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) do { \
 412         (tv) = (value) + TICKS_2_USEC(tcp_rexmit_slop);  \
 413         if ((u_long)(tv) < (u_long)(tvmin)) \
 414                 (tv) = (tvmin); \
 415         if ((u_long)(tv) > (u_long)(tvmax)) \
 416                 (tv) = (tvmax); \
 417 } while (0)
 418
 419 static void
 420 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line);
 421
 422 static int
 423 rack_process_ack(struct mbuf *m, struct tcphdr *th,
 424     struct socket *so, struct tcpcb *tp, struct tcpopt *to,
 425     uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t * ret_val);
 426 static int
 427 rack_process_data(struct mbuf *m, struct tcphdr *th,
 428     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
 429     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt);
 430 static void
 431 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack,
 432    uint32_t th_ack, uint16_t nsegs, uint16_t type, int32_t recovery);
 433 static struct rack_sendmap *rack_alloc(struct tcp_rack *rack);
 434 static struct rack_sendmap *rack_alloc_limit(struct tcp_rack *rack,
 435     uint8_t limit_type);
 436 static struct rack_sendmap *
 437 rack_check_recovery_mode(struct tcpcb *tp,
 438     uint32_t tsused);
 439 static void
 440 rack_cong_signal(struct tcpcb *tp,
 441                  uint32_t type, uint32_t ack);
 442 static void rack_counter_destroy(void);
 443 static int
 444 rack_ctloutput(struct socket *so, struct sockopt *sopt,
 445     struct inpcb *inp, struct tcpcb *tp);
 446 static int32_t rack_ctor(void *mem, int32_t size, void *arg, int32_t how);
 447 static void
 448 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override);
 449 static void
 450 rack_do_segment(struct mbuf *m, struct tcphdr *th,
 451     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
 452     uint8_t iptos);
 453 static void rack_dtor(void *mem, int32_t size, void *arg);
 454 static void
 455 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
 456     uint32_t flex1, uint32_t flex2,
 457     uint32_t flex3, uint32_t flex4,
 458     uint32_t flex5, uint32_t flex6,
 459     uint16_t flex7, uint8_t mod);
 460 static void
 461 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
 462    uint64_t bw_est, uint64_t bw, uint64_t len_time, int method, int line, struct rack_sendmap *rsm);
 463 static struct rack_sendmap *
 464 rack_find_high_nonack(struct tcp_rack *rack,
 465     struct rack_sendmap *rsm);
 466 static struct rack_sendmap *rack_find_lowest_rsm(struct tcp_rack *rack);
 467 static void rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm);
 468 static void rack_fini(struct tcpcb *tp, int32_t tcb_is_purged);
 469 static int
 470 rack_get_sockopt(struct socket *so, struct sockopt *sopt,
 471     struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack);
 472 static void
 473 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
 474                             tcp_seq th_ack, int line);
 475 static uint32_t
 476 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss);
 477 static int32_t rack_handoff_ok(struct tcpcb *tp);
 478 static int32_t rack_init(struct tcpcb *tp);
 479 static void rack_init_sysctls(void);
 480 static void
 481 rack_log_ack(struct tcpcb *tp, struct tcpopt *to,
 482     struct tcphdr *th, int entered_rec, int dup_ack_struck);
 483 static void
 484 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
 485     uint32_t seq_out, uint8_t th_flags, int32_t err, uint64_t ts,
 486     struct rack_sendmap *hintrsm, uint16_t add_flags, struct mbuf *s_mb, uint32_t s_moff);
 487
 488 static void
 489 rack_log_sack_passed(struct tcpcb *tp, struct tcp_rack *rack,
 490     struct rack_sendmap *rsm);
 491 static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm);
 492 static int32_t rack_output(struct tcpcb *tp);
 493
 494 static uint32_t
 495 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack,
 496     struct sackblk *sack, struct tcpopt *to, struct rack_sendmap **prsm,
 497     uint32_t cts, int *moved_two);
 498 static void rack_post_recovery(struct tcpcb *tp, uint32_t th_seq);
 499 static void rack_remxt_tmr(struct tcpcb *tp);
 500 static int
 501 rack_set_sockopt(struct socket *so, struct sockopt *sopt,
 502     struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack);
 503 static void rack_set_state(struct tcpcb *tp, struct tcp_rack *rack);
 504 static int32_t rack_stopall(struct tcpcb *tp);
 505 static void
 506 rack_timer_activate(struct tcpcb *tp, uint32_t timer_type,
 507     uint32_t delta);
 508 static int32_t rack_timer_active(struct tcpcb *tp, uint32_t timer_type);
 509 static void rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line);
 510 static void rack_timer_stop(struct tcpcb *tp, uint32_t timer_type);
 511 static uint32_t
 512 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
 513     struct rack_sendmap *rsm, uint64_t ts, int32_t * lenp, uint16_t add_flag);
 514 static void
 515 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
 516     struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag);
 517 static int
 518 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
 519     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack);
 520 static int32_t tcp_addrack(module_t mod, int32_t type, void *data);
 521 static int
 522 rack_do_close_wait(struct mbuf *m, struct tcphdr *th,
 523     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 524     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 525 static int
 526 rack_do_closing(struct mbuf *m, struct tcphdr *th,
 527     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 528     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 529 static int
 530 rack_do_established(struct mbuf *m, struct tcphdr *th,
 531     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 532     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 533 static int
 534 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th,
 535     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 536     int32_t tlen, uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos);
 537 static int
 538 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th,
 539     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 540     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 541 static int
 542 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th,
