                             tcp_finwait2_timeout :
10922                                             TP_MAXIDLE(tp)));
10923                                 }
10924                                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
10925                         }
10926                 }
10927         }
10928         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
10929            tiwin, thflags, nxt_pkt));
10930 }
10931
10932 /*
10933  * Return value of 1, the TCB is unlocked and most
10934  * likely gone, return value of 0, the TCP is still
10935  * locked.
10936  */
10937 static int
10938 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
10939     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10940     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
10941 {
10942         struct tcp_rack *rack;
10943         int32_t ret_val = 0;
10944         int32_t ourfinisacked = 0;
10945
10946         ctf_calc_rwin(so, tp);
10947         if ((thflags & TH_ACK) &&
10948             (SEQ_LEQ(th->th_ack, tp->snd_una) ||
10949             SEQ_GT(th->th_ack, tp->snd_max))) {
10950                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
10951                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
10952                 return (1);
10953         }
10954         rack = (struct tcp_rack *)tp->t_fb_ptr;
10955         if (IS_FASTOPEN(tp->t_flags)) {
10956                 /*
10957                  * When a TFO connection is in SYN_RECEIVED, the
10958                  * only valid packets are the initial SYN, a
10959                  * retransmit/copy of the initial SYN (possibly with
10960                  * a subset of the original data), a valid ACK, a
10961                  * FIN, or a RST.
10962                  */
10963                 if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
10964                         tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
10965                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
10966                         return (1);
10967                 } else if (thflags & TH_SYN) {
10968                         /* non-initial SYN is ignored */
10969                         if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
10970                             (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
10971                             (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
10972                                 ctf_do_drop(m, NULL);
10973                                 return (0);
10974                         }
10975                 } else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
10976                         ctf_do_drop(m, NULL);
10977                         return (0);
10978                 }
10979         }
10980         if ((thflags & TH_RST) ||
10981             (tp->t_fin_is_rst && (thflags & TH_FIN)))
10982                 return (ctf_process_rst(m, th, so, tp));
10983         /*
10984          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
10985          * it's less than ts_recent, drop it.
10986          */
10987         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
10988             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
10989                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
10990                         return (ret_val);
10991         }
10992         /*
10993          * In the SYN-RECEIVED state, validate that the packet belongs to
10994          * this connection before trimming the data to fit the receive
10995          * window.  Check the sequence number versus IRS since we know the
10996          * sequence numbers haven't wrapped.  This is a partial fix for the
10997          * "LAND" DoS attack.
10998          */
10999         if (SEQ_LT(th->th_seq, tp->irs)) {
11000                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11001                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11002                 return (1);
11003         }
11004         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11005                               &rack->r_ctl.challenge_ack_ts,
11006                               &rack->r_ctl.challenge_ack_cnt)) {
11007                 return (ret_val);
11008         }
11009         /*
11010          * If last ACK falls within this segment's sequence numbers, record
11011          * its timestamp. NOTE: 1) That the test incorporates suggestions
11012          * from the latest proposal of the tcplw@cray.com list (Braden
11013          * 1993/04/26). 2) That updating only on newer timestamps interferes
11014          * with our earlier PAWS tests, so this check should be solely
11015          * predicated on the sequence space of this segment. 3) That we
11016          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11017          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11018          * SEG.Len, This modified check allows us to overcome RFC1323's
11019          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11020          * p.869. In such cases, we can still calculate the RTT correctly
11021          * when RCV.NXT == Last.ACK.Sent.
11022          */
11023         if ((to->to_flags & TOF_TS) != 0 &&
11024             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11025             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11026             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11027                 tp->ts_recent_age = tcp_ts_getticks();
11028                 tp->ts_recent = to->to_tsval;
11029         }
11030         tp->snd_wnd = tiwin;
11031         rack_validate_fo_sendwin_up(tp, rack);
11032         /*
11033          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11034          * is on (half-synchronized state), then queue data for later
11035          * processing; else drop segment and return.
11036          */
11037         if ((thflags & TH_ACK) == 0) {
11038                 if (IS_FASTOPEN(tp->t_flags)) {
11039                         rack_cc_conn_init(tp);
11040                 }
11041                 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11042                     tiwin, thflags, nxt_pkt));
11043         }
11044         KMOD_TCPSTAT_INC(tcps_connects);
11045         soisconnected(so);
11046         /* Do window scaling? */
11047         if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
11048             (TF_RCVD_SCALE | TF_REQ_SCALE)) {
11049                 tp->rcv_scale = tp->request_r_scale;
11050         }
11051         /*
11052          * Make transitions: SYN-RECEIVED  -> ESTABLISHED SYN-RECEIVED* ->
11053          * FIN-WAIT-1
11054          */
11055         tp->t_starttime = ticks;
11056         if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) {
11057                 tcp_fastopen_decrement_counter(tp->t_tfo_pending);
11058                 tp->t_tfo_pending = NULL;
11059         }
11060         if (tp->t_flags & TF_NEEDFIN) {
11061                 tcp_state_change(tp, TCPS_FIN_WAIT_1);
11062                 tp->t_flags &= ~TF_NEEDFIN;
11063         } else {
11064                 tcp_state_change(tp, TCPS_ESTABLISHED);
11065                 TCP_PROBE5(accept__established, NULL, tp,
11066                     mtod(m, const char *), tp, th);
11067                 /*
11068                  * TFO connections call cc_conn_init() during SYN
11069                  * processing.  Calling it again here for such connections
11070                  * is not harmless as it would undo the snd_cwnd reduction
11071                  * that occurs when a TFO SYN|ACK is retransmitted.
11072                  */
11073                 if (!IS_FASTOPEN(tp->t_flags))
11074                         rack_cc_conn_init(tp);
11075         }
11076         /*
11077          * Account for the ACK of our SYN prior to
11078          * regular ACK processing below, except for
11079          * simultaneous SYN, which is handled later.
11080          */
11081         if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN))
11082                 tp->snd_una++;
11083         /*
11084          * If segment contains data or ACK, will call tcp_reass() later; if
11085          * not, do so now to pass queued data to user.
11086          */
11087         if (tlen == 0 && (thflags & TH_FIN) == 0) {
11088                 (void) tcp_reass(tp, (struct tcphdr *)0, NULL, 0,
11089                     (struct mbuf *)0);
11090                 if (tp->t_flags & TF_WAKESOR) {
11091                         tp->t_flags &= ~TF_WAKESOR;
11092                         /* NB: sorwakeup_locked() does an implicit unlock. */
11093                         sorwakeup_locked(so);
11094                 }
11095         }
11096         tp->snd_wl1 = th->th_seq - 1;
11097         /* For syn-recv we need to possibly update the rtt */
11098         if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
11099                 uint32_t t, mcts;
11100
11101                 mcts = tcp_ts_getticks();
11102                 t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
11103                 if (!tp->t_rttlow || tp->t_rttlow > t)
11104                         tp->t_rttlow = t;
11105                 rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 5);
11106                 tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
11107                 tcp_rack_xmit_timer_commit(rack, tp);
11108         }
11109         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11110                 return (ret_val);
11111         }
11112         if (tp->t_state == TCPS_FIN_WAIT_1) {
11113                 /* We could have went to FIN_WAIT_1 (or EST) above */
11114                 /*
11115                  * In FIN_WAIT_1 STATE in addition to the processing for the
11116                  * ESTABLISHED state if our FIN is now acknowledged then
11117                  * enter FIN_WAIT_2.
11118                  */
11119                 if (ourfinisacked) {
11120                         /*
11121                          * If we can't receive any more data, then closing
11122                          * user can proceed. Starting the timer is contrary
11123                          * to the specification, but if we don't get a FIN
11124                          * we'll hang forever.
11125                          *
11126                          * XXXjl: we should release the tp also, and use a
11127                          * compressed state.
11128                          */
11129                         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11130                                 soisdisconnected(so);
11131                                 tcp_timer_activate(tp, TT_2MSL,
11132                                     (tcp_fast_finwait2_recycle ?
11133                                     tcp_finwait2_timeout :
11134                                     TP_MAXIDLE(tp)));
11135                         }
11136                         tcp_state_change(tp, TCPS_FIN_WAIT_2);
11137                 }
11138         }
11139         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11140             tiwin, thflags, nxt_pkt));
11141 }
11142
11143 /*
11144  * Return value of 1, the TCB is unlocked and most
11145  * likely gone, return value of 0, the TCP is still
11146  * locked.
11147  */
11148 static int
11149 rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
11150     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11151     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11152 {
11153         int32_t ret_val = 0;
11154         struct tcp_rack *rack;
11155
11156         /*
11157          * Header prediction: check for the two common cases of a
11158          * uni-directional data xfer.  If the packet has no control flags,
11159          * is in-sequence, the window didn't change and we're not
11160          * retransmitting, it's a candidate.  If the length is zero and the
11161          * ack moved forward, we're the sender side of the xfer.  Just free
11162          * the data acked & wake any higher level process that was blocked
11163          * waiting for space.  If the length is non-zero and the ack didn't
11164          * move, we're the receiver side.  If we're getting packets in-order
11165          * (the reassembly queue is empty), add the data toc The socket
11166          * buffer and note that we need a delayed ack. Make sure that the
11167          * hidden state-flags are also off. Since we check for
11168          * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
11169          */
11170         rack = (struct tcp_rack *)tp->t_fb_ptr;
11171         if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
11172             __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_ACK)) == TH_ACK) &&
11173             __predict_true(SEGQ_EMPTY(tp)) &&
11174             __predict_true(th->th_seq == tp->rcv_nxt)) {
11175                 if (tlen == 0) {
11176                         if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
11177                             tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) {
11178                                 return (0);
11179                         }
11180                 } else {
11181                         if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
11182                             tiwin, nxt_pkt, iptos)) {
11183                                 return (0);
11184                         }
11185                 }
11186         }
11187         ctf_calc_rwin(so, tp);
11188
11189         if ((thflags & TH_RST) ||
11190             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11191                 return (ctf_process_rst(m, th, so, tp));
11192
11193         /*
11194          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11195          * synchronized state.
11196          */
11197         if (thflags & TH_SYN) {
11198                 ctf_challenge_ack(m, th, tp, &ret_val);
11199                 return (ret_val);
11200         }
11201         /*
11202          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11203          * it's less than ts_recent, drop it.
11204          */
11205         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11206             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11207                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11208                         return (ret_val);
11209         }
11210         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11211                               &rack->r_ctl.challenge_ack_ts,
11212                               &rack->r_ctl.challenge_ack_cnt)) {
11213                 return (ret_val);
11214         }
11215         /*
11216          * If last ACK falls within this segment's sequence numbers, record
11217          * its timestamp. NOTE: 1) That the test incorporates suggestions
11218          * from the latest proposal of the tcplw@cray.com list (Braden
11219          * 1993/04/26). 2) That updating only on newer timestamps interferes
11220          * with our earlier PAWS tests, so this check should be solely
11221          * predicated on the sequence space of this segment. 3) That we
11222          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11223          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11224          * SEG.Len, This modified check allows us to overcome RFC1323's
11225          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11226          * p.869. In such cases, we can still calculate the RTT correctly
11227          * when RCV.NXT == Last.ACK.Sent.
11228          */
11229         if ((to->to_flags & TOF_TS) != 0 &&
11230             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11231             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11232             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11233                 tp->ts_recent_age = tcp_ts_getticks();
11234                 tp->ts_recent = to->to_tsval;
11235         }
11236         /*
11237          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11238          * is on (half-synchronized state), then queue data for later
11239          * processing; else drop segment and return.
11240          */
11241         if ((thflags & TH_ACK) == 0) {
11242                 if (tp->t_flags & TF_NEEDSYN) {
11243                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11244                             tiwin, thflags, nxt_pkt));
11245
11246                 } else if (tp->t_flags & TF_ACKNOW) {
11247                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11248                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11249                         return (ret_val);
11250                 } else {
11251                         ctf_do_drop(m, NULL);
11252                         return (0);
11253                 }
11254         }
11255         /*
11256          * Ack processing.
11257          */
11258         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
11259                 return (ret_val);
11260         }
11261         if (sbavail(&so->so_snd)) {
11262                 if (ctf_progress_timeout_check(tp, true)) {
11263                         rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
11264                         tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11265                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11266                         return (1);
11267                 }
11268         }
11269         /* State changes only happen in rack_process_data() */
11270         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11271             tiwin, thflags, nxt_pkt));
11272 }
11273
11274 /*
11275  * Return value of 1, the TCB is unlocked and most
11276  * likely gone, return value of 0, the TCP is still
11277  * locked.
11278  */
11279 static int
11280 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
11281     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11282     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11283 {
11284         int32_t ret_val = 0;
11285         struct tcp_rack *rack;
11286
11287         rack = (struct tcp_rack *)tp->t_fb_ptr;
11288         ctf_calc_rwin(so, tp);
11289         if ((thflags & TH_RST) ||
11290             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11291                 return (ctf_process_rst(m, th, so, tp));
11292         /*
11293          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11294          * synchronized state.
11295          */
11296         if (thflags & TH_SYN) {
11297                 ctf_challenge_ack(m, th, tp, &ret_val);
11298                 return (ret_val);
11299         }
11300         /*
11301          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11302          * it's less than ts_recent, drop it.
11303          */
11304         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11305             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11306                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11307                         return (ret_val);
11308         }
11309         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11310                               &rack->r_ctl.challenge_ack_ts,
11311                               &rack->r_ctl.challenge_ack_cnt)) {
11312                 return (ret_val);
11313         }
11314         /*
11315          * If last ACK falls within this segment's sequence numbers, record
11316          * its timestamp. NOTE: 1) That the test incorporates suggestions
11317          * from the latest proposal of the tcplw@cray.com list (Braden
11318          * 1993/04/26). 2) That updating only on newer timestamps interferes
11319          * with our earlier PAWS tests, so this check should be solely
11320          * predicated on the sequence space of this segment. 3) That we
11321          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11322          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11323          * SEG.Len, This modified check allows us to overcome RFC1323's
11324          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11325          * p.869. In such cases, we can still calculate the RTT correctly
11326          * when RCV.NXT == Last.ACK.Sent.
11327          */
11328         if ((to->to_flags & TOF_TS) != 0 &&
11329             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11330             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11331             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11332                 tp->ts_recent_age = tcp_ts_getticks();
11333                 tp->ts_recent = to->to_tsval;
11334         }
11335         /*
11336          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11337          * is on (half-synchronized state), then queue data for later
11338          * processing; else drop segment and return.
11339          */
11340         if ((thflags & TH_ACK) == 0) {
11341                 if (tp->t_flags & TF_NEEDSYN) {
11342                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11343                             tiwin, thflags, nxt_pkt));
11344
11345                 } else if (tp->t_flags & TF_ACKNOW) {
11346                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11347                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11348                         return (ret_val);
11349                 } else {
11350                         ctf_do_drop(m, NULL);
11351                         return (0);
11352                 }
11353         }
11354         /*
11355          * Ack processing.
11356          */
11357         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
11358                 return (ret_val);
11359         }
11360         if (sbavail(&so->so_snd)) {
11361                 if (ctf_progress_timeout_check(tp, true)) {
11362                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11363                                                 tp, tick, PROGRESS_DROP, __LINE__);
11364                         tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11365                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11366                         return (1);
11367                 }
11368         }
11369         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11370             tiwin, thflags, nxt_pkt));
11371 }
11372
11373 static int
11374 rack_check_data_after_close(struct mbuf *m,
11375     struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so)
11376 {
11377         struct tcp_rack *rack;
11378
11379         rack = (struct tcp_rack *)tp->t_fb_ptr;
11380         if (rack->rc_allow_data_af_clo == 0) {
11381         close_now:
11382                 tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
11383                 /* tcp_close will kill the inp pre-log the Reset */
11384                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
11385                 tp = tcp_close(tp);
11386                 KMOD_TCPSTAT_INC(tcps_rcvafterclose);
11387                 ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen));
11388                 return (1);
11389         }
11390         if (sbavail(&so->so_snd) == 0)
11391                 goto close_now;
11392         /* Ok we allow data that is ignored and a followup reset */
11393         tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
11394         tp->rcv_nxt = th->th_seq + *tlen;
11395         tp->t_flags2 |= TF2_DROP_AF_DATA;
11396         rack->r_wanted_output = 1;
11397         *tlen = 0;
11398         return (0);
11399 }
11400
11401 /*
11402  * Return value of 1, the TCB is unlocked and most
11403  * likely gone, return value of 0, the TCP is still
11404  * locked.
11405  */
11406 static int
11407 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
11408     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11409     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11410 {
11411         int32_t ret_val = 0;
11412         int32_t ourfinisacked = 0;
11413         struct tcp_rack *rack;
11414
11415         rack = (struct tcp_rack *)tp->t_fb_ptr;
11416         ctf_calc_rwin(so, tp);
11417
11418         if ((thflags & TH_RST) ||
11419             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11420                 return (ctf_process_rst(m, th, so, tp));
11421         /*
11422          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11423          * synchronized state.
11424          */
11425         if (thflags & TH_SYN) {
11426                 ctf_challenge_ack(m, th, tp, &ret_val);
11427                 return (ret_val);
11428         }
11429         /*
11430          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11431          * it's less than ts_recent, drop it.
11432          */
11433         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11434             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11435                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11436                         return (ret_val);
11437         }
11438         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11439                               &rack->r_ctl.challenge_ack_ts,
11440                               &rack->r_ctl.challenge_ack_cnt)) {
11441                 return (ret_val);
11442         }
11443         /*
11444          * If new data are received on a connection after the user processes
11445          * are gone, then RST the other end.
11446          */
11447         if ((so->so_state & SS_NOFDREF) && tlen) {
11448                 if (rack_check_data_after_close(m, tp, &tlen, th, so))
11449                         return (1);
11450         }
11451         /*
11452          * If last ACK falls within this segment's sequence numbers, record
11453          * its timestamp. NOTE: 1) That the test incorporates suggestions
11454          * from the latest proposal of the tcplw@cray.com list (Braden
11455          * 1993/04/26). 2) That updating only on newer timestamps interferes
11456          * with our earlier PAWS tests, so this check should be solely
11457          * predicated on the sequence space of this segment. 3) That we
11458          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11459          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11460          * SEG.Len, This modified check allows us to overcome RFC1323's
11461          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11462          * p.869. In such cases, we can still calculate the RTT correctly
11463          * when RCV.NXT == Last.ACK.Sent.
11464          */
11465         if ((to->to_flags & TOF_TS) != 0 &&
11466             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11467             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11468             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11469                 tp->ts_recent_age = tcp_ts_getticks();
11470                 tp->ts_recent = to->to_tsval;
11471         }
11472         /*
11473          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11474          * is on (half-synchronized state), then queue data for later
11475          * processing; else drop segment and return.
11476          */
11477         if ((thflags & TH_ACK) == 0) {
11478                 if (tp->t_flags & TF_NEEDSYN) {
11479                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11480                             tiwin, thflags, nxt_pkt));
11481                 } else if (tp->t_flags & TF_ACKNOW) {
11482                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11483                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11484                         return (ret_val);
11485                 } else {
11486                         ctf_do_drop(m, NULL);
11487                         return (0);
11488                 }
11489         }
11490         /*
11491          * Ack processing.
11492          */
11493         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11494                 return (ret_val);
11495         }
11496         if (ourfinisacked) {
11497                 /*
11498                  * If we can't receive any more data, then closing user can
11499                  * proceed. Starting the timer is contrary to the
11500                  * specification, but if we don't get a FIN we'll hang
11501                  * forever.
11502                  *
11503                  * XXXjl: we should release the tp also, and use a
11504                  * compressed state.
11505                  */
11506                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11507                         soisdisconnected(so);
11508                         tcp_timer_activate(tp, TT_2MSL,
11509                             (tcp_fast_finwait2_recycle ?
11510                             tcp_finwait2_timeout :
11511                             TP_MAXIDLE(tp)));
11512                 }
11513                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
11514         }
11515         if (sbavail(&so->so_snd)) {
11516                 if (ctf_progress_timeout_check(tp, true)) {
11517                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11518                                                 tp, tick, PROGRESS_DROP, __LINE__);
11519                         tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11520                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11521                         return (1);
11522                 }
11523         }
11524         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11525             tiwin, thflags, nxt_pkt));
11526 }
11527
11528 /*
11529  * Return value of 1, the TCB is unlocked and most
11530  * likely gone, return value of 0, the TCP is still
11531  * locked.
11532  */
11533 static int
11534 rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so,
11535     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11536     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11537 {
11538         int32_t ret_val = 0;
11539         int32_t ourfinisacked = 0;
11540         struct tcp_rack *rack;
11541
11542         rack = (struct tcp_rack *)tp->t_fb_ptr;
11543         ctf_calc_rwin(so, tp);
11544
11545         if ((thflags & TH_RST) ||
11546             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11547                 return (ctf_process_rst(m, th, so, tp));
11548         /*
11549          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11550          * synchronized state.
11551          */
11552         if (thflags & TH_SYN) {
11553                 ctf_challenge_ack(m, th, tp, &ret_val);
11554                 return (ret_val);
11555         }
11556         /*
11557          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11558          * it's less than ts_recent, drop it.
11559          */
11560         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11561             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11562                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11563                         return (ret_val);
11564         }
11565         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11566                               &rack->r_ctl.challenge_ack_ts,
11567                               &rack->r_ctl.challenge_ack_cnt)) {
11568                 return (ret_val);
11569         }
11570         /*
11571          * If new data are received on a connection after the user processes
11572          * are gone, then RST the other end.
11573          */
11574         if ((so->so_state & SS_NOFDREF) && tlen) {
11575                 if (rack_check_data_after_close(m, tp, &tlen, th, so))
11576                         return (1);
11577         }
11578         /*
11579          * If last ACK falls within this segment's sequence numbers, record
11580          * its timestamp. NOTE: 1) That the test incorporates suggestions
11581          * from the latest proposal of the tcplw@cray.com list (Braden
11582          * 1993/04/26). 2) That updating only on newer timestamps interferes
11583          * with our earlier PAWS tests, so this check should be solely
11584          * predicated on the sequence space of this segment. 3) That we
11585          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11586          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11587          * SEG.Len, This modified check allows us to overcome RFC1323's
11588          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11589          * p.869. In such cases, we can still calculate the RTT correctly
11590          * when RCV.NXT == Last.ACK.Sent.
11591          */
11592         if ((to->to_flags & TOF_TS) != 0 &&
11593             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11594             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11595             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11596                 tp->ts_recent_age = tcp_ts_getticks();
11597                 tp->ts_recent = to->to_tsval;
11598         }
11599         /*
11600          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11601          * is on (half-synchronized state), then queue data for later
11602          * processing; else drop segment and return.
11603          */
11604         if ((thflags & TH_ACK) == 0) {
11605                 if (tp->t_flags & TF_NEEDSYN) {
11606                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11607                             tiwin, thflags, nxt_pkt));
11608                 } else if (tp->t_flags & TF_ACKNOW) {
11609                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11610                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11611                         return (ret_val);
11612                 } else {
11613                         ctf_do_drop(m, NULL);
11614                         return (0);
11615                 }
11616         }
11617         /*
11618          * Ack processing.
11619          */
11620         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11621                 return (ret_val);
11622         }
11623         if (ourfinisacked) {
11624                 tcp_twstart(tp);
11625                 m_freem(m);
11626                 return (1);
11627         }
11628         if (sbavail(&so->so_snd)) {
11629                 if (ctf_progress_timeout_check(tp, true)) {
11630                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11631                                                 tp, tick, PROGRESS_DROP, __LINE__);
11632                         tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11633                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11634                         return (1);
11635                 }
11636         }
11637         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11638             tiwin, thflags, nxt_pkt));
11639 }
11640
11641 /*
11642  * Return value of 1, the TCB is unlocked and most
11643  * likely gone, return value of 0, the TCP is still
11644  * locked.
11645  */
11646 static int
11647 rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
11648     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11649     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11650 {
11651         int32_t ret_val = 0;
11652         int32_t ourfinisacked = 0;
11653         struct tcp_rack *rack;
11654
11655         rack = (struct tcp_rack *)tp->t_fb_ptr;
11656         ctf_calc_rwin(so, tp);
11657
11658         if ((thflags & TH_RST) ||
11659             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11660                 return (ctf_process_rst(m, th, so, tp));
11661         /*
11662          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11663          * synchronized state.
11664          */
11665         if (thflags & TH_SYN) {
11666                 ctf_challenge_ack(m, th, tp, &ret_val);
11667                 return (ret_val);
11668         }
11669         /*
11670          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11671          * it's less than ts_recent, drop it.
11672          */
11673         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11674             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11675                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11676                         return (ret_val);
11677         }
11678         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11679                               &rack->r_ctl.challenge_ack_ts,
11680                               &rack->r_ctl.challenge_ack_cnt)) {
11681                 return (ret_val);
11682         }
11683         /*
11684          * If new data are received on a connection after the user processes
11685          * are gone, then RST the other end.
11686          */
11687         if ((so->so_state & SS_NOFDREF) && tlen) {
11688                 if (rack_check_data_after_close(m, tp, &tlen, th, so))
11689                         return (1);
11690         }
11691         /*
11692          * If last ACK falls within this segment's sequence numbers, record
11693          * its timestamp. NOTE: 1) That the test incorporates suggestions
11694          * from the latest proposal of the tcplw@cray.com list (Braden
11695          * 1993/04/26). 2) That updating only on newer timestamps interferes
11696          * with our earlier PAWS tests, so this check should be solely
11697          * predicated on the sequence space of this segment. 3) That we
11698          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11699          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11700          * SEG.Len, This modified check allows us to overcome RFC1323's
11701          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11702          * p.869. In such cases, we can still calculate the RTT correctly
11703          * when RCV.NXT == Last.ACK.Sent.
11704          */
11705         if ((to->to_flags & TOF_TS) != 0 &&
11706             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11707             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11708             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11709                 tp->ts_recent_age = tcp_ts_getticks();
11710                 tp->ts_recent = to->to_tsval;
11711         }
11712         /*
11713          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11714          * is on (half-synchronized state), then queue data for later
11715          * processing; else drop segment and return.
11716          */
11717         if ((thflags & TH_ACK) == 0) {
11718                 if (tp->t_flags & TF_NEEDSYN) {
11719                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11720                             tiwin, thflags, nxt_pkt));
11721                 } else if (tp->t_flags & TF_ACKNOW) {
11722                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11723                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11724                         return (ret_val);
11725                 } else {
11726                         ctf_do_drop(m, NULL);
11727                         return (0);
11728                 }
11729         }
11730         /*
11731          * case TCPS_LAST_ACK: Ack processing.
11732          */
11733         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11734                 return (ret_val);
11735         }
11736         if (ourfinisacked) {
11737                 tp = tcp_close(tp);
11738                 ctf_do_drop(m, tp);
11739                 return (1);
11740         }
11741         if (sbavail(&so->so_snd)) {
11742                 if (ctf_progress_timeout_check(tp, true)) {
11743                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11744                                                 tp, tick, PROGRESS_DROP, __LINE__);
11745                         tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11746                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11747                         return (1);
11748                 }
11749         }
11750         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11751             tiwin, thflags, nxt_pkt));
11752 }
11753
11754 /*
11755  * Return value of 1, the TCB is unlocked and most
11756  * likely gone, return value of 0, the TCP is still
11757  * locked.
11758  */
11759 static int
11760 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
11761     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11762     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11763 {
11764         int32_t ret_val = 0;
11765         int32_t ourfinisacked = 0;
11766         struct tcp_rack *rack;
11767
11768         rack = (struct tcp_rack *)tp->t_fb_ptr;
11769         ctf_calc_rwin(so, tp);
11770
11771         /* Reset receive buffer auto scaling when not in bulk receive mode. */
11772         if ((thflags & TH_RST) ||
11773             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11774                 return (ctf_process_rst(m, th, so, tp));
11775         /*
11776          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11777          * synchronized state.
11778          */
11779         if (thflags & TH_SYN) {
11780                 ctf_challenge_ack(m, th, tp, &ret_val);
11781                 return (ret_val);
11782         }
11783         /*
11784          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11785          * it's less than ts_recent, drop it.
11786          */
11787         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11788             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11789                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11790                         return (ret_val);
11791         }
11792         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11793                               &rack->r_ctl.challenge_ack_ts,
11794                               &rack->r_ctl.challenge_ack_cnt)) {
11795                 return (ret_val);
11796         }
11797         /*
11798          * If new data are received on a connection after the user processes
11799          * are gone, then RST the other end.
11800          */
11801         if ((so->so_state & SS_NOFDREF) &&
11802             tlen) {
11803                 if (rack_check_data_after_close(m, tp, &tlen, th, so))
11804                         return (1);
11805         }
11806         /*
11807          * If last ACK falls within this segment's sequence numbers, record
11808          * its timestamp. NOTE: 1) That the test incorporates suggestions
11809          * from the latest proposal of the tcplw@cray.com list (Braden
11810          * 1993/04/26). 2) That updating only on newer timestamps interferes
11811          * with our earlier PAWS tests, so this check should be solely
11812          * predicated on the sequence space of this segment. 3) That we
11813          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11814          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11815          * SEG.Len, This modified check allows us to overcome RFC1323's
11816          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11817          * p.869. In such cases, we can still calculate the RTT correctly
11818          * when RCV.NXT == Last.ACK.Sent.
11819          */
11820         if ((to->to_flags & TOF_TS) != 0 &&
11821             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11822             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11823             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11824                 tp->ts_recent_age = tcp_ts_getticks();
11825                 tp->ts_recent = to->to_tsval;
11826         }
11827         /*
11828          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11829          * is on (half-synchronized state), then queue data for later
11830          * processing; else drop segment and return.
11831          */
11832         if ((thflags & TH_ACK) == 0) {
11833                 if (tp->t_flags & TF_NEEDSYN) {
11834                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11835                             tiwin, thflags, nxt_pkt));
11836                 } else if (tp->t_flags & TF_ACKNOW) {
11837                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11838                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11839                         return (ret_val);
11840                 } else {
11841                         ctf_do_drop(m, NULL);
11842                         return (0);
11843                 }
11844         }
11845         /*
11846          * Ack processing.
11847          */
11848         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11849                 return (ret_val);
11850         }
11851         if (sbavail(&so->so_snd)) {
11852                 if (ctf_progress_timeout_check(tp, true)) {
11853                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11854                                                 tp, tick, PROGRESS_DROP, __LINE__);
11855                         tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11856                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11857                         return (1);
11858                 }
11859         }
11860         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11861             tiwin, thflags, nxt_pkt));
11862 }
11863
11864 static void inline
11865 rack_clear_rate_sample(struct tcp_rack *rack)
11866 {
11867         rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY;
11868         rack->r_ctl.rack_rs.rs_rtt_cnt = 0;
11869         rack->r_ctl.rack_rs.rs_rtt_tot = 0;
11870 }
11871
11872 static void
11873 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override)
11874 {
11875         uint64_t bw_est, rate_wanted;
11876         int chged = 0;
11877         uint32_t user_max, orig_min, orig_max;
11878
11879         orig_min = rack->r_ctl.rc_pace_min_segs;
11880         orig_max = rack->r_ctl.rc_pace_max_segs;
11881         user_max = ctf_fixed_maxseg(tp) * rack->rc_user_set_max_segs;
11882         if (ctf_fixed_maxseg(tp) != rack->r_ctl.rc_pace_min_segs)
11883                 chged = 1;
11884         rack->r_ctl.rc_pace_min_segs = ctf_fixed_maxseg(tp);
11885         if (rack->use_fixed_rate || rack->rc_force_max_seg) {
11886                 if (user_max != rack->r_ctl.rc_pace_max_segs)
11887                         chged = 1;
11888         }
11889         if (rack->rc_force_max_seg) {
11890                 rack->r_ctl.rc_pace_max_segs = user_max;
11891         } else if (rack->use_fixed_rate) {
11892                 bw_est = rack_get_bw(rack);
11893                 if ((rack->r_ctl.crte == NULL) ||
11894                     (bw_est != rack->r_ctl.crte->rate)) {
11895                         rack->r_ctl.rc_pace_max_segs = user_max;
11896                 } else {
11897                         /* We are pacing right at the hardware rate */
11898                         uint32_t segsiz;
11899
11900                         segsiz = min(ctf_fixed_maxseg(tp),
11901                                      rack->r_ctl.rc_pace_min_segs);
11902                         rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(
11903                                                            tp, bw_est, segsiz, 0,
11904                                                            rack->r_ctl.crte, NULL);
11905                 }
11906         } else if (rack->rc_always_pace) {
11907                 if (rack->r_ctl.gp_bw ||
11908 #ifdef NETFLIX_PEAKRATE
11909                     rack->rc_tp->t_maxpeakrate ||
11910 #endif
11911                     rack->r_ctl.init_rate) {
11912                         /* We have a rate of some sort set */
11913                         uint32_t  orig;
11914
11915                         bw_est = rack_get_bw(rack);
11916                         orig = rack->r_ctl.rc_pace_max_segs;
11917                         if (fill_override)
11918                                 rate_wanted = *fill_override;
11919                         else
11920                                 rate_wanted = rack_get_output_bw(rack, bw_est, NULL, NULL);
11921                         if (rate_wanted) {
11922                                 /* We have something */
11923                                 rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack,
11924                                                                                    rate_wanted,
11925                                                                                    ctf_fixed_maxseg(rack->rc_tp));
11926                         } else
11927                                 rack->r_ctl.rc_pace_max_segs = rack->r_ctl.rc_pace_min_segs;
11928                         if (orig != rack->r_ctl.rc_pace_max_segs)
11929                                 chged = 1;
11930                 } else if ((rack->r_ctl.gp_bw == 0) &&
11931                            (rack->r_ctl.rc_pace_max_segs == 0)) {
11932                         /*
11933                          * If we have nothing limit us to bursting
11934                          * out IW sized pieces.
11935                          */
11936                         chged = 1;
11937                         rack->r_ctl.rc_pace_max_segs = rc_init_window(rack);
11938                 }
11939         }
11940         if (rack->r_ctl.rc_pace_max_segs > PACE_MAX_IP_BYTES) {
11941                 chged = 1;
11942                 rack->r_ctl.rc_pace_max_segs = PACE_MAX_IP_BYTES;
11943         }
11944         if (chged)
11945                 rack_log_type_pacing_sizes(tp, rack, orig_min, orig_max, line, 2);
11946 }
11947
11948
11949 static void
11950 rack_init_fsb_block(struct tcpcb *tp, struct tcp_rack *rack)
11951 {
11952 #ifdef INET6
11953         struct ip6_hdr *ip6 = NULL;
11954 #endif
11955 #ifdef INET
11956         struct ip *ip = NULL;
11957 #endif
11958         struct udphdr *udp = NULL;
11959
11960         /* Ok lets fill in the fast block, it can only be used with no IP options! */
11961 #ifdef INET6
11962         if (rack->r_is_v6) {
11963                 rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
11964                 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
11965                 if (tp->t_port) {
11966                         rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
11967                         udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
11968                         udp->uh_sport = htons(V_tcp_udp_tunneling_port);
11969                         udp->uh_dport = tp->t_port;
11970                         rack->r_ctl.fsb.udp = udp;
11971                         rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
11972                 } else
11973                 {
11974                         rack->r_ctl.fsb.th = (struct tcphdr *)(ip6 + 1);
11975                         rack->r_ctl.fsb.udp = NULL;
11976                 }
11977                 tcpip_fillheaders(rack->rc_inp,
11978                                   tp->t_port,
11979                                   ip6, rack->r_ctl.fsb.th);
11980         } else
11981 #endif                          /* INET6 */
11982         {
11983                 rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr);
11984                 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
11985                 if (tp->t_port) {
11986                         rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
11987                         udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
11988                         udp->uh_sport = htons(V_tcp_udp_tunneling_port);
11989                         udp->uh_dport = tp->t_port;
11990                         rack->r_ctl.fsb.udp = udp;
11991                         rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
11992                 } else
11993                 {
11994                         rack->r_ctl.fsb.udp = NULL;
11995                         rack->r_ctl.fsb.th = (struct tcphdr *)(ip + 1);
11996                 }
11997                 tcpip_fillheaders(rack->rc_inp,
11998                                   tp->t_port,
11999                                   ip, rack->r_ctl.fsb.th);
12000         }
12001         rack->r_fsb_inited = 1;
12002 }
12003
12004 static int
12005 rack_init_fsb(struct tcpcb *tp, struct tcp_rack *rack)
12006 {
12007         /*
12008          * Allocate the larger of spaces V6 if available else just
12009          * V4 and include udphdr (overbook)
12010          */
12011 #ifdef INET6
12012         rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + sizeof(struct udphdr);
12013 #else
12014         rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr) + sizeof(struct udphdr);
12015 #endif
12016         rack->r_ctl.fsb.tcp_ip_hdr = malloc(rack->r_ctl.fsb.tcp_ip_hdr_len,
12017                                             M_TCPFSB, M_NOWAIT|M_ZERO);
12018         if (rack->r_ctl.fsb.tcp_ip_hdr == NULL) {
12019                 return (ENOMEM);
12020         }
12021         rack->r_fsb_inited = 0;
12022         return (0);
12023 }
12024
12025 static int
12026 rack_init(struct tcpcb *tp)
12027 {
12028         struct tcp_rack *rack = NULL;
12029         struct rack_sendmap *insret;
12030         uint32_t iwin, snt, us_cts;
12031         int err;
12032
12033         tp->t_fb_ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT);
12034         if (tp->t_fb_ptr == NULL) {
12035                 /*
12036                  * We need to allocate memory but cant. The INP and INP_INFO
12037                  * locks and they are recusive (happens during setup. So a
12038                  * scheme to drop the locks fails :(
12039                  *
12040                  */
12041                 return (ENOMEM);
12042         }
12043         memset(tp->t_fb_ptr, 0, sizeof(struct tcp_rack));
12044
12045         rack = (struct tcp_rack *)tp->t_fb_ptr;
12046         RB_INIT(&rack->r_ctl.rc_mtree);
12047         TAILQ_INIT(&rack->r_ctl.rc_free);
12048         TAILQ_INIT(&rack->r_ctl.rc_tmap);
12049         rack->rc_tp = tp;
12050         rack->rc_inp = tp->t_inpcb;
12051         /* Set the flag */
12052         rack->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
12053         /* Probably not needed but lets be sure */
12054         rack_clear_rate_sample(rack);
12055         /*
12056          * Save off the default values, socket options will poke
12057          * at these if pacing is not on or we have not yet
12058          * reached where pacing is on (gp_ready/fixed enabled).
12059          * When they get set into the CC module (when gp_ready
12060          * is enabled or we enable fixed) then we will set these
12061          * values into the CC and place in here the old values
12062          * so we have a restoral. Then we will set the flag
12063          * rc_pacing_cc_set. That way whenever we turn off pacing
12064          * or switch off this stack, we will know to go restore
12065          * the saved values.