 543     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 544     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 545 static int
 546 rack_do_lastack(struct mbuf *m, struct tcphdr *th,
 547     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 548     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 549 static int
 550 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th,
 551     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 552     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 553 static int
 554 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th,
 555     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 556     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 557 struct rack_sendmap *
 558 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack,
 559     uint32_t tsused);
 560 static void tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt,
 561     uint32_t len, uint32_t us_tim, int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt);
 562 static void
 563      tcp_rack_partialack(struct tcpcb *tp);
 564 static int
 565 rack_set_profile(struct tcp_rack *rack, int prof);
 566 static void
 567 rack_apply_deferred_options(struct tcp_rack *rack);
 568
 569 int32_t rack_clear_counter=0;
 570
 571 static void
 572 rack_set_cc_pacing(struct tcp_rack *rack)
 573 {
 574         struct sockopt sopt;
 575         struct cc_newreno_opts opt;
 576         struct newreno old, *ptr;
 577         struct tcpcb *tp;
 578         int error;
 579
 580         if (rack->rc_pacing_cc_set)
 581                 return;
 582
 583         tp = rack->rc_tp;
 584         if (tp->cc_algo == NULL) {
 585                 /* Tcb is leaving */
 586                 printf("No cc algorithm?\n");
 587                 return;
 588         }
 589         rack->rc_pacing_cc_set = 1;
 590         if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
 591                 /* Not new-reno we can't play games with beta! */
 592                 printf("cc_algo:%s is not NEWRENO:%s\n",
 593                        tp->cc_algo->name, CCALGONAME_NEWRENO);
 594                 goto out;
 595         }
 596         ptr = ((struct newreno *)tp->ccv->cc_data);
 597         if (CC_ALGO(tp)->ctl_output == NULL)  {
 598                 /* Huh, why does new_reno no longer have a set function? */
 599                 printf("no ctl_output for algo:%s\n", tp->cc_algo->name);
 600                 goto out;
 601         }
 602         if (ptr == NULL) {
 603                 /* Just the default values */
 604                 old.beta = V_newreno_beta_ecn;
 605                 old.beta_ecn = V_newreno_beta_ecn;
 606                 old.newreno_flags = 0;
 607         } else {
 608                 old.beta = ptr->beta;
 609                 old.beta_ecn = ptr->beta_ecn;
 610                 old.newreno_flags = ptr->newreno_flags;
 611         }
 612         sopt.sopt_valsize = sizeof(struct cc_newreno_opts);
 613         sopt.sopt_dir = SOPT_SET;
 614         opt.name = CC_NEWRENO_BETA;
 615         opt.val = rack->r_ctl.rc_saved_beta.beta;
 616         error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
 617         if (error)  {
 618                 printf("Error returned by ctl_output %d\n", error);
 619                 goto out;
 620         }
 621         /*
 622          * Hack alert we need to set in our newreno_flags
 623          * so that Abe behavior is also applied.
 624          */
 625         ((struct newreno *)tp->ccv->cc_data)->newreno_flags = CC_NEWRENO_BETA_ECN;
 626         opt.name = CC_NEWRENO_BETA_ECN;
 627         opt.val = rack->r_ctl.rc_saved_beta.beta_ecn;
 628         error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
 629         if (error) {
 630                 printf("Error returned by ctl_output %d\n", error);
 631                 goto out;
 632         }
 633         /* Save off the original values for restoral */
 634         memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno));
 635 out:
 636         if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
 637                 union tcp_log_stackspecific log;
 638                 struct timeval tv;
 639
 640                 ptr = ((struct newreno *)tp->ccv->cc_data);
 641                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
 642                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
 643                 if (ptr) {
 644                         log.u_bbr.flex1 = ptr->beta;
 645                         log.u_bbr.flex2 = ptr->beta_ecn;
 646                         log.u_bbr.flex3 = ptr->newreno_flags;
 647                 }
 648                 log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta;
 649                 log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn;
 650                 log.u_bbr.flex6 = rack->r_ctl.rc_saved_beta.newreno_flags;
 651                 log.u_bbr.flex7 = rack->gp_ready;
 652                 log.u_bbr.flex7 <<= 1;
 653                 log.u_bbr.flex7 |= rack->use_fixed_rate;
 654                 log.u_bbr.flex7 <<= 1;
 655                 log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
 656                 log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
 657                 log.u_bbr.flex8 = 3;
 658                 tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, error,
 659                                0, &log, false, NULL, NULL, 0, &tv);
 660         }
 661 }
 662
 663 static void
 664 rack_undo_cc_pacing(struct tcp_rack *rack)
 665 {
 666         struct newreno old, *ptr;
 667         struct tcpcb *tp;
 668
 669         if (rack->rc_pacing_cc_set == 0)
 670                 return;
 671         tp = rack->rc_tp;
 672         rack->rc_pacing_cc_set = 0;
 673         if (tp->cc_algo == NULL)
 674                 /* Tcb is leaving */
 675                 return;
 676         if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
 677                 /* Not new-reno nothing to do! */
 678                 return;
 679         }
 680         ptr = ((struct newreno *)tp->ccv->cc_data);
 681         if (ptr == NULL) {
 682                 /*
 683                  * This happens at rack_fini() if the
 684                  * cc module gets freed on us. In that
 685                  * case we loose our "new" settings but
 686                  * thats ok, since the tcb is going away anyway.
 687                  */
 688                 return;
 689         }
 690         /* Grab out our set values */
 691         memcpy(&old, ptr, sizeof(struct newreno));
 692         /* Copy back in the original values */
 693         memcpy(ptr, &rack->r_ctl.rc_saved_beta, sizeof(struct newreno));
 694         /* Now save back the values we had set in (for when pacing is restored) */
 695         memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno));
 696         if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
 697                 union tcp_log_stackspecific log;
 698                 struct timeval tv;
 699
 700                 ptr = ((struct newreno *)tp->ccv->cc_data);
 701                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
 702                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
 703                 log.u_bbr.flex1 = ptr->beta;
 704                 log.u_bbr.flex2 = ptr->beta_ecn;
 705                 log.u_bbr.flex3 = ptr->newreno_flags;
 706                 log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta;
 707                 log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn;
 708                 log.u_bbr.flex6 = rack->r_ctl.rc_saved_beta.newreno_flags;
 709                 log.u_bbr.flex7 = rack->gp_ready;
 710                 log.u_bbr.flex7 <<= 1;
 711                 log.u_bbr.flex7 |= rack->use_fixed_rate;
 712                 log.u_bbr.flex7 <<= 1;
 713                 log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
 714                 log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
 715                 log.u_bbr.flex8 = 4;
 716                 tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
 717                                0, &log, false, NULL, NULL, 0, &tv);
 718         }
 719 }
 720
 721 #ifdef NETFLIX_PEAKRATE
 722 static inline void
 723 rack_update_peakrate_thr(struct tcpcb *tp)
 724 {
 725         /* Keep in mind that t_maxpeakrate is in B/s. */
 726         uint64_t peak;
 727         peak = uqmax((tp->t_maxseg * 2),
 728                      (((uint64_t)tp->t_maxpeakrate * (uint64_t)(tp->t_srtt)) / (uint64_t)HPTS_USEC_IN_SEC));
 729         tp->t_peakrate_thr = (uint32_t)uqmin(peak, UINT32_MAX);
 730 }
 731 #endif
 732
 733 static int
 734 sysctl_rack_clear(SYSCTL_HANDLER_ARGS)
 735 {
 736         uint32_t stat;
 737         int32_t error;
 738         int i;
 739
 740         error = SYSCTL_OUT(req, &rack_clear_counter, sizeof(uint32_t));
 741         if (error || req->newptr == NULL)
 742                 return error;
 743
 744         error = SYSCTL_IN(req, &stat, sizeof(uint32_t));
 745         if (error)
 746                 return (error);
 747         if (stat == 1) {
 748 #ifdef INVARIANTS
 749                 printf("Clearing RACK counters\n");
 750 #endif
 751                 counter_u64_zero(rack_badfr);
 752                 counter_u64_zero(rack_badfr_bytes);
 753                 counter_u64_zero(rack_rtm_prr_retran);
 754                 counter_u64_zero(rack_rtm_prr_newdata);
 755                 counter_u64_zero(rack_timestamp_mismatch);
 756                 counter_u64_zero(rack_reorder_seen);
 757                 counter_u64_zero(rack_tlp_tot);
 758                 counter_u64_zero(rack_tlp_newdata);
 759                 counter_u64_zero(rack_tlp_retran);
 760                 counter_u64_zero(rack_tlp_retran_bytes);
 761                 counter_u64_zero(rack_tlp_retran_fail);
 762                 counter_u64_zero(rack_to_tot);
 763                 counter_u64_zero(rack_to_arm_rack);
 764                 counter_u64_zero(rack_to_arm_tlp);
 765                 counter_u64_zero(rack_paced_segments);
 766                 counter_u64_zero(rack_calc_zero);
 767                 counter_u64_zero(rack_calc_nonzero);
 768                 counter_u64_zero(rack_unpaced_segments);
 769                 counter_u64_zero(rack_saw_enobuf);
 770                 counter_u64_zero(rack_saw_enobuf_hw);
 771                 counter_u64_zero(rack_saw_enetunreach);
 772                 counter_u64_zero(rack_per_timer_hole);
 773                 counter_u64_zero(rack_large_ackcmp);
 774                 counter_u64_zero(rack_small_ackcmp);
 775 #ifdef INVARIANTS
 776                 counter_u64_zero(rack_adjust_map_bw);
 777 #endif
 778                 counter_u64_zero(rack_to_alloc_hard);
 779                 counter_u64_zero(rack_to_alloc_emerg);
 780                 counter_u64_zero(rack_sack_proc_all);
 781                 counter_u64_zero(rack_fto_send);
 782                 counter_u64_zero(rack_fto_rsm_send);
 783                 counter_u64_zero(rack_extended_rfo);
 784                 counter_u64_zero(rack_hw_pace_init_fail);
 785                 counter_u64_zero(rack_hw_pace_lost);
 786                 counter_u64_zero(rack_sbsndptr_wrong);
 787                 counter_u64_zero(rack_sbsndptr_right);
 788                 counter_u64_zero(rack_non_fto_send);
 789                 counter_u64_zero(rack_nfto_resend);
 790                 counter_u64_zero(rack_sack_proc_short);
 791                 counter_u64_zero(rack_sack_proc_restart);
 792                 counter_u64_zero(rack_to_alloc);
 793                 counter_u64_zero(rack_to_alloc_limited);
 794                 counter_u64_zero(rack_alloc_limited_conns);
 795                 counter_u64_zero(rack_split_limited);
 796                 for (i = 0; i < MAX_NUM_OF_CNTS; i++) {
 797                         counter_u64_zero(rack_proc_comp_ack[i]);
 798                 }
 799                 counter_u64_zero(rack_multi_single_eq);
 800                 counter_u64_zero(rack_proc_non_comp_ack);
 801                 counter_u64_zero(rack_find_high);
 802                 counter_u64_zero(rack_sack_attacks_detected);
 803                 counter_u64_zero(rack_sack_attacks_reversed);
 804                 counter_u64_zero(rack_sack_used_next_merge);
 805                 counter_u64_zero(rack_sack_used_prev_merge);
 806                 counter_u64_zero(rack_sack_splits);
 807                 counter_u64_zero(rack_sack_skipped_acked);
 808                 counter_u64_zero(rack_ack_total);
 809                 counter_u64_zero(rack_express_sack);
 810                 counter_u64_zero(rack_sack_total);
 811                 counter_u64_zero(rack_move_none);
 812                 counter_u64_zero(rack_move_some);
 813                 counter_u64_zero(rack_used_tlpmethod);
 814                 counter_u64_zero(rack_used_tlpmethod2);
 815                 counter_u64_zero(rack_enter_tlp_calc);
 816                 counter_u64_zero(rack_progress_drops);
 817                 counter_u64_zero(rack_tlp_does_nada);
 818                 counter_u64_zero(rack_try_scwnd);
 819                 counter_u64_zero(rack_collapsed_win);
 820         }
 821         rack_clear_counter = 0;
 822         return (0);
 823 }
 824
 825 static void
 826 rack_init_sysctls(void)
 827 {
 828         int i;
 829         struct sysctl_oid *rack_counters;
 830         struct sysctl_oid *rack_attack;
 831         struct sysctl_oid *rack_pacing;