12066          */
12067         rack->r_ctl.rc_saved_beta.beta = V_newreno_beta_ecn;
12068         rack->r_ctl.rc_saved_beta.beta_ecn = V_newreno_beta_ecn;
12069         /* We want abe like behavior as well */
12070         rack->r_ctl.rc_saved_beta.newreno_flags = CC_NEWRENO_BETA_ECN;
12071         rack->r_ctl.rc_reorder_fade = rack_reorder_fade;
12072         rack->rc_allow_data_af_clo = rack_ignore_data_after_close;
12073         rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh;
12074         if (use_rack_rr)
12075                 rack->use_rack_rr = 1;
12076         if (V_tcp_delack_enabled)
12077                 tp->t_delayed_ack = 1;
12078         else
12079                 tp->t_delayed_ack = 0;
12080 #ifdef TCP_ACCOUNTING
12081         if (rack_tcp_accounting) {
12082                 tp->t_flags2 |= TF2_TCP_ACCOUNTING;
12083         }
12084 #endif
12085         if (rack_enable_shared_cwnd)
12086                 rack->rack_enable_scwnd = 1;
12087         rack->rc_user_set_max_segs = rack_hptsi_segments;
12088         rack->rc_force_max_seg = 0;
12089         if (rack_use_imac_dack)
12090                 rack->rc_dack_mode = 1;
12091         TAILQ_INIT(&rack->r_ctl.opt_list);
12092         rack->r_ctl.rc_reorder_shift = rack_reorder_thresh;
12093         rack->r_ctl.rc_pkt_delay = rack_pkt_delay;
12094         rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp;
12095         rack->r_ctl.rc_lowest_us_rtt = 0xffffffff;
12096         rack->r_ctl.rc_highest_us_rtt = 0;
12097         rack->r_ctl.bw_rate_cap = rack_bw_rate_cap;
12098         rack->r_ctl.timer_slop = TICKS_2_USEC(tcp_rexmit_slop);
12099         if (rack_use_cmp_acks)
12100                 rack->r_use_cmp_ack = 1;
12101         if (rack_disable_prr)
12102                 rack->rack_no_prr = 1;
12103         if (rack_gp_no_rec_chg)
12104                 rack->rc_gp_no_rec_chg = 1;
12105         if (rack_pace_every_seg && tcp_can_enable_pacing()) {
12106                 rack->rc_always_pace = 1;
12107                 if (rack->use_fixed_rate || rack->gp_ready)
12108                         rack_set_cc_pacing(rack);
12109         } else
12110                 rack->rc_always_pace = 0;
12111         if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack)
12112                 rack->r_mbuf_queue = 1;
12113         else
12114                 rack->r_mbuf_queue = 0;
12115         if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
12116                 tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
12117         else
12118                 tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
12119         rack_set_pace_segments(tp, rack, __LINE__, NULL);
12120         if (rack_limits_scwnd)
12121                 rack->r_limit_scw = 1;
12122         else
12123                 rack->r_limit_scw = 0;
12124         rack->rc_labc = V_tcp_abc_l_var;
12125         rack->r_ctl.rc_high_rwnd = tp->snd_wnd;
12126         rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
12127         rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method;
12128         rack->rack_tlp_threshold_use = rack_tlp_threshold_use;
12129         rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr;
12130         rack->r_ctl.rc_min_to = rack_min_to;
12131         microuptime(&rack->r_ctl.act_rcv_time);
12132         rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
12133         rack->r_running_late = 0;
12134         rack->r_running_early = 0;
12135         rack->rc_init_win = rack_default_init_window;
12136         rack->r_ctl.rack_per_of_gp_ss = rack_per_of_gp_ss;
12137         if (rack_hw_up_only)
12138                 rack->r_up_only = 1;
12139         if (rack_do_dyn_mul) {
12140                 /* When dynamic adjustment is on CA needs to start at 100% */
12141                 rack->rc_gp_dyn_mul = 1;
12142                 if (rack_do_dyn_mul >= 100)
12143                         rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
12144         } else
12145                 rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
12146         rack->r_ctl.rack_per_of_gp_rec = rack_per_of_gp_rec;
12147         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
12148         rack->r_ctl.rc_tlp_rxt_last_time = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time);
12149         setup_time_filter_small(&rack->r_ctl.rc_gp_min_rtt, FILTER_TYPE_MIN,
12150                                 rack_probertt_filter_life);
12151         us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
12152         rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
12153         rack->r_ctl.rc_time_of_last_probertt = us_cts;
12154         rack->r_ctl.challenge_ack_ts = tcp_ts_getticks();
12155         rack->r_ctl.rc_time_probertt_starts = 0;
12156         /* We require at least one measurement, even if the sysctl is 0 */
12157         if (rack_req_measurements)
12158                 rack->r_ctl.req_measurements = rack_req_measurements;
12159         else
12160                 rack->r_ctl.req_measurements = 1;
12161         if (rack_enable_hw_pacing)
12162                 rack->rack_hdw_pace_ena = 1;
12163         if (rack_hw_rate_caps)
12164                 rack->r_rack_hw_rate_caps = 1;
12165         /* Do we force on detection? */
12166 #ifdef NETFLIX_EXP_DETECTION
12167         if (tcp_force_detection)
12168                 rack->do_detection = 1;
12169         else
12170 #endif
12171                 rack->do_detection = 0;
12172         if (rack_non_rxt_use_cr)
12173                 rack->rack_rec_nonrxt_use_cr = 1;
12174         err = rack_init_fsb(tp, rack);
12175         if (err) {
12176                 uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12177                 tp->t_fb_ptr = NULL;
12178                 return (err);
12179         }
12180         if (tp->snd_una != tp->snd_max) {
12181                 /* Create a send map for the current outstanding data */
12182                 struct rack_sendmap *rsm;
12183
12184                 rsm = rack_alloc(rack);
12185                 if (rsm == NULL) {
12186                         uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12187                         tp->t_fb_ptr = NULL;
12188                         return (ENOMEM);
12189                 }
12190                 rsm->r_no_rtt_allowed = 1;
12191                 rsm->r_tim_lastsent[0] = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
12192                 rsm->r_rtr_cnt = 1;
12193                 rsm->r_rtr_bytes = 0;
12194                 if (tp->t_flags & TF_SENTFIN) {
12195                         rsm->r_end = tp->snd_max - 1;
12196                         rsm->r_flags |= RACK_HAS_FIN;
12197                 } else {
12198                         rsm->r_end = tp->snd_max;
12199                 }
12200                 if (tp->snd_una == tp->iss) {
12201                         /* The data space is one beyond snd_una */
12202                         rsm->r_flags |= RACK_HAS_SYN;
12203                         rsm->r_start = tp->iss;
12204                         rsm->r_end = rsm->r_start + (tp->snd_max - tp->snd_una);
12205                 } else
12206                         rsm->r_start = tp->snd_una;
12207                 rsm->r_dupack = 0;
12208                 if (rack->rc_inp->inp_socket->so_snd.sb_mb != NULL) {
12209                         rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd, 0, &rsm->soff);
12210                         if (rsm->m)
12211                                 rsm->orig_m_len = rsm->m->m_len;
12212                         else
12213                                 rsm->orig_m_len = 0;
12214                 } else {
12215                         /*
12216                          * This can happen if we have a stand-alone FIN or
12217                          *  SYN.
12218                          */
12219                         rsm->m = NULL;
12220                         rsm->orig_m_len = 0;
12221                         rsm->soff = 0;
12222                 }
12223                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12224 #ifdef INVARIANTS
12225                 if (insret != NULL) {
12226                         panic("Insert in rb tree fails ret:%p rack:%p rsm:%p",
12227                               insret, rack, rsm);
12228                 }
12229 #endif
12230                 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
12231                 rsm->r_in_tmap = 1;
12232         }
12233         /*
12234          * Timers in Rack are kept in microseconds so lets
12235          * convert any initial incoming variables
12236          * from ticks into usecs. Note that we
12237          * also change the values of t_srtt and t_rttvar, if
12238          * they are non-zero. They are kept with a 5
12239          * bit decimal so we have to carefully convert
12240          * these to get the full precision.
12241          */
12242         rack_convert_rtts(tp);
12243         tp->t_rttlow = TICKS_2_USEC(tp->t_rttlow);
12244         if (rack_def_profile)
12245                 rack_set_profile(rack, rack_def_profile);
12246         /* Cancel the GP measurement in progress */
12247         tp->t_flags &= ~TF_GPUTINPROG;
12248         if (SEQ_GT(tp->snd_max, tp->iss))
12249                 snt = tp->snd_max - tp->iss;
12250         else
12251                 snt = 0;
12252         iwin = rc_init_window(rack);
12253         if (snt < iwin) {
12254                 /* We are not past the initial window
12255                  * so we need to make sure cwnd is
12256                  * correct.
12257                  */
12258                 if (tp->snd_cwnd < iwin)
12259                         tp->snd_cwnd = iwin;
12260                 /*
12261                  * If we are within the initial window
12262                  * we want ssthresh to be unlimited. Setting
12263                  * it to the rwnd (which the default stack does
12264                  * and older racks) is not really a good idea
12265                  * since we want to be in SS and grow both the
12266                  * cwnd and the rwnd (via dynamic rwnd growth). If
12267                  * we set it to the rwnd then as the peer grows its
12268                  * rwnd we will be stuck in CA and never hit SS.
12269                  *
12270                  * Its far better to raise it up high (this takes the
12271                  * risk that there as been a loss already, probably
12272                  * we should have an indicator in all stacks of loss
12273                  * but we don't), but considering the normal use this
12274                  * is a risk worth taking. The consequences of not
12275                  * hitting SS are far worse than going one more time
12276                  * into it early on (before we have sent even a IW).
12277                  * It is highly unlikely that we will have had a loss
12278                  * before getting the IW out.
12279                  */
12280                 tp->snd_ssthresh = 0xffffffff;
12281         }
12282         rack_stop_all_timers(tp);
12283         /* Lets setup the fsb block */
12284         rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
12285         rack_log_rtt_shrinks(rack,  us_cts,  tp->t_rxtcur,
12286                              __LINE__, RACK_RTTS_INIT);
12287         return (0);
12288 }
12289
12290 static int
12291 rack_handoff_ok(struct tcpcb *tp)
12292 {
12293         if ((tp->t_state == TCPS_CLOSED) ||
12294             (tp->t_state == TCPS_LISTEN)) {
12295                 /* Sure no problem though it may not stick */
12296                 return (0);
12297         }
12298         if ((tp->t_state == TCPS_SYN_SENT) ||
12299             (tp->t_state == TCPS_SYN_RECEIVED)) {
12300                 /*
12301                  * We really don't know if you support sack,
12302                  * you have to get to ESTAB or beyond to tell.
12303                  */
12304                 return (EAGAIN);
12305         }
12306         if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) > 1)) {
12307                 /*
12308                  * Rack will only send a FIN after all data is acknowledged.
12309                  * So in this case we have more data outstanding. We can't
12310                  * switch stacks until either all data and only the FIN
12311                  * is left (in which case rack_init() now knows how
12312                  * to deal with that) <or> all is acknowledged and we
12313                  * are only left with incoming data, though why you
12314                  * would want to switch to rack after all data is acknowledged
12315                  * I have no idea (rrs)!
12316                  */
12317                 return (EAGAIN);
12318         }
12319         if ((tp->t_flags & TF_SACK_PERMIT) || rack_sack_not_required){
12320                 return (0);
12321         }
12322         /*
12323          * If we reach here we don't do SACK on this connection so we can
12324          * never do rack.
12325          */
12326         return (EINVAL);
12327 }
12328
12329
12330 static void
12331 rack_fini(struct tcpcb *tp, int32_t tcb_is_purged)
12332 {
12333         int ack_cmp = 0;
12334
12335         if (tp->t_fb_ptr) {
12336                 struct tcp_rack *rack;
12337                 struct rack_sendmap *rsm, *nrsm, *rm;
12338
12339                 rack = (struct tcp_rack *)tp->t_fb_ptr;
12340                 if (tp->t_in_pkt) {
12341                         /*
12342                          * It is unsafe to process the packets since a
12343                          * reset may be lurking in them (its rare but it
12344                          * can occur). If we were to find a RST, then we
12345                          * would end up dropping the connection and the
12346                          * INP lock, so when we return the caller (tcp_usrreq)
12347                          * will blow up when it trys to unlock the inp.
12348                          */
12349                         struct mbuf *save, *m;
12350
12351                         m = tp->t_in_pkt;
12352                         tp->t_in_pkt = NULL;
12353                         tp->t_tail_pkt = NULL;
12354                         while (m) {
12355                                 save = m->m_nextpkt;
12356                                 m->m_nextpkt = NULL;
12357                                 m_freem(m);
12358                                 m = save;
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         int nsegs;
13456         uint32_t us_cts;
13457         /*
13458          * tv passed from common code is from either M_TSTMP_LRO or
13459          * tcp_get_usecs() if no LRO m_pkthdr timestamp is present.
13460          */
13461         if (m->m_flags & M_ACKCMP) {
13462                 return (rack_do_compressed_ack_processing(tp, so, m, nxt_pkt, tv));
13463         }
13464         if (m->m_flags & M_ACKCMP) {
13465                 panic("Impossible reach m has ackcmp? m:%p tp:%p", m, tp);
13466         }
13467         nsegs = m->m_pkthdr.lro_nsegs;
13468         counter_u64_add(rack_proc_non_comp_ack, 1);
13469         thflags = th->th_flags;
13470 #ifdef TCP_ACCOUNTING
13471         sched_pin();
13472         if (thflags & TH_ACK)
13473                 ts_val = get_cyclecount();
13474 #endif
13475         cts = tcp_tv_to_usectick(tv);
13476         rack = (struct tcp_rack *)tp->t_fb_ptr;
13477
13478         if ((m->m_flags & M_TSTMP) ||
13479             (m->m_flags & M_TSTMP_LRO)) {
13480                 mbuf_tstmp2timespec(m, &ts);
13481                 rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
13482                 rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
13483         } else
13484                 rack->r_ctl.act_rcv_time = *tv;
13485         kern_prefetch(rack, &prev_state);
13486         prev_state = 0;
13487         /*
13488          * Unscale the window into a 32-bit value. For the SYN_SENT state
13489          * the scale is zero.
13490          */
13491         tiwin = th->th_win << tp->snd_scale;
13492         /*
13493          * Parse options on any incoming segment.
13494          */
13495         memset(&to, 0, sizeof(to));
13496         tcp_dooptions(&to, (u_char *)(th + 1),
13497             (th->th_off << 2) - sizeof(struct tcphdr),
13498             (thflags & TH_SYN) ? TO_SYN : 0);
13499 #ifdef TCP_ACCOUNTING
13500         if (thflags & TH_ACK) {
13501                 /*
13502                  * We have a tradeoff here. We can either do what we are
13503                  * doing i.e. pinning to this CPU and then doing the accounting
13504                  * <or> we could do a critical enter, setup the rdtsc and cpu
13505                  * as in below, and then validate we are on the same CPU on
13506                  * exit. I have choosen to not do the critical enter since
13507                  * that often will gain you a context switch, and instead lock
13508                  * us (line above this if) to the same CPU with sched_pin(). This
13509                  * means we may be context switched out for a higher priority
13510                  * interupt but we won't be moved to another CPU.
13511                  *
13512                  * If this occurs (which it won't very often since we most likely
13513                  * are running this code in interupt context and only a higher
13514                  * priority will bump us ... clock?) we will falsely add in
13515                  * to the time the interupt processing time plus the ack processing
13516                  * time. This is ok since its a rare event.
13517                  */
13518                 ack_val_set = tcp_do_ack_accounting(tp, th, &to, tiwin,
13519                                                     ctf_fixed_maxseg(tp));
13520         }
13521 #endif
13522         NET_EPOCH_ASSERT();
13523         INP_WLOCK_ASSERT(tp->t_inpcb);
13524         KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
13525             __func__));
13526         KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
13527             __func__));
13528         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
13529                 union tcp_log_stackspecific log;
13530                 struct timeval ltv;
13531 #ifdef NETFLIX_HTTP_LOGGING
13532                 struct http_sendfile_track *http_req;
13533
13534                 if (SEQ_GT(th->th_ack, tp->snd_una)) {
13535                         http_req = tcp_http_find_req_for_seq(tp, (th->th_ack-1));
13536                 } else {
13537                         http_req = tcp_http_find_req_for_seq(tp, th->th_ack);
13538                 }
13539 #endif
13540                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
13541                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
13542                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
13543                 if (rack->rack_no_prr == 0)
13544                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
13545                 else
13546                         log.u_bbr.flex1 = 0;
13547                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
13548                 log.u_bbr.use_lt_bw <<= 1;
13549                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
13550                 log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
13551                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
13552                 log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
13553                 log.u_bbr.flex3 = m->m_flags;
13554                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
13555                 log.u_bbr.lost = thflags;
13556                 log.u_bbr.pacing_gain = 0x1;
13557 #ifdef TCP_ACCOUNTING
13558                 log.u_bbr.cwnd_gain = ack_val_set;
13559 #endif
13560                 log.u_bbr.flex7 = 2;
13561                 if (m->m_flags & M_TSTMP) {
13562                         /* Record the hardware timestamp if present */
13563                         mbuf_tstmp2timespec(m, &ts);
13564                         ltv.tv_sec = ts.tv_sec;
13565                         ltv.tv_usec = ts.tv_nsec / 1000;
13566                         log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
13567                 } else if (m->m_flags & M_TSTMP_LRO) {
13568                         /* Record the LRO the arrival timestamp */
13569                         mbuf_tstmp2timespec(m, &ts);
13570                         ltv.tv_sec = ts.tv_sec;
13571                         ltv.tv_usec = ts.tv_nsec / 1000;
13572                         log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
13573                 }
13574                 log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
13575                 /* Log the rcv time */
13576                 log.u_bbr.delRate = m->m_pkthdr.rcv_tstmp;
13577 #ifdef NETFLIX_HTTP_LOGGING
13578                 log.u_bbr.applimited = tp->t_http_closed;
13579                 log.u_bbr.applimited <<= 8;
13580                 log.u_bbr.applimited |= tp->t_http_open;
13581                 log.u_bbr.applimited <<= 8;
13582                 log.u_bbr.applimited |= tp->t_http_req;
13583                 if (http_req) {
13584                         /* Copy out any client req info */
13585                         /* seconds */
13586                         log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC);
13587                         /* useconds */
13588                         log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC);
13589                         log.u_bbr.rttProp = http_req->timestamp;
13590                         log.u_bbr.cur_del_rate = http_req->start;
13591                         if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) {
13592                                 log.u_bbr.flex8 |= 1;
13593                         } else {
13594                                 log.u_bbr.flex8 |= 2;
13595                                 log.u_bbr.bw_inuse = http_req->end;
13596                         }
13597                         log.u_bbr.flex6 = http_req->start_seq;
13598                         if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) {
13599                                 log.u_bbr.flex8 |= 4;
13600                                 log.u_bbr.epoch = http_req->end_seq;
13601                         }
13602                 }
13603 #endif
13604                 TCP_LOG_EVENTP(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0,
13605                     tlen, &log, true, &ltv);
13606         }
13607         if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) {
13608                 way_out = 4;
13609                 retval = 0;
13610                 m_freem(m);
13611                 goto done_with_input;
13612         }
13613         /*
13614          * If a segment with the ACK-bit set arrives in the SYN-SENT state
13615          * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9.
13616          */
13617         if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) &&
13618             (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) {
13619                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13620                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13621 #ifdef TCP_ACCOUNTING
13622                 sched_unpin();
13623 #endif
13624                 return (1);
13625         }
13626
13627         /*
13628          * Parse options on any incoming segment.
13629          */
13630         tcp_dooptions(&to, (u_char *)(th + 1),
13631             (th->th_off << 2) - sizeof(struct tcphdr),
13632             (thflags & TH_SYN) ? TO_SYN : 0);
13633
13634         /*
13635          * If timestamps were negotiated during SYN/ACK and a
13636          * segment without a timestamp is received, silently drop
13637          * the segment, unless it is a RST segment or missing timestamps are
13638          * tolerated.
13639          * See section 3.2 of RFC 7323.
13640          */
13641         if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) &&
13642             ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) {
13643                 way_out = 5;
13644                 retval = 0;
13645                 m_freem(m);
13646                 goto done_with_input;
13647         }
13648
13649         /*
13650          * Segment received on connection. Reset idle time and keep-alive
13651          * timer. XXX: This should be done after segment validation to
13652          * ignore broken/spoofed segs.
13653          */
13654         if  (tp->t_idle_reduce &&
13655              (tp->snd_max == tp->snd_una) &&
13656              ((ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
13657                 counter_u64_add(rack_input_idle_reduces, 1);
13658                 rack_cc_after_idle(rack, tp);
13659         }
13660         tp->t_rcvtime = ticks;
13661 #ifdef STATS
13662         stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
13663 #endif
13664         if (tiwin > rack->r_ctl.rc_high_rwnd)
13665                 rack->r_ctl.rc_high_rwnd = tiwin;
13666         /*
13667          * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move
13668          * this to occur after we've validated the segment.
13669          */
13670         if (tp->t_flags2 & TF2_ECN_PERMIT) {
13671                 if (thflags & TH_CWR) {
13672                         tp->t_flags2 &= ~TF2_ECN_SND_ECE;
13673                         tp->t_flags |= TF_ACKNOW;
13674                 }
13675                 switch (iptos & IPTOS_ECN_MASK) {
13676                 case IPTOS_ECN_CE:
13677                         tp->t_flags2 |= TF2_ECN_SND_ECE;
13678                         KMOD_TCPSTAT_INC(tcps_ecn_ce);
13679                         break;
13680                 case IPTOS_ECN_ECT0:
13681                         KMOD_TCPSTAT_INC(tcps_ecn_ect0);
13682                         break;
13683                 case IPTOS_ECN_ECT1:
13684                         KMOD_TCPSTAT_INC(tcps_ecn_ect1);
13685                         break;
13686                 }
13687
13688                 /* Process a packet differently from RFC3168. */
13689                 cc_ecnpkt_handler(tp, th, iptos);
13690
13691                 /* Congestion experienced. */
13692                 if (thflags & TH_ECE) {
13693                         rack_cong_signal(tp, CC_ECN, th->th_ack);
13694                 }
13695         }
13696
13697         /*
13698          * If echoed timestamp is later than the current time, fall back to
13699          * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
13700          * were used when this connection was established.
13701          */
13702         if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
13703                 to.to_tsecr -= tp->ts_offset;
13704                 if (TSTMP_GT(to.to_tsecr, cts))
13705                         to.to_tsecr = 0;
13706         }
13707
13708         /*
13709          * If its the first time in we need to take care of options and
13710          * verify we can do SACK for rack!
13711          */
13712         if (rack->r_state == 0) {
13713                 /* Should be init'd by rack_init() */
13714                 KASSERT(rack->rc_inp != NULL,
13715                     ("%s: rack->rc_inp unexpectedly NULL", __func__));
13716                 if (rack->rc_inp == NULL) {
13717                         rack->rc_inp = tp->t_inpcb;
13718                 }
13719
13720                 /*
13721                  * Process options only when we get SYN/ACK back. The SYN
13722                  * case for incoming connections is handled in tcp_syncache.
13723                  * According to RFC1323 the window field in a SYN (i.e., a
13724                  * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
13725                  * this is traditional behavior, may need to be cleaned up.
13726                  */
13727                 if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
13728                         /* Handle parallel SYN for ECN */
13729                         if (!(thflags & TH_ACK) &&
13730                             ((thflags & (TH_CWR | TH_ECE)) == (TH_CWR | TH_ECE)) &&
13731                             ((V_tcp_do_ecn == 1) || (V_tcp_do_ecn == 2))) {
13732                                 tp->t_flags2 |= TF2_ECN_PERMIT;
13733                                 tp->t_flags2 |= TF2_ECN_SND_ECE;
13734                                 TCPSTAT_INC(tcps_ecn_shs);
13735                         }
13736                         if ((to.to_flags & TOF_SCALE) &&
13737                             (tp->t_flags & TF_REQ_SCALE)) {
13738                                 tp->t_flags |= TF_RCVD_SCALE;
13739                                 tp->snd_scale = to.to_wscale;
13740                         } else
13741                                 tp->t_flags &= ~TF_REQ_SCALE;
13742                         /*
13743                          * Initial send window.  It will be updated with the
13744                          * next incoming segment to the scaled value.
13745                          */
13746                         tp->snd_wnd = th->th_win;
13747                         rack_validate_fo_sendwin_up(tp, rack);
13748                         if ((to.to_flags & TOF_TS) &&
13749                             (tp->t_flags & TF_REQ_TSTMP)) {
13750                                 tp->t_flags |= TF_RCVD_TSTMP;
13751                                 tp->ts_recent = to.to_tsval;
13752                                 tp->ts_recent_age = cts;
13753                         } else
13754                                 tp->t_flags &= ~TF_REQ_TSTMP;
13755                         if (to.to_flags & TOF_MSS) {
13756                                 tcp_mss(tp, to.to_mss);
13757                         }
13758                         if ((tp->t_flags & TF_SACK_PERMIT) &&
13759                             (to.to_flags & TOF_SACKPERM) == 0)
13760                                 tp->t_flags &= ~TF_SACK_PERMIT;
13761                         if (IS_FASTOPEN(tp->t_flags)) {
13762                                 if (to.to_flags & TOF_FASTOPEN) {
13763                                         uint16_t mss;
13764
13765                                         if (to.to_flags & TOF_MSS)
13766                                                 mss = to.to_mss;
13767                                         else
13768                                                 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0)
13769                                                         mss = TCP6_MSS;
13770                                                 else
13771                                                         mss = TCP_MSS;
13772                                         tcp_fastopen_update_cache(tp, mss,
13773                                             to.to_tfo_len, to.to_tfo_cookie);
13774                                 } else
13775                                         tcp_fastopen_disable_path(tp);
13776                         }
13777                 }
13778                 /*
13779                  * At this point we are at the initial call. Here we decide
13780                  * if we are doing RACK or not. We do this by seeing if
13781                  * TF_SACK_PERMIT is set and the sack-not-required is clear.
13782                  * The code now does do dup-ack counting so if you don't
13783                  * switch back you won't get rack & TLP, but you will still
13784                  * get this stack.
13785                  */
13786
13787                 if ((rack_sack_not_required == 0) &&
13788                     ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
13789                         tcp_switch_back_to_default(tp);
13790                         (*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen,
13791                             tlen, iptos);
13792 #ifdef TCP_ACCOUNTING
13793                         sched_unpin();
13794 #endif
13795                         return (1);
13796                 }
13797                 tcp_set_hpts(tp->t_inpcb);
13798                 sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack);
13799         }
13800         if (thflags & TH_FIN)
13801                 tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN);
13802         us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
13803         if ((rack->rc_gp_dyn_mul) &&
13804             (rack->use_fixed_rate == 0) &&
13805             (rack->rc_always_pace)) {
13806                 /* Check in on probertt */
13807                 rack_check_probe_rtt(rack, us_cts);
13808         }
13809         if (rack->forced_ack) {
13810                 uint32_t us_rtt;
13811
13812                 /*
13813                  * A persist or keep-alive was forced out, update our
13814                  * min rtt time. Note we do not worry about lost
13815                  * retransmissions since KEEP-ALIVES and persists
13816                  * are usually way long on times of sending (though
13817                  * if we were really paranoid or worried we could
13818                  * at least use timestamps if available to validate).
13819                  */
13820                 rack->forced_ack = 0;
13821                 us_rtt = us_cts - rack->r_ctl.forced_ack_ts;
13822                 if (us_rtt == 0)
13823                         us_rtt = 1;
13824                 rack_log_rtt_upd(tp, rack, us_rtt, 0, NULL, 3);
13825                 rack_apply_updated_usrtt(rack, us_rtt, us_cts);
13826         }
13827         /*
13828          * This is the one exception case where we set the rack state
13829          * always. All other times (timers etc) we must have a rack-state
13830          * set (so we assure we have done the checks above for SACK).
13831          */
13832         rack->r_ctl.rc_rcvtime = cts;
13833         if (rack->r_state != tp->t_state)
13834                 rack_set_state(tp, rack);
13835         if (SEQ_GT(th->th_ack, tp->snd_una) &&
13836             (rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree)) != NULL)
13837                 kern_prefetch(rsm, &prev_state);
13838         prev_state = rack->r_state;
13839         rack_clear_rate_sample(rack);
13840         retval = (*rack->r_substate) (m, th, so,
13841             tp, &to, drop_hdrlen,
13842             tlen, tiwin, thflags, nxt_pkt, iptos);
13843 #ifdef INVARIANTS
13844         if ((retval == 0) &&
13845             (tp->t_inpcb == NULL)) {
13846                 panic("retval:%d tp:%p t_inpcb:NULL state:%d",
13847                     retval, tp, prev_state);
13848         }
13849 #endif
13850         if (retval == 0) {
13851                 /*
13852                  * If retval is 1 the tcb is unlocked and most likely the tp
13853                  * is gone.
13854                  */
13855                 INP_WLOCK_ASSERT(tp->t_inpcb);
13856                 if ((rack->rc_gp_dyn_mul) &&
13857                     (rack->rc_always_pace) &&
13858                     (rack->use_fixed_rate == 0) &&
13859                     rack->in_probe_rtt &&
13860                     (rack->r_ctl.rc_time_probertt_starts == 0)) {
13861                         /*
13862                          * If we are going for target, lets recheck before
13863                          * we output.
13864                          */
13865                         rack_check_probe_rtt(rack, us_cts);
13866                 }
13867                 if (rack->set_pacing_done_a_iw == 0) {
13868                         /* How much has been acked? */
13869                         if ((tp->snd_una - tp->iss) > (ctf_fixed_maxseg(tp) * 10)) {
13870                                 /* We have enough to set in the pacing segment size */
13871                                 rack->set_pacing_done_a_iw = 1;
13872                                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
13873                         }
13874                 }
13875                 tcp_rack_xmit_timer_commit(rack, tp);
13876 #ifdef TCP_ACCOUNTING
13877                 /*
13878                  * If we set the ack_val_se to what ack processing we are doing
13879                  * we also want to track how many cycles we burned. Note
13880                  * the bits after tcp_output we let be "free". This is because
13881                  * we are also tracking the tcp_output times as well. Note the
13882                  * use of 0xf here since we only have 11 counter (0 - 0xa) and
13883                  * 0xf cannot be returned and is what we initialize it too to
13884                  * indicate we are not doing the tabulations.
13885                  */
13886                 if (ack_val_set != 0xf) {
13887                         uint64_t crtsc;
13888
13889                         crtsc = get_cyclecount();
13890                         counter_u64_add(tcp_proc_time[ack_val_set] , (crtsc - ts_val));
13891                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13892                                 tp->tcp_proc_time[ack_val_set] += (crtsc - ts_val);
13893                         }
13894                 }
13895 #endif
13896                 if (nxt_pkt == 0) {
13897                         if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
13898 do_output_now:
13899                                 did_out = 1;
13900                                 (void)tp->t_fb->tfb_tcp_output(tp);
13901                         }
13902                         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
13903                         rack_free_trim(rack);
13904                 }
13905                 if ((nxt_pkt == 0) &&
13906                     ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
13907                     (SEQ_GT(tp->snd_max, tp->snd_una) ||
13908                      (tp->t_flags & TF_DELACK) ||
13909                      ((V_tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
13910                       (tp->t_state <= TCPS_CLOSING)))) {
13911                         /* We could not send (probably in the hpts but stopped the timer earlier)? */
13912                         if ((tp->snd_max == tp->snd_una) &&
13913                             ((tp->t_flags & TF_DELACK) == 0) &&
13914                             (rack->rc_inp->inp_in_hpts) &&
13915                             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
13916                                 /* keep alive not needed if we are hptsi output yet */
13917                                 ;
13918                         } else {
13919                                 int late = 0;
13920                                 if (rack->rc_inp->inp_in_hpts) {
13921                                         if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
13922                                                 us_cts = tcp_get_usecs(NULL);
13923                                                 if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
13924                                                         rack->r_early = 1;
13925                                                         rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
13926                                                 } else
13927                                                         late = 1;
13928                                                 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
13929                                         }
13930                                         tcp_hpts_remove(tp->t_inpcb, HPTS_REMOVE_OUTPUT);
13931                                 }
13932                                 if (late && (did_out == 0)) {
13933                                         /*
13934                                          * We are late in the sending
13935                                          * and we did not call the output
13936                                          * (this probably should not happen).
13937                                          */
13938                                         goto do_output_now;
13939                                 }
13940                                 rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
13941                         }
13942                         way_out = 1;
13943                 } else if (nxt_pkt == 0) {
13944                         /* Do we have the correct timer running? */
13945                         rack_timer_audit(tp, rack, &so->so_snd);
13946                         way_out = 2;
13947                 }
13948         done_with_input:
13949                 rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out, max(1, nsegs));
13950                 if (did_out)
13951                         rack->r_wanted_output = 0;
13952 #ifdef INVARIANTS
13953                 if (tp->t_inpcb == NULL) {
13954                         panic("OP:%d retval:%d tp:%p t_inpcb:NULL state:%d",
13955                               did_out,
13956                               retval, tp, prev_state);
13957                 }
13958 #endif
13959 #ifdef TCP_ACCOUNTING
13960         } else {
13961                 /*
13962                  * Track the time (see above).
13963                  */
13964                 if (ack_val_set != 0xf) {
13965                         uint64_t crtsc;
13966
13967                         crtsc = get_cyclecount();
13968                         counter_u64_add(tcp_proc_time[ack_val_set] , (crtsc - ts_val));
13969                         /*
13970                          * Note we *DO NOT* increment the per-tcb counters since
13971                          * in the else the TP may be gone!!
13972                          */
13973                 }
13974 #endif
13975         }
13976 #ifdef TCP_ACCOUNTING
13977         sched_unpin();
13978 #endif
13979         return (retval);
13980 }
13981
13982 void
13983 rack_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
13984     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
13985 {
13986         struct timeval tv;
13987
13988         /* First lets see if we have old packets */
13989         if (tp->t_in_pkt) {
13990                 if (ctf_do_queued_segments(so, tp, 1)) {
13991                         m_freem(m);
13992                         return;
13993                 }
13994         }
13995         if (m->m_flags & M_TSTMP_LRO) {
13996                 tv.tv_sec = m->m_pkthdr.rcv_tstmp /1000000000;
13997                 tv.tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000)/1000;
13998         } else {
13999                 /* Should not be should we kassert instead? */
14000                 tcp_get_usecs(&tv);
14001         }
14002         if (rack_do_segment_nounlock(m, th, so, tp,
14003                                      drop_hdrlen, tlen, iptos, 0, &tv) == 0) {
14004                 INP_WUNLOCK(tp->t_inpcb);
14005         }
14006 }
14007
14008 struct rack_sendmap *
14009 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused)
14010 {
14011         struct rack_sendmap *rsm = NULL;
14012         int32_t idx;
14013         uint32_t srtt = 0, thresh = 0, ts_low = 0;
14014
14015         /* Return the next guy to be re-transmitted */
14016         if (RB_EMPTY(&rack->r_ctl.rc_mtree)) {
14017                 return (NULL);
14018         }
14019         if (tp->t_flags & TF_SENTFIN) {
14020                 /* retran the end FIN? */
14021                 return (NULL);
14022         }
14023         /* ok lets look at this one */
14024         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
14025         if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) {
14026                 goto check_it;
14027         }
14028         rsm = rack_find_lowest_rsm(rack);
14029         if (rsm == NULL) {
14030                 return (NULL);
14031         }
14032 check_it:
14033         if (((rack->rc_tp->t_flags & TF_SACK_PERMIT) == 0) &&
14034             (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
14035                 /*
14036                  * No sack so we automatically do the 3 strikes and
14037                  * retransmit (no rack timer would be started).
14038                  */
14039
14040                 return (rsm);
14041         }
14042         if (rsm->r_flags & RACK_ACKED) {
14043                 return (NULL);
14044         }
14045         if (((rsm->r_flags & RACK_SACK_PASSED) == 0) &&
14046             (rsm->r_dupack < DUP_ACK_THRESHOLD)) {
14047                 /* Its not yet ready */
14048                 return (NULL);
14049         }
14050         srtt = rack_grab_rtt(tp, rack);
14051         idx = rsm->r_rtr_cnt - 1;
14052         ts_low = (uint32_t)rsm->r_tim_lastsent[idx];
14053         thresh = rack_calc_thresh_rack(rack, srtt, tsused);
14054         if ((tsused == ts_low) ||
14055             (TSTMP_LT(tsused, ts_low))) {
14056                 /* No time since sending */
14057                 return (NULL);
14058         }
14059         if ((tsused - ts_low) < thresh) {
14060                 /* It has not been long enough yet */
14061                 return (NULL);
14062         }
14063         if ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
14064             ((rsm->r_flags & RACK_SACK_PASSED) &&
14065              (rack->sack_attack_disable == 0))) {
14066                 /*
14067                  * We have passed the dup-ack threshold <or>
14068                  * a SACK has indicated this is missing.
14069                  * Note that if you are a declared attacker
14070                  * it is only the dup-ack threshold that
14071                  * will cause retransmits.
14072                  */
14073                 /* log retransmit reason */
14074                 rack_log_retran_reason(rack, rsm, (tsused - ts_low), thresh, 1);
14075                 rack->r_fast_output = 0;
14076                 return (rsm);
14077         }
14078         return (NULL);
14079 }
14080
14081 static void
14082 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
14083                            uint64_t bw_est, uint64_t bw, uint64_t len_time, int method,
14084                            int line, struct rack_sendmap *rsm)
14085 {
14086         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
14087                 union tcp_log_stackspecific log;
14088                 struct timeval tv;
14089
14090                 memset(&log, 0, sizeof(log));
14091                 log.u_bbr.flex1 = slot;
14092                 log.u_bbr.flex2 = len;
14093                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_min_segs;
14094                 log.u_bbr.flex4 = rack->r_ctl.rc_pace_max_segs;
14095                 log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ss;
14096                 log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_ca;
14097                 log.u_bbr.use_lt_bw = rack->rc_ack_can_sendout_data;
14098                 log.u_bbr.use_lt_bw <<= 1;
14099                 log.u_bbr.use_lt_bw |= rack->r_late;
14100                 log.u_bbr.use_lt_bw <<= 1;
14101                 log.u_bbr.use_lt_bw |= rack->r_early;
14102                 log.u_bbr.use_lt_bw <<= 1;
14103                 log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
14104                 log.u_bbr.use_lt_bw <<= 1;
14105                 log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
14106                 log.u_bbr.use_lt_bw <<= 1;
14107                 log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
14108                 log.u_bbr.use_lt_bw <<= 1;
14109                 log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
14110                 log.u_bbr.use_lt_bw <<= 1;
14111                 log.u_bbr.use_lt_bw |= rack->gp_ready;
14112                 log.u_bbr.pkt_epoch = line;
14113                 log.u_bbr.epoch = rack->r_ctl.rc_agg_delayed;
14114                 log.u_bbr.lt_epoch = rack->r_ctl.rc_agg_early;
14115                 log.u_bbr.applimited = rack->r_ctl.rack_per_of_gp_rec;
14116                 log.u_bbr.bw_inuse = bw_est;
14117                 log.u_bbr.delRate = bw;
14118                 if (rack->r_ctl.gp_bw == 0)
14119                         log.u_bbr.cur_del_rate = 0;
14120                 else
14121                         log.u_bbr.cur_del_rate = rack_get_bw(rack);
14122                 log.u_bbr.rttProp = len_time;
14123                 log.u_bbr.pkts_out = rack->r_ctl.rc_rack_min_rtt;
14124                 log.u_bbr.lost = rack->r_ctl.rc_probertt_sndmax_atexit;
14125                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
14126                 if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) {
14127                         /* We are in slow start */
14128                         log.u_bbr.flex7 = 1;
14129                 } else {
14130                         /* we are on congestion avoidance */
14131                         log.u_bbr.flex7 = 0;
14132                 }
14133                 log.u_bbr.flex8 = method;
14134                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14135                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14136                 log.u_bbr.cwnd_gain = rack->rc_gp_saw_rec;
14137                 log.u_bbr.cwnd_gain <<= 1;
14138                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
14139                 log.u_bbr.cwnd_gain <<= 1;
14140                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
14141                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
14142                     &rack->rc_inp->inp_socket->so_rcv,
14143                     &rack->rc_inp->inp_socket->so_snd,
14144                     BBR_LOG_HPTSI_CALC, 0,
14145                     0, &log, false, &tv);
14146         }
14147 }
14148
14149 static uint32_t
14150 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss)
14151 {
14152         uint32_t new_tso, user_max;
14153
14154         user_max = rack->rc_user_set_max_segs * mss;
14155         if (rack->rc_force_max_seg) {
14156                 return (user_max);
14157         }
14158         if (rack->use_fixed_rate &&
14159             ((rack->r_ctl.crte == NULL) ||
14160              (bw != rack->r_ctl.crte->rate))) {
14161                 /* Use the user mss since we are not exactly matched */
14162                 return (user_max);
14163         }
14164         new_tso = tcp_get_pacing_burst_size(rack->rc_tp, bw, mss, rack_pace_one_seg, rack->r_ctl.crte, NULL);
14165         if (new_tso > user_max)
14166                 new_tso = user_max;
14167         return (new_tso);
14168 }
14169
14170 static int32_t
14171 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)
14172 {
14173         uint64_t lentim, fill_bw;
14174
14175         /* Lets first see if we are full, if so continue with normal rate */
14176         rack->r_via_fill_cw = 0;
14177         if (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.cwnd_to_use)
14178                 return (slot);
14179         if ((ctf_outstanding(rack->rc_tp) + (segsiz-1)) > rack->rc_tp->snd_wnd)
14180                 return (slot);
14181         if (rack->r_ctl.rc_last_us_rtt == 0)
14182                 return (slot);
14183         if (rack->rc_pace_fill_if_rttin_range &&
14184             (rack->r_ctl.rc_last_us_rtt >=
14185              (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack->rtt_limit_mul))) {
14186                 /* The rtt is huge, N * smallest, lets not fill */
14187                 return (slot);
14188         }
14189         /*
14190          * first lets calculate the b/w based on the last us-rtt
14191          * and the sndwnd.
14192          */
14193         fill_bw = rack->r_ctl.cwnd_to_use;
14194         /* Take the rwnd if its smaller */
14195         if (fill_bw > rack->rc_tp->snd_wnd)
14196                 fill_bw = rack->rc_tp->snd_wnd;
14197         if (rack->r_fill_less_agg) {
14198                 /*
14199                  * Now take away the inflight (this will reduce our
14200                  * aggressiveness and yeah, if we get that much out in 1RTT
14201                  * we will have had acks come back and still be behind).
14202                  */
14203                 fill_bw -= ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14204         }
14205         /* Now lets make it into a b/w */
14206         fill_bw *= (uint64_t)HPTS_USEC_IN_SEC;
14207         fill_bw /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
14208         /* We are below the min b/w */
14209         if (non_paced)
14210                 *rate_wanted = fill_bw;
14211         if ((fill_bw < RACK_MIN_BW) || (fill_bw < *rate_wanted))
14212                 return (slot);
14213         if (rack->r_ctl.bw_rate_cap && (fill_bw > rack->r_ctl.bw_rate_cap))
14214                 fill_bw = rack->r_ctl.bw_rate_cap;
14215         rack->r_via_fill_cw = 1;
14216         if (rack->r_rack_hw_rate_caps &&
14217             (rack->r_ctl.crte != NULL)) {
14218                 uint64_t high_rate;
14219
14220                 high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
14221                 if (fill_bw > high_rate) {
14222                         /* We are capping bw at the highest rate table entry */
14223                         if (*rate_wanted > high_rate) {
14224                                 /* The original rate was also capped */
14225                                 rack->r_via_fill_cw = 0;
14226                         }
14227                         rack_log_hdwr_pacing(rack,
14228                                              fill_bw, high_rate, __LINE__,
14229                                              0, 3);
14230                         fill_bw = high_rate;
14231                         if (capped)
14232                                 *capped = 1;
14233                 }
14234         } else if ((rack->r_ctl.crte == NULL) &&
14235                    (rack->rack_hdrw_pacing == 0) &&
14236                    (rack->rack_hdw_pace_ena) &&
14237                    rack->r_rack_hw_rate_caps &&
14238                    (rack->rack_attempt_hdwr_pace == 0) &&
14239                    (rack->rc_inp->inp_route.ro_nh != NULL) &&
14240                    (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
14241                 /*
14242                  * Ok we may have a first attempt that is greater than our top rate
14243                  * lets check.
14244                  */
14245                 uint64_t high_rate;
14246
14247                 high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
14248                 if (high_rate) {
14249                         if (fill_bw > high_rate) {
14250                                 fill_bw = high_rate;
14251                                 if (capped)
14252                                         *capped = 1;
14253                         }
14254                 }
14255         }
14256         /*
14257          * Ok fill_bw holds our mythical b/w to fill the cwnd
14258          * in a rtt, what does that time wise equate too?
14259          */
14260         lentim = (uint64_t)(len) * (uint64_t)HPTS_USEC_IN_SEC;
14261         lentim /= fill_bw;
14262         *rate_wanted = fill_bw;
14263         if (non_paced || (lentim < slot)) {
14264                 rack_log_pacing_delay_calc(rack, len, slot, fill_bw,
14265                                            0, lentim, 12, __LINE__, NULL);
14266                 return ((int32_t)lentim);
14267         } else
14268                 return (slot);
14269 }
14270
14271 static int32_t
14272 rack_get_pacing_delay(struct tcp_rack *rack, struct tcpcb *tp, uint32_t len, struct rack_sendmap *rsm, uint32_t segsiz)
14273 {
14274         struct rack_sendmap *lrsm;
14275         int32_t slot = 0;
14276         int can_start_hw_pacing = 1;
14277         int err;
14278
14279         if (rack->rc_always_pace == 0) {
14280                 /*
14281                  * We use the most optimistic possible cwnd/srtt for
14282                  * sending calculations. This will make our
14283                  * calculation anticipate getting more through
14284                  * quicker then possible. But thats ok we don't want
14285                  * the peer to have a gap in data sending.
14286                  */
14287                 uint32_t srtt, cwnd, tr_perms = 0;
14288                 int32_t reduce = 0;
14289
14290         old_method:
14291                 /*
14292                  * We keep no precise pacing with the old method
14293                  * instead we use the pacer to mitigate bursts.
14294                  */
14295                 if (rack->r_ctl.rc_rack_min_rtt)
14296                         srtt = rack->r_ctl.rc_rack_min_rtt;
14297                 else
14298                         srtt = max(tp->t_srtt, 1);
14299                 if (rack->r_ctl.rc_rack_largest_cwnd)
14300                         cwnd = rack->r_ctl.rc_rack_largest_cwnd;
14301                 else
14302                         cwnd = rack->r_ctl.cwnd_to_use;
14303                 /* Inflate cwnd by 1000 so srtt of usecs is in ms */
14304                 tr_perms = (cwnd * 1000) / srtt;
14305                 if (tr_perms == 0) {
14306                         tr_perms = ctf_fixed_maxseg(tp);
14307                 }
14308                 /*
14309                  * Calculate how long this will take to drain, if
14310                  * the calculation comes out to zero, thats ok we
14311                  * will use send_a_lot to possibly spin around for
14312                  * more increasing tot_len_this_send to the point
14313                  * that its going to require a pace, or we hit the
14314                  * cwnd. Which in that case we are just waiting for
14315                  * a ACK.
14316                  */
14317                 slot = len / tr_perms;
14318                 /* Now do we reduce the time so we don't run dry? */
14319                 if (slot && rack_slot_reduction) {
14320                         reduce = (slot / rack_slot_reduction);
14321                         if (reduce < slot) {
14322                                 slot -= reduce;
14323                         } else
14324                                 slot = 0;
14325                 }
14326                 slot *= HPTS_USEC_IN_MSEC;
14327                 if (rsm == NULL) {
14328                         /*
14329                          * We always consider ourselves app limited with old style
14330                          * that are not retransmits. This could be the initial
14331                          * measurement, but thats ok its all setup and specially
14332                          * handled. If another send leaks out, then that too will
14333                          * be mark app-limited.
14334                          */
14335                         lrsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
14336                         if (lrsm && ((lrsm->r_flags & RACK_APP_LIMITED) == 0)) {
14337                                 rack->r_ctl.rc_first_appl = lrsm;
14338                                 lrsm->r_flags |= RACK_APP_LIMITED;
14339                                 rack->r_ctl.rc_app_limited_cnt++;
14340                         }
14341                 }
14342                 if (rack->rc_pace_to_cwnd) {
14343                         uint64_t rate_wanted = 0;
14344
14345                         slot = pace_to_fill_cwnd(rack, slot, len, segsiz, NULL, &rate_wanted, 1);
14346                         rack->rc_ack_can_sendout_data = 1;
14347                         rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, 0, 0, 14, __LINE__, NULL);
14348                 } else
14349                         rack_log_pacing_delay_calc(rack, len, slot, tr_perms, reduce, 0, 7, __LINE__, NULL);
14350         } else {
14351                 uint64_t bw_est, res, lentim, rate_wanted;
14352                 uint32_t orig_val, srtt, segs, oh;
14353                 int capped = 0;
14354                 int prev_fill;
14355
14356                 if ((rack->r_rr_config == 1) && rsm) {
14357                         return (rack->r_ctl.rc_min_to);
14358                 }
14359                 if (rack->use_fixed_rate) {
14360                         rate_wanted = bw_est = rack_get_fixed_pacing_bw(rack);
14361                 } else if ((rack->r_ctl.init_rate == 0) &&
14362 #ifdef NETFLIX_PEAKRATE
14363                            (rack->rc_tp->t_maxpeakrate == 0) &&
14364 #endif
14365                            (rack->r_ctl.gp_bw == 0)) {
14366                         /* no way to yet do an estimate */
14367                         bw_est = rate_wanted = 0;
14368                 } else {
14369                         bw_est = rack_get_bw(rack);
14370                         rate_wanted = rack_get_output_bw(rack, bw_est, rsm, &capped);
14371                 }
14372                 if ((bw_est == 0) || (rate_wanted == 0) ||
14373                     ((rack->gp_ready == 0) && (rack->use_fixed_rate == 0))) {
14374                         /*
14375                          * No way yet to make a b/w estimate or
14376                          * our raise is set incorrectly.
14377                          */
14378                         goto old_method;
14379                 }
14380                 /* We need to account for all the overheads */
14381                 segs = (len + segsiz - 1) / segsiz;
14382                 /*
14383                  * We need the diff between 1514 bytes (e-mtu with e-hdr)
14384                  * and how much data we put in each packet. Yes this
14385                  * means we may be off if we are larger than 1500 bytes
14386                  * or smaller. But this just makes us more conservative.
14387                  */
14388                 if (rack_hw_rate_min &&
14389                     (bw_est < rack_hw_rate_min))
14390                         can_start_hw_pacing = 0;
14391                 if (ETHERNET_SEGMENT_SIZE > segsiz)
14392                         oh = ETHERNET_SEGMENT_SIZE - segsiz;
14393                 else
14394                         oh = 0;
14395                 segs *= oh;
14396                 lentim = (uint64_t)(len + segs) * (uint64_t)HPTS_USEC_IN_SEC;
14397                 res = lentim / rate_wanted;
14398                 slot = (uint32_t)res;
14399                 orig_val = rack->r_ctl.rc_pace_max_segs;
14400                 if (rack->r_ctl.crte == NULL) {
14401                         /*
14402                          * Only do this if we are not hardware pacing
14403                          * since if we are doing hw-pacing below we will
14404                          * set make a call after setting up or changing
14405                          * the rate.
14406                          */
14407                         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
14408                 } else if (rack->rc_inp->inp_snd_tag == NULL) {
14409                         /*
14410                          * We lost our rate somehow, this can happen
14411                          * if the interface changed underneath us.
14412                          */
14413                         tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
14414                         rack->r_ctl.crte = NULL;
14415                         /* Lets re-allow attempting to setup pacing */
14416                         rack->rack_hdrw_pacing = 0;
14417                         rack->rack_attempt_hdwr_pace = 0;
14418                         rack_log_hdwr_pacing(rack,
14419                                              rate_wanted, bw_est, __LINE__,
14420                                              0, 6);
14421                 }
14422                 /* Did we change the TSO size, if so log it */
14423                 if (rack->r_ctl.rc_pace_max_segs != orig_val)
14424                         rack_log_pacing_delay_calc(rack, len, slot, orig_val, 0, 0, 15, __LINE__, NULL);
14425                 prev_fill = rack->r_via_fill_cw;
14426                 if ((rack->rc_pace_to_cwnd) &&
14427                     (capped == 0) &&
14428                     (rack->use_fixed_rate == 0) &&
14429                     (rack->in_probe_rtt == 0) &&
14430                     (IN_FASTRECOVERY(rack->rc_tp->t_flags) == 0)) {
14431                         /*
14432                          * We want to pace at our rate *or* faster to
14433                          * fill the cwnd to the max if its not full.
14434                          */
14435                         slot = pace_to_fill_cwnd(rack, slot, (len+segs), segsiz, &capped, &rate_wanted, 0);
14436                 }
14437                 if ((rack->rc_inp->inp_route.ro_nh != NULL) &&
14438                     (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
14439                         if ((rack->rack_hdw_pace_ena) &&
14440                             (can_start_hw_pacing > 0) &&
14441                             (rack->rack_hdrw_pacing == 0) &&
14442                             (rack->rack_attempt_hdwr_pace == 0)) {
14443                                 /*
14444                                  * Lets attempt to turn on hardware pacing
14445                                  * if we can.
14446                                  */
14447                                 rack->rack_attempt_hdwr_pace = 1;
14448                                 rack->r_ctl.crte = tcp_set_pacing_rate(rack->rc_tp,
14449                                                                        rack->rc_inp->inp_route.ro_nh->nh_ifp,
14450                                                                        rate_wanted,
14451                                                                        RS_PACING_GEQ,
14452                                                                        &err, &rack->r_ctl.crte_prev_rate);
14453                                 if (rack->r_ctl.crte) {
14454                                         rack->rack_hdrw_pacing = 1;
14455                                         rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted, segsiz,
14456                                                                                                  0, rack->r_ctl.crte,
14457                                                                                                  NULL);
14458                                         rack_log_hdwr_pacing(rack,
14459                                                              rate_wanted, rack->r_ctl.crte->rate, __LINE__,
14460                                                              err, 0);
14461                                         rack->r_ctl.last_hw_bw_req = rate_wanted;
14462                                 } else {
14463                                         counter_u64_add(rack_hw_pace_init_fail, 1);
14464                                 }
14465                         } else if (rack->rack_hdrw_pacing &&
14466                                    (rack->r_ctl.last_hw_bw_req != rate_wanted)) {
14467                                 /* Do we need to adjust our rate? */
14468                                 const struct tcp_hwrate_limit_table *nrte;
14469
14470                                 if (rack->r_up_only &&
14471                                     (rate_wanted < rack->r_ctl.crte->rate)) {
14472                                         /**
14473                                          * We have four possible states here
14474                                          * having to do with the previous time
14475                                          * and this time.
14476                                          *   previous  |  this-time
14477                                          * A)     0      |     0   -- fill_cw not in the picture
14478                                          * B)     1      |     0   -- we were doing a fill-cw but now are not
14479                                          * C)     1      |     1   -- all rates from fill_cw
14480                                          * D)     0      |     1   -- we were doing non-fill and now we are filling
14481                                          *
14482                                          * For case A, C and D we don't allow a drop. But for
14483                                          * case B where we now our on our steady rate we do
14484                                          * allow a drop.
14485                                          *
14486                                          */
14487                                         if (!((prev_fill == 1) && (rack->r_via_fill_cw == 0)))
14488                                                 goto done_w_hdwr;
14489                                 }
14490                                 if ((rate_wanted > rack->r_ctl.crte->rate) ||
14491                                     (rate_wanted <= rack->r_ctl.crte_prev_rate)) {
14492                                         if (rack_hw_rate_to_low &&
14493                                             (bw_est < rack_hw_rate_to_low)) {
14494                                                 /*
14495                                                  * The pacing rate is too low for hardware, but
14496                                                  * do allow hardware pacing to be restarted.
14497                                                  */
14498                                                 rack_log_hdwr_pacing(rack,
14499                                                              bw_est, rack->r_ctl.crte->rate, __LINE__,
14500                                                              0, 5);
14501                                                 tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
14502                                                 rack->r_ctl.crte = NULL;
14503                                                 rack->rack_attempt_hdwr_pace = 0;
14504                                                 rack->rack_hdrw_pacing = 0;
14505                                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
14506                                                 goto done_w_hdwr;
14507                                         }
14508                                         nrte = tcp_chg_pacing_rate(rack->r_ctl.crte,
14509                                                                    rack->rc_tp,
14510                                                                    rack->rc_inp->inp_route.ro_nh->nh_ifp,
14511                                                                    rate_wanted,
14512                                                                    RS_PACING_GEQ,
14513                                                                    &err, &rack->r_ctl.crte_prev_rate);
14514                                         if (nrte == NULL) {
14515                                                 /* Lost the rate */
14516                                                 rack->rack_hdrw_pacing = 0;
14517                                                 rack->r_ctl.crte = NULL;
14518                                                 rack_log_hdwr_pacing(rack,
14519                                                                      rate_wanted, 0, __LINE__,
14520                                                                      err, 1);
14521                                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
14522                                                 counter_u64_add(rack_hw_pace_lost, 1);
14523                                         } else if (nrte != rack->r_ctl.crte) {
14524                                                 rack->r_ctl.crte = nrte;
14525                                                 rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted,
14526                                                                                                          segsiz, 0,
14527                                                                                                          rack->r_ctl.crte,
14528                                                                                                          NULL);
14529                                                 rack_log_hdwr_pacing(rack,
14530                                                                      rate_wanted, rack->r_ctl.crte->rate, __LINE__,
14531                                                                      err, 2);
14532                                                 rack->r_ctl.last_hw_bw_req = rate_wanted;
14533                                         }
14534                                 } else {
14535                                         /* We just need to adjust the segment size */
14536                                         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
14537                                         rack_log_hdwr_pacing(rack,
14538                                                              rate_wanted, rack->r_ctl.crte->rate, __LINE__,
14539                                                              0, 4);
14540                                         rack->r_ctl.last_hw_bw_req = rate_wanted;
14541                                 }
14542                         }
14543                 }
14544                 if ((rack->r_ctl.crte != NULL) &&
14545                     (rack->r_ctl.crte->rate == rate_wanted)) {
14546                         /*
14547                          * We need to add a extra if the rates
14548                          * are exactly matched. The idea is
14549                          * we want the software to make sure the
14550                          * queue is empty before adding more, this
14551                          * gives us N MSS extra pace times where
14552                          * N is our sysctl
14553                          */
14554                         slot += (rack->r_ctl.crte->time_between * rack_hw_pace_extra_slots);
14555                 }
14556 done_w_hdwr:
14557                 if (rack_limit_time_with_srtt &&
14558                     (rack->use_fixed_rate == 0) &&
14559 #ifdef NETFLIX_PEAKRATE
14560                     (rack->rc_tp->t_maxpeakrate == 0) &&
14561 #endif
14562                     (rack->rack_hdrw_pacing == 0)) {
14563                         /*
14564                          * Sanity check, we do not allow the pacing delay
14565                          * to be longer than the SRTT of the path. If it is
14566                          * a slow path, then adding a packet should increase
14567                          * the RTT and compensate for this i.e. the srtt will
14568                          * be greater so the allowed pacing time will be greater.
14569                          *
14570                          * Note this restriction is not for where a peak rate
14571                          * is set, we are doing fixed pacing or hardware pacing.
14572                          */
14573                         if (rack->rc_tp->t_srtt)
14574                                 srtt = rack->rc_tp->t_srtt;
14575                         else
14576                                 srtt = RACK_INITIAL_RTO * HPTS_USEC_IN_MSEC;    /* its in ms convert */
14577                         if (srtt < slot) {
14578                                 rack_log_pacing_delay_calc(rack, srtt, slot, rate_wanted, bw_est, lentim, 99, __LINE__, NULL);
14579                                 slot = srtt;
14580                         }
14581                 }
14582                 rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, bw_est, lentim, 2, __LINE__, rsm);
14583         }
14584         if (rack->r_ctl.crte && (rack->r_ctl.crte->rs_num_enobufs > 0)) {
14585                 /*
14586                  * If this rate is seeing enobufs when it
14587                  * goes to send then either the nic is out
14588                  * of gas or we are mis-estimating the time
14589                  * somehow and not letting the queue empty
14590                  * completely. Lets add to the pacing time.
14591                  */
14592                 int hw_boost_delay;
14593
14594                 hw_boost_delay = rack->r_ctl.crte->time_between * rack_enobuf_hw_boost_mult;
14595                 if (hw_boost_delay > rack_enobuf_hw_max)
14596                         hw_boost_delay = rack_enobuf_hw_max;
14597                 else if (hw_boost_delay < rack_enobuf_hw_min)
14598                         hw_boost_delay = rack_enobuf_hw_min;
14599                 slot += hw_boost_delay;
14600         }
14601         if (slot)
14602                 counter_u64_add(rack_calc_nonzero, 1);
14603         else
14604                 counter_u64_add(rack_calc_zero, 1);
14605         return (slot);
14606 }
14607
14608 static void
14609 rack_start_gp_measurement(struct tcpcb *tp, struct tcp_rack *rack,
14610     tcp_seq startseq, uint32_t sb_offset)
14611 {
14612         struct rack_sendmap *my_rsm = NULL;
14613         struct rack_sendmap fe;
14614
14615         if (tp->t_state < TCPS_ESTABLISHED) {
14616                 /*
14617                  * We don't start any measurements if we are
14618                  * not at least established.
14619                  */
14620                 return;
14621         }
14622         tp->t_flags |= TF_GPUTINPROG;
14623         rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
14624         rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
14625         tp->gput_seq = startseq;
14626         rack->app_limited_needs_set = 0;
14627         if (rack->in_probe_rtt)
14628                 rack->measure_saw_probe_rtt = 1;
14629         else if ((rack->measure_saw_probe_rtt) &&
14630                  (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
14631                 rack->measure_saw_probe_rtt = 0;
14632         if (rack->rc_gp_filled)
14633                 tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
14634         else {
14635                 /* Special case initial measurement */
14636                 struct timeval tv;
14637
14638                 tp->gput_ts = tcp_get_usecs(&tv);
14639                 rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
14640         }
14641         /*
14642          * We take a guess out into the future,
14643          * if we have no measurement and no
14644          * initial rate, we measure the first
14645          * initial-windows worth of data to
14646          * speed up getting some GP measurement and
14647          * thus start pacing.
14648          */
14649         if ((rack->rc_gp_filled == 0) && (rack->r_ctl.init_rate == 0)) {
14650                 rack->app_limited_needs_set = 1;
14651                 tp->gput_ack = startseq + max(rc_init_window(rack),
14652                                               (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
14653                 rack_log_pacing_delay_calc(rack,
14654                                            tp->gput_seq,
14655                                            tp->gput_ack,
14656                                            0,
14657                                            tp->gput_ts,
14658                                            rack->r_ctl.rc_app_limited_cnt,
14659                                            9,
14660                                            __LINE__, NULL);
14661                 return;
14662         }
14663         if (sb_offset) {
14664                 /*
14665                  * We are out somewhere in the sb
14666                  * can we use the already outstanding data?
14667                  */
14668
14669                 if (rack->r_ctl.rc_app_limited_cnt == 0) {
14670                         /*
14671                          * Yes first one is good and in this case
14672                          * the tp->gput_ts is correctly set based on
14673                          * the last ack that arrived (no need to
14674                          * set things up when an ack comes in).
14675                          */
14676                         my_rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
14677                         if ((my_rsm == NULL) ||
14678                             (my_rsm->r_rtr_cnt != 1)) {
14679                                 /* retransmission? */
14680                                 goto use_latest;
14681                         }
14682                 } else {
14683                         if (rack->r_ctl.rc_first_appl == NULL) {
14684                                 /*
14685                                  * If rc_first_appl is NULL
14686                                  * then the cnt should be 0.
14687                                  * This is probably an error, maybe
14688                                  * a KASSERT would be approprate.
14689                                  */
14690                                 goto use_latest;
14691                         }
14692                         /*
14693                          * If we have a marker pointer to the last one that is
14694                          * app limited we can use that, but we need to set
14695                          * things up so that when it gets ack'ed we record
14696                          * the ack time (if its not already acked).
14697                          */
14698                         rack->app_limited_needs_set = 1;
14699                         /*
14700                          * We want to get to the rsm that is either
14701                          * next with space i.e. over 1 MSS or the one
14702                          * after that (after the app-limited).
14703                          */
14704                         my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
14705                                          rack->r_ctl.rc_first_appl);
14706                         if (my_rsm) {
14707                                 if ((my_rsm->r_end - my_rsm->r_start) <= ctf_fixed_maxseg(tp))
14708                                         /* Have to use the next one */
14709                                         my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
14710                                                          my_rsm);
14711                                 else {
14712                                         /* Use after the first MSS of it is acked */
14713                                         tp->gput_seq = my_rsm->r_start + ctf_fixed_maxseg(tp);
14714                                         goto start_set;
14715                                 }
14716                         }
14717                         if ((my_rsm == NULL) ||
14718                             (my_rsm->r_rtr_cnt != 1)) {
14719                                 /*
14720                                  * Either its a retransmit or
14721                                  * the last is the app-limited one.
14722                                  */
14723                                 goto use_latest;
14724                         }
14725                 }
14726                 tp->gput_seq = my_rsm->r_start;
14727 start_set:
14728                 if (my_rsm->r_flags & RACK_ACKED) {
14729                         /*
14730                          * This one has been acked use the arrival ack time
14731                          */
14732                         tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
14733                         rack->app_limited_needs_set = 0;
14734                 }
14735                 rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)];
14736                 tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
14737                 rack_log_pacing_delay_calc(rack,
14738                                            tp->gput_seq,
14739                                            tp->gput_ack,
14740                                            (uint64_t)my_rsm,
14741                                            tp->gput_ts,
14742                                            rack->r_ctl.rc_app_limited_cnt,
14743                                            9,
14744                                            __LINE__, NULL);
14745                 return;
14746         }
14747
14748 use_latest:
14749         /*
14750          * We don't know how long we may have been
14751          * idle or if this is the first-send. Lets
14752          * setup the flag so we will trim off
14753          * the first ack'd data so we get a true
14754          * measurement.
14755          */
14756         rack->app_limited_needs_set = 1;
14757         tp->gput_ack = startseq + rack_get_measure_window(tp, rack);
14758         /* Find this guy so we can pull the send time */
14759         fe.r_start = startseq;
14760         my_rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
14761         if (my_rsm) {
14762                 rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)];
14763                 if (my_rsm->r_flags & RACK_ACKED) {
14764                         /*
14765                          * Unlikely since its probably what was
14766                          * just transmitted (but I am paranoid).
14767                          */
14768                         tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
14769                         rack->app_limited_needs_set = 0;
14770                 }
14771                 if (SEQ_LT(my_rsm->r_start, tp->gput_seq)) {
14772                         /* This also is unlikely */
14773                         tp->gput_seq = my_rsm->r_start;
14774                 }
14775         } else {
14776                 /*
14777                  * TSNH unless we have some send-map limit,
14778                  * and even at that it should not be hitting
14779                  * that limit (we should have stopped sending).
14780                  */
14781                 struct timeval tv;
14782
14783                 microuptime(&tv);
14784                 rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
14785         }
14786         rack_log_pacing_delay_calc(rack,
14787                                    tp->gput_seq,
14788                                    tp->gput_ack,
14789                                    (uint64_t)my_rsm,
14790                                    tp->gput_ts,
14791                                    rack->r_ctl.rc_app_limited_cnt,
14792                                    9, __LINE__, NULL);
14793 }
14794
14795 static inline uint32_t
14796 rack_what_can_we_send(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t cwnd_to_use,
14797     uint32_t avail, int32_t sb_offset)
14798 {
14799         uint32_t len;
14800         uint32_t sendwin;
14801
14802         if (tp->snd_wnd > cwnd_to_use)
14803                 sendwin = cwnd_to_use;
14804         else
14805                 sendwin = tp->snd_wnd;
14806         if (ctf_outstanding(tp) >= tp->snd_wnd) {
14807                 /* We never want to go over our peers rcv-window */
14808                 len = 0;
14809         } else {
14810                 uint32_t flight;
14811
14812                 flight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
14813                 if (flight >= sendwin) {
14814                         /*
14815                          * We have in flight what we are allowed by cwnd (if
14816                          * it was rwnd blocking it would have hit above out
14817                          * >= tp->snd_wnd).
14818                          */
14819                         return (0);
14820                 }
14821                 len = sendwin - flight;
14822                 if ((len + ctf_outstanding(tp)) > tp->snd_wnd) {
14823                         /* We would send too much (beyond the rwnd) */
14824                         len = tp->snd_wnd - ctf_outstanding(tp);
14825                 }
14826                 if ((len + sb_offset) > avail) {
14827                         /*
14828                          * We don't have that much in the SB, how much is
14829                          * there?
14830                          */
14831                         len = avail - sb_offset;
14832                 }
14833         }
14834         return (len);
14835 }
14836
14837 static void
14838 rack_log_fsb(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t flags,
14839              unsigned ipoptlen, int32_t orig_len, int32_t len, int error,
14840              int rsm_is_null, int optlen, int line, uint16_t mode)
14841 {
14842         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
14843                 union tcp_log_stackspecific log;
14844                 struct timeval tv;
14845
14846                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
14847                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
14848                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
14849                 log.u_bbr.flex1 = error;
14850                 log.u_bbr.flex2 = flags;
14851                 log.u_bbr.flex3 = rsm_is_null;
14852                 log.u_bbr.flex4 = ipoptlen;
14853                 log.u_bbr.flex5 = tp->rcv_numsacks;
14854                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
14855                 log.u_bbr.flex7 = optlen;
14856                 log.u_bbr.flex8 = rack->r_fsb_inited;
14857                 log.u_bbr.applimited = rack->r_fast_output;
14858                 log.u_bbr.bw_inuse = rack_get_bw(rack);
14859                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
14860                 log.u_bbr.cwnd_gain = mode;
14861                 log.u_bbr.pkts_out = orig_len;
14862                 log.u_bbr.lt_epoch = len;
14863                 log.u_bbr.delivered = line;
14864                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14865                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14866                 tcp_log_event_(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_FSB, 0,
14867                                len, &log, false, NULL, NULL, 0, &tv);
14868         }
14869 }
14870
14871
14872 static struct mbuf *
14873 rack_fo_base_copym(struct mbuf *the_m, uint32_t the_off, int32_t *plen,
14874                    struct rack_fast_send_blk *fsb,
14875                    int32_t seglimit, int32_t segsize)
14876 {
14877 #ifdef KERN_TLS
14878         struct ktls_session *tls, *ntls;
14879         struct mbuf *start;
14880 #endif
14881         struct mbuf *m, *n, **np, *smb;
14882         struct mbuf *top;
14883         int32_t off, soff;
14884         int32_t len = *plen;
14885         int32_t fragsize;
14886         int32_t len_cp = 0;
14887         uint32_t mlen, frags;
14888
14889         soff = off = the_off;
14890         smb = m = the_m;
14891         np = &top;
14892         top = NULL;
14893 #ifdef KERN_TLS
14894         if (hw_tls && (m->m_flags & M_EXTPG))
14895                 tls = m->m_epg_tls;
14896         else
14897                 tls = NULL;
14898         start = m;
14899 #endif
14900         while (len > 0) {
14901                 if (m == NULL) {
14902                         *plen = len_cp;
14903                         break;
14904                 }
14905 #ifdef KERN_TLS
14906                 if (hw_tls) {
14907                         if (m->m_flags & M_EXTPG)
14908                                 ntls = m->m_epg_tls;
14909                         else
14910                                 ntls = NULL;
14911
14912                         /*
14913                          * Avoid mixing TLS records with handshake
14914                          * data or TLS records from different
14915                          * sessions.
14916                          */
14917                         if (tls != ntls) {
14918                                 MPASS(m != start);
14919                                 *plen = len_cp;
14920                                 break;
14921                         }
14922                 }
14923 #endif
14924                 mlen = min(len, m->m_len - off);
14925                 if (seglimit) {
14926                         /*
14927                          * For M_EXTPG mbufs, add 3 segments
14928                          * + 1 in case we are crossing page boundaries
14929                          * + 2 in case the TLS hdr/trailer are used
14930                          * It is cheaper to just add the segments
14931                          * than it is to take the cache miss to look
14932                          * at the mbuf ext_pgs state in detail.
14933                          */
14934                         if (m->m_flags & M_EXTPG) {
14935                                 fragsize = min(segsize, PAGE_SIZE);
14936                                 frags = 3;
14937                         } else {
14938                                 fragsize = segsize;
14939                                 frags = 0;
14940                         }
14941
14942                         /* Break if we really can't fit anymore. */
14943                         if ((frags + 1) >= seglimit) {
14944                                 *plen = len_cp;
14945                                 break;
14946                         }
14947
14948                         /*
14949                          * Reduce size if you can't copy the whole
14950                          * mbuf. If we can't copy the whole mbuf, also
14951                          * adjust len so the loop will end after this
14952                          * mbuf.
14953                          */
14954                         if ((frags + howmany(mlen, fragsize)) >= seglimit) {
14955                                 mlen = (seglimit - frags - 1) * fragsize;
14956                                 len = mlen;
14957                                 *plen = len_cp + len;
14958                         }
14959                         frags += howmany(mlen, fragsize);
14960                         if (frags == 0)
14961                                 frags++;
14962                         seglimit -= frags;
14963                         KASSERT(seglimit > 0,
14964                             ("%s: seglimit went too low", __func__));
14965                 }
14966                 n = m_get(M_NOWAIT, m->m_type);
14967                 *np = n;
14968                 if (n == NULL)
14969                         goto nospace;
14970                 n->m_len = mlen;
14971                 soff += mlen;
14972                 len_cp += n->m_len;
14973                 if (m->m_flags & (M_EXT|M_EXTPG)) {
14974                         n->m_data = m->m_data + off;
14975                         mb_dupcl(n, m);
14976                 } else {
14977                         bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
14978                             (u_int)n->m_len);
14979                 }
14980                 len -= n->m_len;
14981                 off = 0;
14982                 m = m->m_next;
14983                 np = &n->m_next;
14984                 if (len || (soff == smb->m_len)) {
14985                         /*
14986                          * We have more so we move forward  or
14987                          * we have consumed the entire mbuf and
14988                          * len has fell to 0.
14989                          */
14990                         soff = 0;
14991                         smb = m;
14992                 }
14993
14994         }
14995         if (fsb != NULL) {
14996                 fsb->m = smb;
14997                 fsb->off = soff;
14998                 if (smb) {
14999                         /*
15000                          * Save off the size of the mbuf. We do
15001                          * this so that we can recognize when it
15002                          * has been trimmed by sbcut() as acks
15003                          * come in.
15004                          */
15005                         fsb->o_m_len = smb->m_len;
15006                 } else {
15007                         /*
15008                          * This is the case where the next mbuf went to NULL. This
15009                          * means with this copy we have sent everything in the sb.
15010                          * In theory we could clear the fast_output flag, but lets
15011                          * not since its possible that we could get more added
15012                          * and acks that call the extend function which would let
15013                          * us send more.
15014                          */
15015                         fsb->o_m_len = 0;
15016                 }
15017         }
15018         return (top);
15019 nospace:
15020         if (top)
15021                 m_freem(top);
15022         return (NULL);
15023
15024 }
15025
15026 /*
15027  * This is a copy of m_copym(), taking the TSO segment size/limit
15028  * constraints into account, and advancing the sndptr as it goes.
15029  */
15030 static struct mbuf *
15031 rack_fo_m_copym(struct tcp_rack *rack, int32_t *plen,
15032                 int32_t seglimit, int32_t segsize, struct mbuf **s_mb, int *s_soff)
15033 {
15034         struct mbuf *m, *n;
15035         int32_t soff;
15036
15037         soff = rack->r_ctl.fsb.off;
15038         m = rack->r_ctl.fsb.m;
15039         if (rack->r_ctl.fsb.o_m_len != m->m_len) {
15040                 /*
15041                  * The mbuf had the front of it chopped off by an ack
15042                  * we need to adjust the soff/off by that difference.
15043                  */
15044                 uint32_t delta;
15045
15046                 delta = rack->r_ctl.fsb.o_m_len - m->m_len;
15047                 soff -= delta;
15048         }
15049         KASSERT(soff >= 0, ("%s, negative off %d", __FUNCTION__, soff));
15050         KASSERT(*plen >= 0, ("%s, negative len %d", __FUNCTION__, *plen));
15051         KASSERT(soff < m->m_len, ("%s rack:%p len:%u m:%p m->m_len:%u < off?",
15052                                  __FUNCTION__,
15053                                  rack, *plen, m, m->m_len));
15054         /* Save off the right location before we copy and advance */
15055         *s_soff = soff;
15056         *s_mb = rack->r_ctl.fsb.m;
15057         n = rack_fo_base_copym(m, soff, plen,
15058                                &rack->r_ctl.fsb,
15059                                seglimit, segsize);
15060         return (n);
15061 }
15062
15063 static int
15064 rack_fast_rsm_output(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm,
15065                      uint64_t ts_val, uint32_t cts, uint32_t ms_cts, struct timeval *tv, int len)
15066 {
15067         /*
15068          * Enter the fast retransmit path. We are given that a sched_pin is
15069          * in place (if accounting is compliled in) and the cycle count taken
15070          * at the entry is in the ts_val. The concept her is that the rsm
15071          * now holds the mbuf offsets and such so we can directly transmit
15072          * without a lot of overhead, the len field is already set for
15073          * us to prohibit us from sending too much (usually its 1MSS).