 832         struct sysctl_oid *rack_timely;
 833         struct sysctl_oid *rack_timers;
 834         struct sysctl_oid *rack_tlp;
 835         struct sysctl_oid *rack_misc;
 836         struct sysctl_oid *rack_measure;
 837         struct sysctl_oid *rack_probertt;
 838         struct sysctl_oid *rack_hw_pacing;
 839
 840         rack_attack = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 841             SYSCTL_CHILDREN(rack_sysctl_root),
 842             OID_AUTO,
 843             "sack_attack",
 844             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 845             "Rack Sack Attack Counters and Controls");
 846         rack_counters = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 847             SYSCTL_CHILDREN(rack_sysctl_root),
 848             OID_AUTO,
 849             "stats",
 850             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 851             "Rack Counters");
 852         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 853             SYSCTL_CHILDREN(rack_sysctl_root),
 854             OID_AUTO, "rate_sample_method", CTLFLAG_RW,
 855             &rack_rate_sample_method , USE_RTT_LOW,
 856             "What method should we use for rate sampling 0=high, 1=low ");
 857         /* Probe rtt related controls */
 858         rack_probertt = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 859             SYSCTL_CHILDREN(rack_sysctl_root),
 860             OID_AUTO,
 861             "probertt",
 862             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 863             "ProbeRTT related Controls");
 864         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 865             SYSCTL_CHILDREN(rack_probertt),
 866             OID_AUTO, "exit_per_hpb", CTLFLAG_RW,
 867             &rack_atexit_prtt_hbp, 130,
 868             "What percentage above goodput do we clamp CA/SS to at exit on high-BDP path 110%");
 869         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 870             SYSCTL_CHILDREN(rack_probertt),
 871             OID_AUTO, "exit_per_nonhpb", CTLFLAG_RW,
 872             &rack_atexit_prtt, 130,
 873             "What percentage above goodput do we clamp CA/SS to at exit on a non high-BDP path 100%");
 874         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 875             SYSCTL_CHILDREN(rack_probertt),
 876             OID_AUTO, "gp_per_mul", CTLFLAG_RW,
 877             &rack_per_of_gp_probertt, 60,
 878             "What percentage of goodput do we pace at in probertt");
 879         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 880             SYSCTL_CHILDREN(rack_probertt),
 881             OID_AUTO, "gp_per_reduce", CTLFLAG_RW,
 882             &rack_per_of_gp_probertt_reduce, 10,
 883             "What percentage of goodput do we reduce every gp_srtt");
 884         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 885             SYSCTL_CHILDREN(rack_probertt),
 886             OID_AUTO, "gp_per_low", CTLFLAG_RW,
 887             &rack_per_of_gp_lowthresh, 40,
 888             "What percentage of goodput do we allow the multiplier to fall to");
 889         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 890             SYSCTL_CHILDREN(rack_probertt),
 891             OID_AUTO, "time_between", CTLFLAG_RW,
 892             & rack_time_between_probertt, 96000000,
 893             "How many useconds between the lowest rtt falling must past before we enter probertt");
 894         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 895             SYSCTL_CHILDREN(rack_probertt),
 896             OID_AUTO, "safety", CTLFLAG_RW,
 897             &rack_probe_rtt_safety_val, 2000000,
 898             "If not zero, provides a maximum usecond that you can stay in probertt (2sec = 2000000)");
 899         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 900             SYSCTL_CHILDREN(rack_probertt),
 901             OID_AUTO, "sets_cwnd", CTLFLAG_RW,
 902             &rack_probe_rtt_sets_cwnd, 0,
 903             "Do we set the cwnd too (if always_lower is on)");
 904         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 905             SYSCTL_CHILDREN(rack_probertt),
 906             OID_AUTO, "maxdrainsrtts", CTLFLAG_RW,
 907             &rack_max_drain_wait, 2,
 908             "Maximum number of gp_srtt's to hold in drain waiting for flight to reach goal");
 909         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 910             SYSCTL_CHILDREN(rack_probertt),
 911             OID_AUTO, "mustdrainsrtts", CTLFLAG_RW,
 912             &rack_must_drain, 1,
 913             "We must drain this many gp_srtt's waiting for flight to reach goal");
 914         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 915             SYSCTL_CHILDREN(rack_probertt),
 916             OID_AUTO, "goal_use_min_entry", CTLFLAG_RW,
 917             &rack_probertt_use_min_rtt_entry, 1,
 918             "Should we use the min-rtt to calculate the goal rtt (else gp_srtt) at entry");
 919         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 920             SYSCTL_CHILDREN(rack_probertt),
 921             OID_AUTO, "goal_use_min_exit", CTLFLAG_RW,
 922             &rack_probertt_use_min_rtt_exit, 0,
 923             "How to set cwnd at exit, 0 - dynamic, 1 - use min-rtt, 2 - use curgprtt, 3 - entry gp-rtt");
 924         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 925             SYSCTL_CHILDREN(rack_probertt),
 926             OID_AUTO, "length_div", CTLFLAG_RW,
 927             &rack_probertt_gpsrtt_cnt_div, 0,
 928             "How many recent goodput srtt periods plus hold tim does probertt last (bottom of fraction)");
 929         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 930             SYSCTL_CHILDREN(rack_probertt),
 931             OID_AUTO, "length_mul", CTLFLAG_RW,
 932             &rack_probertt_gpsrtt_cnt_mul, 0,
 933             "How many recent goodput srtt periods plus hold tim does probertt last (top of fraction)");
 934         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 935             SYSCTL_CHILDREN(rack_probertt),
 936             OID_AUTO, "holdtim_at_target", CTLFLAG_RW,
 937             &rack_min_probertt_hold, 200000,
 938             "What is the minimum time we hold probertt at target");
 939         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 940             SYSCTL_CHILDREN(rack_probertt),
 941             OID_AUTO, "filter_life", CTLFLAG_RW,
 942             &rack_probertt_filter_life, 10000000,
 943             "What is the time for the filters life in useconds");
 944         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 945             SYSCTL_CHILDREN(rack_probertt),
 946             OID_AUTO, "lower_within", CTLFLAG_RW,
 947             &rack_probertt_lower_within, 10,
 948             "If the rtt goes lower within this percentage of the time, go into probe-rtt");
 949         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 950             SYSCTL_CHILDREN(rack_probertt),
 951             OID_AUTO, "must_move", CTLFLAG_RW,