15074          */
15075         struct ip *ip = NULL;
15076         struct udphdr *udp = NULL;
15077         struct tcphdr *th = NULL;
15078         struct mbuf *m = NULL;
15079         struct inpcb *inp;
15080         uint8_t *cpto;
15081         struct tcp_log_buffer *lgb;
15082 #ifdef TCP_ACCOUNTING
15083         uint64_t crtsc;
15084         int cnt_thru = 1;
15085 #endif
15086         int doing_tlp = 0;
15087         struct tcpopt to;
15088         u_char opt[TCP_MAXOLEN];
15089         uint32_t hdrlen, optlen;
15090         int32_t slot, segsiz, max_val, tso = 0, error, flags, ulen = 0;
15091         uint32_t us_cts;
15092         uint32_t if_hw_tsomaxsegcount = 0, startseq;
15093         uint32_t if_hw_tsomaxsegsize;
15094
15095 #ifdef INET6
15096         struct ip6_hdr *ip6 = NULL;
15097
15098         if (rack->r_is_v6) {
15099                 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
15100                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
15101         } else
15102 #endif                          /* INET6 */
15103         {
15104                 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
15105                 hdrlen = sizeof(struct tcpiphdr);
15106         }
15107         if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
15108                 goto failed;
15109         }
15110         if (rsm->r_flags & RACK_TLP)
15111                 doing_tlp = 1;
15112         startseq = rsm->r_start;
15113         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
15114         inp = rack->rc_inp;
15115         to.to_flags = 0;
15116         flags = tcp_outflags[tp->t_state];
15117         if (flags & (TH_SYN|TH_RST)) {
15118                 goto failed;
15119         }
15120         if (rsm->r_flags & RACK_HAS_FIN) {
15121                 /* We can't send a FIN here */
15122                 goto failed;
15123         }
15124         if (flags & TH_FIN) {
15125                 /* We never send a FIN */
15126                 flags &= ~TH_FIN;
15127         }
15128         if (tp->t_flags & TF_RCVD_TSTMP) {
15129                 to.to_tsval = ms_cts + tp->ts_offset;
15130                 to.to_tsecr = tp->ts_recent;
15131                 to.to_flags = TOF_TS;
15132         }
15133         optlen = tcp_addoptions(&to, opt);
15134         hdrlen += optlen;
15135         udp = rack->r_ctl.fsb.udp;
15136         if (udp)
15137                 hdrlen += sizeof(struct udphdr);
15138         if (rack->r_ctl.rc_pace_max_segs)
15139                 max_val = rack->r_ctl.rc_pace_max_segs;
15140         else if (rack->rc_user_set_max_segs)
15141                 max_val = rack->rc_user_set_max_segs * segsiz;
15142         else
15143                 max_val = len;
15144         if ((tp->t_flags & TF_TSO) &&
15145             V_tcp_do_tso &&
15146             (len > segsiz) &&
15147             (tp->t_port == 0))
15148                 tso = 1;
15149 #ifdef INET6
15150         if (MHLEN < hdrlen + max_linkhdr)
15151                 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
15152         else
15153 #endif
15154                 m = m_gethdr(M_NOWAIT, MT_DATA);
15155         if (m == NULL)
15156                 goto failed;
15157         m->m_data += max_linkhdr;
15158         m->m_len = hdrlen;
15159         th = rack->r_ctl.fsb.th;
15160         /* Establish the len to send */
15161         if (len > max_val)
15162                 len = max_val;
15163         if ((tso) && (len + optlen > tp->t_maxseg)) {
15164                 uint32_t if_hw_tsomax;
15165                 int32_t max_len;
15166
15167                 /* extract TSO information */
15168                 if_hw_tsomax = tp->t_tsomax;
15169                 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
15170                 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
15171                 /*
15172                  * Check if we should limit by maximum payload
15173                  * length:
15174                  */
15175                 if (if_hw_tsomax != 0) {
15176                         /* compute maximum TSO length */
15177                         max_len = (if_hw_tsomax - hdrlen -
15178                                    max_linkhdr);
15179                         if (max_len <= 0) {
15180                                 goto failed;
15181                         } else if (len > max_len) {
15182                                 len = max_len;
15183                         }
15184                 }
15185                 if (len <= segsiz) {
15186                         /*
15187                          * In case there are too many small fragments don't
15188                          * use TSO:
15189                          */
15190                         tso = 0;
15191                 }
15192         } else {
15193                 tso = 0;
15194         }
15195         if ((tso == 0) && (len > segsiz))
15196                 len = segsiz;
15197         us_cts = tcp_get_usecs(tv);
15198         if ((len == 0) ||
15199             (len <= MHLEN - hdrlen - max_linkhdr)) {
15200                 goto failed;
15201         }
15202         th->th_seq = htonl(rsm->r_start);
15203         th->th_ack = htonl(tp->rcv_nxt);
15204         /*
15205          * The PUSH bit should only be applied
15206          * if the full retransmission is made. If
15207          * we are sending less than this is the
15208          * left hand edge and should not have
15209          * the PUSH bit.
15210          */
15211         if ((rsm->r_flags & RACK_HAD_PUSH) &&
15212             (len == (rsm->r_end - rsm->r_start)))
15213                 flags |= TH_PUSH;
15214         th->th_flags = flags;
15215         th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
15216         if (th->th_win == 0) {
15217                 tp->t_sndzerowin++;
15218                 tp->t_flags |= TF_RXWIN0SENT;
15219         } else
15220                 tp->t_flags &= ~TF_RXWIN0SENT;
15221         if (rsm->r_flags & RACK_TLP) {
15222                 /*
15223                  * TLP should not count in retran count, but
15224                  * in its own bin
15225                  */
15226                 counter_u64_add(rack_tlp_retran, 1);
15227                 counter_u64_add(rack_tlp_retran_bytes, len);
15228         } else {
15229                 tp->t_sndrexmitpack++;
15230                 KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
15231                 KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
15232         }
15233 #ifdef STATS
15234         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
15235                                  len);
15236 #endif
15237         if (rsm->m == NULL)
15238                 goto failed;
15239         if (rsm->orig_m_len != rsm->m->m_len) {
15240                 /* Fix up the orig_m_len and possibly the mbuf offset */
15241                 rack_adjust_orig_mlen(rsm);
15242         }
15243         m->m_next = rack_fo_base_copym(rsm->m, rsm->soff, &len, NULL, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize);
15244         if (len <= segsiz) {
15245                 /*
15246                  * Must have ran out of mbufs for the copy
15247                  * shorten it to no longer need tso. Lets
15248                  * not put on sendalot since we are low on
15249                  * mbufs.
15250                  */
15251                 tso = 0;
15252         }
15253         if ((m->m_next == NULL) || (len <= 0)){
15254                 goto failed;
15255         }
15256         if (udp) {
15257                 if (rack->r_is_v6)
15258                         ulen = hdrlen + len - sizeof(struct ip6_hdr);
15259                 else
15260                         ulen = hdrlen + len - sizeof(struct ip);
15261                 udp->uh_ulen = htons(ulen);
15262         }
15263         m->m_pkthdr.rcvif = (struct ifnet *)0;
15264         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
15265 #ifdef INET6
15266         if (rack->r_is_v6) {
15267                 if (tp->t_port) {
15268                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
15269                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15270                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
15271                         th->th_sum = htons(0);
15272                         UDPSTAT_INC(udps_opackets);
15273                 } else {
15274                         m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
15275                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15276                         th->th_sum = in6_cksum_pseudo(ip6,
15277                                                       sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
15278                                                       0);
15279                 }
15280         }
15281 #endif
15282 #if defined(INET6) && defined(INET)
15283         else
15284 #endif
15285 #ifdef INET
15286         {
15287                 if (tp->t_port) {
15288                         m->m_pkthdr.csum_flags = CSUM_UDP;
15289                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15290                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
15291                                                 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
15292                         th->th_sum = htons(0);
15293                         UDPSTAT_INC(udps_opackets);
15294                 } else {
15295                         m->m_pkthdr.csum_flags = CSUM_TCP;
15296                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15297                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
15298                                                ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
15299                                                                         IPPROTO_TCP + len + optlen));
15300                 }
15301                 /* IP version must be set here for ipv4/ipv6 checking later */
15302                 KASSERT(ip->ip_v == IPVERSION,
15303                         ("%s: IP version incorrect: %d", __func__, ip->ip_v));
15304         }
15305 #endif
15306         if (tso) {
15307                 KASSERT(len > tp->t_maxseg - optlen,
15308                         ("%s: len <= tso_segsz tp:%p", __func__, tp));
15309                 m->m_pkthdr.csum_flags |= CSUM_TSO;
15310                 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
15311         }
15312 #ifdef INET6
15313         if (rack->r_is_v6) {
15314                 ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
15315                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
15316                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
15317                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15318                 else
15319                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15320         }
15321 #endif
15322 #if defined(INET) && defined(INET6)
15323         else
15324 #endif
15325 #ifdef INET
15326         {
15327                 ip->ip_len = htons(m->m_pkthdr.len);
15328                 ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
15329                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
15330                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15331                         if (tp->t_port == 0 || len < V_tcp_minmss) {
15332                                 ip->ip_off |= htons(IP_DF);
15333                         }
15334                 } else {
15335                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15336                 }
15337         }
15338 #endif
15339         /* Time to copy in our header */
15340         cpto = mtod(m, uint8_t *);
15341         memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
15342         th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
15343         if (optlen) {
15344                 bcopy(opt, th + 1, optlen);
15345                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
15346         } else {
15347                 th->th_off = sizeof(struct tcphdr) >> 2;
15348         }
15349         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
15350                 union tcp_log_stackspecific log;
15351
15352                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15353                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
15354                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
15355                 if (rack->rack_no_prr)
15356                         log.u_bbr.flex1 = 0;
15357                 else
15358                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
15359                 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
15360                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
15361                 log.u_bbr.flex4 = max_val;
15362                 log.u_bbr.flex5 = 0;
15363                 /* Save off the early/late values */
15364                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
15365                 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
15366                 log.u_bbr.bw_inuse = rack_get_bw(rack);
15367                 log.u_bbr.flex8 = 1;
15368                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
15369                 log.u_bbr.flex7 = 55;
15370                 log.u_bbr.pkts_out = tp->t_maxseg;
15371                 log.u_bbr.timeStamp = cts;
15372                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15373                 log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
15374                 log.u_bbr.delivered = 0;
15375                 lgb = tcp_log_event_(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
15376                                      len, &log, false, NULL, NULL, 0, tv);
15377         } else
15378                 lgb = NULL;
15379 #ifdef INET6
15380         if (rack->r_is_v6) {
15381                 error = ip6_output(m, NULL,
15382                                    &inp->inp_route6,
15383                                    0, NULL, NULL, inp);
15384         }
15385 #endif
15386 #if defined(INET) && defined(INET6)
15387         else
15388 #endif
15389 #ifdef INET
15390         {
15391                 error = ip_output(m, NULL,
15392                                   &inp->inp_route,
15393                                   0, 0, inp);
15394         }
15395 #endif
15396         m = NULL;
15397         if (lgb) {
15398                 lgb->tlb_errno = error;
15399                 lgb = NULL;
15400         }
15401         if (error) {
15402                 goto failed;
15403         }
15404         rack_log_output(tp, &to, len, rsm->r_start, flags, error, rack_to_usec_ts(tv),
15405                         rsm, RACK_SENT_FP, rsm->m, rsm->soff);
15406         if (doing_tlp && (rack->fast_rsm_hack == 0)) {
15407                 rack->rc_tlp_in_progress = 1;
15408                 rack->r_ctl.rc_tlp_cnt_out++;
15409         }
15410         if (error == 0)
15411                 tcp_account_for_send(tp, len, 1, doing_tlp);
15412         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
15413         rack->forced_ack = 0;   /* If we send something zap the FA flag */
15414         if (IN_FASTRECOVERY(tp->t_flags) && rsm)
15415                 rack->r_ctl.retran_during_recovery += len;
15416         {
15417                 int idx;
15418
15419                 idx = (len / segsiz) + 3;
15420                 if (idx >= TCP_MSS_ACCT_ATIMER)
15421                         counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
15422                 else
15423                         counter_u64_add(rack_out_size[idx], 1);
15424         }
15425         if (tp->t_rtttime == 0) {
15426                 tp->t_rtttime = ticks;
15427                 tp->t_rtseq = startseq;
15428                 KMOD_TCPSTAT_INC(tcps_segstimed);
15429         }
15430         counter_u64_add(rack_fto_rsm_send, 1);
15431         if (error && (error == ENOBUFS)) {
15432                 slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
15433                 if (rack->rc_enobuf < 0x7f)
15434                         rack->rc_enobuf++;
15435                 if (slot < (10 * HPTS_USEC_IN_MSEC))
15436                         slot = 10 * HPTS_USEC_IN_MSEC;
15437         } else
15438                 slot = rack_get_pacing_delay(rack, tp, len, NULL, segsiz);
15439         if ((slot == 0) ||
15440             (rack->rc_always_pace == 0) ||
15441             (rack->r_rr_config == 1)) {
15442                 /*
15443                  * We have no pacing set or we
15444                  * are using old-style rack or
15445                  * we are overriden to use the old 1ms pacing.
15446                  */
15447                 slot = rack->r_ctl.rc_min_to;
15448         }
15449         rack_start_hpts_timer(rack, tp, cts, slot, len, 0);
15450         if (rack->r_must_retran) {
15451                 rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
15452                 if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
15453                         /*
15454                          * We have retransmitted all we need.
15455                          */
15456                         rack->r_must_retran = 0;
15457                         rack->r_ctl.rc_out_at_rto = 0;
15458                 }
15459         }
15460 #ifdef TCP_ACCOUNTING
15461         crtsc = get_cyclecount();
15462         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15463                 tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
15464         }
15465         counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], cnt_thru);
15466         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15467                 tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
15468         }
15469         counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
15470         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15471                 tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((len + segsiz - 1) / segsiz);
15472         }
15473         counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((len + segsiz - 1) / segsiz));
15474         sched_unpin();
15475 #endif
15476         return (0);
15477 failed:
15478         if (m)
15479                 m_free(m);
15480         return (-1);
15481 }
15482
15483 static void
15484 rack_sndbuf_autoscale(struct tcp_rack *rack)
15485 {
15486         /*
15487          * Automatic sizing of send socket buffer.  Often the send buffer
15488          * size is not optimally adjusted to the actual network conditions
15489          * at hand (delay bandwidth product).  Setting the buffer size too
15490          * small limits throughput on links with high bandwidth and high
15491          * delay (eg. trans-continental/oceanic links).  Setting the
15492          * buffer size too big consumes too much real kernel memory,
15493          * especially with many connections on busy servers.
15494          *
15495          * The criteria to step up the send buffer one notch are:
15496          *  1. receive window of remote host is larger than send buffer
15497          *     (with a fudge factor of 5/4th);
15498          *  2. send buffer is filled to 7/8th with data (so we actually
15499          *     have data to make use of it);
15500          *  3. send buffer fill has not hit maximal automatic size;
15501          *  4. our send window (slow start and cogestion controlled) is
15502          *     larger than sent but unacknowledged data in send buffer.
15503          *
15504          * Note that the rack version moves things much faster since
15505          * we want to avoid hitting cache lines in the rack_fast_output()
15506          * path so this is called much less often and thus moves
15507          * the SB forward by a percentage.
15508          */
15509         struct socket *so;
15510         struct tcpcb *tp;
15511         uint32_t sendwin, scaleup;
15512
15513         tp = rack->rc_tp;
15514         so = rack->rc_inp->inp_socket;
15515         sendwin = min(rack->r_ctl.cwnd_to_use, tp->snd_wnd);
15516         if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
15517                 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
15518                     sbused(&so->so_snd) >=
15519                     (so->so_snd.sb_hiwat / 8 * 7) &&
15520                     sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
15521                     sendwin >= (sbused(&so->so_snd) -
15522                     (tp->snd_nxt - tp->snd_una))) {
15523                         if (rack_autosndbuf_inc)
15524                                 scaleup = (rack_autosndbuf_inc * so->so_snd.sb_hiwat) / 100;
15525                         else
15526                                 scaleup = V_tcp_autosndbuf_inc;
15527                         if (scaleup < V_tcp_autosndbuf_inc)
15528                                 scaleup = V_tcp_autosndbuf_inc;
15529                         scaleup += so->so_snd.sb_hiwat;
15530                         if (scaleup > V_tcp_autosndbuf_max)
15531                                 scaleup = V_tcp_autosndbuf_max;
15532                         if (!sbreserve_locked(&so->so_snd, scaleup, so, curthread))
15533                                 so->so_snd.sb_flags &= ~SB_AUTOSIZE;
15534                 }
15535         }
15536 }
15537
15538 static int
15539 rack_fast_output(struct tcpcb *tp, struct tcp_rack *rack, uint64_t ts_val,
15540                  uint32_t cts, uint32_t ms_cts, struct timeval *tv, long tot_len, int *send_err)
15541 {
15542         /*
15543          * Enter to do fast output. We are given that the sched_pin is
15544          * in place (if accounting is compiled in) and the cycle count taken
15545          * at entry is in place in ts_val. The idea here is that
15546          * we know how many more bytes needs to be sent (presumably either
15547          * during pacing or to fill the cwnd and that was greater than
15548          * the max-burst). We have how much to send and all the info we
15549          * need to just send.
15550          */
15551         struct ip *ip = NULL;
15552         struct udphdr *udp = NULL;
15553         struct tcphdr *th = NULL;
15554         struct mbuf *m, *s_mb;
15555         struct inpcb *inp;
15556         uint8_t *cpto;
15557         struct tcp_log_buffer *lgb;
15558 #ifdef TCP_ACCOUNTING
15559         uint64_t crtsc;
15560 #endif
15561         struct tcpopt to;
15562         u_char opt[TCP_MAXOLEN];
15563         uint32_t hdrlen, optlen;
15564         int cnt_thru = 1;
15565         int32_t slot, segsiz, len, max_val, tso = 0, sb_offset, error, flags, ulen = 0;
15566         uint32_t us_cts, s_soff;
15567         uint32_t if_hw_tsomaxsegcount = 0, startseq;
15568         uint32_t if_hw_tsomaxsegsize;
15569         uint16_t add_flag = RACK_SENT_FP;
15570 #ifdef INET6
15571         struct ip6_hdr *ip6 = NULL;
15572
15573         if (rack->r_is_v6) {
15574                 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
15575                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
15576         } else
15577 #endif                          /* INET6 */
15578         {
15579                 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
15580                 hdrlen = sizeof(struct tcpiphdr);
15581         }
15582         if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
15583                 m = NULL;
15584                 goto failed;
15585         }
15586         startseq = tp->snd_max;
15587         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
15588         inp = rack->rc_inp;
15589         len = rack->r_ctl.fsb.left_to_send;
15590         to.to_flags = 0;
15591         flags = rack->r_ctl.fsb.tcp_flags;
15592         if (tp->t_flags & TF_RCVD_TSTMP) {
15593                 to.to_tsval = ms_cts + tp->ts_offset;
15594                 to.to_tsecr = tp->ts_recent;
15595                 to.to_flags = TOF_TS;
15596         }
15597         optlen = tcp_addoptions(&to, opt);
15598         hdrlen += optlen;
15599         udp = rack->r_ctl.fsb.udp;
15600         if (udp)
15601                 hdrlen += sizeof(struct udphdr);
15602         if (rack->r_ctl.rc_pace_max_segs)
15603                 max_val = rack->r_ctl.rc_pace_max_segs;
15604         else if (rack->rc_user_set_max_segs)
15605                 max_val = rack->rc_user_set_max_segs * segsiz;
15606         else
15607                 max_val = len;
15608         if ((tp->t_flags & TF_TSO) &&
15609             V_tcp_do_tso &&
15610             (len > segsiz) &&
15611             (tp->t_port == 0))
15612                 tso = 1;
15613 again:
15614 #ifdef INET6
15615         if (MHLEN < hdrlen + max_linkhdr)
15616                 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
15617         else
15618 #endif
15619                 m = m_gethdr(M_NOWAIT, MT_DATA);
15620         if (m == NULL)
15621                 goto failed;
15622         m->m_data += max_linkhdr;
15623         m->m_len = hdrlen;
15624         th = rack->r_ctl.fsb.th;
15625         /* Establish the len to send */
15626         if (len > max_val)
15627                 len = max_val;
15628         if ((tso) && (len + optlen > tp->t_maxseg)) {
15629                 uint32_t if_hw_tsomax;
15630                 int32_t max_len;
15631
15632                 /* extract TSO information */
15633                 if_hw_tsomax = tp->t_tsomax;
15634                 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
15635                 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
15636                 /*
15637                  * Check if we should limit by maximum payload
15638                  * length:
15639                  */
15640                 if (if_hw_tsomax != 0) {
15641                         /* compute maximum TSO length */
15642                         max_len = (if_hw_tsomax - hdrlen -
15643                                    max_linkhdr);
15644                         if (max_len <= 0) {
15645                                 goto failed;
15646                         } else if (len > max_len) {
15647                                 len = max_len;
15648                         }
15649                 }
15650                 if (len <= segsiz) {
15651                         /*
15652                          * In case there are too many small fragments don't
15653                          * use TSO:
15654                          */
15655                         tso = 0;
15656                 }
15657         } else {
15658                 tso = 0;
15659         }
15660         if ((tso == 0) && (len > segsiz))
15661                 len = segsiz;
15662         us_cts = tcp_get_usecs(tv);
15663         if ((len == 0) ||
15664             (len <= MHLEN - hdrlen - max_linkhdr)) {
15665                 goto failed;
15666         }
15667         sb_offset = tp->snd_max - tp->snd_una;
15668         th->th_seq = htonl(tp->snd_max);
15669         th->th_ack = htonl(tp->rcv_nxt);
15670         th->th_flags = flags;
15671         th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
15672         if (th->th_win == 0) {
15673                 tp->t_sndzerowin++;
15674                 tp->t_flags |= TF_RXWIN0SENT;
15675         } else
15676                 tp->t_flags &= ~TF_RXWIN0SENT;
15677         tp->snd_up = tp->snd_una;       /* drag it along, its deprecated */
15678         KMOD_TCPSTAT_INC(tcps_sndpack);
15679         KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
15680 #ifdef STATS
15681         stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
15682                                  len);
15683 #endif
15684         if (rack->r_ctl.fsb.m == NULL)
15685                 goto failed;
15686
15687         /* s_mb and s_soff are saved for rack_log_output */
15688         m->m_next = rack_fo_m_copym(rack, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, &s_mb, &s_soff);
15689         if (len <= segsiz) {
15690                 /*
15691                  * Must have ran out of mbufs for the copy
15692                  * shorten it to no longer need tso. Lets
15693                  * not put on sendalot since we are low on
15694                  * mbufs.
15695                  */
15696                 tso = 0;
15697         }
15698         if (rack->r_ctl.fsb.rfo_apply_push &&
15699             (len == rack->r_ctl.fsb.left_to_send)) {
15700                 th->th_flags |= TH_PUSH;
15701                 add_flag |= RACK_HAD_PUSH;
15702         }
15703         if ((m->m_next == NULL) || (len <= 0)){
15704                 goto failed;
15705         }
15706         if (udp) {
15707                 if (rack->r_is_v6)
15708                         ulen = hdrlen + len - sizeof(struct ip6_hdr);
15709                 else
15710                         ulen = hdrlen + len - sizeof(struct ip);
15711                 udp->uh_ulen = htons(ulen);
15712         }
15713         m->m_pkthdr.rcvif = (struct ifnet *)0;
15714         if (tp->t_state == TCPS_ESTABLISHED &&
15715             (tp->t_flags2 & TF2_ECN_PERMIT)) {
15716                 /*
15717                  * If the peer has ECN, mark data packets with ECN capable
15718                  * transmission (ECT). Ignore pure ack packets,
15719                  * retransmissions.
15720                  */
15721                 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max)) {
15722 #ifdef INET6
15723                         if (rack->r_is_v6)
15724                                 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
15725                         else
15726 #endif
15727                                 ip->ip_tos |= IPTOS_ECN_ECT0;
15728                         KMOD_TCPSTAT_INC(tcps_ecn_ect0);
15729                         /*
15730                          * Reply with proper ECN notifications.
15731                          * Only set CWR on new data segments.
15732                          */
15733                         if (tp->t_flags2 & TF2_ECN_SND_CWR) {
15734                                 flags |= TH_CWR;
15735                                 tp->t_flags2 &= ~TF2_ECN_SND_CWR;
15736                         }
15737                 }
15738                 if (tp->t_flags2 & TF2_ECN_SND_ECE)
15739                         flags |= TH_ECE;
15740         }
15741         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
15742 #ifdef INET6
15743         if (rack->r_is_v6) {
15744                 if (tp->t_port) {
15745                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
15746                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15747                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
15748                         th->th_sum = htons(0);
15749                         UDPSTAT_INC(udps_opackets);
15750                 } else {
15751                         m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
15752                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15753                         th->th_sum = in6_cksum_pseudo(ip6,
15754                                                       sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
15755                                                       0);
15756                 }
15757         }
15758 #endif
15759 #if defined(INET6) && defined(INET)
15760         else
15761 #endif
15762 #ifdef INET
15763         {
15764                 if (tp->t_port) {
15765                         m->m_pkthdr.csum_flags = CSUM_UDP;
15766                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15767                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
15768                                                 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
15769                         th->th_sum = htons(0);
15770                         UDPSTAT_INC(udps_opackets);
15771                 } else {
15772                         m->m_pkthdr.csum_flags = CSUM_TCP;
15773                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15774                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
15775                                                ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
15776                                                                         IPPROTO_TCP + len + optlen));
15777                 }
15778                 /* IP version must be set here for ipv4/ipv6 checking later */
15779                 KASSERT(ip->ip_v == IPVERSION,
15780                         ("%s: IP version incorrect: %d", __func__, ip->ip_v));
15781         }
15782 #endif
15783         if (tso) {
15784                 KASSERT(len > tp->t_maxseg - optlen,
15785                         ("%s: len <= tso_segsz tp:%p", __func__, tp));
15786                 m->m_pkthdr.csum_flags |= CSUM_TSO;
15787                 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
15788         }
15789 #ifdef INET6
15790         if (rack->r_is_v6) {
15791                 ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
15792                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
15793                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
15794                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15795                 else
15796                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15797         }
15798 #endif
15799 #if defined(INET) && defined(INET6)
15800         else
15801 #endif
15802 #ifdef INET
15803         {
15804                 ip->ip_len = htons(m->m_pkthdr.len);
15805                 ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
15806                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
15807                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15808                         if (tp->t_port == 0 || len < V_tcp_minmss) {
15809                                 ip->ip_off |= htons(IP_DF);
15810                         }
15811                 } else {
15812                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15813                 }
15814         }
15815 #endif
15816         /* Time to copy in our header */
15817         cpto = mtod(m, uint8_t *);
15818         memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
15819         th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
15820         if (optlen) {
15821                 bcopy(opt, th + 1, optlen);
15822                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
15823         } else {
15824                 th->th_off = sizeof(struct tcphdr) >> 2;
15825         }
15826         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
15827                 union tcp_log_stackspecific log;
15828
15829                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15830                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
15831                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
15832                 if (rack->rack_no_prr)
15833                         log.u_bbr.flex1 = 0;
15834                 else
15835                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
15836                 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
15837                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
15838                 log.u_bbr.flex4 = max_val;
15839                 log.u_bbr.flex5 = 0;
15840                 /* Save off the early/late values */
15841                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
15842                 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
15843                 log.u_bbr.bw_inuse = rack_get_bw(rack);
15844                 log.u_bbr.flex8 = 0;
15845                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
15846                 log.u_bbr.flex7 = 44;
15847                 log.u_bbr.pkts_out = tp->t_maxseg;
15848                 log.u_bbr.timeStamp = cts;
15849                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15850                 log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
15851                 log.u_bbr.delivered = 0;
15852                 lgb = tcp_log_event_(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
15853                                      len, &log, false, NULL, NULL, 0, tv);
15854         } else
15855                 lgb = NULL;
15856 #ifdef INET6
15857         if (rack->r_is_v6) {
15858                 error = ip6_output(m, NULL,
15859                                    &inp->inp_route6,
15860                                    0, NULL, NULL, inp);
15861         }
15862 #endif
15863 #if defined(INET) && defined(INET6)
15864         else
15865 #endif
15866 #ifdef INET
15867         {
15868                 error = ip_output(m, NULL,
15869                                   &inp->inp_route,
15870                                   0, 0, inp);
15871         }
15872 #endif
15873         if (lgb) {
15874                 lgb->tlb_errno = error;
15875                 lgb = NULL;
15876         }
15877         if (error) {
15878                 *send_err = error;
15879                 m = NULL;
15880                 goto failed;
15881         }
15882         rack_log_output(tp, &to, len, tp->snd_max, flags, error, rack_to_usec_ts(tv),
15883                         NULL, add_flag, s_mb, s_soff);
15884         m = NULL;
15885         if (tp->snd_una == tp->snd_max) {
15886                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
15887                 rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
15888                 tp->t_acktime = ticks;
15889         }
15890         if (error == 0)
15891                 tcp_account_for_send(tp, len, 0, 0);
15892
15893         rack->forced_ack = 0;   /* If we send something zap the FA flag */
15894         tot_len += len;
15895         if ((tp->t_flags & TF_GPUTINPROG) == 0)
15896                 rack_start_gp_measurement(tp, rack, tp->snd_max, sb_offset);
15897         tp->snd_max += len;
15898         tp->snd_nxt = tp->snd_max;
15899         {
15900                 int idx;
15901
15902                 idx = (len / segsiz) + 3;
15903                 if (idx >= TCP_MSS_ACCT_ATIMER)
15904                         counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
15905                 else
15906                         counter_u64_add(rack_out_size[idx], 1);
15907         }
15908         if (len <= rack->r_ctl.fsb.left_to_send)
15909                 rack->r_ctl.fsb.left_to_send -= len;
15910         else
15911                 rack->r_ctl.fsb.left_to_send = 0;
15912         if (rack->r_ctl.fsb.left_to_send < segsiz) {
15913                 rack->r_fast_output = 0;
15914                 rack->r_ctl.fsb.left_to_send = 0;
15915                 /* At the end of fast_output scale up the sb */
15916                 SOCKBUF_LOCK(&rack->rc_inp->inp_socket->so_snd);
15917                 rack_sndbuf_autoscale(rack);
15918                 SOCKBUF_UNLOCK(&rack->rc_inp->inp_socket->so_snd);
15919         }
15920         if (tp->t_rtttime == 0) {
15921                 tp->t_rtttime = ticks;
15922                 tp->t_rtseq = startseq;
15923                 KMOD_TCPSTAT_INC(tcps_segstimed);
15924         }
15925         if ((rack->r_ctl.fsb.left_to_send >= segsiz) &&
15926             (max_val > len) &&
15927             (tso == 0)) {
15928                 max_val -= len;
15929                 len = segsiz;
15930                 th = rack->r_ctl.fsb.th;
15931                 cnt_thru++;
15932                 goto again;
15933         }
15934         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
15935         counter_u64_add(rack_fto_send, 1);
15936         slot = rack_get_pacing_delay(rack, tp, tot_len, NULL, segsiz);
15937         rack_start_hpts_timer(rack, tp, cts, slot, tot_len, 0);
15938 #ifdef TCP_ACCOUNTING
15939         crtsc = get_cyclecount();
15940         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15941                 tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
15942         }
15943         counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], cnt_thru);
15944         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15945                 tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
15946         }
15947         counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
15948         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15949                 tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len + segsiz - 1) / segsiz);
15950         }
15951         counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len + segsiz - 1) / segsiz));
15952         sched_unpin();
15953 #endif
15954         return (0);
15955 failed:
15956         if (m)
15957                 m_free(m);
15958         rack->r_fast_output = 0;
15959         return (-1);
15960 }
15961
15962 static int
15963 rack_output(struct tcpcb *tp)
15964 {
15965         struct socket *so;
15966         uint32_t recwin;
15967         uint32_t sb_offset, s_moff = 0;
15968         int32_t len, flags, error = 0;
15969         struct mbuf *m, *s_mb = NULL;
15970         struct mbuf *mb;
15971         uint32_t if_hw_tsomaxsegcount = 0;
15972         uint32_t if_hw_tsomaxsegsize;
15973         int32_t segsiz, minseg;
15974         long tot_len_this_send = 0;
15975 #ifdef INET
15976         struct ip *ip = NULL;
15977 #endif
15978 #ifdef TCPDEBUG
15979         struct ipovly *ipov = NULL;
15980 #endif
15981         struct udphdr *udp = NULL;
15982         struct tcp_rack *rack;
15983         struct tcphdr *th;
15984         uint8_t pass = 0;
15985         uint8_t mark = 0;
15986         uint8_t wanted_cookie = 0;
15987         u_char opt[TCP_MAXOLEN];
15988         unsigned ipoptlen, optlen, hdrlen, ulen=0;
15989         uint32_t rack_seq;
15990
15991 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
15992         unsigned ipsec_optlen = 0;
15993
15994 #endif
15995         int32_t idle, sendalot;
15996         int32_t sub_from_prr = 0;
15997         volatile int32_t sack_rxmit;
15998         struct rack_sendmap *rsm = NULL;
15999         int32_t tso, mtu;
16000         struct tcpopt to;
16001         int32_t slot = 0;
16002         int32_t sup_rack = 0;
16003         uint32_t cts, ms_cts, delayed, early;
16004         uint16_t add_flag = RACK_SENT_SP;
16005         uint8_t hpts_calling,  doing_tlp = 0;
16006         uint32_t cwnd_to_use, pace_max_seg;
16007         int32_t do_a_prefetch = 0;
16008         int32_t prefetch_rsm = 0;
16009         int32_t orig_len = 0;
16010         struct timeval tv;
16011         int32_t prefetch_so_done = 0;
16012         struct tcp_log_buffer *lgb;
16013         struct inpcb *inp;
16014         struct sockbuf *sb;
16015         uint64_t ts_val = 0;
16016 #ifdef TCP_ACCOUNTING
16017         uint64_t crtsc;
16018 #endif
16019 #ifdef INET6
16020         struct ip6_hdr *ip6 = NULL;
16021         int32_t isipv6;
16022 #endif
16023         uint8_t filled_all = 0;
16024         bool hw_tls = false;
16025
16026         /* setup and take the cache hits here */
16027         rack = (struct tcp_rack *)tp->t_fb_ptr;
16028 #ifdef TCP_ACCOUNTING
16029         sched_pin();
16030         ts_val = get_cyclecount();
16031 #endif
16032         hpts_calling = rack->rc_inp->inp_hpts_calls;
16033         NET_EPOCH_ASSERT();
16034         INP_WLOCK_ASSERT(rack->rc_inp);
16035 #ifdef TCP_OFFLOAD
16036         if (tp->t_flags & TF_TOE) {
16037 #ifdef TCP_ACCOUNTING
16038                 sched_unpin();
16039 #endif
16040                 return (tcp_offload_output(tp));
16041         }
16042 #endif
16043         /*
16044          * For TFO connections in SYN_RECEIVED, only allow the initial
16045          * SYN|ACK and those sent by the retransmit timer.
16046          */
16047         if (IS_FASTOPEN(tp->t_flags) &&
16048             (tp->t_state == TCPS_SYN_RECEIVED) &&
16049             SEQ_GT(tp->snd_max, tp->snd_una) &&    /* initial SYN|ACK sent */
16050             (rack->r_ctl.rc_resend == NULL)) {         /* not a retransmit */
16051 #ifdef TCP_ACCOUNTING
16052                 sched_unpin();
16053 #endif
16054                 return (0);
16055         }
16056 #ifdef INET6
16057         if (rack->r_state) {
16058                 /* Use the cache line loaded if possible */
16059                 isipv6 = rack->r_is_v6;
16060         } else {
16061                 isipv6 = (rack->rc_inp->inp_vflag & INP_IPV6) != 0;
16062         }
16063 #endif
16064         early = 0;
16065         cts = tcp_get_usecs(&tv);
16066         ms_cts = tcp_tv_to_mssectick(&tv);
16067         if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
16068             rack->rc_inp->inp_in_hpts) {
16069                 /*
16070                  * We are on the hpts for some timer but not hptsi output.
16071                  * Remove from the hpts unconditionally.
16072                  */
16073                 rack_timer_cancel(tp, rack, cts, __LINE__);
16074         }
16075         /* Are we pacing and late? */
16076         if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16077             TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to)) {
16078                 /* We are delayed */
16079                 delayed = cts - rack->r_ctl.rc_last_output_to;
16080         } else {
16081                 delayed = 0;
16082         }
16083         /* Do the timers, which may override the pacer */
16084         if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
16085                 if (rack_process_timers(tp, rack, cts, hpts_calling)) {
16086                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1);
16087 #ifdef TCP_ACCOUNTING
16088                         sched_unpin();
16089 #endif
16090                         return (0);
16091                 }
16092         }
16093         if (rack->rc_in_persist) {
16094                 if (rack->rc_inp->inp_in_hpts == 0) {
16095                         /* Timer is not running */
16096                         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
16097                 }
16098 #ifdef TCP_ACCOUNTING
16099                 sched_unpin();
16100 #endif
16101                 return (0);
16102         }
16103         if ((rack->r_timer_override) ||
16104             (rack->rc_ack_can_sendout_data) ||
16105             (delayed) ||
16106             (tp->t_state < TCPS_ESTABLISHED)) {
16107                 rack->rc_ack_can_sendout_data = 0;
16108                 if (rack->rc_inp->inp_in_hpts)
16109                         tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
16110         } else if (rack->rc_inp->inp_in_hpts) {
16111                 /*
16112                  * On the hpts you can't pass even if ACKNOW is on, we will
16113                  * when the hpts fires.
16114                  */
16115 #ifdef TCP_ACCOUNTING
16116                 crtsc = get_cyclecount();
16117                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16118                         tp->tcp_proc_time[SND_BLOCKED] += (crtsc - ts_val);
16119                 }
16120                 counter_u64_add(tcp_proc_time[SND_BLOCKED], (crtsc - ts_val));
16121                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16122                         tp->tcp_cnt_counters[SND_BLOCKED]++;
16123                 }
16124                 counter_u64_add(tcp_cnt_counters[SND_BLOCKED], 1);
16125                 sched_unpin();
16126 #endif
16127                 counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1);
16128                 return (0);
16129         }
16130         rack->rc_inp->inp_hpts_calls = 0;
16131         /* Finish out both pacing early and late accounting */
16132         if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16133             TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
16134                 early = rack->r_ctl.rc_last_output_to - cts;
16135         } else
16136                 early = 0;
16137         if (delayed) {
16138                 rack->r_ctl.rc_agg_delayed += delayed;
16139                 rack->r_late = 1;
16140         } else if (early) {
16141                 rack->r_ctl.rc_agg_early += early;
16142                 rack->r_early = 1;
16143         }
16144         /* Now that early/late accounting is done turn off the flag */
16145         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
16146         rack->r_wanted_output = 0;
16147         rack->r_timer_override = 0;
16148         if ((tp->t_state != rack->r_state) &&
16149             TCPS_HAVEESTABLISHED(tp->t_state)) {
16150                 rack_set_state(tp, rack);
16151         }
16152         if ((rack->r_fast_output) &&
16153             (tp->rcv_numsacks == 0)) {
16154                 int ret;
16155
16156                 error = 0;
16157                 ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
16158                 if (ret >= 0)
16159                         return(ret);
16160                 else if (error) {
16161                         inp = rack->rc_inp;
16162                         so = inp->inp_socket;
16163                         sb = &so->so_snd;
16164                         goto nomore;
16165                 }
16166         }
16167         inp = rack->rc_inp;
16168         /*
16169          * For TFO connections in SYN_SENT or SYN_RECEIVED,
16170          * only allow the initial SYN or SYN|ACK and those sent
16171          * by the retransmit timer.
16172          */
16173         if (IS_FASTOPEN(tp->t_flags) &&
16174             ((tp->t_state == TCPS_SYN_RECEIVED) ||
16175              (tp->t_state == TCPS_SYN_SENT)) &&
16176             SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
16177             (tp->t_rxtshift == 0)) {              /* not a retransmit */
16178                 cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
16179                 so = inp->inp_socket;
16180                 sb = &so->so_snd;
16181                 goto just_return_nolock;
16182         }
16183         /*
16184          * Determine length of data that should be transmitted, and flags
16185          * that will be used. If there is some data or critical controls
16186          * (SYN, RST) to send, then transmit; otherwise, investigate
16187          * further.
16188          */
16189         idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
16190         if (tp->t_idle_reduce) {
16191                 if (idle && ((ticks - tp->t_rcvtime) >= tp->t_rxtcur))
16192                         rack_cc_after_idle(rack, tp);
16193         }
16194         tp->t_flags &= ~TF_LASTIDLE;
16195         if (idle) {
16196                 if (tp->t_flags & TF_MORETOCOME) {
16197                         tp->t_flags |= TF_LASTIDLE;
16198                         idle = 0;
16199                 }
16200         }
16201         if ((tp->snd_una == tp->snd_max) &&
16202             rack->r_ctl.rc_went_idle_time &&
16203             TSTMP_GT(cts, rack->r_ctl.rc_went_idle_time)) {
16204                 idle = cts - rack->r_ctl.rc_went_idle_time;
16205                 if (idle > rack_min_probertt_hold) {
16206                         /* Count as a probe rtt */
16207                         if (rack->in_probe_rtt == 0) {
16208                                 rack->r_ctl.rc_lower_rtt_us_cts = cts;
16209                                 rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
16210                                 rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
16211                                 rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
16212                         } else {
16213                                 rack_exit_probertt(rack, cts);
16214                         }
16215                 }
16216                 idle = 0;
16217         }
16218         if (rack_use_fsb && (rack->r_fsb_inited == 0))
16219                 rack_init_fsb_block(tp, rack);
16220 again:
16221         /*
16222          * If we've recently taken a timeout, snd_max will be greater than
16223          * snd_nxt.  There may be SACK information that allows us to avoid
16224          * resending already delivered data.  Adjust snd_nxt accordingly.
16225          */
16226         sendalot = 0;
16227         cts = tcp_get_usecs(&tv);
16228         ms_cts = tcp_tv_to_mssectick(&tv);
16229         tso = 0;
16230         mtu = 0;
16231         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
16232         minseg = segsiz;
16233         if (rack->r_ctl.rc_pace_max_segs == 0)
16234                 pace_max_seg = rack->rc_user_set_max_segs * segsiz;
16235         else
16236                 pace_max_seg = rack->r_ctl.rc_pace_max_segs;
16237         sb_offset = tp->snd_max - tp->snd_una;
16238         cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
16239         flags = tcp_outflags[tp->t_state];
16240         while (rack->rc_free_cnt < rack_free_cache) {
16241                 rsm = rack_alloc(rack);
16242                 if (rsm == NULL) {
16243                         if (inp->inp_hpts_calls)
16244                                 /* Retry in a ms */
16245                                 slot = (1 * HPTS_USEC_IN_MSEC);
16246                         so = inp->inp_socket;
16247                         sb = &so->so_snd;
16248                         goto just_return_nolock;
16249                 }
16250                 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_tnext);
16251                 rack->rc_free_cnt++;
16252                 rsm = NULL;
16253         }
16254         if (inp->inp_hpts_calls)
16255                 inp->inp_hpts_calls = 0;
16256         sack_rxmit = 0;
16257         len = 0;
16258         rsm = NULL;
16259         if (flags & TH_RST) {
16260                 SOCKBUF_LOCK(&inp->inp_socket->so_snd);
16261                 so = inp->inp_socket;
16262                 sb = &so->so_snd;
16263                 goto send;
16264         }
16265         if (rack->r_ctl.rc_resend) {
16266                 /* Retransmit timer */
16267                 rsm = rack->r_ctl.rc_resend;
16268                 rack->r_ctl.rc_resend = NULL;
16269                 rsm->r_flags &= ~RACK_TLP;
16270                 len = rsm->r_end - rsm->r_start;
16271                 sack_rxmit = 1;
16272                 sendalot = 0;
16273                 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16274                         ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16275                          __func__, __LINE__,
16276                          rsm->r_start, tp->snd_una, tp, rack, rsm));
16277                 sb_offset = rsm->r_start - tp->snd_una;
16278                 if (len >= segsiz)
16279                         len = segsiz;
16280         } else if ((rsm = tcp_rack_output(tp, rack, cts)) != NULL) {
16281                 /* We have a retransmit that takes precedence */
16282                 rsm->r_flags &= ~RACK_TLP;
16283                 if ((!IN_FASTRECOVERY(tp->t_flags)) &&
16284                     ((tp->t_flags & TF_WASFRECOVERY) == 0)) {
16285                         /* Enter recovery if not induced by a time-out */
16286                         rack->r_ctl.rc_rsm_start = rsm->r_start;
16287                         rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
16288                         rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
16289                         rack_cong_signal(tp, CC_NDUPACK, tp->snd_una);
16290                 }
16291 #ifdef INVARIANTS
16292                 if (SEQ_LT(rsm->r_start, tp->snd_una)) {
16293                         panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n",
16294                               tp, rack, rsm, rsm->r_start, tp->snd_una);
16295                 }
16296 #endif
16297                 len = rsm->r_end - rsm->r_start;
16298                 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16299                         ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16300                          __func__, __LINE__,
16301                          rsm->r_start, tp->snd_una, tp, rack, rsm));
16302                 sb_offset = rsm->r_start - tp->snd_una;
16303                 sendalot = 0;
16304                 if (len >= segsiz)
16305                         len = segsiz;
16306                 if (len > 0) {
16307                         sack_rxmit = 1;
16308                         KMOD_TCPSTAT_INC(tcps_sack_rexmits);
16309                         KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes,
16310                             min(len, segsiz));
16311                         counter_u64_add(rack_rtm_prr_retran, 1);
16312                 }
16313         } else if (rack->r_ctl.rc_tlpsend) {
16314                 /* Tail loss probe */
16315                 long cwin;
16316                 long tlen;
16317
16318                 doing_tlp = 1;
16319                 /*
16320                  * Check if we can do a TLP with a RACK'd packet
16321                  * this can happen if we are not doing the rack
16322                  * cheat and we skipped to a TLP and it
16323                  * went off.
16324                  */
16325                 rsm = rack->r_ctl.rc_tlpsend;
16326                 rsm->r_flags |= RACK_TLP;
16327
16328                 rack->r_ctl.rc_tlpsend = NULL;
16329                 sack_rxmit = 1;
16330                 tlen = rsm->r_end - rsm->r_start;
16331                 if (tlen > segsiz)
16332                         tlen = segsiz;
16333                 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16334                         ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16335                          __func__, __LINE__,
16336                          rsm->r_start, tp->snd_una, tp, rack, rsm));
16337                 sb_offset = rsm->r_start - tp->snd_una;
16338                 cwin = min(tp->snd_wnd, tlen);
16339                 len = cwin;
16340         }
16341         if (rack->r_must_retran &&
16342             (rsm == NULL)) {
16343                 /*
16344                  * Non-Sack and we had a RTO or MTU change, we
16345                  * need to retransmit until we reach
16346                  * the former snd_max (rack->r_ctl.rc_snd_max_at_rto).
16347                  */
16348                 if (SEQ_GT(tp->snd_max, tp->snd_una)) {
16349                         int sendwin, flight;
16350
16351                         sendwin = min(tp->snd_wnd, tp->snd_cwnd);
16352                         flight = ctf_flight_size(tp, rack->r_ctl.rc_out_at_rto);
16353                         if (flight >= sendwin) {
16354                                 so = inp->inp_socket;
16355                                 sb = &so->so_snd;
16356                                 goto just_return_nolock;
16357                         }
16358                         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
16359                         KASSERT(rsm != NULL, ("rsm is NULL rack:%p r_must_retran set", rack));
16360                         if (rsm == NULL) {
16361                                 /* TSNH */
16362                                 rack->r_must_retran = 0;
16363                                 rack->r_ctl.rc_out_at_rto = 0;
16364                                 rack->r_must_retran = 0;
16365                                 so = inp->inp_socket;
16366                                 sb = &so->so_snd;
16367                                 goto just_return_nolock;
16368                         }
16369                         sack_rxmit = 1;
16370                         len = rsm->r_end - rsm->r_start;
16371                         sendalot = 0;
16372                         sb_offset = rsm->r_start - tp->snd_una;
16373                         if (len >= segsiz)
16374                                 len = segsiz;
16375                 } else {
16376                         /* We must be done if there is nothing outstanding */
16377                         rack->r_must_retran = 0;
16378                         rack->r_ctl.rc_out_at_rto = 0;
16379                 }
16380         }
16381         /*
16382          * Enforce a connection sendmap count limit if set
16383          * as long as we are not retransmiting.
16384          */
16385         if ((rsm == NULL) &&
16386             (rack->do_detection == 0) &&
16387             (V_tcp_map_entries_limit > 0) &&
16388             (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
16389                 counter_u64_add(rack_to_alloc_limited, 1);
16390                 if (!rack->alloc_limit_reported) {
16391                         rack->alloc_limit_reported = 1;
16392                         counter_u64_add(rack_alloc_limited_conns, 1);
16393                 }
16394                 so = inp->inp_socket;
16395                 sb = &so->so_snd;
16396                 goto just_return_nolock;
16397         }
16398         if (rsm && (rsm->r_flags & RACK_HAS_FIN)) {
16399                 /* we are retransmitting the fin */
16400                 len--;
16401                 if (len) {
16402                         /*
16403                          * When retransmitting data do *not* include the
16404                          * FIN. This could happen from a TLP probe.
16405                          */
16406                         flags &= ~TH_FIN;
16407                 }
16408         }
16409 #ifdef INVARIANTS
16410         /* For debugging */
16411         rack->r_ctl.rc_rsm_at_retran = rsm;
16412 #endif
16413         if (rsm && rack->r_fsb_inited && rack_use_rsm_rfo &&
16414             ((rsm->r_flags & RACK_HAS_FIN) == 0)) {
16415                 int ret;
16416
16417                 ret = rack_fast_rsm_output(tp, rack, rsm, ts_val, cts, ms_cts, &tv, len);
16418                 if (ret == 0)
16419                         return (0);
16420         }
16421         so = inp->inp_socket;
16422         sb = &so->so_snd;
16423         if (do_a_prefetch == 0) {
16424                 kern_prefetch(sb, &do_a_prefetch);
16425                 do_a_prefetch = 1;
16426         }
16427 #ifdef NETFLIX_SHARED_CWND
16428         if ((tp->t_flags2 & TF2_TCP_SCWND_ALLOWED) &&
16429             rack->rack_enable_scwnd) {
16430                 /* We are doing cwnd sharing */
16431                 if (rack->gp_ready &&
16432                     (rack->rack_attempted_scwnd == 0) &&
16433                     (rack->r_ctl.rc_scw == NULL) &&
16434                     tp->t_lib) {
16435                         /* The pcbid is in, lets make an attempt */
16436                         counter_u64_add(rack_try_scwnd, 1);
16437                         rack->rack_attempted_scwnd = 1;
16438                         rack->r_ctl.rc_scw = tcp_shared_cwnd_alloc(tp,
16439                                                                    &rack->r_ctl.rc_scw_index,
16440                                                                    segsiz);
16441                 }
16442                 if (rack->r_ctl.rc_scw &&
16443                     (rack->rack_scwnd_is_idle == 1) &&
16444                     sbavail(&so->so_snd)) {
16445                         /* we are no longer out of data */
16446                         tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
16447                         rack->rack_scwnd_is_idle = 0;
16448                 }
16449                 if (rack->r_ctl.rc_scw) {
16450                         /* First lets update and get the cwnd */
16451                         rack->r_ctl.cwnd_to_use = cwnd_to_use = tcp_shared_cwnd_update(rack->r_ctl.rc_scw,
16452                                                                     rack->r_ctl.rc_scw_index,
16453                                                                     tp->snd_cwnd, tp->snd_wnd, segsiz);
16454                 }
16455         }
16456 #endif
16457         /*
16458          * Get standard flags, and add SYN or FIN if requested by 'hidden'
16459          * state flags.
16460          */
16461         if (tp->t_flags & TF_NEEDFIN)
16462                 flags |= TH_FIN;
16463         if (tp->t_flags & TF_NEEDSYN)
16464                 flags |= TH_SYN;
16465         if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
16466                 void *end_rsm;
16467                 end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
16468                 if (end_rsm)
16469                         kern_prefetch(end_rsm, &prefetch_rsm);
16470                 prefetch_rsm = 1;
16471         }
16472         SOCKBUF_LOCK(sb);
16473         /*
16474          * If snd_nxt == snd_max and we have transmitted a FIN, the
16475          * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a
16476          * negative length.  This can also occur when TCP opens up its
16477          * congestion window while receiving additional duplicate acks after
16478          * fast-retransmit because TCP will reset snd_nxt to snd_max after
16479          * the fast-retransmit.
16480          *
16481          * In the normal retransmit-FIN-only case, however, snd_nxt will be
16482          * set to snd_una, the sb_offset will be 0, and the length may wind
16483          * up 0.
16484          *
16485          * If sack_rxmit is true we are retransmitting from the scoreboard
16486          * in which case len is already set.
16487          */
16488         if ((sack_rxmit == 0) &&
16489             (TCPS_HAVEESTABLISHED(tp->t_state) || IS_FASTOPEN(tp->t_flags))) {
16490                 uint32_t avail;
16491
16492                 avail = sbavail(sb);
16493                 if (SEQ_GT(tp->snd_nxt, tp->snd_una) && avail)
16494                         sb_offset = tp->snd_nxt - tp->snd_una;
16495                 else
16496                         sb_offset = 0;
16497                 if ((IN_FASTRECOVERY(tp->t_flags) == 0) || rack->rack_no_prr) {
16498                         if (rack->r_ctl.rc_tlp_new_data) {
16499                                 /* TLP is forcing out new data */
16500                                 if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) {
16501                                         rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset);
16502                                 }
16503                                 if ((rack->r_ctl.rc_tlp_new_data + sb_offset) > tp->snd_wnd) {
16504                                         if (tp->snd_wnd > sb_offset)
16505                                                 len = tp->snd_wnd - sb_offset;
16506                                         else
16507                                                 len = 0;
16508                                 } else {
16509                                         len = rack->r_ctl.rc_tlp_new_data;
16510                                 }
16511                                 rack->r_ctl.rc_tlp_new_data = 0;
16512                                 doing_tlp = 1;
16513                         }  else {
16514                                 len = rack_what_can_we_send(tp, rack, cwnd_to_use, avail, sb_offset);
16515                         }
16516                         if ((rack->r_ctl.crte == NULL) && IN_FASTRECOVERY(tp->t_flags) && (len > segsiz)) {
16517                                 /*
16518                                  * For prr=off, we need to send only 1 MSS
16519                                  * at a time. We do this because another sack could
16520                                  * be arriving that causes us to send retransmits and
16521                                  * we don't want to be on a long pace due to a larger send
16522                                  * that keeps us from sending out the retransmit.
16523                                  */
16524                                 len = segsiz;
16525                         }
16526                 } else {
16527                         uint32_t outstanding;
16528                         /*
16529                          * We are inside of a Fast recovery episode, this
16530                          * is caused by a SACK or 3 dup acks. At this point
16531                          * we have sent all the retransmissions and we rely
16532                          * on PRR to dictate what we will send in the form of
16533                          * new data.
16534                          */
16535
16536                         outstanding = tp->snd_max - tp->snd_una;
16537                         if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) {
16538                                 if (tp->snd_wnd > outstanding) {
16539                                         len = tp->snd_wnd - outstanding;
16540                                         /* Check to see if we have the data */
16541                                         if ((sb_offset + len) > avail) {
16542                                                 /* It does not all fit */
16543                                                 if (avail > sb_offset)
16544                                                         len = avail - sb_offset;
16545                                                 else
16546                                                         len = 0;
16547                                         }
16548                                 } else {
16549                                         len = 0;
16550                                 }
16551                         } else if (avail > sb_offset) {
16552                                 len = avail - sb_offset;
16553                         } else {
16554                                 len = 0;
16555                         }
16556                         if (len > 0) {
16557                                 if (len > rack->r_ctl.rc_prr_sndcnt) {
16558                                         len = rack->r_ctl.rc_prr_sndcnt;
16559                                 }
16560                                 if (len > 0) {
16561                                         sub_from_prr = 1;
16562                                         counter_u64_add(rack_rtm_prr_newdata, 1);
16563                                 }
16564                         }
16565                         if (len > segsiz) {
16566                                 /*
16567                                  * We should never send more than a MSS when
16568                                  * retransmitting or sending new data in prr
16569                                  * mode unless the override flag is on. Most
16570                                  * likely the PRR algorithm is not going to
16571                                  * let us send a lot as well :-)
16572                                  */
16573                                 if (rack->r_ctl.rc_prr_sendalot == 0) {
16574                                         len = segsiz;
16575                                 }
16576                         } else if (len < segsiz) {
16577                                 /*
16578                                  * Do we send any? The idea here is if the
16579                                  * send empty's the socket buffer we want to
16580                                  * do it. However if not then lets just wait
16581                                  * for our prr_sndcnt to get bigger.
16582                                  */
16583                                 long leftinsb;
16584
16585                                 leftinsb = sbavail(sb) - sb_offset;
16586                                 if (leftinsb > len) {
16587                                         /* This send does not empty the sb */
16588                                         len = 0;
16589                                 }
16590                         }
16591                 }
16592         } else if (!TCPS_HAVEESTABLISHED(tp->t_state)) {
16593                 /*
16594                  * If you have not established
16595                  * and are not doing FAST OPEN
16596                  * no data please.
16597                  */
16598                 if ((sack_rxmit == 0) &&
16599                     (!IS_FASTOPEN(tp->t_flags))){
16600                         len = 0;
16601                         sb_offset = 0;
16602                 }
16603         }
16604         if (prefetch_so_done == 0) {
16605                 kern_prefetch(so, &prefetch_so_done);
16606                 prefetch_so_done = 1;
16607         }
16608         /*
16609          * Lop off SYN bit if it has already been sent.  However, if this is
16610          * SYN-SENT state and if segment contains data and if we don't know
16611          * that foreign host supports TAO, suppress sending segment.
16612          */
16613         if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una) &&
16614             ((sack_rxmit == 0) && (tp->t_rxtshift == 0))) {
16615                 /*
16616                  * When sending additional segments following a TFO SYN|ACK,
16617                  * do not include the SYN bit.
16618                  */
16619                 if (IS_FASTOPEN(tp->t_flags) &&
16620                     (tp->t_state == TCPS_SYN_RECEIVED))
16621                         flags &= ~TH_SYN;
16622         }
16623         /*
16624          * Be careful not to send data and/or FIN on SYN segments. This
16625          * measure is needed to prevent interoperability problems with not
16626          * fully conformant TCP implementations.
16627          */
16628         if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
16629                 len = 0;
16630                 flags &= ~TH_FIN;
16631         }
16632         /*
16633          * On TFO sockets, ensure no data is sent in the following cases:
16634          *
16635          *  - When retransmitting SYN|ACK on a passively-created socket
16636          *
16637          *  - When retransmitting SYN on an actively created socket
16638          *
16639          *  - When sending a zero-length cookie (cookie request) on an
16640          *    actively created socket
16641          *
16642          *  - When the socket is in the CLOSED state (RST is being sent)
16643          */
16644         if (IS_FASTOPEN(tp->t_flags) &&
16645             (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
16646              ((tp->t_state == TCPS_SYN_SENT) &&
16647               (tp->t_tfo_client_cookie_len == 0)) ||
16648              (flags & TH_RST))) {
16649                 sack_rxmit = 0;
16650                 len = 0;
16651         }
16652         /* Without fast-open there should never be data sent on a SYN */
16653         if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags))) {
16654                 tp->snd_nxt = tp->iss;
16655                 len = 0;
16656         }
16657         if ((len > segsiz) && (tcp_dsack_block_exists(tp))) {
16658                 /* We only send 1 MSS if we have a DSACK block */
16659                 add_flag |= RACK_SENT_W_DSACK;
16660                 len = segsiz;
16661         }
16662         orig_len = len;
16663         if (len <= 0) {
16664                 /*
16665                  * If FIN has been sent but not acked, but we haven't been
16666                  * called to retransmit, len will be < 0.  Otherwise, window
16667                  * shrank after we sent into it.  If window shrank to 0,
16668                  * cancel pending retransmit, pull snd_nxt back to (closed)
16669                  * window, and set the persist timer if it isn't already
16670                  * going.  If the window didn't close completely, just wait
16671                  * for an ACK.
16672                  *
16673                  * We also do a general check here to ensure that we will
16674                  * set the persist timer when we have data to send, but a
16675                  * 0-byte window. This makes sure the persist timer is set
16676                  * even if the packet hits one of the "goto send" lines
16677                  * below.
16678                  */
16679                 len = 0;
16680                 if ((tp->snd_wnd == 0) &&
16681                     (TCPS_HAVEESTABLISHED(tp->t_state)) &&
16682                     (tp->snd_una == tp->snd_max) &&
16683                     (sb_offset < (int)sbavail(sb))) {
16684                         rack_enter_persist(tp, rack, cts);
16685                 }
16686         } else if ((rsm == NULL) &&
16687                    (doing_tlp == 0) &&
16688                    (len < pace_max_seg)) {
16689                 /*
16690                  * We are not sending a maximum sized segment for
16691                  * some reason. Should we not send anything (think
16692                  * sws or persists)?
16693                  */
16694                 if ((tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
16695                     (TCPS_HAVEESTABLISHED(tp->t_state)) &&
16696                     (len < minseg) &&
16697                     (len < (int)(sbavail(sb) - sb_offset))) {
16698                         /*
16699                          * Here the rwnd is less than
16700                          * the minimum pacing size, this is not a retransmit,
16701                          * we are established and
16702                          * the send is not the last in the socket buffer
16703                          * we send nothing, and we may enter persists
16704                          * if nothing is outstanding.
16705                          */
16706                         len = 0;
16707                         if (tp->snd_max == tp->snd_una) {
16708                                 /*
16709                                  * Nothing out we can
16710                                  * go into persists.
16711                                  */
16712                                 rack_enter_persist(tp, rack, cts);
16713                         }
16714                      } else if ((cwnd_to_use >= max(minseg, (segsiz * 4))) &&
16715                            (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
16716                            (len < (int)(sbavail(sb) - sb_offset)) &&
16717                            (len < minseg)) {
16718                         /*
16719                          * Here we are not retransmitting, and
16720                          * the cwnd is not so small that we could
16721                          * not send at least a min size (rxt timer
16722                          * not having gone off), We have 2 segments or
16723                          * more already in flight, its not the tail end
16724                          * of the socket buffer  and the cwnd is blocking
16725                          * us from sending out a minimum pacing segment size.
16726                          * Lets not send anything.
16727                          */
16728                         len = 0;
16729                 } else if (((tp->snd_wnd - ctf_outstanding(tp)) <
16730                             min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
16731                            (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
16732                            (len < (int)(sbavail(sb) - sb_offset)) &&
16733                            (TCPS_HAVEESTABLISHED(tp->t_state))) {
16734                         /*
16735                          * Here we have a send window but we have
16736                          * filled it up and we can't send another pacing segment.
16737                          * We also have in flight more than 2 segments
16738                          * and we are not completing the sb i.e. we allow
16739                          * the last bytes of the sb to go out even if
16740                          * its not a full pacing segment.
16741                          */
16742                         len = 0;
16743                 } else if ((rack->r_ctl.crte != NULL) &&
16744                            (tp->snd_wnd >= (pace_max_seg * max(1, rack_hw_rwnd_factor))) &&
16745                            (cwnd_to_use >= (pace_max_seg + (4 * segsiz))) &&
16746                            (ctf_flight_size(tp, rack->r_ctl.rc_sacked) >= (2 * segsiz)) &&
16747                            (len < (int)(sbavail(sb) - sb_offset))) {
16748                         /*
16749                          * Here we are doing hardware pacing, this is not a TLP,
16750                          * we are not sending a pace max segment size, there is rwnd
16751                          * room to send at least N pace_max_seg, the cwnd is greater
16752                          * than or equal to a full pacing segments plus 4 mss and we have 2 or
16753                          * more segments in flight and its not the tail of the socket buffer.
16754                          *
16755                          * We don't want to send instead we need to get more ack's in to
16756                          * allow us to send a full pacing segment. Normally, if we are pacing
16757                          * about the right speed, we should have finished our pacing
16758                          * send as most of the acks have come back if we are at the
16759                          * right rate. This is a bit fuzzy since return path delay
16760                          * can delay the acks, which is why we want to make sure we
16761                          * have cwnd space to have a bit more than a max pace segments in flight.
16762                          *
16763                          * If we have not gotten our acks back we are pacing at too high a
16764                          * rate delaying will not hurt and will bring our GP estimate down by
16765                          * injecting the delay. If we don't do this we will send
16766                          * 2 MSS out in response to the acks being clocked in which
16767                          * defeats the point of hw-pacing (i.e. to help us get
16768                          * larger TSO's out).
16769                          */
16770                         len = 0;
16771
16772                 }
16773
16774         }
16775         /* len will be >= 0 after this point. */
16776         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
16777         rack_sndbuf_autoscale(rack);
16778         /*
16779          * Decide if we can use TCP Segmentation Offloading (if supported by
16780          * hardware).
16781          *
16782          * TSO may only be used if we are in a pure bulk sending state.  The
16783          * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
16784          * options prevent using TSO.  With TSO the TCP header is the same
16785          * (except for the sequence number) for all generated packets.  This
16786          * makes it impossible to transmit any options which vary per
16787          * generated segment or packet.
16788          *
16789          * IPv4 handling has a clear separation of ip options and ip header
16790          * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
16791          * the right thing below to provide length of just ip options and thus
16792          * checking for ipoptlen is enough to decide if ip options are present.
16793          */
16794         ipoptlen = 0;
16795 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
16796         /*
16797          * Pre-calculate here as we save another lookup into the darknesses
16798          * of IPsec that way and can actually decide if TSO is ok.
16799          */
16800 #ifdef INET6
16801         if (isipv6 && IPSEC_ENABLED(ipv6))
16802                 ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb);
16803 #ifdef INET
16804         else
16805 #endif
16806 #endif                          /* INET6 */
16807 #ifdef INET
16808                 if (IPSEC_ENABLED(ipv4))
16809                         ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb);
16810 #endif                          /* INET */
16811 #endif
16812
16813 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
16814         ipoptlen += ipsec_optlen;
16815 #endif
16816         if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > segsiz &&
16817             (tp->t_port == 0) &&
16818             ((tp->t_flags & TF_SIGNATURE) == 0) &&
16819             tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
16820             ipoptlen == 0)
16821                 tso = 1;
16822         {
16823                 uint32_t outstanding;
16824
16825                 outstanding = tp->snd_max - tp->snd_una;
16826                 if (tp->t_flags & TF_SENTFIN) {
16827                         /*
16828                          * If we sent a fin, snd_max is 1 higher than
16829                          * snd_una
16830                          */
16831                         outstanding--;
16832                 }
16833                 if (sack_rxmit) {
16834                         if ((rsm->r_flags & RACK_HAS_FIN) == 0)
16835                                 flags &= ~TH_FIN;
16836                 } else {
16837                         if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
16838                                    sbused(sb)))
16839                                 flags &= ~TH_FIN;
16840                 }
16841         }
16842         recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
16843             (long)TCP_MAXWIN << tp->rcv_scale);
16844
16845         /*
16846          * Sender silly window avoidance.   We transmit under the following
16847          * conditions when len is non-zero:
16848          *
16849          * - We have a full segment (or more with TSO) - This is the last
16850          * buffer in a write()/send() and we are either idle or running
16851          * NODELAY - we've timed out (e.g. persist timer) - we have more
16852          * then 1/2 the maximum send window's worth of data (receiver may be
16853          * limited the window size) - we need to retransmit
16854          */
16855         if (len) {
16856                 if (len >= segsiz) {
16857                         goto send;
16858                 }
16859                 /*
16860                  * NOTE! on localhost connections an 'ack' from the remote
16861                  * end may occur synchronously with the output and cause us
16862                  * to flush a buffer queued with moretocome.  XXX
16863                  *
16864                  */
16865                 if (!(tp->t_flags & TF_MORETOCOME) &&   /* normal case */
16866                     (idle || (tp->t_flags & TF_NODELAY)) &&
16867                     ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
16868                     (tp->t_flags & TF_NOPUSH) == 0) {
16869                         pass = 2;
16870                         goto send;
16871                 }
16872                 if ((tp->snd_una == tp->snd_max) && len) {      /* Nothing outstanding */
16873                         pass = 22;
16874                         goto send;
16875                 }
16876                 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) {
16877                         pass = 4;
16878                         goto send;
16879                 }
16880                 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { /* retransmit case */
16881                         pass = 5;
16882                         goto send;
16883                 }
16884                 if (sack_rxmit) {
16885                         pass = 6;
16886                         goto send;
16887                 }
16888                 if (((tp->snd_wnd - ctf_outstanding(tp)) < segsiz) &&
16889                     (ctf_outstanding(tp) < (segsiz * 2))) {
16890                         /*
16891                          * We have less than two MSS outstanding (delayed ack)
16892                          * and our rwnd will not let us send a full sized
16893                          * MSS. Lets go ahead and let this small segment
16894                          * out because we want to try to have at least two
16895                          * packets inflight to not be caught by delayed ack.
16896                          */
16897                         pass = 12;
16898                         goto send;
16899                 }
16900         }
16901         /*
16902          * Sending of standalone window updates.
16903          *
16904          * Window updates are important when we close our window due to a
16905          * full socket buffer and are opening it again after the application
16906          * reads data from it.  Once the window has opened again and the
16907          * remote end starts to send again the ACK clock takes over and
16908          * provides the most current window information.
16909          *
16910          * We must avoid the silly window syndrome whereas every read from
16911          * the receive buffer, no matter how small, causes a window update
16912          * to be sent.  We also should avoid sending a flurry of window
16913          * updates when the socket buffer had queued a lot of data and the
16914          * application is doing small reads.
16915          *
16916          * Prevent a flurry of pointless window updates by only sending an
16917          * update when we can increase the advertized window by more than
16918          * 1/4th of the socket buffer capacity.  When the buffer is getting
16919          * full or is very small be more aggressive and send an update
16920          * whenever we can increase by two mss sized segments. In all other
16921          * situations the ACK's to new incoming data will carry further
16922          * window increases.
16923          *
16924          * Don't send an independent window update if a delayed ACK is
16925          * pending (it will get piggy-backed on it) or the remote side
16926          * already has done a half-close and won't send more data.  Skip
16927          * this if the connection is in T/TCP half-open state.
16928          */
16929         if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
16930             !(tp->t_flags & TF_DELACK) &&
16931             !TCPS_HAVERCVDFIN(tp->t_state)) {
16932                 /*
16933                  * "adv" is the amount we could increase the window, taking
16934                  * into account that we are limited by TCP_MAXWIN <<
16935                  * tp->rcv_scale.
16936                  */
16937                 int32_t adv;
16938                 int oldwin;
16939
16940                 adv = recwin;
16941                 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
16942                         oldwin = (tp->rcv_adv - tp->rcv_nxt);
16943                         if (adv > oldwin)
16944                             adv -= oldwin;
16945                         else {
16946                                 /* We can't increase the window */
16947                                 adv = 0;
16948                         }
16949                 } else
16950                         oldwin = 0;
16951
16952                 /*
16953                  * If the new window size ends up being the same as or less
16954                  * than the old size when it is scaled, then don't force
16955                  * a window update.
16956                  */
16957                 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
16958                         goto dontupdate;
16959
16960                 if (adv >= (int32_t)(2 * segsiz) &&
16961                     (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
16962                      recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) ||
16963                      so->so_rcv.sb_hiwat <= 8 * segsiz)) {
16964                         pass = 7;
16965                         goto send;
16966                 }
16967                 if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) {
16968                         pass = 23;
16969                         goto send;
16970                 }
16971         }
16972 dontupdate:
16973
16974         /*
16975          * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
16976          * is also a catch-all for the retransmit timer timeout case.
16977          */
16978         if (tp->t_flags & TF_ACKNOW) {
16979                 pass = 8;
16980                 goto send;
16981         }
16982         if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) {
16983                 pass = 9;
16984                 goto send;
16985         }
16986         /*
16987          * If our state indicates that FIN should be sent and we have not
16988          * yet done so, then we need to send.
16989          */
16990         if ((flags & TH_FIN) &&
16991             (tp->snd_nxt == tp->snd_una)) {
16992                 pass = 11;
16993                 goto send;
16994         }
16995         /*
16996          * No reason to send a segment, just return.
16997          */
16998 just_return:
16999         SOCKBUF_UNLOCK(sb);
17000 just_return_nolock:
17001         {
17002                 int app_limited = CTF_JR_SENT_DATA;
17003
17004                 if (tot_len_this_send > 0) {
17005                         /* Make sure snd_nxt is up to max */
17006                         rack->r_ctl.fsb.recwin = recwin;
17007                         slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, NULL, segsiz);
17008                         if ((error == 0) &&
17009                             rack_use_rfo &&
17010                             ((flags & (TH_SYN|TH_FIN)) == 0) &&
17011                             (ipoptlen == 0) &&
17012                             (tp->snd_nxt == tp->snd_max) &&
17013                             (tp->rcv_numsacks == 0) &&
17014                             rack->r_fsb_inited &&
17015                             TCPS_HAVEESTABLISHED(tp->t_state) &&
17016                             (rack->r_must_retran == 0) &&
17017                             ((tp->t_flags & TF_NEEDFIN) == 0) &&
17018                             (len > 0) && (orig_len > 0) &&
17019                             (orig_len > len) &&
17020                             ((orig_len - len) >= segsiz) &&
17021                             ((optlen == 0) ||
17022                              ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
17023                                 /* We can send at least one more MSS using our fsb */
17024
17025                                 rack->r_fast_output = 1;
17026                                 rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
17027                                 rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
17028                                 rack->r_ctl.fsb.tcp_flags = flags;
17029                                 rack->r_ctl.fsb.left_to_send = orig_len - len;
17030                                 KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
17031                                         ("rack:%p left_to_send:%u sbavail:%u out:%u",
17032                                         rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
17033                                          (tp->snd_max - tp->snd_una)));
17034                                 if (rack->r_ctl.fsb.left_to_send < segsiz)
17035                                         rack->r_fast_output = 0;
17036                                 else {
17037                                         if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
17038                                                 rack->r_ctl.fsb.rfo_apply_push = 1;
17039                                         else
17040                                                 rack->r_ctl.fsb.rfo_apply_push = 0;
17041                                 }
17042                         } else
17043                                 rack->r_fast_output = 0;
17044
17045
17046                         rack_log_fsb(rack, tp, so, flags,
17047                                      ipoptlen, orig_len, len, 0,
17048                                      1, optlen, __LINE__, 1);
17049                         if (SEQ_GT(tp->snd_max, tp->snd_nxt))
17050                                 tp->snd_nxt = tp->snd_max;
17051                 } else {
17052                         int end_window = 0;
17053                         uint32_t seq = tp->gput_ack;
17054
17055                         rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
17056                         if (rsm) {
17057                                 /*
17058                                  * Mark the last sent that we just-returned (hinting
17059                                  * that delayed ack may play a role in any rtt measurement).
17060                                  */
17061                                 rsm->r_just_ret = 1;
17062                         }
17063                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1);
17064                         rack->r_ctl.rc_agg_delayed = 0;
17065                         rack->r_early = 0;
17066                         rack->r_late = 0;
17067                         rack->r_ctl.rc_agg_early = 0;
17068                         if ((ctf_outstanding(tp) +
17069                              min(max(segsiz, (rack->r_ctl.rc_high_rwnd/2)),
17070                                  minseg)) >= tp->snd_wnd) {
17071                                 /* We are limited by the rwnd */
17072                                 app_limited = CTF_JR_RWND_LIMITED;
17073                                 if (IN_FASTRECOVERY(tp->t_flags))
17074                                     rack->r_ctl.rc_prr_sndcnt = 0;
17075                         } else if (ctf_outstanding(tp) >= sbavail(sb)) {
17076                                 /* We are limited by whats available -- app limited */
17077                                 app_limited = CTF_JR_APP_LIMITED;
17078                                 if (IN_FASTRECOVERY(tp->t_flags))
17079                                     rack->r_ctl.rc_prr_sndcnt = 0;
17080                         } else if ((idle == 0) &&
17081                                    ((tp->t_flags & TF_NODELAY) == 0) &&
17082                                    ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
17083                                    (len < segsiz)) {
17084                                 /*
17085                                  * No delay is not on and the
17086                                  * user is sending less than 1MSS. This
17087                                  * brings out SWS avoidance so we
17088                                  * don't send. Another app-limited case.