 952             &rack_min_rtt_movement, 250,
 953             "How much is the minimum movement in rtt to count as a drop for probertt purposes");
 954         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 955             SYSCTL_CHILDREN(rack_probertt),
 956             OID_AUTO, "clear_is_cnts", CTLFLAG_RW,
 957             &rack_probertt_clear_is, 1,
 958             "Do we clear I/S counts on exiting probe-rtt");
 959         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 960             SYSCTL_CHILDREN(rack_probertt),
 961             OID_AUTO, "hbp_extra_drain", CTLFLAG_RW,
 962             &rack_max_drain_hbp, 1,
 963             "How many extra drain gpsrtt's do we get in highly buffered paths");
 964         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 965             SYSCTL_CHILDREN(rack_probertt),
 966             OID_AUTO, "hbp_threshold", CTLFLAG_RW,
 967             &rack_hbp_thresh, 3,
 968             "We are highly buffered if min_rtt_seen / max_rtt_seen > this-threshold");
 969         /* Pacing related sysctls */
 970         rack_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 971             SYSCTL_CHILDREN(rack_sysctl_root),
 972             OID_AUTO,
 973             "pacing",
 974             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 975             "Pacing related Controls");
 976         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 977             SYSCTL_CHILDREN(rack_pacing),
 978             OID_AUTO, "max_pace_over", CTLFLAG_RW,
 979             &rack_max_per_above, 30,
 980             "What is the maximum allowable percentage that we can pace above (so 30 = 130% of our goal)");
 981         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 982             SYSCTL_CHILDREN(rack_pacing),
 983             OID_AUTO, "pace_to_one", CTLFLAG_RW,
 984             &rack_pace_one_seg, 0,
 985             "Do we allow low b/w pacing of 1MSS instead of two");
 986         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 987             SYSCTL_CHILDREN(rack_pacing),
 988             OID_AUTO, "limit_wsrtt", CTLFLAG_RW,
 989             &rack_limit_time_with_srtt, 0,
 990             "Do we limit pacing time based on srtt");
 991         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 992             SYSCTL_CHILDREN(rack_pacing),
 993             OID_AUTO, "init_win", CTLFLAG_RW,
 994             &rack_default_init_window, 0,
 995             "Do we have a rack initial window 0 = system default");
 996         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 997             SYSCTL_CHILDREN(rack_pacing),
 998             OID_AUTO, "gp_per_ss", CTLFLAG_RW,
 999             &rack_per_of_gp_ss, 250,
1000             "If non zero, what percentage of goodput to pace at in slow start");
1001         SYSCTL_ADD_U16(&rack_sysctl_ctx,
1002             SYSCTL_CHILDREN(rack_pacing),
1003             OID_AUTO, "gp_per_ca", CTLFLAG_RW,
1004             &rack_per_of_gp_ca, 150,
1005             "If non zero, what percentage of goodput to pace at in congestion avoidance");
1006         SYSCTL_ADD_U16(&rack_sysctl_ctx,
1007             SYSCTL_CHILDREN(rack_pacing),
1008             OID_AUTO, "gp_per_rec", CTLFLAG_RW,
1009             &rack_per_of_gp_rec, 200,
1010             "If non zero, what percentage of goodput to pace at in recovery");
1011         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1012             SYSCTL_CHILDREN(rack_pacing),
1013             OID_AUTO, "pace_max_seg", CTLFLAG_RW,
1014             &rack_hptsi_segments, 40,
1015             "What size is the max for TSO segments in pacing and burst mitigation");
1016         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1017             SYSCTL_CHILDREN(rack_pacing),
1018             OID_AUTO, "burst_reduces", CTLFLAG_RW,
1019             &rack_slot_reduction, 4,
1020             "When doing only burst mitigation what is the reduce divisor");
1021         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1022             SYSCTL_CHILDREN(rack_sysctl_root),
1023             OID_AUTO, "use_pacing", CTLFLAG_RW,
1024             &rack_pace_every_seg, 0,
1025             "If set we use pacing, if clear we use only the original burst mitigation");
1026         SYSCTL_ADD_U64(&rack_sysctl_ctx,
1027             SYSCTL_CHILDREN(rack_pacing),
1028             OID_AUTO, "rate_cap", CTLFLAG_RW,
1029             &rack_bw_rate_cap, 0,
1030             "If set we apply this value to the absolute rate cap used by pacing");
1031         SYSCTL_ADD_U8(&rack_sysctl_ctx,
1032             SYSCTL_CHILDREN(rack_sysctl_root),
1033             OID_AUTO, "req_measure_cnt", CTLFLAG_RW,
1034             &rack_req_measurements, 1,
1035             "If doing dynamic pacing, how many measurements must be in before we start pacing?");
1036         /* Hardware pacing */
1037         rack_hw_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1038             SYSCTL_CHILDREN(rack_sysctl_root),
1039             OID_AUTO,
1040             "hdwr_pacing",
1041             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1042             "Pacing related Controls");
1043         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1044             SYSCTL_CHILDREN(rack_hw_pacing),
1045             OID_AUTO, "rwnd_factor", CTLFLAG_RW,
1046             &rack_hw_rwnd_factor, 2,
1047             "How many times does snd_wnd need to be bigger than pace_max_seg so we will hold off and get more acks?");
1048         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1049             SYSCTL_CHILDREN(rack_hw_pacing),
1050             OID_AUTO, "pace_enobuf_mult", CTLFLAG_RW,
1051             &rack_enobuf_hw_boost_mult, 2,
1052             "By how many time_betweens should we boost the pacing time if we see a ENOBUFS?");
1053         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1054             SYSCTL_CHILDREN(rack_hw_pacing),
1055             OID_AUTO, "pace_enobuf_max", CTLFLAG_RW,
1056             &rack_enobuf_hw_max, 2,
1057             "What is the max boost the pacing time if we see a ENOBUFS?");
1058         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1059             SYSCTL_CHILDREN(rack_hw_pacing),
1060             OID_AUTO, "pace_enobuf_min", CTLFLAG_RW,
1061             &rack_enobuf_hw_min, 2,
1062             "What is the min boost the pacing time if we see a ENOBUFS?");
1063         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1064             SYSCTL_CHILDREN(rack_hw_pacing),
1065             OID_AUTO, "enable", CTLFLAG_RW,
1066             &rack_enable_hw_pacing, 0,
1067             "Should RACK attempt to use hw pacing?");
1068         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1069             SYSCTL_CHILDREN(rack_hw_pacing),
1070             OID_AUTO, "rate_cap", CTLFLAG_RW,
1071             &rack_hw_rate_caps, 1,
1072             "Does the highest hardware pacing rate cap the rate we will send at??");
1073         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1074             SYSCTL_CHILDREN(rack_hw_pacing),