17089                                  */
17090                                 app_limited = CTF_JR_APP_LIMITED;
17091                         } else if (tp->t_flags & TF_NOPUSH) {
17092                                 /*
17093                                  * The user has requested no push of
17094                                  * the last segment and we are
17095                                  * at the last segment. Another app
17096                                  * limited case.
17097                                  */
17098                                 app_limited = CTF_JR_APP_LIMITED;
17099                         } else if ((ctf_outstanding(tp) + minseg) > cwnd_to_use) {
17100                                 /* Its the cwnd */
17101                                 app_limited = CTF_JR_CWND_LIMITED;
17102                         } else if (IN_FASTRECOVERY(tp->t_flags) &&
17103                                    (rack->rack_no_prr == 0) &&
17104                                    (rack->r_ctl.rc_prr_sndcnt < segsiz)) {
17105                                 app_limited = CTF_JR_PRR;
17106                         } else {
17107                                 /* Now why here are we not sending? */
17108 #ifdef NOW
17109 #ifdef INVARIANTS
17110                                 panic("rack:%p hit JR_ASSESSING case cwnd_to_use:%u?", rack, cwnd_to_use);
17111 #endif
17112 #endif
17113                                 app_limited = CTF_JR_ASSESSING;
17114                         }
17115                         /*
17116                          * App limited in some fashion, for our pacing GP
17117                          * measurements we don't want any gap (even cwnd).
17118                          * Close  down the measurement window.
17119                          */
17120                         if (rack_cwnd_block_ends_measure &&
17121                             ((app_limited == CTF_JR_CWND_LIMITED) ||
17122                              (app_limited == CTF_JR_PRR))) {
17123                                 /*
17124                                  * The reason we are not sending is
17125                                  * the cwnd (or prr). We have been configured
17126                                  * to end the measurement window in
17127                                  * this case.
17128                                  */
17129                                 end_window = 1;
17130                         } else if (rack_rwnd_block_ends_measure &&
17131                                    (app_limited == CTF_JR_RWND_LIMITED)) {
17132                                 /*
17133                                  * We are rwnd limited and have been
17134                                  * configured to end the measurement
17135                                  * window in this case.
17136                                  */
17137                                 end_window = 1;
17138                         } else if (app_limited == CTF_JR_APP_LIMITED) {
17139                                 /*
17140                                  * A true application limited period, we have
17141                                  * ran out of data.
17142                                  */
17143                                 end_window = 1;
17144                         } else if (app_limited == CTF_JR_ASSESSING) {
17145                                 /*
17146                                  * In the assessing case we hit the end of
17147                                  * the if/else and had no known reason
17148                                  * This will panic us under invariants..
17149                                  *
17150                                  * If we get this out in logs we need to
17151                                  * investagate which reason we missed.
17152                                  */
17153                                 end_window = 1;
17154                         }
17155                         if (end_window) {
17156                                 uint8_t log = 0;
17157
17158                                 if ((tp->t_flags & TF_GPUTINPROG) &&
17159                                     SEQ_GT(tp->gput_ack, tp->snd_max)) {
17160                                         /* Mark the last packet has app limited */
17161                                         tp->gput_ack = tp->snd_max;
17162                                         log = 1;
17163                                 }
17164                                 rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
17165                                 if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
17166                                         if (rack->r_ctl.rc_app_limited_cnt == 0)
17167                                                 rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
17168                                         else {
17169                                                 /*
17170                                                  * Go out to the end app limited and mark
17171                                                  * this new one as next and move the end_appl up
17172                                                  * to this guy.
17173                                                  */
17174                                                 if (rack->r_ctl.rc_end_appl)
17175                                                         rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
17176                                                 rack->r_ctl.rc_end_appl = rsm;
17177                                         }
17178                                         rsm->r_flags |= RACK_APP_LIMITED;
17179                                         rack->r_ctl.rc_app_limited_cnt++;
17180                                 }
17181                                 if (log)
17182                                         rack_log_pacing_delay_calc(rack,
17183                                                                    rack->r_ctl.rc_app_limited_cnt, seq,
17184                                                                    tp->gput_ack, 0, 0, 4, __LINE__, NULL);
17185                         }
17186                 }
17187                 if (slot) {
17188                         /* set the rack tcb into the slot N */
17189                         counter_u64_add(rack_paced_segments, 1);
17190                 } else if (tot_len_this_send) {
17191                         counter_u64_add(rack_unpaced_segments, 1);
17192                 }
17193                 /* Check if we need to go into persists or not */
17194                 if ((tp->snd_max == tp->snd_una) &&
17195                     TCPS_HAVEESTABLISHED(tp->t_state) &&
17196                     sbavail(sb) &&
17197                     (sbavail(sb) > tp->snd_wnd) &&
17198                     (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg))) {
17199                         /* Yes lets make sure to move to persist before timer-start */
17200                         rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
17201                 }
17202                 rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, sup_rack);
17203                 rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling, app_limited, cwnd_to_use);
17204         }
17205 #ifdef NETFLIX_SHARED_CWND
17206         if ((sbavail(sb) == 0) &&
17207             rack->r_ctl.rc_scw) {
17208                 tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
17209                 rack->rack_scwnd_is_idle = 1;
17210         }
17211 #endif
17212 #ifdef TCP_ACCOUNTING
17213         if (tot_len_this_send > 0) {
17214                 crtsc = get_cyclecount();
17215                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17216                         tp->tcp_cnt_counters[SND_OUT_DATA]++;
17217                 }
17218                 counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], 1);
17219                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17220                         tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
17221                 }
17222                 counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
17223                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17224                         tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) / segsiz);
17225                 }
17226                 counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len_this_send + segsiz - 1) / segsiz));
17227         } else {
17228                 crtsc = get_cyclecount();
17229                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17230                         tp->tcp_cnt_counters[SND_LIMITED]++;
17231                 }
17232                 counter_u64_add(tcp_cnt_counters[SND_LIMITED], 1);
17233                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17234                         tp->tcp_proc_time[SND_LIMITED] += (crtsc - ts_val);
17235                 }
17236                 counter_u64_add(tcp_proc_time[SND_LIMITED], (crtsc - ts_val));
17237         }
17238         sched_unpin();
17239 #endif
17240         return (0);
17241
17242 send:
17243         if (rsm || sack_rxmit)
17244                 counter_u64_add(rack_nfto_resend, 1);
17245         else
17246                 counter_u64_add(rack_non_fto_send, 1);
17247         if ((flags & TH_FIN) &&
17248             sbavail(sb)) {
17249                 /*
17250                  * We do not transmit a FIN
17251                  * with data outstanding. We
17252                  * need to make it so all data
17253                  * is acked first.
17254                  */
17255                 flags &= ~TH_FIN;
17256         }
17257         /* Enforce stack imposed max seg size if we have one */
17258         if (rack->r_ctl.rc_pace_max_segs &&
17259             (len > rack->r_ctl.rc_pace_max_segs)) {
17260                 mark = 1;
17261                 len = rack->r_ctl.rc_pace_max_segs;
17262         }
17263         SOCKBUF_LOCK_ASSERT(sb);
17264         if (len > 0) {
17265                 if (len >= segsiz)
17266                         tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
17267                 else
17268                         tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
17269         }
17270         /*
17271          * Before ESTABLISHED, force sending of initial options unless TCP
17272          * set not to do any options. NOTE: we assume that the IP/TCP header
17273          * plus TCP options always fit in a single mbuf, leaving room for a
17274          * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr)
17275          * + optlen <= MCLBYTES
17276          */
17277         optlen = 0;
17278 #ifdef INET6
17279         if (isipv6)
17280                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
17281         else
17282 #endif
17283                 hdrlen = sizeof(struct tcpiphdr);
17284
17285         /*
17286          * Compute options for segment. We only have to care about SYN and
17287          * established connection segments.  Options for SYN-ACK segments
17288          * are handled in TCP syncache.
17289          */
17290         to.to_flags = 0;
17291         if ((tp->t_flags & TF_NOOPT) == 0) {
17292                 /* Maximum segment size. */
17293                 if (flags & TH_SYN) {
17294                         tp->snd_nxt = tp->iss;
17295                         to.to_mss = tcp_mssopt(&inp->inp_inc);
17296                         if (tp->t_port)
17297                                 to.to_mss -= V_tcp_udp_tunneling_overhead;
17298                         to.to_flags |= TOF_MSS;
17299
17300                         /*
17301                          * On SYN or SYN|ACK transmits on TFO connections,
17302                          * only include the TFO option if it is not a
17303                          * retransmit, as the presence of the TFO option may
17304                          * have caused the original SYN or SYN|ACK to have
17305                          * been dropped by a middlebox.
17306                          */
17307                         if (IS_FASTOPEN(tp->t_flags) &&
17308                             (tp->t_rxtshift == 0)) {
17309                                 if (tp->t_state == TCPS_SYN_RECEIVED) {
17310                                         to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
17311                                         to.to_tfo_cookie =
17312                                                 (u_int8_t *)&tp->t_tfo_cookie.server;
17313                                         to.to_flags |= TOF_FASTOPEN;
17314                                         wanted_cookie = 1;
17315                                 } else if (tp->t_state == TCPS_SYN_SENT) {
17316                                         to.to_tfo_len =
17317                                                 tp->t_tfo_client_cookie_len;
17318                                         to.to_tfo_cookie =
17319                                                 tp->t_tfo_cookie.client;
17320                                         to.to_flags |= TOF_FASTOPEN;
17321                                         wanted_cookie = 1;
17322                                         /*
17323                                          * If we wind up having more data to
17324                                          * send with the SYN than can fit in
17325                                          * one segment, don't send any more
17326                                          * until the SYN|ACK comes back from
17327                                          * the other end.
17328                                          */
17329                                         sendalot = 0;
17330                                 }
17331                         }
17332                 }
17333                 /* Window scaling. */
17334                 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
17335                         to.to_wscale = tp->request_r_scale;
17336                         to.to_flags |= TOF_SCALE;
17337                 }
17338                 /* Timestamps. */
17339                 if ((tp->t_flags & TF_RCVD_TSTMP) ||
17340                     ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
17341                         to.to_tsval = ms_cts + tp->ts_offset;
17342                         to.to_tsecr = tp->ts_recent;
17343                         to.to_flags |= TOF_TS;
17344                 }
17345                 /* Set receive buffer autosizing timestamp. */
17346                 if (tp->rfbuf_ts == 0 &&
17347                     (so->so_rcv.sb_flags & SB_AUTOSIZE))
17348                         tp->rfbuf_ts = tcp_ts_getticks();
17349                 /* Selective ACK's. */
17350                 if (tp->t_flags & TF_SACK_PERMIT) {
17351                         if (flags & TH_SYN)
17352                                 to.to_flags |= TOF_SACKPERM;
17353                         else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
17354                                  tp->rcv_numsacks > 0) {
17355                                 to.to_flags |= TOF_SACK;
17356                                 to.to_nsacks = tp->rcv_numsacks;
17357                                 to.to_sacks = (u_char *)tp->sackblks;
17358                         }
17359                 }
17360 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
17361                 /* TCP-MD5 (RFC2385). */
17362                 if (tp->t_flags & TF_SIGNATURE)
17363                         to.to_flags |= TOF_SIGNATURE;
17364 #endif                          /* TCP_SIGNATURE */
17365
17366                 /* Processing the options. */
17367                 hdrlen += optlen = tcp_addoptions(&to, opt);
17368                 /*
17369                  * If we wanted a TFO option to be added, but it was unable
17370                  * to fit, ensure no data is sent.
17371                  */
17372                 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
17373                     !(to.to_flags & TOF_FASTOPEN))
17374                         len = 0;
17375         }
17376         if (tp->t_port) {
17377                 if (V_tcp_udp_tunneling_port == 0) {
17378                         /* The port was removed?? */
17379                         SOCKBUF_UNLOCK(&so->so_snd);
17380 #ifdef TCP_ACCOUNTING
17381                         crtsc = get_cyclecount();
17382                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17383                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
17384                         }
17385                         counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
17386                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17387                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
17388                         }
17389                         counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
17390                         sched_unpin();
17391 #endif
17392                         return (EHOSTUNREACH);
17393                 }
17394                 hdrlen += sizeof(struct udphdr);
17395         }
17396 #ifdef INET6
17397         if (isipv6)
17398                 ipoptlen = ip6_optlen(tp->t_inpcb);
17399         else
17400 #endif
17401                 if (tp->t_inpcb->inp_options)
17402                         ipoptlen = tp->t_inpcb->inp_options->m_len -
17403                                 offsetof(struct ipoption, ipopt_list);
17404                 else
17405                         ipoptlen = 0;
17406 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
17407         ipoptlen += ipsec_optlen;
17408 #endif
17409
17410         /*
17411          * Adjust data length if insertion of options will bump the packet
17412          * length beyond the t_maxseg length. Clear the FIN bit because we
17413          * cut off the tail of the segment.
17414          */
17415         if (len + optlen + ipoptlen > tp->t_maxseg) {
17416                 if (tso) {
17417                         uint32_t if_hw_tsomax;
17418                         uint32_t moff;
17419                         int32_t max_len;
17420
17421                         /* extract TSO information */
17422                         if_hw_tsomax = tp->t_tsomax;
17423                         if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
17424                         if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
17425                         KASSERT(ipoptlen == 0,
17426                                 ("%s: TSO can't do IP options", __func__));
17427
17428                         /*
17429                          * Check if we should limit by maximum payload
17430                          * length:
17431                          */
17432                         if (if_hw_tsomax != 0) {
17433                                 /* compute maximum TSO length */
17434                                 max_len = (if_hw_tsomax - hdrlen -
17435                                            max_linkhdr);
17436                                 if (max_len <= 0) {
17437                                         len = 0;
17438                                 } else if (len > max_len) {
17439                                         sendalot = 1;
17440                                         len = max_len;
17441                                         mark = 2;
17442                                 }
17443                         }
17444                         /*
17445                          * Prevent the last segment from being fractional
17446                          * unless the send sockbuf can be emptied:
17447                          */
17448                         max_len = (tp->t_maxseg - optlen);
17449                         if ((sb_offset + len) < sbavail(sb)) {
17450                                 moff = len % (u_int)max_len;
17451                                 if (moff != 0) {
17452                                         mark = 3;
17453                                         len -= moff;
17454                                 }
17455                         }
17456                         /*
17457                          * In case there are too many small fragments don't
17458                          * use TSO:
17459                          */
17460                         if (len <= segsiz) {
17461                                 mark = 4;
17462                                 tso = 0;
17463                         }
17464                         /*
17465                          * Send the FIN in a separate segment after the bulk
17466                          * sending is done. We don't trust the TSO
17467                          * implementations to clear the FIN flag on all but
17468                          * the last segment.
17469                          */
17470                         if (tp->t_flags & TF_NEEDFIN) {
17471                                 sendalot = 4;
17472                         }
17473                 } else {
17474                         mark = 5;
17475                         if (optlen + ipoptlen >= tp->t_maxseg) {
17476                                 /*
17477                                  * Since we don't have enough space to put
17478                                  * the IP header chain and the TCP header in
17479                                  * one packet as required by RFC 7112, don't
17480                                  * send it. Also ensure that at least one
17481                                  * byte of the payload can be put into the
17482                                  * TCP segment.
17483                                  */
17484                                 SOCKBUF_UNLOCK(&so->so_snd);
17485                                 error = EMSGSIZE;
17486                                 sack_rxmit = 0;
17487                                 goto out;
17488                         }
17489                         len = tp->t_maxseg - optlen - ipoptlen;
17490                         sendalot = 5;
17491                 }
17492         } else {
17493                 tso = 0;
17494                 mark = 6;
17495         }
17496         KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
17497                 ("%s: len > IP_MAXPACKET", __func__));
17498 #ifdef DIAGNOSTIC
17499 #ifdef INET6
17500         if (max_linkhdr + hdrlen > MCLBYTES)
17501 #else
17502                 if (max_linkhdr + hdrlen > MHLEN)
17503 #endif
17504                         panic("tcphdr too big");
17505 #endif
17506
17507         /*
17508          * This KASSERT is here to catch edge cases at a well defined place.
17509          * Before, those had triggered (random) panic conditions further
17510          * down.
17511          */
17512         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
17513         if ((len == 0) &&
17514             (flags & TH_FIN) &&
17515             (sbused(sb))) {
17516                 /*
17517                  * We have outstanding data, don't send a fin by itself!.
17518                  */
17519                 goto just_return;
17520         }
17521         /*
17522          * Grab a header mbuf, attaching a copy of data to be transmitted,
17523          * and initialize the header from the template for sends on this
17524          * connection.
17525          */
17526         hw_tls = (sb->sb_flags & SB_TLS_IFNET) != 0;
17527         if (len) {
17528                 uint32_t max_val;
17529                 uint32_t moff;
17530
17531                 if (rack->r_ctl.rc_pace_max_segs)
17532                         max_val = rack->r_ctl.rc_pace_max_segs;
17533                 else if (rack->rc_user_set_max_segs)
17534                         max_val = rack->rc_user_set_max_segs * segsiz;
17535                 else
17536                         max_val = len;
17537                 /*
17538                  * We allow a limit on sending with hptsi.
17539                  */
17540                 if (len > max_val) {
17541                         mark = 7;
17542                         len = max_val;
17543                 }
17544 #ifdef INET6
17545                 if (MHLEN < hdrlen + max_linkhdr)
17546                         m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
17547                 else
17548 #endif
17549                         m = m_gethdr(M_NOWAIT, MT_DATA);
17550
17551                 if (m == NULL) {
17552                         SOCKBUF_UNLOCK(sb);
17553                         error = ENOBUFS;
17554                         sack_rxmit = 0;
17555                         goto out;
17556                 }
17557                 m->m_data += max_linkhdr;
17558                 m->m_len = hdrlen;
17559
17560                 /*
17561                  * Start the m_copy functions from the closest mbuf to the
17562                  * sb_offset in the socket buffer chain.
17563                  */
17564                 mb = sbsndptr_noadv(sb, sb_offset, &moff);
17565                 s_mb = mb;
17566                 s_moff = moff;
17567                 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
17568                         m_copydata(mb, moff, (int)len,
17569                                    mtod(m, caddr_t)+hdrlen);
17570                         if (SEQ_LT(tp->snd_nxt, tp->snd_max))
17571                                 sbsndptr_adv(sb, mb, len);
17572                         m->m_len += len;
17573                 } else {
17574                         struct sockbuf *msb;
17575
17576                         if (SEQ_LT(tp->snd_nxt, tp->snd_max))
17577                                 msb = NULL;
17578                         else
17579                                 msb = sb;
17580                         m->m_next = tcp_m_copym(
17581                                 mb, moff, &len,
17582                                 if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb,
17583                                 ((rsm == NULL) ? hw_tls : 0)
17584 #ifdef NETFLIX_COPY_ARGS
17585                                 , &filled_all
17586 #endif
17587                                 );
17588                         if (len <= (tp->t_maxseg - optlen)) {
17589                                 /*
17590                                  * Must have ran out of mbufs for the copy
17591                                  * shorten it to no longer need tso. Lets
17592                                  * not put on sendalot since we are low on
17593                                  * mbufs.
17594                                  */
17595                                 tso = 0;
17596                         }
17597                         if (m->m_next == NULL) {
17598                                 SOCKBUF_UNLOCK(sb);
17599                                 (void)m_free(m);
17600                                 error = ENOBUFS;
17601                                 sack_rxmit = 0;
17602                                 goto out;
17603                         }
17604                 }
17605                 if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
17606                         if (rsm && (rsm->r_flags & RACK_TLP)) {
17607                                 /*
17608                                  * TLP should not count in retran count, but
17609                                  * in its own bin
17610                                  */
17611                                 counter_u64_add(rack_tlp_retran, 1);
17612                                 counter_u64_add(rack_tlp_retran_bytes, len);
17613                         } else {
17614                                 tp->t_sndrexmitpack++;
17615                                 KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
17616                                 KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
17617                         }
17618 #ifdef STATS
17619                         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
17620                                                  len);
17621 #endif
17622                 } else {
17623                         KMOD_TCPSTAT_INC(tcps_sndpack);
17624                         KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
17625 #ifdef STATS
17626                         stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
17627                                                  len);
17628 #endif
17629                 }
17630                 /*
17631                  * If we're sending everything we've got, set PUSH. (This
17632                  * will keep happy those implementations which only give
17633                  * data to the user when a buffer fills or a PUSH comes in.)
17634                  */
17635                 if (sb_offset + len == sbused(sb) &&
17636                     sbused(sb) &&
17637                     !(flags & TH_SYN)) {
17638                         flags |= TH_PUSH;
17639                         add_flag |= RACK_HAD_PUSH;
17640                 }
17641
17642                 SOCKBUF_UNLOCK(sb);
17643         } else {
17644                 SOCKBUF_UNLOCK(sb);
17645                 if (tp->t_flags & TF_ACKNOW)
17646                         KMOD_TCPSTAT_INC(tcps_sndacks);
17647                 else if (flags & (TH_SYN | TH_FIN | TH_RST))
17648                         KMOD_TCPSTAT_INC(tcps_sndctrl);
17649                 else
17650                         KMOD_TCPSTAT_INC(tcps_sndwinup);
17651
17652                 m = m_gethdr(M_NOWAIT, MT_DATA);
17653                 if (m == NULL) {
17654                         error = ENOBUFS;
17655                         sack_rxmit = 0;
17656                         goto out;
17657                 }
17658 #ifdef INET6
17659                 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
17660                     MHLEN >= hdrlen) {
17661                         M_ALIGN(m, hdrlen);
17662                 } else
17663 #endif
17664                         m->m_data += max_linkhdr;
17665                 m->m_len = hdrlen;
17666         }
17667         SOCKBUF_UNLOCK_ASSERT(sb);
17668         m->m_pkthdr.rcvif = (struct ifnet *)0;
17669 #ifdef MAC
17670         mac_inpcb_create_mbuf(inp, m);
17671 #endif
17672         if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) &&  rack->r_fsb_inited) {
17673 #ifdef INET6
17674                 if (isipv6)
17675                         ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
17676                 else
17677 #endif                          /* INET6 */
17678                         ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
17679                 th = rack->r_ctl.fsb.th;
17680                 udp = rack->r_ctl.fsb.udp;
17681                 if (udp) {
17682 #ifdef INET6
17683                         if (isipv6)
17684                                 ulen = hdrlen + len - sizeof(struct ip6_hdr);
17685                         else
17686 #endif                          /* INET6 */
17687                                 ulen = hdrlen + len - sizeof(struct ip);
17688                         udp->uh_ulen = htons(ulen);
17689                 }
17690         } else {
17691 #ifdef INET6
17692                 if (isipv6) {
17693                         ip6 = mtod(m, struct ip6_hdr *);
17694                         if (tp->t_port) {
17695                                 udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
17696                                 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
17697                                 udp->uh_dport = tp->t_port;
17698                                 ulen = hdrlen + len - sizeof(struct ip6_hdr);
17699                                 udp->uh_ulen = htons(ulen);
17700                                 th = (struct tcphdr *)(udp + 1);
17701                         } else
17702                                 th = (struct tcphdr *)(ip6 + 1);
17703                         tcpip_fillheaders(inp, tp->t_port, ip6, th);
17704                 } else
17705 #endif                          /* INET6 */
17706                 {
17707                         ip = mtod(m, struct ip *);
17708 #ifdef TCPDEBUG
17709                         ipov = (struct ipovly *)ip;
17710 #endif
17711                         if (tp->t_port) {
17712                                 udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
17713                                 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
17714                                 udp->uh_dport = tp->t_port;
17715                                 ulen = hdrlen + len - sizeof(struct ip);
17716                                 udp->uh_ulen = htons(ulen);
17717                                 th = (struct tcphdr *)(udp + 1);
17718                         } else
17719                                 th = (struct tcphdr *)(ip + 1);
17720                         tcpip_fillheaders(inp, tp->t_port, ip, th);
17721                 }
17722         }
17723         /*
17724          * Fill in fields, remembering maximum advertised window for use in
17725          * delaying messages about window sizes. If resending a FIN, be sure
17726          * not to use a new sequence number.
17727          */
17728         if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
17729             tp->snd_nxt == tp->snd_max)
17730                 tp->snd_nxt--;
17731         /*
17732          * If we are starting a connection, send ECN setup SYN packet. If we
17733          * are on a retransmit, we may resend those bits a number of times
17734          * as per RFC 3168.
17735          */
17736         if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) {
17737                 if (tp->t_rxtshift >= 1) {
17738                         if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
17739                                 flags |= TH_ECE | TH_CWR;
17740                 } else
17741                         flags |= TH_ECE | TH_CWR;
17742         }
17743         /* Handle parallel SYN for ECN */
17744         if ((tp->t_state == TCPS_SYN_RECEIVED) &&
17745             (tp->t_flags2 & TF2_ECN_SND_ECE)) {
17746                 flags |= TH_ECE;
17747                 tp->t_flags2 &= ~TF2_ECN_SND_ECE;
17748         }
17749         if (TCPS_HAVEESTABLISHED(tp->t_state) &&
17750             (tp->t_flags2 & TF2_ECN_PERMIT)) {
17751                 /*
17752                  * If the peer has ECN, mark data packets with ECN capable
17753                  * transmission (ECT). Ignore pure ack packets,
17754                  * retransmissions.
17755                  */
17756                 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
17757                     (sack_rxmit == 0)) {
17758 #ifdef INET6
17759                         if (isipv6)
17760                                 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
17761                         else
17762 #endif
17763                                 ip->ip_tos |= IPTOS_ECN_ECT0;
17764                         KMOD_TCPSTAT_INC(tcps_ecn_ect0);
17765                         /*
17766                          * Reply with proper ECN notifications.
17767                          * Only set CWR on new data segments.
17768                          */
17769                         if (tp->t_flags2 & TF2_ECN_SND_CWR) {
17770                                 flags |= TH_CWR;
17771                                 tp->t_flags2 &= ~TF2_ECN_SND_CWR;
17772                         }
17773                 }
17774                 if (tp->t_flags2 & TF2_ECN_SND_ECE)
17775                         flags |= TH_ECE;
17776         }
17777         /*
17778          * If we are doing retransmissions, then snd_nxt will not reflect
17779          * the first unsent octet.  For ACK only packets, we do not want the
17780          * sequence number of the retransmitted packet, we want the sequence
17781          * number of the next unsent octet.  So, if there is no data (and no
17782          * SYN or FIN), use snd_max instead of snd_nxt when filling in
17783          * ti_seq.  But if we are in persist state, snd_max might reflect
17784          * one byte beyond the right edge of the window, so use snd_nxt in
17785          * that case, since we know we aren't doing a retransmission.
17786          * (retransmit and persist are mutually exclusive...)
17787          */
17788         if (sack_rxmit == 0) {
17789                 if (len || (flags & (TH_SYN | TH_FIN))) {
17790                         th->th_seq = htonl(tp->snd_nxt);
17791                         rack_seq = tp->snd_nxt;
17792                 } else {
17793                         th->th_seq = htonl(tp->snd_max);
17794                         rack_seq = tp->snd_max;
17795                 }
17796         } else {
17797                 th->th_seq = htonl(rsm->r_start);
17798                 rack_seq = rsm->r_start;
17799         }
17800         th->th_ack = htonl(tp->rcv_nxt);
17801         th->th_flags = flags;
17802         /*
17803          * Calculate receive window.  Don't shrink window, but avoid silly
17804          * window syndrome.
17805          * If a RST segment is sent, advertise a window of zero.
17806          */
17807         if (flags & TH_RST) {
17808                 recwin = 0;
17809         } else {
17810                 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
17811                     recwin < (long)segsiz) {
17812                         recwin = 0;
17813                 }
17814                 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
17815                     recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
17816                         recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
17817         }
17818
17819         /*
17820          * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
17821          * <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK> case is
17822          * handled in syncache.
17823          */
17824         if (flags & TH_SYN)
17825                 th->th_win = htons((u_short)
17826                                    (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
17827         else {
17828                 /* Avoid shrinking window with window scaling. */
17829                 recwin = roundup2(recwin, 1 << tp->rcv_scale);
17830                 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
17831         }
17832         /*
17833          * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
17834          * window.  This may cause the remote transmitter to stall.  This
17835          * flag tells soreceive() to disable delayed acknowledgements when
17836          * draining the buffer.  This can occur if the receiver is
17837          * attempting to read more data than can be buffered prior to
17838          * transmitting on the connection.
17839          */
17840         if (th->th_win == 0) {
17841                 tp->t_sndzerowin++;
17842                 tp->t_flags |= TF_RXWIN0SENT;
17843         } else
17844                 tp->t_flags &= ~TF_RXWIN0SENT;
17845         tp->snd_up = tp->snd_una;       /* drag it along, its deprecated */
17846         /* Now are we using fsb?, if so copy the template data to the mbuf */
17847         if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) && rack->r_fsb_inited) {
17848                 uint8_t *cpto;
17849
17850                 cpto = mtod(m, uint8_t *);
17851                 memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
17852                 /*
17853                  * We have just copied in:
17854                  * IP/IP6
17855                  * <optional udphdr>
17856                  * tcphdr (no options)
17857                  *
17858                  * We need to grab the correct pointers into the mbuf
17859                  * for both the tcp header, and possibly the udp header (if tunneling).
17860                  * We do this by using the offset in the copy buffer and adding it
17861                  * to the mbuf base pointer (cpto).
17862                  */
17863 #ifdef INET6
17864                 if (isipv6)
17865                         ip6 = mtod(m, struct ip6_hdr *);
17866                 else
17867 #endif                          /* INET6 */
17868                         ip = mtod(m, struct ip *);
17869                 th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
17870                 /* If we have a udp header lets set it into the mbuf as well */
17871                 if (udp)
17872                         udp = (struct udphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.udp - rack->r_ctl.fsb.tcp_ip_hdr));
17873         }
17874 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
17875         if (to.to_flags & TOF_SIGNATURE) {
17876                 /*
17877                  * Calculate MD5 signature and put it into the place
17878                  * determined before.
17879                  * NOTE: since TCP options buffer doesn't point into
17880                  * mbuf's data, calculate offset and use it.
17881                  */
17882                 if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
17883                                                        (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
17884                         /*
17885                          * Do not send segment if the calculation of MD5
17886                          * digest has failed.
17887                          */
17888                         goto out;
17889                 }
17890         }
17891 #endif
17892         if (optlen) {
17893                 bcopy(opt, th + 1, optlen);
17894                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
17895         }
17896         /*
17897          * Put TCP length in extended header, and then checksum extended
17898          * header and data.
17899          */
17900         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
17901 #ifdef INET6
17902         if (isipv6) {
17903                 /*
17904                  * ip6_plen is not need to be filled now, and will be filled
17905                  * in ip6_output.
17906                  */
17907                 if (tp->t_port) {
17908                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
17909                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
17910                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
17911                         th->th_sum = htons(0);
17912                         UDPSTAT_INC(udps_opackets);
17913                 } else {
17914                         m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
17915                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
17916                         th->th_sum = in6_cksum_pseudo(ip6,
17917                                                       sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
17918                                                       0);
17919                 }
17920         }
17921 #endif
17922 #if defined(INET6) && defined(INET)
17923         else
17924 #endif
17925 #ifdef INET
17926         {
17927                 if (tp->t_port) {
17928                         m->m_pkthdr.csum_flags = CSUM_UDP;
17929                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
17930                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
17931                                                 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
17932                         th->th_sum = htons(0);
17933                         UDPSTAT_INC(udps_opackets);
17934                 } else {
17935                         m->m_pkthdr.csum_flags = CSUM_TCP;
17936                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
17937                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
17938                                                ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
17939                                                                         IPPROTO_TCP + len + optlen));
17940                 }
17941                 /* IP version must be set here for ipv4/ipv6 checking later */
17942                 KASSERT(ip->ip_v == IPVERSION,
17943                         ("%s: IP version incorrect: %d", __func__, ip->ip_v));
17944         }
17945 #endif
17946         /*
17947          * Enable TSO and specify the size of the segments. The TCP pseudo
17948          * header checksum is always provided. XXX: Fixme: This is currently
17949          * not the case for IPv6.
17950          */
17951         if (tso) {
17952                 KASSERT(len > tp->t_maxseg - optlen,
17953                         ("%s: len <= tso_segsz", __func__));
17954                 m->m_pkthdr.csum_flags |= CSUM_TSO;
17955                 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
17956         }
17957         KASSERT(len + hdrlen == m_length(m, NULL),
17958                 ("%s: mbuf chain different than expected: %d + %u != %u",
17959                  __func__, len, hdrlen, m_length(m, NULL)));
17960
17961 #ifdef TCP_HHOOK
17962         /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
17963         hhook_run_tcp_est_out(tp, th, &to, len, tso);
17964 #endif
17965         /* We're getting ready to send; log now. */
17966         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
17967                 union tcp_log_stackspecific log;
17968
17969                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
17970                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
17971                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
17972                 if (rack->rack_no_prr)
17973                         log.u_bbr.flex1 = 0;
17974                 else
17975                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
17976                 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
17977                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
17978                 log.u_bbr.flex4 = orig_len;
17979                 if (filled_all)
17980                         log.u_bbr.flex5 = 0x80000000;
17981                 else
17982                         log.u_bbr.flex5 = 0;
17983                 /* Save off the early/late values */
17984                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
17985                 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
17986                 log.u_bbr.bw_inuse = rack_get_bw(rack);
17987                 if (rsm || sack_rxmit) {
17988                         if (doing_tlp)
17989                                 log.u_bbr.flex8 = 2;
17990                         else
17991                                 log.u_bbr.flex8 = 1;
17992                 } else {
17993                         log.u_bbr.flex8 = 0;
17994                 }
17995                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
17996                 log.u_bbr.flex7 = mark;
17997                 log.u_bbr.flex7 <<= 8;
17998                 log.u_bbr.flex7 |= pass;
17999                 log.u_bbr.pkts_out = tp->t_maxseg;
18000                 log.u_bbr.timeStamp = cts;
18001                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18002                 log.u_bbr.lt_epoch = cwnd_to_use;
18003                 log.u_bbr.delivered = sendalot;
18004                 lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
18005                                      len, &log, false, NULL, NULL, 0, &tv);
18006         } else
18007                 lgb = NULL;
18008
18009         /*
18010          * Fill in IP length and desired time to live and send to IP level.
18011          * There should be a better way to handle ttl and tos; we could keep
18012          * them in the template, but need a way to checksum without them.
18013          */
18014         /*
18015          * m->m_pkthdr.len should have been set before cksum calcuration,
18016          * because in6_cksum() need it.
18017          */
18018 #ifdef INET6
18019         if (isipv6) {
18020                 /*
18021                  * we separately set hoplimit for every segment, since the
18022                  * user might want to change the value via setsockopt. Also,
18023                  * desired default hop limit might be changed via Neighbor
18024                  * Discovery.
18025                  */
18026                 rack->r_ctl.fsb.hoplimit = ip6->ip6_hlim = in6_selecthlim(inp, NULL);
18027
18028                 /*
18029                  * Set the packet size here for the benefit of DTrace
18030                  * probes. ip6_output() will set it properly; it's supposed
18031                  * to include the option header lengths as well.
18032                  */
18033                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
18034
18035                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
18036                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18037                 else
18038                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18039
18040                 if (tp->t_state == TCPS_SYN_SENT)
18041                         TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
18042
18043                 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
18044                 /* TODO: IPv6 IP6TOS_ECT bit on */
18045                 error = ip6_output(m,
18046 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18047                                    inp->in6p_outputopts,
18048 #else
18049                                    NULL,
18050 #endif
18051                                    &inp->inp_route6,
18052                                    ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0),
18053                                    NULL, NULL, inp);
18054
18055                 if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
18056                         mtu = inp->inp_route6.ro_nh->nh_mtu;
18057         }
18058 #endif                          /* INET6 */
18059 #if defined(INET) && defined(INET6)
18060         else
18061 #endif
18062 #ifdef INET
18063         {
18064                 ip->ip_len = htons(m->m_pkthdr.len);
18065 #ifdef INET6
18066                 if (inp->inp_vflag & INP_IPV6PROTO)
18067                         ip->ip_ttl = in6_selecthlim(inp, NULL);
18068 #endif                          /* INET6 */
18069                 rack->r_ctl.fsb.hoplimit = ip->ip_ttl;
18070                 /*
18071                  * If we do path MTU discovery, then we set DF on every
18072                  * packet. This might not be the best thing to do according
18073                  * to RFC3390 Section 2. However the tcp hostcache migitates
18074                  * the problem so it affects only the first tcp connection
18075                  * with a host.
18076                  *
18077                  * NB: Don't set DF on small MTU/MSS to have a safe
18078                  * fallback.
18079                  */
18080                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
18081                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18082                         if (tp->t_port == 0 || len < V_tcp_minmss) {
18083                                 ip->ip_off |= htons(IP_DF);
18084                         }
18085                 } else {
18086                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18087                 }
18088
18089                 if (tp->t_state == TCPS_SYN_SENT)
18090                         TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
18091
18092                 TCP_PROBE5(send, NULL, tp, ip, tp, th);
18093
18094                 error = ip_output(m,
18095 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18096                                   inp->inp_options,
18097 #else
18098                                   NULL,
18099 #endif
18100                                   &inp->inp_route,
18101                                   ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0,
18102                                   inp);
18103                 if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
18104                         mtu = inp->inp_route.ro_nh->nh_mtu;
18105         }
18106 #endif                          /* INET */
18107
18108 out:
18109         if (lgb) {
18110                 lgb->tlb_errno = error;
18111                 lgb = NULL;
18112         }
18113         /*
18114          * In transmit state, time the transmission and arrange for the
18115          * retransmit.  In persist state, just set snd_max.
18116          */
18117         if (error == 0) {
18118                 tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp);
18119                 rack->forced_ack = 0;   /* If we send something zap the FA flag */
18120                 if (rsm && (doing_tlp == 0)) {
18121                         /* Set we retransmitted */
18122                         rack->rc_gp_saw_rec = 1;
18123                 } else {
18124                         if (cwnd_to_use > tp->snd_ssthresh) {
18125                                 /* Set we sent in CA */
18126                                 rack->rc_gp_saw_ca = 1;
18127                         } else {
18128                                 /* Set we sent in SS */
18129                                 rack->rc_gp_saw_ss = 1;
18130                         }
18131                 }
18132                 if (TCPS_HAVEESTABLISHED(tp->t_state) &&
18133                     (tp->t_flags & TF_SACK_PERMIT) &&
18134                     tp->rcv_numsacks > 0)
18135                         tcp_clean_dsack_blocks(tp);
18136                 tot_len_this_send += len;
18137                 if (len == 0)
18138                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1);
18139                 else if (len == 1) {
18140                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1);
18141                 } else if (len > 1) {
18142                         int idx;
18143
18144                         idx = (len / segsiz) + 3;
18145                         if (idx >= TCP_MSS_ACCT_ATIMER)
18146                                 counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
18147                         else
18148                                 counter_u64_add(rack_out_size[idx], 1);
18149                 }
18150         }
18151         if ((rack->rack_no_prr == 0) &&
18152             sub_from_prr &&
18153             (error == 0)) {
18154                 if (rack->r_ctl.rc_prr_sndcnt >= len)
18155                         rack->r_ctl.rc_prr_sndcnt -= len;
18156                 else
18157                         rack->r_ctl.rc_prr_sndcnt = 0;
18158         }
18159         sub_from_prr = 0;
18160         if (doing_tlp && (rsm == NULL)) {
18161                 /* New send doing a TLP */
18162                 add_flag |= RACK_TLP;
18163         }
18164         rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error,
18165                         rack_to_usec_ts(&tv),
18166                         rsm, add_flag, s_mb, s_moff);
18167
18168
18169         if ((error == 0) &&
18170             (len > 0) &&
18171             (tp->snd_una == tp->snd_max))
18172                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
18173         {
18174                 tcp_seq startseq = tp->snd_nxt;
18175
18176                 /* Track our lost count */
18177                 if (rsm && (doing_tlp == 0))
18178                         rack->r_ctl.rc_loss_count += rsm->r_end - rsm->r_start;
18179                 /*
18180                  * Advance snd_nxt over sequence space of this segment.
18181                  */
18182                 if (error)
18183                         /* We don't log or do anything with errors */
18184                         goto nomore;
18185                 if (doing_tlp == 0) {
18186                         if (rsm == NULL) {
18187                                 /*
18188                                  * Not a retransmission of some
18189                                  * sort, new data is going out so
18190                                  * clear our TLP count and flag.
18191                                  */
18192                                 rack->rc_tlp_in_progress = 0;
18193                                 rack->r_ctl.rc_tlp_cnt_out = 0;
18194                         }
18195                 } else {
18196                         /*
18197                          * We have just sent a TLP, mark that it is true
18198                          * and make sure our in progress is set so we
18199                          * continue to check the count.
18200                          */
18201                         rack->rc_tlp_in_progress = 1;
18202                         rack->r_ctl.rc_tlp_cnt_out++;
18203                 }
18204                 if (flags & (TH_SYN | TH_FIN)) {
18205                         if (flags & TH_SYN)
18206                                 tp->snd_nxt++;
18207                         if (flags & TH_FIN) {
18208                                 tp->snd_nxt++;
18209                                 tp->t_flags |= TF_SENTFIN;
18210                         }
18211                 }
18212                 /* In the ENOBUFS case we do *not* update snd_max */
18213                 if (sack_rxmit)
18214                         goto nomore;
18215
18216                 tp->snd_nxt += len;
18217                 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
18218                         if (tp->snd_una == tp->snd_max) {
18219                                 /*
18220                                  * Update the time we just added data since
18221                                  * none was outstanding.
18222                                  */
18223                                 rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
18224                                 tp->t_acktime = ticks;
18225                         }
18226                         tp->snd_max = tp->snd_nxt;
18227                         /*
18228                          * Time this transmission if not a retransmission and
18229                          * not currently timing anything.
18230                          * This is only relevant in case of switching back to
18231                          * the base stack.
18232                          */
18233                         if (tp->t_rtttime == 0) {
18234                                 tp->t_rtttime = ticks;
18235                                 tp->t_rtseq = startseq;
18236                                 KMOD_TCPSTAT_INC(tcps_segstimed);
18237                         }
18238                         if (len &&
18239                             ((tp->t_flags & TF_GPUTINPROG) == 0))
18240                                 rack_start_gp_measurement(tp, rack, startseq, sb_offset);
18241                 }
18242                 /*
18243                  * If we are doing FO we need to update the mbuf position and subtract
18244                  * this happens when the peer sends us duplicate information and
18245                  * we thus want to send a DSACK.
18246                  *
18247                  * XXXRRS: This brings to mind a ?, when we send a DSACK block is TSO
18248                  * turned off? If not then we are going to echo multiple DSACK blocks
18249                  * out (with the TSO), which we should not be doing.
18250                  */
18251                 if (rack->r_fast_output && len) {
18252                         if (rack->r_ctl.fsb.left_to_send > len)
18253                                 rack->r_ctl.fsb.left_to_send -= len;
18254                         else
18255                                 rack->r_ctl.fsb.left_to_send = 0;
18256                         if (rack->r_ctl.fsb.left_to_send < segsiz)
18257                                 rack->r_fast_output = 0;
18258                         if (rack->r_fast_output) {
18259                                 rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
18260                                 rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
18261                         }
18262                 }
18263         }
18264 nomore:
18265         if (error) {
18266                 rack->r_ctl.rc_agg_delayed = 0;
18267                 rack->r_early = 0;
18268                 rack->r_late = 0;
18269                 rack->r_ctl.rc_agg_early = 0;
18270                 SOCKBUF_UNLOCK_ASSERT(sb);      /* Check gotos. */
18271                 /*
18272                  * Failures do not advance the seq counter above. For the
18273                  * case of ENOBUFS we will fall out and retry in 1ms with
18274                  * the hpts. Everything else will just have to retransmit
18275                  * with the timer.
18276                  *
18277                  * In any case, we do not want to loop around for another
18278                  * send without a good reason.
18279                  */
18280                 sendalot = 0;
18281                 switch (error) {
18282                 case EPERM:
18283                         tp->t_softerror = error;
18284 #ifdef TCP_ACCOUNTING
18285                         crtsc = get_cyclecount();
18286                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18287                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
18288                         }
18289                         counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
18290                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18291                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
18292                         }
18293                         counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
18294                         sched_unpin();
18295 #endif
18296                         return (error);
18297                 case ENOBUFS:
18298                         /*
18299                          * Pace us right away to retry in a some
18300                          * time
18301                          */
18302                         slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
18303                         if (rack->rc_enobuf < 0x7f)
18304                                 rack->rc_enobuf++;
18305                         if (slot < (10 * HPTS_USEC_IN_MSEC))
18306                                 slot = 10 * HPTS_USEC_IN_MSEC;
18307                         if (rack->r_ctl.crte != NULL) {
18308                                 counter_u64_add(rack_saw_enobuf_hw, 1);
18309                                 tcp_rl_log_enobuf(rack->r_ctl.crte);
18310                         }
18311                         counter_u64_add(rack_saw_enobuf, 1);
18312                         goto enobufs;
18313                 case EMSGSIZE:
18314                         /*
18315                          * For some reason the interface we used initially
18316                          * to send segments changed to another or lowered
18317                          * its MTU. If TSO was active we either got an
18318                          * interface without TSO capabilits or TSO was
18319                          * turned off. If we obtained mtu from ip_output()
18320                          * then update it and try again.
18321                          */
18322                         if (tso)
18323                                 tp->t_flags &= ~TF_TSO;
18324                         if (mtu != 0) {
18325                                 tcp_mss_update(tp, -1, mtu, NULL, NULL);
18326                                 goto again;
18327                         }
18328                         slot = 10 * HPTS_USEC_IN_MSEC;
18329                         rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
18330 #ifdef TCP_ACCOUNTING
18331                         crtsc = get_cyclecount();
18332                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18333                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
18334                         }
18335                         counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
18336                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18337                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
18338                         }
18339                         counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
18340                         sched_unpin();
18341 #endif
18342                         return (error);
18343                 case ENETUNREACH:
18344                         counter_u64_add(rack_saw_enetunreach, 1);
18345                 case EHOSTDOWN:
18346                 case EHOSTUNREACH:
18347                 case ENETDOWN:
18348                         if (TCPS_HAVERCVDSYN(tp->t_state)) {
18349                                 tp->t_softerror = error;
18350                         }
18351                         /* FALLTHROUGH */
18352                 default:
18353                         slot = 10 * HPTS_USEC_IN_MSEC;
18354                         rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
18355 #ifdef TCP_ACCOUNTING
18356                         crtsc = get_cyclecount();
18357                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18358                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
18359                         }
18360                         counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
18361                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18362                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
18363                         }
18364                         counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
18365                         sched_unpin();
18366 #endif
18367                         return (error);
18368                 }
18369         } else {
18370                 rack->rc_enobuf = 0;
18371                 if (IN_FASTRECOVERY(tp->t_flags) && rsm)
18372                         rack->r_ctl.retran_during_recovery += len;
18373         }
18374         KMOD_TCPSTAT_INC(tcps_sndtotal);
18375
18376         /*
18377          * Data sent (as far as we can tell). If this advertises a larger
18378          * window than any other segment, then remember the size of the
18379          * advertised window. Any pending ACK has now been sent.
18380          */
18381         if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
18382                 tp->rcv_adv = tp->rcv_nxt + recwin;
18383
18384         tp->last_ack_sent = tp->rcv_nxt;
18385         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
18386 enobufs:
18387         if (sendalot) {
18388                 /* Do we need to turn off sendalot? */
18389                 if (rack->r_ctl.rc_pace_max_segs &&
18390                     (tot_len_this_send >= rack->r_ctl.rc_pace_max_segs)) {
18391                         /* We hit our max. */
18392                         sendalot = 0;
18393                 } else if ((rack->rc_user_set_max_segs) &&
18394                            (tot_len_this_send >= (rack->rc_user_set_max_segs * segsiz))) {
18395                         /* We hit the user defined max */
18396                         sendalot = 0;
18397                 }
18398         }
18399         if ((error == 0) && (flags & TH_FIN))
18400                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN);
18401         if (flags & TH_RST) {
18402                 /*
18403                  * We don't send again after sending a RST.
18404                  */
18405                 slot = 0;
18406                 sendalot = 0;
18407                 if (error == 0)
18408                         tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
18409         } else if ((slot == 0) && (sendalot == 0) && tot_len_this_send) {
18410                 /*
18411                  * Get our pacing rate, if an error
18412                  * occurred in sending (ENOBUF) we would
18413                  * hit the else if with slot preset. Other
18414                  * errors return.
18415                  */
18416                 slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, rsm, segsiz);
18417         }
18418         if (rsm &&
18419             (rsm->r_flags & RACK_HAS_SYN) == 0 &&
18420             rack->use_rack_rr) {
18421                 /* Its a retransmit and we use the rack cheat? */
18422                 if ((slot == 0) ||
18423                     (rack->rc_always_pace == 0) ||
18424                     (rack->r_rr_config == 1)) {
18425                         /*
18426                          * We have no pacing set or we
18427                          * are using old-style rack or
18428                          * we are overriden to use the old 1ms pacing.
18429                          */
18430                         slot = rack->r_ctl.rc_min_to;
18431                 }
18432         }
18433         /* We have sent clear the flag */
18434         rack->r_ent_rec_ns = 0;
18435         if (rack->r_must_retran) {
18436                 if (rsm) {
18437                         rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
18438                         if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
18439                                 /*
18440                                  * We have retransmitted all.
18441                                  */
18442                                 rack->r_must_retran = 0;
18443                                 rack->r_ctl.rc_out_at_rto = 0;
18444                         }
18445                 } else if (SEQ_GEQ(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
18446                         /*
18447                          * Sending new data will also kill
18448                          * the loop.
18449                          */
18450                         rack->r_must_retran = 0;
18451                         rack->r_ctl.rc_out_at_rto = 0;
18452                 }
18453         }
18454         rack->r_ctl.fsb.recwin = recwin;
18455         if ((tp->t_flags & (TF_WASCRECOVERY|TF_WASFRECOVERY)) &&
18456             SEQ_GT(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
18457                 /*
18458                  * We hit an RTO and now have past snd_max at the RTO
18459                  * clear all the WAS flags.
18460                  */
18461                 tp->t_flags &= ~(TF_WASCRECOVERY|TF_WASFRECOVERY);
18462         }
18463         if (slot) {
18464                 /* set the rack tcb into the slot N */
18465                 counter_u64_add(rack_paced_segments, 1);
18466                 if ((error == 0) &&
18467                     rack_use_rfo &&
18468                     ((flags & (TH_SYN|TH_FIN)) == 0) &&
18469                     (rsm == NULL) &&
18470                     (tp->snd_nxt == tp->snd_max) &&
18471                     (ipoptlen == 0) &&
18472                     (tp->rcv_numsacks == 0) &&
18473                     rack->r_fsb_inited &&
18474                     TCPS_HAVEESTABLISHED(tp->t_state) &&
18475                     (rack->r_must_retran == 0) &&
18476                     ((tp->t_flags & TF_NEEDFIN) == 0) &&
18477                     (len > 0) && (orig_len > 0) &&
18478                     (orig_len > len) &&
18479                     ((orig_len - len) >= segsiz) &&
18480                     ((optlen == 0) ||
18481                      ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
18482                         /* We can send at least one more MSS using our fsb */
18483
18484                         rack->r_fast_output = 1;
18485                         rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
18486                         rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
18487                         rack->r_ctl.fsb.tcp_flags = flags;
18488                         rack->r_ctl.fsb.left_to_send = orig_len - len;
18489                         KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
18490                                 ("rack:%p left_to_send:%u sbavail:%u out:%u",
18491                                  rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
18492                                  (tp->snd_max - tp->snd_una)));
18493                         if (rack->r_ctl.fsb.left_to_send < segsiz)
18494                                 rack->r_fast_output = 0;
18495                         else {
18496                                 if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
18497                                         rack->r_ctl.fsb.rfo_apply_push = 1;
18498                                 else
18499                                         rack->r_ctl.fsb.rfo_apply_push = 0;
18500                         }
18501                 } else
18502                         rack->r_fast_output = 0;
18503                 rack_log_fsb(rack, tp, so, flags,
18504                              ipoptlen, orig_len, len, error,
18505                              (rsm == NULL), optlen, __LINE__, 2);
18506         } else if (sendalot) {
18507                 int ret;
18508
18509                 if (len)
18510                         counter_u64_add(rack_unpaced_segments, 1);
18511                 sack_rxmit = 0;
18512                 if ((error == 0) &&
18513                     rack_use_rfo &&
18514                     ((flags & (TH_SYN|TH_FIN)) == 0) &&
18515                     (rsm == NULL) &&
18516                     (ipoptlen == 0) &&
18517                     (tp->rcv_numsacks == 0) &&
18518                     (tp->snd_nxt == tp->snd_max) &&
18519                     (rack->r_must_retran == 0) &&
18520                     rack->r_fsb_inited &&
18521                     TCPS_HAVEESTABLISHED(tp->t_state) &&
18522                     ((tp->t_flags & TF_NEEDFIN) == 0) &&
18523                     (len > 0) && (orig_len > 0) &&
18524                     (orig_len > len) &&
18525                     ((orig_len - len) >= segsiz) &&
18526                     ((optlen == 0) ||
18527                      ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
18528                         /* we can use fast_output for more */
18529
18530                         rack->r_fast_output = 1;
18531                         rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
18532                         rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
18533                         rack->r_ctl.fsb.tcp_flags = flags;
18534                         rack->r_ctl.fsb.left_to_send = orig_len - len;
18535                         KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
18536                                 ("rack:%p left_to_send:%u sbavail:%u out:%u",
18537                                  rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
18538                                  (tp->snd_max - tp->snd_una)));
18539                         if (rack->r_ctl.fsb.left_to_send < segsiz) {
18540                                 rack->r_fast_output = 0;
18541                         }
18542                         if (rack->r_fast_output) {
18543                                 if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
18544                                         rack->r_ctl.fsb.rfo_apply_push = 1;
18545                                 else
18546                                         rack->r_ctl.fsb.rfo_apply_push = 0;
18547                                 rack_log_fsb(rack, tp, so, flags,
18548                                              ipoptlen, orig_len, len, error,
18549                                              (rsm == NULL), optlen, __LINE__, 3);
18550                                 error = 0;
18551                                 ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
18552                                 if (ret >= 0)
18553                                         return (ret);
18554                                 else if (error)
18555                                         goto nomore;
18556
18557                         }
18558                 }
18559                 goto again;
18560         } else if (len) {
18561                 counter_u64_add(rack_unpaced_segments, 1);
18562         }
18563         /* Assure when we leave that snd_nxt will point to top */
18564         if (SEQ_GT(tp->snd_max, tp->snd_nxt))
18565                 tp->snd_nxt = tp->snd_max;
18566         rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, 0);
18567 #ifdef TCP_ACCOUNTING
18568         crtsc = get_cyclecount() - ts_val;
18569         if (tot_len_this_send) {
18570                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18571                         tp->tcp_cnt_counters[SND_OUT_DATA]++;
18572                 }
18573                 counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], 1);
18574                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18575                         tp->tcp_proc_time[SND_OUT_DATA] += crtsc;
18576                 }
18577                 counter_u64_add(tcp_proc_time[SND_OUT_DATA], crtsc);
18578                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18579                         tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) /segsiz);
18580                 }
18581                 counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len_this_send + segsiz - 1) /segsiz));
18582         } else {
18583                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18584                         tp->tcp_cnt_counters[SND_OUT_ACK]++;
18585                 }
18586                 counter_u64_add(tcp_cnt_counters[SND_OUT_ACK], 1);
18587                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18588                         tp->tcp_proc_time[SND_OUT_ACK] += crtsc;
18589                 }
18590                 counter_u64_add(tcp_proc_time[SND_OUT_ACK], crtsc);
18591         }
18592         sched_unpin();
18593 #endif
18594         if (error == ENOBUFS)
18595                 error = 0;
18596         return (error);
18597 }
18598
18599 static void
18600 rack_update_seg(struct tcp_rack *rack)
18601 {
18602         uint32_t orig_val;
18603
18604         orig_val = rack->r_ctl.rc_pace_max_segs;
18605         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
18606         if (orig_val != rack->r_ctl.rc_pace_max_segs)
18607                 rack_log_pacing_delay_calc(rack, 0, 0, orig_val, 0, 0, 15, __LINE__, NULL);
18608 }
18609
18610 static void
18611 rack_mtu_change(struct tcpcb *tp)
18612 {
18613         /*
18614          * The MSS may have changed
18615          */
18616         struct tcp_rack *rack;
18617
18618         rack = (struct tcp_rack *)tp->t_fb_ptr;
18619         if (rack->r_ctl.rc_pace_min_segs != ctf_fixed_maxseg(tp)) {
18620                 /*
18621                  * The MTU has changed we need to resend everything
18622                  * since all we have sent is lost. We first fix
18623                  * up the mtu though.
18624                  */
18625                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
18626                 /* We treat this like a full retransmit timeout without the cwnd adjustment */
18627                 rack_remxt_tmr(tp);
18628                 rack->r_fast_output = 0;
18629                 rack->r_ctl.rc_out_at_rto = ctf_flight_size(tp,
18630                                                 rack->r_ctl.rc_sacked);
18631                 rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
18632                 rack->r_must_retran = 1;
18633
18634         }
18635         sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
18636         /* We don't use snd_nxt to retransmit */
18637         tp->snd_nxt = tp->snd_max;
18638 }
18639
18640 static int
18641 rack_set_profile(struct tcp_rack *rack, int prof)
18642 {
18643         int err = EINVAL;
18644         if (prof == 1) {
18645                 /* pace_always=1 */
18646                 if (rack->rc_always_pace == 0) {
18647                         if (tcp_can_enable_pacing() == 0)
18648                                 return (EBUSY);
18649                 }
18650                 rack->rc_always_pace = 1;
18651                 if (rack->use_fixed_rate || rack->gp_ready)
18652                         rack_set_cc_pacing(rack);
18653                 rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
18654                 rack->rack_attempt_hdwr_pace = 0;
18655                 /* cmpack=1 */
18656                 if (rack_use_cmp_acks)
18657                         rack->r_use_cmp_ack = 1;
18658                 if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
18659                     rack->r_use_cmp_ack)
18660                         rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
18661                 /* scwnd=1 */
18662                 rack->rack_enable_scwnd = 1;
18663                 /* dynamic=100 */
18664                 rack->rc_gp_dyn_mul = 1;
18665                 /* gp_inc_ca */
18666                 rack->r_ctl.rack_per_of_gp_ca = 100;
18667                 /* rrr_conf=3 */
18668                 rack->r_rr_config = 3;
18669                 /* npush=2 */
18670                 rack->r_ctl.rc_no_push_at_mrtt = 2;
18671                 /* fillcw=1 */
18672                 rack->rc_pace_to_cwnd = 1;
18673                 rack->rc_pace_fill_if_rttin_range = 0;
18674                 rack->rtt_limit_mul = 0;
18675                 /* noprr=1 */
18676                 rack->rack_no_prr = 1;
18677                 /* lscwnd=1 */
18678                 rack->r_limit_scw = 1;
18679                 /* gp_inc_rec */
18680                 rack->r_ctl.rack_per_of_gp_rec = 90;
18681                 err = 0;
18682
18683         } else if (prof == 3) {
18684                 /* Same as profile one execept fill_cw becomes 2 (less aggressive set) */
18685                 /* pace_always=1 */
18686                 if (rack->rc_always_pace == 0) {
18687                         if (tcp_can_enable_pacing() == 0)
18688                                 return (EBUSY);
18689                 }
18690                 rack->rc_always_pace = 1;
18691                 if (rack->use_fixed_rate || rack->gp_ready)
18692                         rack_set_cc_pacing(rack);
18693                 rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
18694                 rack->rack_attempt_hdwr_pace = 0;
18695                 /* cmpack=1 */
18696                 if (rack_use_cmp_acks)
18697                         rack->r_use_cmp_ack = 1;
18698                 if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
18699                     rack->r_use_cmp_ack)
18700                         rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
18701                 /* scwnd=1 */
18702                 rack->rack_enable_scwnd = 1;
18703                 /* dynamic=100 */
18704                 rack->rc_gp_dyn_mul = 1;
18705                 /* gp_inc_ca */
18706                 rack->r_ctl.rack_per_of_gp_ca = 100;
18707                 /* rrr_conf=3 */
18708                 rack->r_rr_config = 3;
18709                 /* npush=2 */
18710                 rack->r_ctl.rc_no_push_at_mrtt = 2;
18711                 /* fillcw=2 */
18712                 rack->rc_pace_to_cwnd = 1;
18713                 rack->r_fill_less_agg = 1;
18714                 rack->rc_pace_fill_if_rttin_range = 0;
18715                 rack->rtt_limit_mul = 0;
18716                 /* noprr=1 */
18717                 rack->rack_no_prr = 1;
18718                 /* lscwnd=1 */
18719                 rack->r_limit_scw = 1;
18720                 /* gp_inc_rec */
18721                 rack->r_ctl.rack_per_of_gp_rec = 90;
18722                 err = 0;
18723
18724
18725         } else if (prof == 2) {
18726                 /* cmpack=1 */
18727                 if (rack->rc_always_pace == 0) {
18728                         if (tcp_can_enable_pacing() == 0)
18729                                 return (EBUSY);
18730                 }
18731                 rack->rc_always_pace = 1;
18732                 if (rack->use_fixed_rate || rack->gp_ready)
18733                         rack_set_cc_pacing(rack);
18734                 rack->r_use_cmp_ack = 1;
18735                 if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
18736                         rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
18737                 /* pace_always=1 */
18738                 rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
18739                 /* scwnd=1 */
18740                 rack->rack_enable_scwnd = 1;
18741                 /* dynamic=100 */
18742                 rack->rc_gp_dyn_mul = 1;
18743                 rack->r_ctl.rack_per_of_gp_ca = 100;
18744                 /* rrr_conf=3 */
18745                 rack->r_rr_config = 3;
18746                 /* npush=2 */
18747                 rack->r_ctl.rc_no_push_at_mrtt = 2;
18748                 /* fillcw=1 */
18749                 rack->rc_pace_to_cwnd = 1;
18750                 rack->rc_pace_fill_if_rttin_range = 0;
18751                 rack->rtt_limit_mul = 0;
18752                 /* noprr=1 */
18753                 rack->rack_no_prr = 1;
18754                 /* lscwnd=0 */
18755                 rack->r_limit_scw = 0;
18756                 err = 0;
18757         } else if (prof == 0) {
18758                 /* This changes things back to the default settings */
18759                 err = 0;
18760                 if (rack->rc_always_pace) {
18761                         tcp_decrement_paced_conn();
18762                         rack_undo_cc_pacing(rack);
18763                         rack->rc_always_pace = 0;
18764                 }
18765                 if (rack_pace_every_seg && tcp_can_enable_pacing()) {
18766                         rack->rc_always_pace = 1;
18767                         if (rack->use_fixed_rate || rack->gp_ready)
18768                                 rack_set_cc_pacing(rack);
18769                 } else
18770                         rack->rc_always_pace = 0;
18771                 if (rack_use_cmp_acks)
18772                         rack->r_use_cmp_ack = 1;
18773                 else
18774                         rack->r_use_cmp_ack = 0;
18775                 if (rack_disable_prr)
18776                         rack->rack_no_prr = 1;
18777                 else
18778                         rack->rack_no_prr = 0;
18779                 if (rack_gp_no_rec_chg)
18780                         rack->rc_gp_no_rec_chg = 1;
18781                 else
18782                         rack->rc_gp_no_rec_chg = 0;
18783                 if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack) {
18784                         rack->r_mbuf_queue = 1;
18785                         if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
18786                                 rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
18787                         rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
18788                 } else {
18789                         rack->r_mbuf_queue = 0;
18790                         rack->rc_inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
18791                 }
18792                 if (rack_enable_shared_cwnd)
18793                         rack->rack_enable_scwnd = 1;
18794                 else
18795                         rack->rack_enable_scwnd = 0;
18796                 if (rack_do_dyn_mul) {
18797                         /* When dynamic adjustment is on CA needs to start at 100% */
18798                         rack->rc_gp_dyn_mul = 1;
18799                         if (rack_do_dyn_mul >= 100)
18800                                 rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
18801                 } else {
18802                         rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
18803                         rack->rc_gp_dyn_mul = 0;
18804                 }
18805                 rack->r_rr_config = 0;
18806                 rack->r_ctl.rc_no_push_at_mrtt = 0;
18807                 rack->rc_pace_to_cwnd = 0;
18808                 rack->rc_pace_fill_if_rttin_range = 0;
18809                 rack->rtt_limit_mul = 0;
18810
18811                 if (rack_enable_hw_pacing)
18812                         rack->rack_hdw_pace_ena = 1;
18813                 else
18814                         rack->rack_hdw_pace_ena = 0;
18815                 if (rack_disable_prr)
18816                         rack->rack_no_prr = 1;
18817                 else
18818                         rack->rack_no_prr = 0;
18819                 if (rack_limits_scwnd)
18820                         rack->r_limit_scw  = 1;
18821                 else
18822                         rack->r_limit_scw  = 0;
18823                 err = 0;
18824         }
18825         return (err);
18826 }
18827
18828 static int
18829 rack_add_deferred_option(struct tcp_rack *rack, int sopt_name, uint64_t loptval)
18830 {
18831         struct deferred_opt_list *dol;
18832
18833         dol = malloc(sizeof(struct deferred_opt_list),
18834                      M_TCPFSB, M_NOWAIT|M_ZERO);
18835         if (dol == NULL) {
18836                 /*
18837                  * No space yikes -- fail out..
18838                  */
18839                 return (0);
18840         }
18841         dol->optname = sopt_name;
18842         dol->optval = loptval;
18843         TAILQ_INSERT_TAIL(&rack->r_ctl.opt_list, dol, next);
18844         return (1);
18845 }
18846
18847 static int
18848 rack_process_option(struct tcpcb *tp, struct tcp_rack *rack, int sopt_name,
18849                     uint32_t optval, uint64_t loptval)
18850 {
18851         struct epoch_tracker et;
18852         struct sockopt sopt;
18853         struct cc_newreno_opts opt;
18854         uint64_t val;
18855         int error = 0;
18856         uint16_t ca, ss;
18857
18858         switch (sopt_name) {
18859
18860         case TCP_RACK_PACING_BETA:
18861                 RACK_OPTS_INC(tcp_rack_beta);
18862                 if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
18863                         /* This only works for newreno. */
18864                         error = EINVAL;
18865                         break;
18866                 }
18867                 if (rack->rc_pacing_cc_set) {
18868                         /*
18869                          * Set them into the real CC module
18870                          * whats in the rack pcb is the old values
18871                          * to be used on restoral/
18872                          */
18873                         sopt.sopt_dir = SOPT_SET;
18874                         opt.name = CC_NEWRENO_BETA;
18875                         opt.val = optval;
18876                         if (CC_ALGO(tp)->ctl_output != NULL)
18877                                 error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
18878                         else {
18879                                 error = ENOENT;
18880                                 break;
18881                         }
18882                 } else {
18883                         /*
18884                          * Not pacing yet so set it into our local
18885                          * rack pcb storage.
18886                          */
18887                         rack->r_ctl.rc_saved_beta.beta = optval;
18888                 }
18889                 break;
18890         case TCP_RACK_TIMER_SLOP:
18891                 RACK_OPTS_INC(tcp_rack_timer_slop);
18892                 rack->r_ctl.timer_slop = optval;
18893                 if (rack->rc_tp->t_srtt) {
18894                         /*
18895                          * If we have an SRTT lets update t_rxtcur
18896                          * to have the new slop.
18897                          */
18898                         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
18899                                            rack_rto_min, rack_rto_max,
18900                                            rack->r_ctl.timer_slop);
18901                 }
18902                 break;
18903         case TCP_RACK_PACING_BETA_ECN:
18904                 RACK_OPTS_INC(tcp_rack_beta_ecn);
18905                 if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
18906                         /* This only works for newreno. */
18907                         error = EINVAL;
18908                         break;
18909                 }
18910                 if (rack->rc_pacing_cc_set) {
18911                         /*
18912                          * Set them into the real CC module
18913                          * whats in the rack pcb is the old values
18914                          * to be used on restoral/
18915                          */
18916                         sopt.sopt_dir = SOPT_SET;
18917                         opt.name = CC_NEWRENO_BETA_ECN;
18918                         opt.val = optval;
18919                         if (CC_ALGO(tp)->ctl_output != NULL)
18920                                 error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
18921                         else
18922                                 error = ENOENT;
18923                 } else {
18924                         /*
18925                          * Not pacing yet so set it into our local
18926                          * rack pcb storage.
18927                          */
18928                         rack->r_ctl.rc_saved_beta.beta_ecn = optval;
18929                         rack->r_ctl.rc_saved_beta.newreno_flags = CC_NEWRENO_BETA_ECN;
18930                 }
18931                 break;
18932         case TCP_DEFER_OPTIONS:
18933                 RACK_OPTS_INC(tcp_defer_opt);
18934                 if (optval) {
18935                         if (rack->gp_ready) {
18936                                 /* Too late */
18937                                 error = EINVAL;
18938                                 break;
18939                         }
18940                         rack->defer_options = 1;
18941                 } else
18942                         rack->defer_options = 0;
18943                 break;
18944         case TCP_RACK_MEASURE_CNT:
18945                 RACK_OPTS_INC(tcp_rack_measure_cnt);
18946                 if (optval && (optval <= 0xff)) {
18947                         rack->r_ctl.req_measurements = optval;
18948                 } else
18949                         error = EINVAL;
18950                 break;
18951         case TCP_REC_ABC_VAL:
18952                 RACK_OPTS_INC(tcp_rec_abc_val);
18953                 if (optval > 0)
18954                         rack->r_use_labc_for_rec = 1;
18955                 else
18956                         rack->r_use_labc_for_rec = 0;
18957                 break;
18958         case TCP_RACK_ABC_VAL:
18959                 RACK_OPTS_INC(tcp_rack_abc_val);
18960                 if ((optval > 0) && (optval < 255))
18961                         rack->rc_labc = optval;
18962                 else
18963                         error = EINVAL;
18964                 break;
18965         case TCP_HDWR_UP_ONLY:
18966                 RACK_OPTS_INC(tcp_pacing_up_only);
18967                 if (optval)
18968                         rack->r_up_only = 1;
18969                 else
18970                         rack->r_up_only = 0;
18971                 break;
18972         case TCP_PACING_RATE_CAP:
18973                 RACK_OPTS_INC(tcp_pacing_rate_cap);
18974                 rack->r_ctl.bw_rate_cap = loptval;
18975                 break;
18976         case TCP_RACK_PROFILE:
18977                 RACK_OPTS_INC(tcp_profile);
18978                 error = rack_set_profile(rack, optval);
18979                 break;
18980         case TCP_USE_CMP_ACKS:
18981                 RACK_OPTS_INC(tcp_use_cmp_acks);
18982                 if ((optval == 0) && (rack->rc_inp->inp_flags2 & INP_MBUF_ACKCMP)) {
18983                         /* You can't turn it off once its on! */
18984                         error = EINVAL;
18985                 } else if ((optval == 1) && (rack->r_use_cmp_ack == 0)) {
18986                         rack->r_use_cmp_ack = 1;
18987                         rack->r_mbuf_queue = 1;
18988                         tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
18989                 }
18990                 if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
18991                         rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
18992                 break;
18993         case TCP_SHARED_CWND_TIME_LIMIT:
18994                 RACK_OPTS_INC(tcp_lscwnd);
18995                 if (optval)
18996                         rack->r_limit_scw = 1;
18997                 else
18998                         rack->r_limit_scw = 0;
18999                 break;
19000         case TCP_RACK_PACE_TO_FILL:
19001                 RACK_OPTS_INC(tcp_fillcw);
19002                 if (optval == 0)
19003                         rack->rc_pace_to_cwnd = 0;
19004                 else {
19005                         rack->rc_pace_to_cwnd = 1;
19006                         if (optval > 1)
19007                                 rack->r_fill_less_agg = 1;
19008                 }
19009                 if ((optval >= rack_gp_rtt_maxmul) &&
19010                     rack_gp_rtt_maxmul &&
19011                     (optval < 0xf)) {
19012                         rack->rc_pace_fill_if_rttin_range = 1;
19013                         rack->rtt_limit_mul = optval;
19014                 } else {
19015                         rack->rc_pace_fill_if_rttin_range = 0;
19016                         rack->rtt_limit_mul = 0;
19017                 }
19018                 break;
19019         case TCP_RACK_NO_PUSH_AT_MAX:
19020                 RACK_OPTS_INC(tcp_npush);
19021                 if (optval == 0)
19022                         rack->r_ctl.rc_no_push_at_mrtt = 0;
19023                 else if (optval < 0xff)
19024                         rack->r_ctl.rc_no_push_at_mrtt = optval;
19025                 else
19026                         error = EINVAL;
19027                 break;
19028         case TCP_SHARED_CWND_ENABLE:
19029                 RACK_OPTS_INC(tcp_rack_scwnd);
19030                 if (optval == 0)
19031                         rack->rack_enable_scwnd = 0;
19032                 else
19033                         rack->rack_enable_scwnd = 1;
19034                 break;
19035         case TCP_RACK_MBUF_QUEUE:
19036                 /* Now do we use the LRO mbuf-queue feature */
19037                 RACK_OPTS_INC(tcp_rack_mbufq);
19038                 if (optval || rack->r_use_cmp_ack)
19039                         rack->r_mbuf_queue = 1;
19040                 else
19041                         rack->r_mbuf_queue = 0;
19042                 if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
19043                         tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19044                 else
19045                         tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19046                 break;
19047         case TCP_RACK_NONRXT_CFG_RATE:
19048                 RACK_OPTS_INC(tcp_rack_cfg_rate);
19049                 if (optval == 0)
19050                         rack->rack_rec_nonrxt_use_cr = 0;
19051                 else
19052                         rack->rack_rec_nonrxt_use_cr = 1;
19053                 break;
19054         case TCP_NO_PRR:
19055                 RACK_OPTS_INC(tcp_rack_noprr);
19056                 if (optval == 0)
19057                         rack->rack_no_prr = 0;
19058                 else if (optval == 1)
19059                         rack->rack_no_prr = 1;
19060                 else if (optval == 2)
19061                         rack->no_prr_addback = 1;
19062                 else
19063                         error = EINVAL;
19064                 break;
19065         case TCP_TIMELY_DYN_ADJ:
19066                 RACK_OPTS_INC(tcp_timely_dyn);
19067                 if (optval == 0)
19068                         rack->rc_gp_dyn_mul = 0;
19069                 else {
19070                         rack->rc_gp_dyn_mul = 1;
19071                         if (optval >= 100) {
19072                                 /*
19073                                  * If the user sets something 100 or more
19074                                  * its the gp_ca value.
19075                                  */
19076                                 rack->r_ctl.rack_per_of_gp_ca  = optval;
19077                         }
19078                 }
19079                 break;
19080         case TCP_RACK_DO_DETECTION:
19081                 RACK_OPTS_INC(tcp_rack_do_detection);
19082                 if (optval == 0)
19083                         rack->do_detection = 0;
19084                 else
19085                         rack->do_detection = 1;
19086                 break;
19087         case TCP_RACK_TLP_USE:
19088                 if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) {
19089                         error = EINVAL;
19090                         break;
19091                 }
19092                 RACK_OPTS_INC(tcp_tlp_use);
19093                 rack->rack_tlp_threshold_use = optval;
19094                 break;
19095         case TCP_RACK_TLP_REDUCE:
19096                 /* RACK TLP cwnd reduction (bool) */
19097                 RACK_OPTS_INC(tcp_rack_tlp_reduce);
19098                 rack->r_ctl.rc_tlp_cwnd_reduce = optval;
19099                 break;
19100         /*  Pacing related ones */
19101         case TCP_RACK_PACE_ALWAYS:
19102                 /*
19103                  * zero is old rack method, 1 is new
19104                  * method using a pacing rate.
19105                  */
19106                 RACK_OPTS_INC(tcp_rack_pace_always);
19107                 if (optval > 0) {
19108                         if (rack->rc_always_pace) {
19109                                 error = EALREADY;
19110                                 break;
19111                         } else if (tcp_can_enable_pacing()) {
19112                                 rack->rc_always_pace = 1;
19113                                 if (rack->use_fixed_rate || rack->gp_ready)
19114                                         rack_set_cc_pacing(rack);
19115                         }
19116                         else {
19117                                 error = ENOSPC;
19118                                 break;
19119                         }
19120                 } else {
19121                         if (rack->rc_always_pace) {
19122                                 tcp_decrement_paced_conn();
19123                                 rack->rc_always_pace = 0;
19124                                 rack_undo_cc_pacing(rack);
19125                         }
19126                 }
19127                 if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
19128                         tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19129                 else
19130                         tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19131                 /* A rate may be set irate or other, if so set seg size */
19132                 rack_update_seg(rack);
19133                 break;
19134         case TCP_BBR_RACK_INIT_RATE:
19135                 RACK_OPTS_INC(tcp_initial_rate);
19136                 val = optval;
19137                 /* Change from kbits per second to bytes per second */
19138                 val *= 1000;
19139                 val /= 8;
19140                 rack->r_ctl.init_rate = val;
19141                 if (rack->rc_init_win != rack_default_init_window) {
19142                         uint32_t win, snt;
19143
19144                         /*
19145                          * Options don't always get applied
19146                          * in the order you think. So in order
19147                          * to assure we update a cwnd we need
19148                          * to check and see if we are still
19149                          * where we should raise the cwnd.