1075             OID_AUTO, "rate_min", CTLFLAG_RW,
1076             &rack_hw_rate_min, 0,
1077             "Do we need a minimum estimate of this many bytes per second in order to engage hw pacing?");
1078         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1079             SYSCTL_CHILDREN(rack_hw_pacing),
1080             OID_AUTO, "rate_to_low", CTLFLAG_RW,
1081             &rack_hw_rate_to_low, 0,
1082             "If we fall below this rate, dis-engage hw pacing?");
1083         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1084             SYSCTL_CHILDREN(rack_hw_pacing),
1085             OID_AUTO, "up_only", CTLFLAG_RW,
1086             &rack_hw_up_only, 1,
1087             "Do we allow hw pacing to lower the rate selected?");
1088         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1089             SYSCTL_CHILDREN(rack_hw_pacing),
1090             OID_AUTO, "extra_mss_precise", CTLFLAG_RW,
1091             &rack_hw_pace_extra_slots, 2,
1092             "If the rates between software and hardware match precisely how many extra time_betweens do we get?");
1093         rack_timely = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1094             SYSCTL_CHILDREN(rack_sysctl_root),
1095             OID_AUTO,
1096             "timely",
1097             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1098             "Rack Timely RTT Controls");
1099         /* Timely based GP dynmics */
1100         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1101             SYSCTL_CHILDREN(rack_timely),
1102             OID_AUTO, "upper", CTLFLAG_RW,
1103             &rack_gp_per_bw_mul_up, 2,
1104             "Rack timely upper range for equal b/w (in percentage)");
1105         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1106             SYSCTL_CHILDREN(rack_timely),
1107             OID_AUTO, "lower", CTLFLAG_RW,
1108             &rack_gp_per_bw_mul_down, 4,
1109             "Rack timely lower range for equal b/w (in percentage)");
1110         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1111             SYSCTL_CHILDREN(rack_timely),
1112             OID_AUTO, "rtt_max_mul", CTLFLAG_RW,
1113             &rack_gp_rtt_maxmul, 3,
1114             "Rack timely multipler of lowest rtt for rtt_max");
1115         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1116             SYSCTL_CHILDREN(rack_timely),
1117             OID_AUTO, "rtt_min_div", CTLFLAG_RW,
1118             &rack_gp_rtt_mindiv, 4,
1119             "Rack timely divisor used for rtt + (rtt * mul/divisor) for check for lower rtt");
1120         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1121             SYSCTL_CHILDREN(rack_timely),
1122             OID_AUTO, "rtt_min_mul", CTLFLAG_RW,
1123             &rack_gp_rtt_minmul, 1,
1124             "Rack timely multiplier used for rtt + (rtt * mul/divisor) for check for lower rtt");
1125         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1126             SYSCTL_CHILDREN(rack_timely),
1127             OID_AUTO, "decrease", CTLFLAG_RW,
1128             &rack_gp_decrease_per, 20,
1129             "Rack timely decrease percentage of our GP multiplication factor");
1130         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1131             SYSCTL_CHILDREN(rack_timely),
1132             OID_AUTO, "increase", CTLFLAG_RW,
1133             &rack_gp_increase_per, 2,
1134             "Rack timely increase perentage of our GP multiplication factor");
1135         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1136             SYSCTL_CHILDREN(rack_timely),
1137             OID_AUTO, "lowerbound", CTLFLAG_RW,
1138             &rack_per_lower_bound, 50,
1139             "Rack timely lowest percentage we allow GP multiplier to fall to");
1140         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1141             SYSCTL_CHILDREN(rack_timely),
1142             OID_AUTO, "upperboundss", CTLFLAG_RW,
1143             &rack_per_upper_bound_ss, 0,
1144             "Rack timely higest percentage we allow GP multiplier in SS to raise to (0 is no upperbound)");
1145         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1146             SYSCTL_CHILDREN(rack_timely),
1147             OID_AUTO, "upperboundca", CTLFLAG_RW,
1148             &rack_per_upper_bound_ca, 0,
1149             "Rack timely higest percentage we allow GP multiplier to CA raise to (0 is no upperbound)");
1150         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1151             SYSCTL_CHILDREN(rack_timely),
1152             OID_AUTO, "dynamicgp", CTLFLAG_RW,
1153             &rack_do_dyn_mul, 0,
1154             "Rack timely do we enable dynmaic timely goodput by default");
1155         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1156             SYSCTL_CHILDREN(rack_timely),
1157             OID_AUTO, "no_rec_red", CTLFLAG_RW,
1158             &rack_gp_no_rec_chg, 1,
1159             "Rack timely do we prohibit the recovery multiplier from being lowered");
1160         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1161             SYSCTL_CHILDREN(rack_timely),
1162             OID_AUTO, "red_clear_cnt", CTLFLAG_RW,
1163             &rack_timely_dec_clear, 6,
1164             "Rack timely what threshold do we count to before another boost during b/w decent");
1165         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1166             SYSCTL_CHILDREN(rack_timely),
1167             OID_AUTO, "max_push_rise", CTLFLAG_RW,
1168             &rack_timely_max_push_rise, 3,
1169             "Rack timely how many times do we push up with b/w increase");
1170         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1171             SYSCTL_CHILDREN(rack_timely),
1172             OID_AUTO, "max_push_drop", CTLFLAG_RW,
1173             &rack_timely_max_push_drop, 3,
1174             "Rack timely how many times do we push back on b/w decent");
1175         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1176             SYSCTL_CHILDREN(rack_timely),
1177             OID_AUTO, "min_segs", CTLFLAG_RW,
1178             &rack_timely_min_segs, 4,
1179             "Rack timely when setting the cwnd what is the min num segments");
1180         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1181             SYSCTL_CHILDREN(rack_timely),
1182             OID_AUTO, "noback_max", CTLFLAG_RW,
1183             &rack_use_max_for_nobackoff, 0,
1184             "Rack timely when deciding if to backoff on a loss, do we use under max rtt else min");
1185         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1186             SYSCTL_CHILDREN(rack_timely),
1187             OID_AUTO, "interim_timely_only", CTLFLAG_RW,
1188             &rack_timely_int_timely_only, 0,
1189             "Rack timely when doing interim timely's do we only do timely (no b/w consideration)");
1190         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1191             SYSCTL_CHILDREN(rack_timely),
1192             OID_AUTO, "nonstop", CTLFLAG_RW,
1193             &rack_timely_no_stopping, 0,
1194             "Rack timely don't stop increase");