19150                          */
19151                         win = rc_init_window(rack);
19152                         if (SEQ_GT(tp->snd_max, tp->iss))
19153                                 snt = tp->snd_max - tp->iss;
19154                         else
19155                                 snt = 0;
19156                         if ((snt < win) &&
19157                             (tp->snd_cwnd < win))
19158                                 tp->snd_cwnd = win;
19159                 }
19160                 if (rack->rc_always_pace)
19161                         rack_update_seg(rack);
19162                 break;
19163         case TCP_BBR_IWINTSO:
19164                 RACK_OPTS_INC(tcp_initial_win);
19165                 if (optval && (optval <= 0xff)) {
19166                         uint32_t win, snt;
19167
19168                         rack->rc_init_win = optval;
19169                         win = rc_init_window(rack);
19170                         if (SEQ_GT(tp->snd_max, tp->iss))
19171                                 snt = tp->snd_max - tp->iss;
19172                         else
19173                                 snt = 0;
19174                         if ((snt < win) &&
19175                             (tp->t_srtt |
19176 #ifdef NETFLIX_PEAKRATE
19177                              tp->t_maxpeakrate |
19178 #endif
19179                              rack->r_ctl.init_rate)) {
19180                                 /*
19181                                  * We are not past the initial window
19182                                  * and we have some bases for pacing,
19183                                  * so we need to possibly adjust up
19184                                  * the cwnd. Note even if we don't set
19185                                  * the cwnd, its still ok to raise the rc_init_win
19186                                  * which can be used coming out of idle when we
19187                                  * would have a rate.
19188                                  */
19189                                 if (tp->snd_cwnd < win)
19190                                         tp->snd_cwnd = win;
19191                         }
19192                         if (rack->rc_always_pace)
19193                                 rack_update_seg(rack);
19194                 } else
19195                         error = EINVAL;
19196                 break;
19197         case TCP_RACK_FORCE_MSEG:
19198                 RACK_OPTS_INC(tcp_rack_force_max_seg);
19199                 if (optval)
19200                         rack->rc_force_max_seg = 1;
19201                 else
19202                         rack->rc_force_max_seg = 0;
19203                 break;
19204         case TCP_RACK_PACE_MAX_SEG:
19205                 /* Max segments size in a pace in bytes */
19206                 RACK_OPTS_INC(tcp_rack_max_seg);
19207                 rack->rc_user_set_max_segs = optval;
19208                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
19209                 break;
19210         case TCP_RACK_PACE_RATE_REC:
19211                 /* Set the fixed pacing rate in Bytes per second ca */
19212                 RACK_OPTS_INC(tcp_rack_pace_rate_rec);
19213                 rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19214                 if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
19215                         rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19216                 if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
19217                         rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19218                 rack->use_fixed_rate = 1;
19219                 if (rack->rc_always_pace)
19220                         rack_set_cc_pacing(rack);
19221                 rack_log_pacing_delay_calc(rack,
19222                                            rack->r_ctl.rc_fixed_pacing_rate_ss,
19223                                            rack->r_ctl.rc_fixed_pacing_rate_ca,
19224                                            rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19225                                            __LINE__, NULL);
19226                 break;
19227
19228         case TCP_RACK_PACE_RATE_SS:
19229                 /* Set the fixed pacing rate in Bytes per second ca */
19230                 RACK_OPTS_INC(tcp_rack_pace_rate_ss);
19231                 rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19232                 if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
19233                         rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19234                 if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
19235                         rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19236                 rack->use_fixed_rate = 1;
19237                 if (rack->rc_always_pace)
19238                         rack_set_cc_pacing(rack);
19239                 rack_log_pacing_delay_calc(rack,
19240                                            rack->r_ctl.rc_fixed_pacing_rate_ss,
19241                                            rack->r_ctl.rc_fixed_pacing_rate_ca,
19242                                            rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19243                                            __LINE__, NULL);
19244                 break;
19245
19246         case TCP_RACK_PACE_RATE_CA:
19247                 /* Set the fixed pacing rate in Bytes per second ca */
19248                 RACK_OPTS_INC(tcp_rack_pace_rate_ca);
19249                 rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19250                 if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
19251                         rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19252                 if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
19253                         rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19254                 rack->use_fixed_rate = 1;
19255                 if (rack->rc_always_pace)
19256                         rack_set_cc_pacing(rack);
19257                 rack_log_pacing_delay_calc(rack,
19258                                            rack->r_ctl.rc_fixed_pacing_rate_ss,
19259                                            rack->r_ctl.rc_fixed_pacing_rate_ca,
19260                                            rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19261                                            __LINE__, NULL);
19262                 break;
19263         case TCP_RACK_GP_INCREASE_REC:
19264                 RACK_OPTS_INC(tcp_gp_inc_rec);
19265                 rack->r_ctl.rack_per_of_gp_rec = optval;
19266                 rack_log_pacing_delay_calc(rack,
19267                                            rack->r_ctl.rack_per_of_gp_ss,
19268                                            rack->r_ctl.rack_per_of_gp_ca,
19269                                            rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
19270                                            __LINE__, NULL);
19271                 break;
19272         case TCP_RACK_GP_INCREASE_CA:
19273                 RACK_OPTS_INC(tcp_gp_inc_ca);
19274                 ca = optval;
19275                 if (ca < 100) {
19276                         /*
19277                          * We don't allow any reduction
19278                          * over the GP b/w.
19279                          */
19280                         error = EINVAL;
19281                         break;
19282                 }
19283                 rack->r_ctl.rack_per_of_gp_ca = ca;
19284                 rack_log_pacing_delay_calc(rack,
19285                                            rack->r_ctl.rack_per_of_gp_ss,
19286                                            rack->r_ctl.rack_per_of_gp_ca,
19287                                            rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
19288                                            __LINE__, NULL);
19289                 break;
19290         case TCP_RACK_GP_INCREASE_SS:
19291                 RACK_OPTS_INC(tcp_gp_inc_ss);
19292                 ss = optval;
19293                 if (ss < 100) {
19294                         /*
19295                          * We don't allow any reduction
19296                          * over the GP b/w.
19297                          */
19298                         error = EINVAL;
19299                         break;
19300                 }
19301                 rack->r_ctl.rack_per_of_gp_ss = ss;
19302                 rack_log_pacing_delay_calc(rack,
19303                                            rack->r_ctl.rack_per_of_gp_ss,
19304                                            rack->r_ctl.rack_per_of_gp_ca,
19305                                            rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
19306                                            __LINE__, NULL);
19307                 break;
19308         case TCP_RACK_RR_CONF:
19309                 RACK_OPTS_INC(tcp_rack_rrr_no_conf_rate);
19310                 if (optval && optval <= 3)
19311                         rack->r_rr_config = optval;
19312                 else
19313                         rack->r_rr_config = 0;
19314                 break;
19315         case TCP_HDWR_RATE_CAP:
19316                 RACK_OPTS_INC(tcp_hdwr_rate_cap);
19317                 if (optval) {
19318                         if (rack->r_rack_hw_rate_caps == 0)
19319                                 rack->r_rack_hw_rate_caps = 1;
19320                         else
19321                                 error = EALREADY;
19322                 } else {
19323                         rack->r_rack_hw_rate_caps = 0;
19324                 }
19325                 break;
19326         case TCP_BBR_HDWR_PACE:
19327                 RACK_OPTS_INC(tcp_hdwr_pacing);
19328                 if (optval){
19329                         if (rack->rack_hdrw_pacing == 0) {
19330                                 rack->rack_hdw_pace_ena = 1;
19331                                 rack->rack_attempt_hdwr_pace = 0;
19332                         } else
19333                                 error = EALREADY;
19334                 } else {
19335                         rack->rack_hdw_pace_ena = 0;
19336 #ifdef RATELIMIT
19337                         if (rack->r_ctl.crte != NULL) {
19338                                 rack->rack_hdrw_pacing = 0;
19339                                 rack->rack_attempt_hdwr_pace = 0;
19340                                 tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
19341                                 rack->r_ctl.crte = NULL;
19342                         }
19343 #endif
19344                 }
19345                 break;
19346         /*  End Pacing related ones */
19347         case TCP_RACK_PRR_SENDALOT:
19348                 /* Allow PRR to send more than one seg */
19349                 RACK_OPTS_INC(tcp_rack_prr_sendalot);
19350                 rack->r_ctl.rc_prr_sendalot = optval;
19351                 break;
19352         case TCP_RACK_MIN_TO:
19353                 /* Minimum time between rack t-o's in ms */
19354                 RACK_OPTS_INC(tcp_rack_min_to);
19355                 rack->r_ctl.rc_min_to = optval;
19356                 break;
19357         case TCP_RACK_EARLY_SEG:
19358                 /* If early recovery max segments */
19359                 RACK_OPTS_INC(tcp_rack_early_seg);
19360                 rack->r_ctl.rc_early_recovery_segs = optval;
19361                 break;
19362         case TCP_RACK_REORD_THRESH:
19363                 /* RACK reorder threshold (shift amount) */
19364                 RACK_OPTS_INC(tcp_rack_reord_thresh);
19365                 if ((optval > 0) && (optval < 31))
19366                         rack->r_ctl.rc_reorder_shift = optval;
19367                 else
19368                         error = EINVAL;
19369                 break;
19370         case TCP_RACK_REORD_FADE:
19371                 /* Does reordering fade after ms time */
19372                 RACK_OPTS_INC(tcp_rack_reord_fade);
19373                 rack->r_ctl.rc_reorder_fade = optval;
19374                 break;
19375         case TCP_RACK_TLP_THRESH:
19376                 /* RACK TLP theshold i.e. srtt+(srtt/N) */
19377                 RACK_OPTS_INC(tcp_rack_tlp_thresh);
19378                 if (optval)
19379                         rack->r_ctl.rc_tlp_threshold = optval;
19380                 else
19381                         error = EINVAL;
19382                 break;
19383         case TCP_BBR_USE_RACK_RR:
19384                 RACK_OPTS_INC(tcp_rack_rr);
19385                 if (optval)
19386                         rack->use_rack_rr = 1;
19387                 else
19388                         rack->use_rack_rr = 0;
19389                 break;
19390         case TCP_FAST_RSM_HACK:
19391                 RACK_OPTS_INC(tcp_rack_fastrsm_hack);
19392                 if (optval)
19393                         rack->fast_rsm_hack = 1;
19394                 else
19395                         rack->fast_rsm_hack = 0;
19396                 break;
19397         case TCP_RACK_PKT_DELAY:
19398                 /* RACK added ms i.e. rack-rtt + reord + N */
19399                 RACK_OPTS_INC(tcp_rack_pkt_delay);
19400                 rack->r_ctl.rc_pkt_delay = optval;
19401                 break;
19402         case TCP_DELACK:
19403                 RACK_OPTS_INC(tcp_rack_delayed_ack);
19404                 if (optval == 0)
19405                         tp->t_delayed_ack = 0;
19406                 else
19407                         tp->t_delayed_ack = 1;
19408                 if (tp->t_flags & TF_DELACK) {
19409                         tp->t_flags &= ~TF_DELACK;
19410                         tp->t_flags |= TF_ACKNOW;
19411                         NET_EPOCH_ENTER(et);
19412                         rack_output(tp);
19413                         NET_EPOCH_EXIT(et);
19414                 }
19415                 break;
19416
19417         case TCP_BBR_RACK_RTT_USE:
19418                 RACK_OPTS_INC(tcp_rack_rtt_use);
19419                 if ((optval != USE_RTT_HIGH) &&
19420                     (optval != USE_RTT_LOW) &&
19421                     (optval != USE_RTT_AVG))
19422                         error = EINVAL;
19423                 else
19424                         rack->r_ctl.rc_rate_sample_method = optval;
19425                 break;
19426         case TCP_DATA_AFTER_CLOSE:
19427                 RACK_OPTS_INC(tcp_data_after_close);
19428                 if (optval)
19429                         rack->rc_allow_data_af_clo = 1;
19430                 else
19431                         rack->rc_allow_data_af_clo = 0;
19432                 break;
19433         default:
19434                 break;
19435         }
19436 #ifdef NETFLIX_STATS
19437         tcp_log_socket_option(tp, sopt_name, optval, error);
19438 #endif
19439         return (error);
19440 }
19441
19442
19443 static void
19444 rack_apply_deferred_options(struct tcp_rack *rack)
19445 {
19446         struct deferred_opt_list *dol, *sdol;
19447         uint32_t s_optval;
19448
19449         TAILQ_FOREACH_SAFE(dol, &rack->r_ctl.opt_list, next, sdol) {
19450                 TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
19451                 /* Disadvantage of deferal is you loose the error return */
19452                 s_optval = (uint32_t)dol->optval;
19453                 (void)rack_process_option(rack->rc_tp, rack, dol->optname, s_optval, dol->optval);
19454                 free(dol, M_TCPDO);
19455         }
19456 }
19457
19458 static int
19459 rack_pru_options(struct tcpcb *tp, int flags)
19460 {
19461         if (flags & PRUS_OOB)
19462                 return (EOPNOTSUPP);
19463         return (0);
19464 }
19465
19466 static struct tcp_function_block __tcp_rack = {
19467         .tfb_tcp_block_name = __XSTRING(STACKNAME),
19468         .tfb_tcp_output = rack_output,
19469         .tfb_do_queued_segments = ctf_do_queued_segments,
19470         .tfb_do_segment_nounlock = rack_do_segment_nounlock,
19471         .tfb_tcp_do_segment = rack_do_segment,
19472         .tfb_tcp_ctloutput = rack_ctloutput,
19473         .tfb_tcp_fb_init = rack_init,
19474         .tfb_tcp_fb_fini = rack_fini,
19475         .tfb_tcp_timer_stop_all = rack_stopall,
19476         .tfb_tcp_timer_activate = rack_timer_activate,
19477         .tfb_tcp_timer_active = rack_timer_active,
19478         .tfb_tcp_timer_stop = rack_timer_stop,
19479         .tfb_tcp_rexmit_tmr = rack_remxt_tmr,
19480         .tfb_tcp_handoff_ok = rack_handoff_ok,
19481         .tfb_tcp_mtu_chg = rack_mtu_change,
19482         .tfb_pru_options = rack_pru_options,
19483
19484 };
19485
19486 /*
19487  * rack_ctloutput() must drop the inpcb lock before performing copyin on
19488  * socket option arguments.  When it re-acquires the lock after the copy, it
19489  * has to revalidate that the connection is still valid for the socket
19490  * option.
19491  */
19492 static int
19493 rack_set_sockopt(struct socket *so, struct sockopt *sopt,
19494     struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack)
19495 {
19496         uint64_t loptval;
19497         int32_t error = 0, optval;
19498
19499         switch (sopt->sopt_name) {
19500         case TCP_RACK_TLP_REDUCE:               /*  URL:tlp_reduce */
19501         /*  Pacing related ones */
19502         case TCP_RACK_PACE_ALWAYS:              /*  URL:pace_always */
19503         case TCP_BBR_RACK_INIT_RATE:            /*  URL:irate */
19504         case TCP_BBR_IWINTSO:                   /*  URL:tso_iwin */
19505         case TCP_RACK_PACE_MAX_SEG:             /*  URL:pace_max_seg */
19506         case TCP_RACK_FORCE_MSEG:               /*  URL:force_max_seg */
19507         case TCP_RACK_PACE_RATE_CA:             /*  URL:pr_ca */
19508         case TCP_RACK_PACE_RATE_SS:             /*  URL:pr_ss*/
19509         case TCP_RACK_PACE_RATE_REC:            /*  URL:pr_rec */
19510         case TCP_RACK_GP_INCREASE_CA:           /*  URL:gp_inc_ca */
19511         case TCP_RACK_GP_INCREASE_SS:           /*  URL:gp_inc_ss */
19512         case TCP_RACK_GP_INCREASE_REC:          /*  URL:gp_inc_rec */
19513         case TCP_RACK_RR_CONF:                  /*  URL:rrr_conf */
19514         case TCP_BBR_HDWR_PACE:                 /*  URL:hdwrpace */
19515         case TCP_HDWR_RATE_CAP:                 /*  URL: hdwrcap boolean */
19516         case TCP_PACING_RATE_CAP:               /*  URL:cap-- used by side-channel */
19517         case TCP_HDWR_UP_ONLY:                  /*  URL:uponly -- hardware pacing  boolean */
19518        /* End pacing related */
19519         case TCP_FAST_RSM_HACK:                 /*  URL:frsm_hack */
19520         case TCP_DELACK:                        /*  URL:delack (in base TCP i.e. tcp_hints along with cc etc ) */
19521         case TCP_RACK_PRR_SENDALOT:             /*  URL:prr_sendalot */
19522         case TCP_RACK_MIN_TO:                   /*  URL:min_to */
19523         case TCP_RACK_EARLY_SEG:                /*  URL:early_seg */
19524         case TCP_RACK_REORD_THRESH:             /*  URL:reord_thresh */
19525         case TCP_RACK_REORD_FADE:               /*  URL:reord_fade */
19526         case TCP_RACK_TLP_THRESH:               /*  URL:tlp_thresh */
19527         case TCP_RACK_PKT_DELAY:                /*  URL:pkt_delay */
19528         case TCP_RACK_TLP_USE:                  /*  URL:tlp_use */
19529         case TCP_BBR_RACK_RTT_USE:              /*  URL:rttuse */
19530         case TCP_BBR_USE_RACK_RR:               /*  URL:rackrr */
19531         case TCP_RACK_DO_DETECTION:             /*  URL:detect */
19532         case TCP_NO_PRR:                        /*  URL:noprr */
19533         case TCP_TIMELY_DYN_ADJ:                /*  URL:dynamic */
19534         case TCP_DATA_AFTER_CLOSE:              /*  no URL */
19535         case TCP_RACK_NONRXT_CFG_RATE:          /*  URL:nonrxtcr */
19536         case TCP_SHARED_CWND_ENABLE:            /*  URL:scwnd */
19537         case TCP_RACK_MBUF_QUEUE:               /*  URL:mqueue */
19538         case TCP_RACK_NO_PUSH_AT_MAX:           /*  URL:npush */
19539         case TCP_RACK_PACE_TO_FILL:             /*  URL:fillcw */
19540         case TCP_SHARED_CWND_TIME_LIMIT:        /*  URL:lscwnd */
19541         case TCP_RACK_PROFILE:                  /*  URL:profile */
19542         case TCP_USE_CMP_ACKS:                  /*  URL:cmpack */
19543         case TCP_RACK_ABC_VAL:                  /*  URL:labc */
19544         case TCP_REC_ABC_VAL:                   /*  URL:reclabc */
19545         case TCP_RACK_MEASURE_CNT:              /*  URL:measurecnt */
19546         case TCP_DEFER_OPTIONS:                 /*  URL:defer */
19547         case TCP_RACK_PACING_BETA:              /*  URL:pacing_beta */
19548         case TCP_RACK_PACING_BETA_ECN:          /*  URL:pacing_beta_ecn */
19549         case TCP_RACK_TIMER_SLOP:               /*  URL:timer_slop */
19550                 break;
19551         default:
19552                 /* Filter off all unknown options to the base stack */
19553                 return (tcp_default_ctloutput(so, sopt, inp, tp));
19554                 break;
19555         }
19556         INP_WUNLOCK(inp);
19557         if (sopt->sopt_name == TCP_PACING_RATE_CAP) {
19558                 error = sooptcopyin(sopt, &loptval, sizeof(loptval), sizeof(loptval));
19559                 /*
19560                  * We truncate it down to 32 bits for the socket-option trace this
19561                  * means rates > 34Gbps won't show right, but thats probably ok.
19562                  */
19563                 optval = (uint32_t)loptval;
19564         } else {
19565                 error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
19566                 /* Save it in 64 bit form too */
19567                 loptval = optval;
19568         }
19569         if (error)
19570                 return (error);
19571         INP_WLOCK(inp);
19572         if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
19573                 INP_WUNLOCK(inp);
19574                 return (ECONNRESET);
19575         }
19576         if (tp->t_fb != &__tcp_rack) {
19577                 INP_WUNLOCK(inp);
19578                 return (ENOPROTOOPT);
19579         }
19580         if (rack->defer_options && (rack->gp_ready == 0) &&
19581             (sopt->sopt_name != TCP_DEFER_OPTIONS) &&
19582             (sopt->sopt_name != TCP_RACK_PACING_BETA) &&
19583             (sopt->sopt_name != TCP_RACK_PACING_BETA_ECN) &&
19584             (sopt->sopt_name != TCP_RACK_MEASURE_CNT)) {
19585                 /* Options are beind deferred */
19586                 if (rack_add_deferred_option(rack, sopt->sopt_name, loptval)) {
19587                         INP_WUNLOCK(inp);
19588                         return (0);
19589                 } else {
19590                         /* No memory to defer, fail */
19591                         INP_WUNLOCK(inp);
19592                         return (ENOMEM);
19593                 }
19594         }
19595         error = rack_process_option(tp, rack, sopt->sopt_name, optval, loptval);
19596         INP_WUNLOCK(inp);
19597         return (error);
19598 }
19599
19600 static void
19601 rack_fill_info(struct tcpcb *tp, struct tcp_info *ti)
19602 {
19603
19604         INP_WLOCK_ASSERT(tp->t_inpcb);
19605         bzero(ti, sizeof(*ti));
19606
19607         ti->tcpi_state = tp->t_state;
19608         if ((tp->t_flags & TF_REQ_TSTMP) && (tp->t_flags & TF_RCVD_TSTMP))
19609                 ti->tcpi_options |= TCPI_OPT_TIMESTAMPS;
19610         if (tp->t_flags & TF_SACK_PERMIT)
19611                 ti->tcpi_options |= TCPI_OPT_SACK;
19612         if ((tp->t_flags & TF_REQ_SCALE) && (tp->t_flags & TF_RCVD_SCALE)) {
19613                 ti->tcpi_options |= TCPI_OPT_WSCALE;
19614                 ti->tcpi_snd_wscale = tp->snd_scale;
19615                 ti->tcpi_rcv_wscale = tp->rcv_scale;
19616         }
19617         if (tp->t_flags2 & TF2_ECN_PERMIT)
19618                 ti->tcpi_options |= TCPI_OPT_ECN;
19619         if (tp->t_flags & TF_FASTOPEN)
19620                 ti->tcpi_options |= TCPI_OPT_TFO;
19621         /* still kept in ticks is t_rcvtime */
19622         ti->tcpi_last_data_recv = ((uint32_t)ticks - tp->t_rcvtime) * tick;
19623         /* Since we hold everything in precise useconds this is easy */
19624         ti->tcpi_rtt = tp->t_srtt;
19625         ti->tcpi_rttvar = tp->t_rttvar;
19626         ti->tcpi_rto = tp->t_rxtcur;
19627         ti->tcpi_snd_ssthresh = tp->snd_ssthresh;
19628         ti->tcpi_snd_cwnd = tp->snd_cwnd;
19629         /*
19630          * FreeBSD-specific extension fields for tcp_info.
19631          */
19632         ti->tcpi_rcv_space = tp->rcv_wnd;
19633         ti->tcpi_rcv_nxt = tp->rcv_nxt;
19634         ti->tcpi_snd_wnd = tp->snd_wnd;
19635         ti->tcpi_snd_bwnd = 0;          /* Unused, kept for compat. */
19636         ti->tcpi_snd_nxt = tp->snd_nxt;
19637         ti->tcpi_snd_mss = tp->t_maxseg;
19638         ti->tcpi_rcv_mss = tp->t_maxseg;
19639         ti->tcpi_snd_rexmitpack = tp->t_sndrexmitpack;
19640         ti->tcpi_rcv_ooopack = tp->t_rcvoopack;
19641         ti->tcpi_snd_zerowin = tp->t_sndzerowin;
19642 #ifdef NETFLIX_STATS
19643         ti->tcpi_total_tlp = tp->t_sndtlppack;
19644         ti->tcpi_total_tlp_bytes = tp->t_sndtlpbyte;
19645         memcpy(&ti->tcpi_rxsyninfo, &tp->t_rxsyninfo, sizeof(struct tcpsyninfo));
19646 #endif
19647 #ifdef TCP_OFFLOAD
19648         if (tp->t_flags & TF_TOE) {
19649                 ti->tcpi_options |= TCPI_OPT_TOE;
19650                 tcp_offload_tcp_info(tp, ti);
19651         }
19652 #endif
19653 }
19654
19655 static int
19656 rack_get_sockopt(struct socket *so, struct sockopt *sopt,
19657     struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack)
19658 {
19659         int32_t error, optval;
19660         uint64_t val, loptval;
19661         struct  tcp_info ti;
19662         /*
19663          * Because all our options are either boolean or an int, we can just
19664          * pull everything into optval and then unlock and copy. If we ever
19665          * add a option that is not a int, then this will have quite an
19666          * impact to this routine.
19667          */
19668         error = 0;
19669         switch (sopt->sopt_name) {
19670         case TCP_INFO:
19671                 /* First get the info filled */
19672                 rack_fill_info(tp, &ti);
19673                 /* Fix up the rtt related fields if needed */
19674                 INP_WUNLOCK(inp);
19675                 error = sooptcopyout(sopt, &ti, sizeof ti);
19676                 return (error);
19677         /*
19678          * Beta is the congestion control value for NewReno that influences how
19679          * much of a backoff happens when loss is detected. It is normally set
19680          * to 50 for 50% i.e. the cwnd is reduced to 50% of its previous value
19681          * when you exit recovery.
19682          */
19683         case TCP_RACK_PACING_BETA:
19684                 if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0)
19685                         error = EINVAL;
19686                 else if (rack->rc_pacing_cc_set == 0)
19687                         optval = rack->r_ctl.rc_saved_beta.beta;
19688                 else {
19689                         /*
19690                          * Reach out into the CC data and report back what
19691                          * I have previously set. Yeah it looks hackish but
19692                          * we don't want to report the saved values.
19693                          */
19694                         if (tp->ccv->cc_data)
19695                                 optval = ((struct newreno *)tp->ccv->cc_data)->beta;
19696                         else
19697                                 error = EINVAL;
19698                 }
19699                 break;
19700                 /*
19701                  * Beta_ecn is the congestion control value for NewReno that influences how
19702                  * much of a backoff happens when a ECN mark is detected. It is normally set
19703                  * to 80 for 80% i.e. the cwnd is reduced by 20% of its previous value when
19704                  * you exit recovery. Note that classic ECN has a beta of 50, it is only
19705                  * ABE Ecn that uses this "less" value, but we do too with pacing :)
19706                  */
19707
19708         case TCP_RACK_PACING_BETA_ECN:
19709                 if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0)
19710                         error = EINVAL;
19711                 else if (rack->rc_pacing_cc_set == 0)
19712                         optval = rack->r_ctl.rc_saved_beta.beta_ecn;
19713                 else {
19714                         /*
19715                          * Reach out into the CC data and report back what
19716                          * I have previously set. Yeah it looks hackish but
19717                          * we don't want to report the saved values.
19718                          */
19719                         if (tp->ccv->cc_data)
19720                                 optval = ((struct newreno *)tp->ccv->cc_data)->beta_ecn;
19721                         else
19722                                 error = EINVAL;
19723                 }
19724                 break;
19725         case TCP_FAST_RSM_HACK:
19726                 optval = rack->fast_rsm_hack;
19727                 break;
19728         case TCP_DEFER_OPTIONS:
19729                 optval = rack->defer_options;
19730                 break;
19731         case TCP_RACK_MEASURE_CNT:
19732                 optval = rack->r_ctl.req_measurements;
19733                 break;
19734         case TCP_REC_ABC_VAL:
19735                 optval = rack->r_use_labc_for_rec;
19736                 break;
19737         case TCP_RACK_ABC_VAL:
19738                 optval = rack->rc_labc;
19739                 break;
19740         case TCP_HDWR_UP_ONLY:
19741                 optval= rack->r_up_only;
19742                 break;
19743         case TCP_PACING_RATE_CAP:
19744                 loptval = rack->r_ctl.bw_rate_cap;
19745                 break;
19746         case TCP_RACK_PROFILE:
19747                 /* You cannot retrieve a profile, its write only */
19748                 error = EINVAL;
19749                 break;
19750         case TCP_USE_CMP_ACKS:
19751                 optval = rack->r_use_cmp_ack;
19752                 break;
19753         case TCP_RACK_PACE_TO_FILL:
19754                 optval = rack->rc_pace_to_cwnd;
19755                 if (optval && rack->r_fill_less_agg)
19756                         optval++;
19757                 break;
19758         case TCP_RACK_NO_PUSH_AT_MAX:
19759                 optval = rack->r_ctl.rc_no_push_at_mrtt;
19760                 break;
19761         case TCP_SHARED_CWND_ENABLE:
19762                 optval = rack->rack_enable_scwnd;
19763                 break;
19764         case TCP_RACK_NONRXT_CFG_RATE:
19765                 optval = rack->rack_rec_nonrxt_use_cr;
19766                 break;
19767         case TCP_NO_PRR:
19768                 if (rack->rack_no_prr  == 1)
19769                         optval = 1;
19770                 else if (rack->no_prr_addback == 1)
19771                         optval = 2;
19772                 else
19773                         optval = 0;
19774                 break;
19775         case TCP_RACK_DO_DETECTION:
19776                 optval = rack->do_detection;
19777                 break;
19778         case TCP_RACK_MBUF_QUEUE:
19779                 /* Now do we use the LRO mbuf-queue feature */
19780                 optval = rack->r_mbuf_queue;
19781                 break;
19782         case TCP_TIMELY_DYN_ADJ:
19783                 optval = rack->rc_gp_dyn_mul;
19784                 break;
19785         case TCP_BBR_IWINTSO:
19786                 optval = rack->rc_init_win;
19787                 break;
19788         case TCP_RACK_TLP_REDUCE:
19789                 /* RACK TLP cwnd reduction (bool) */
19790                 optval = rack->r_ctl.rc_tlp_cwnd_reduce;
19791                 break;
19792         case TCP_BBR_RACK_INIT_RATE:
19793                 val = rack->r_ctl.init_rate;
19794                 /* convert to kbits per sec */
19795                 val *= 8;
19796                 val /= 1000;
19797                 optval = (uint32_t)val;
19798                 break;
19799         case TCP_RACK_FORCE_MSEG:
19800                 optval = rack->rc_force_max_seg;
19801                 break;
19802         case TCP_RACK_PACE_MAX_SEG:
19803                 /* Max segments in a pace */
19804                 optval = rack->rc_user_set_max_segs;
19805                 break;
19806         case TCP_RACK_PACE_ALWAYS:
19807                 /* Use the always pace method */
19808                 optval = rack->rc_always_pace;
19809                 break;
19810         case TCP_RACK_PRR_SENDALOT:
19811                 /* Allow PRR to send more than one seg */
19812                 optval = rack->r_ctl.rc_prr_sendalot;
19813                 break;
19814         case TCP_RACK_MIN_TO:
19815                 /* Minimum time between rack t-o's in ms */
19816                 optval = rack->r_ctl.rc_min_to;
19817                 break;
19818         case TCP_RACK_EARLY_SEG:
19819                 /* If early recovery max segments */
19820                 optval = rack->r_ctl.rc_early_recovery_segs;
19821                 break;
19822         case TCP_RACK_REORD_THRESH:
19823                 /* RACK reorder threshold (shift amount) */
19824                 optval = rack->r_ctl.rc_reorder_shift;
19825                 break;
19826         case TCP_RACK_REORD_FADE:
19827                 /* Does reordering fade after ms time */
19828                 optval = rack->r_ctl.rc_reorder_fade;
19829                 break;
19830         case TCP_BBR_USE_RACK_RR:
19831                 /* Do we use the rack cheat for rxt */
19832                 optval = rack->use_rack_rr;
19833                 break;
19834         case TCP_RACK_RR_CONF:
19835                 optval = rack->r_rr_config;
19836                 break;
19837         case TCP_HDWR_RATE_CAP:
19838                 optval = rack->r_rack_hw_rate_caps;
19839                 break;
19840         case TCP_BBR_HDWR_PACE:
19841                 optval = rack->rack_hdw_pace_ena;
19842                 break;
19843         case TCP_RACK_TLP_THRESH:
19844                 /* RACK TLP theshold i.e. srtt+(srtt/N) */
19845                 optval = rack->r_ctl.rc_tlp_threshold;
19846                 break;
19847         case TCP_RACK_PKT_DELAY:
19848                 /* RACK added ms i.e. rack-rtt + reord + N */
19849                 optval = rack->r_ctl.rc_pkt_delay;
19850                 break;
19851         case TCP_RACK_TLP_USE:
19852                 optval = rack->rack_tlp_threshold_use;
19853                 break;
19854         case TCP_RACK_PACE_RATE_CA:
19855                 optval = rack->r_ctl.rc_fixed_pacing_rate_ca;
19856                 break;
19857         case TCP_RACK_PACE_RATE_SS:
19858                 optval = rack->r_ctl.rc_fixed_pacing_rate_ss;
19859                 break;
19860         case TCP_RACK_PACE_RATE_REC:
19861                 optval = rack->r_ctl.rc_fixed_pacing_rate_rec;
19862                 break;
19863         case TCP_RACK_GP_INCREASE_SS:
19864                 optval = rack->r_ctl.rack_per_of_gp_ca;
19865                 break;
19866         case TCP_RACK_GP_INCREASE_CA:
19867                 optval = rack->r_ctl.rack_per_of_gp_ss;
19868                 break;
19869         case TCP_BBR_RACK_RTT_USE:
19870                 optval = rack->r_ctl.rc_rate_sample_method;
19871                 break;
19872         case TCP_DELACK:
19873                 optval = tp->t_delayed_ack;
19874                 break;
19875         case TCP_DATA_AFTER_CLOSE:
19876                 optval = rack->rc_allow_data_af_clo;
19877                 break;
19878         case TCP_SHARED_CWND_TIME_LIMIT:
19879                 optval = rack->r_limit_scw;
19880                 break;
19881         case TCP_RACK_TIMER_SLOP:
19882                 optval = rack->r_ctl.timer_slop;
19883                 break;
19884         default:
19885                 return (tcp_default_ctloutput(so, sopt, inp, tp));
19886                 break;
19887         }
19888         INP_WUNLOCK(inp);
19889         if (error == 0) {
19890                 if (TCP_PACING_RATE_CAP)
19891                         error = sooptcopyout(sopt, &loptval, sizeof loptval);
19892                 else
19893                         error = sooptcopyout(sopt, &optval, sizeof optval);
19894         }
19895         return (error);
19896 }
19897
19898 static int
19899 rack_ctloutput(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp)
19900 {
19901         int32_t error = EINVAL;
19902         struct tcp_rack *rack;
19903
19904         rack = (struct tcp_rack *)tp->t_fb_ptr;
19905         if (rack == NULL) {
19906                 /* Huh? */
19907                 goto out;
19908         }
19909         if (sopt->sopt_dir == SOPT_SET) {
19910                 return (rack_set_sockopt(so, sopt, inp, tp, rack));
19911         } else if (sopt->sopt_dir == SOPT_GET) {
19912                 return (rack_get_sockopt(so, sopt, inp, tp, rack));
19913         }
19914 out:
19915         INP_WUNLOCK(inp);
19916         return (error);
19917 }
19918
19919 static const char *rack_stack_names[] = {
19920         __XSTRING(STACKNAME),
19921 #ifdef STACKALIAS
19922         __XSTRING(STACKALIAS),
19923 #endif
19924 };
19925
19926 static int
19927 rack_ctor(void *mem, int32_t size, void *arg, int32_t how)
19928 {
19929         memset(mem, 0, size);
19930         return (0);
19931 }
19932
19933 static void
19934 rack_dtor(void *mem, int32_t size, void *arg)
19935 {
19936
19937 }
19938
19939 static bool rack_mod_inited = false;
19940
19941 static int
19942 tcp_addrack(module_t mod, int32_t type, void *data)
19943 {
19944         int32_t err = 0;
19945         int num_stacks;
19946
19947         switch (type) {
19948         case MOD_LOAD:
19949                 rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
19950                     sizeof(struct rack_sendmap),
19951                     rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0);
19952
19953                 rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
19954                     sizeof(struct tcp_rack),
19955                     rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
19956
19957                 sysctl_ctx_init(&rack_sysctl_ctx);
19958                 rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
19959                     SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
19960                     OID_AUTO,
19961 #ifdef STACKALIAS
19962                     __XSTRING(STACKALIAS),
19963 #else
19964                     __XSTRING(STACKNAME),
19965 #endif
19966                     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
19967                     "");
19968                 if (rack_sysctl_root == NULL) {
19969                         printf("Failed to add sysctl node\n");
19970                         err = EFAULT;
19971                         goto free_uma;
19972                 }
19973                 rack_init_sysctls();
19974                 num_stacks = nitems(rack_stack_names);
19975                 err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK,
19976                     rack_stack_names, &num_stacks);
19977                 if (err) {
19978                         printf("Failed to register %s stack name for "
19979                             "%s module\n", rack_stack_names[num_stacks],
19980                             __XSTRING(MODNAME));
19981                         sysctl_ctx_free(&rack_sysctl_ctx);
19982 free_uma:
19983                         uma_zdestroy(rack_zone);
19984                         uma_zdestroy(rack_pcb_zone);
19985                         rack_counter_destroy();
19986                         printf("Failed to register rack module -- err:%d\n", err);
19987                         return (err);
19988                 }
19989                 tcp_lro_reg_mbufq();
19990                 rack_mod_inited = true;
19991                 break;
19992         case MOD_QUIESCE:
19993                 err = deregister_tcp_functions(&__tcp_rack, true, false);
19994                 break;
19995         case MOD_UNLOAD:
19996                 err = deregister_tcp_functions(&__tcp_rack, false, true);
19997                 if (err == EBUSY)
19998                         break;
19999                 if (rack_mod_inited) {
20000                         uma_zdestroy(rack_zone);
20001                         uma_zdestroy(rack_pcb_zone);
20002                         sysctl_ctx_free(&rack_sysctl_ctx);
20003                         rack_counter_destroy();
20004                         rack_mod_inited = false;
20005                 }
20006                 tcp_lro_dereg_mbufq();
20007                 err = 0;
20008                 break;
20009         default:
20010                 return (EOPNOTSUPP);
20011         }
20012         return (err);
20013 }
20014
20015 static moduledata_t tcp_rack = {
20016         .name = __XSTRING(MODNAME),
20017         .evhand = tcp_addrack,
20018         .priv = 0
20019 };
20020
20021 MODULE_VERSION(MODNAME, 1);
20022 DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
20023 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);