1195         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1196             SYSCTL_CHILDREN(rack_timely),
1197             OID_AUTO, "dec_raise_thresh", CTLFLAG_RW,
1198             &rack_down_raise_thresh, 100,
1199             "If the CA or SS is below this threshold raise on the first 3 b/w lowers (0=always)");
1200         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1201             SYSCTL_CHILDREN(rack_timely),
1202             OID_AUTO, "bottom_drag_segs", CTLFLAG_RW,
1203             &rack_req_segs, 1,
1204             "Bottom dragging if not these many segments outstanding and room");
1205
1206         /* TLP and Rack related parameters */
1207         rack_tlp = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1208             SYSCTL_CHILDREN(rack_sysctl_root),
1209             OID_AUTO,
1210             "tlp",
1211             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1212             "TLP and Rack related Controls");
1213         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1214             SYSCTL_CHILDREN(rack_tlp),
1215             OID_AUTO, "use_rrr", CTLFLAG_RW,
1216             &use_rack_rr, 1,
1217             "Do we use Rack Rapid Recovery");
1218         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1219             SYSCTL_CHILDREN(rack_tlp),
1220             OID_AUTO, "post_rec_labc", CTLFLAG_RW,
1221             &rack_max_abc_post_recovery, 2,
1222             "Since we do early recovery, do we override the l_abc to a value, if so what?");
1223         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1224             SYSCTL_CHILDREN(rack_tlp),
1225             OID_AUTO, "nonrxt_use_cr", CTLFLAG_RW,
1226             &rack_non_rxt_use_cr, 0,
1227             "Do we use ss/ca rate if in recovery we are transmitting a new data chunk");
1228         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1229             SYSCTL_CHILDREN(rack_tlp),
1230             OID_AUTO, "tlpmethod", CTLFLAG_RW,
1231             &rack_tlp_threshold_use, TLP_USE_TWO_ONE,
1232             "What method do we do for TLP time calc 0=no-de-ack-comp, 1=ID, 2=2.1, 3=2.2");
1233         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1234             SYSCTL_CHILDREN(rack_tlp),
1235             OID_AUTO, "limit", CTLFLAG_RW,
1236             &rack_tlp_limit, 2,
1237             "How many TLP's can be sent without sending new data");
1238         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1239             SYSCTL_CHILDREN(rack_tlp),
1240             OID_AUTO, "use_greater", CTLFLAG_RW,
1241             &rack_tlp_use_greater, 1,
1242             "Should we use the rack_rtt time if its greater than srtt");
1243         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1244             SYSCTL_CHILDREN(rack_tlp),
1245             OID_AUTO, "tlpminto", CTLFLAG_RW,
1246             &rack_tlp_min, 10000,
1247             "TLP minimum timeout per the specification (in microseconds)");
1248         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1249             SYSCTL_CHILDREN(rack_tlp),
1250             OID_AUTO, "send_oldest", CTLFLAG_RW,
1251             &rack_always_send_oldest, 0,
1252             "Should we always send the oldest TLP and RACK-TLP");
1253         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1254             SYSCTL_CHILDREN(rack_tlp),
1255             OID_AUTO, "rack_tlimit", CTLFLAG_RW,
1256             &rack_limited_retran, 0,
1257             "How many times can a rack timeout drive out sends");
1258         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1259             SYSCTL_CHILDREN(rack_tlp),
1260             OID_AUTO, "tlp_cwnd_flag", CTLFLAG_RW,
1261             &rack_lower_cwnd_at_tlp, 0,
1262             "When a TLP completes a retran should we enter recovery");
1263         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1264             SYSCTL_CHILDREN(rack_tlp),
1265             OID_AUTO, "reorder_thresh", CTLFLAG_RW,
1266             &rack_reorder_thresh, 2,
1267             "What factor for rack will be added when seeing reordering (shift right)");
1268         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1269             SYSCTL_CHILDREN(rack_tlp),
1270             OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW,
1271             &rack_tlp_thresh, 1,
1272             "What divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)");
1273         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1274             SYSCTL_CHILDREN(rack_tlp),
1275             OID_AUTO, "reorder_fade", CTLFLAG_RW,
1276             &rack_reorder_fade, 60000000,
1277             "Does reorder detection fade, if so how many microseconds (0 means never)");
1278         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1279             SYSCTL_CHILDREN(rack_tlp),
1280             OID_AUTO, "pktdelay", CTLFLAG_RW,
1281             &rack_pkt_delay, 1000,
1282             "Extra RACK time (in microseconds) besides reordering thresh");
1283
1284         /* Timer related controls */
1285         rack_timers = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1286             SYSCTL_CHILDREN(rack_sysctl_root),
1287             OID_AUTO,
1288             "timers",
1289             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1290             "Timer related controls");
1291         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1292             SYSCTL_CHILDREN(rack_timers),
1293             OID_AUTO, "persmin", CTLFLAG_RW,
1294             &rack_persist_min, 250000,
1295             "What is the minimum time in microseconds between persists");
1296         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1297             SYSCTL_CHILDREN(rack_timers),
1298             OID_AUTO, "persmax", CTLFLAG_RW,
1299             &rack_persist_max, 2000000,
1300             "What is the largest delay in microseconds between persists");
1301         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1302             SYSCTL_CHILDREN(rack_timers),
1303             OID_AUTO, "delayed_ack", CTLFLAG_RW,
1304             &rack_delayed_ack_time, 40000,
1305             "Delayed ack time (40ms in microseconds)");
1306         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1307             SYSCTL_CHILDREN(rack_timers),
1308             OID_AUTO, "minrto", CTLFLAG_RW,
1309             &rack_rto_min, 30000,
1310             "Minimum RTO in microseconds -- set with caution below 1000 due to TLP");
1311         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1312             SYSCTL_CHILDREN(rack_timers),
1313             OID_AUTO, "maxrto", CTLFLAG_RW,
1314             &rack_rto_max, 4000000,
1315             "Maxiumum RTO in microseconds -- should be at least as large as min_rto");
1316         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1317             SYSCTL_CHILDREN(rack_timers),
1318             OID_AUTO, "minto", CTLFLAG_RW,
1319             &rack_min_to, 1000,
1320             "Minimum rack timeout in microseconds");
1321         /* Measure controls */
1322         rack_measure = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1323             SYSCTL_CHILDREN(rack_sysctl_root),
1324             OID_AUTO,
1325             "measure",
1326             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1327             "Measure related controls");
1328         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1329             SYSCTL_CHILDREN(rack_measure),
1330             OID_AUTO, "wma_divisor", CTLFLAG_RW,
1331             &rack_wma_divisor, 8,
1332             "When doing b/w calculation what is the  divisor for the WMA");
1333         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1334             SYSCTL_CHILDREN(rack_measure),
1335             OID_AUTO, "end_cwnd", CTLFLAG_RW,
1336             &rack_cwnd_block_ends_measure, 0,
1337             "Does a cwnd just-return end the measurement window (app limited)");
1338         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1339             SYSCTL_CHILDREN(rack_measure),
1340             OID_AUTO, "end_rwnd", CTLFLAG_RW,
1341             &rack_rwnd_block_ends_measure, 0,
1342             "Does an rwnd just-return end the measurement window (app limited -- not persists)");
1343         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1344             SYSCTL_CHILDREN(rack_measure),
1345             OID_AUTO, "min_target", CTLFLAG_RW,
1346             &rack_def_data_window, 20,
1347             "What is the minimum target window (in mss) for a GP measurements");
1348         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1349             SYSCTL_CHILDREN(rack_measure),
1350             OID_AUTO, "goal_bdp", CTLFLAG_RW,
1351             &rack_goal_bdp, 2,
1352             "What is the goal BDP to measure");
1353         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1354             SYSCTL_CHILDREN(rack_measure),
1355             OID_AUTO, "min_srtts", CTLFLAG_RW,
1356             &rack_min_srtts, 1,
1357             "What is the goal BDP to measure");
1358         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1359             SYSCTL_CHILDREN(rack_measure),
1360             OID_AUTO, "min_measure_tim", CTLFLAG_RW,
1361             &rack_min_measure_usec, 0,
1362             "What is the Minimum time time for a measurement if 0, this is off");
1363         /* Misc rack controls */
1364         rack_misc = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1365             SYSCTL_CHILDREN(rack_sysctl_root),
1366             OID_AUTO,
1367             "misc",
1368             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1369             "Misc related controls");
1370 #ifdef TCP_ACCOUNTING
1371         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1372             SYSCTL_CHILDREN(rack_misc),
1373             OID_AUTO, "tcp_acct", CTLFLAG_RW,
1374             &rack_tcp_accounting, 0,
1375             "Should we turn on TCP accounting for all rack sessions?");
1376 #endif
1377         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1378             SYSCTL_CHILDREN(rack_misc),
1379             OID_AUTO, "prr_addback_max", CTLFLAG_RW,
1380             &rack_prr_addbackmax, 2,
1381             "What is the maximum number of MSS we allow to be added back if prr can't send all its data?");
1382         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1383             SYSCTL_CHILDREN(rack_misc),
1384             OID_AUTO, "stats_gets_ms", CTLFLAG_RW,
1385             &rack_stats_gets_ms_rtt, 1,
1386             "What do we feed the stats framework (1 = ms_rtt, 0 = us_rtt, 2 = ms_rtt from hdwr, > 2 usec rtt from hdwr)?");
1387         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1388             SYSCTL_CHILDREN(rack_misc),
1389             OID_AUTO, "clientlowbuf", CTLFLAG_RW,
1390             &rack_client_low_buf, 0,
1391             "Client low buffer level (below this we are more aggressive in DGP exiting recovery (0 = off)?");
1392         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1393             SYSCTL_CHILDREN(rack_misc),
1394             OID_AUTO, "defprofile", CTLFLAG_RW,
1395             &rack_def_profile, 0,
1396             "Should RACK use a default profile (0=no, num == profile num)?");
1397         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1398             SYSCTL_CHILDREN(rack_misc),
1399             OID_AUTO, "cmpack", CTLFLAG_RW,
1400             &rack_use_cmp_acks, 1,
1401             "Should RACK have LRO send compressed acks");
1402         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1403             SYSCTL_CHILDREN(rack_misc),
1404             OID_AUTO, "fsb", CTLFLAG_RW,
1405             &rack_use_fsb, 1,
1406             "Should RACK use the fast send block?");
1407         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1408             SYSCTL_CHILDREN(rack_misc),
1409             OID_AUTO, "rfo", CTLFLAG_RW,
1410             &rack_use_rfo, 1,
1411             "Should RACK use rack_fast_output()?");
1412         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1413             SYSCTL_CHILDREN(rack_misc),
1414             OID_AUTO, "rsmrfo", CTLFLAG_RW,
1415             &rack_use_rsm_rfo, 1,
1416             "Should RACK use rack_fast_rsm_output()?");
1417         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1418             SYSCTL_CHILDREN(rack_misc),
1419             OID_AUTO, "shared_cwnd", CTLFLAG_RW,
1420             &rack_enable_shared_cwnd, 1,
1421             "Should RACK try to use the shared cwnd on connections where allowed");
1422         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1423             SYSCTL_CHILDREN(rack_misc),
1424             OID_AUTO, "limits_on_scwnd", CTLFLAG_RW,
1425             &rack_limits_scwnd, 1,
1426             "Should RACK place low end time limits on the shared cwnd feature");
1427         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1428             SYSCTL_CHILDREN(rack_misc),
1429             OID_AUTO, "non_paced_lro_queue", CTLFLAG_RW,
1430             &rack_enable_mqueue_for_nonpaced, 0,
1431             "Should RACK use mbuf queuing for non-paced connections");
1432         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1433             SYSCTL_CHILDREN(rack_misc),
1434             OID_AUTO, "iMac_dack", CTLFLAG_RW,
1435             &rack_use_imac_dack, 0,
1436             "Should RACK try to emulate iMac delayed ack");
1437         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1438             SYSCTL_CHILDREN(rack_misc),
1439             OID_AUTO, "no_prr", CTLFLAG_RW,
1440             &rack_disable_prr, 0,
1441             "Should RACK not use prr and only pace (must have pacing on)");
1442         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1443             SYSCTL_CHILDREN(rack_misc),
1444             OID_AUTO, "bb_verbose", CTLFLAG_RW,
1445             &rack_verbose_logging, 0,
1446             "Should RACK black box logging be verbose");
1447         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1448             SYSCTL_CHILDREN(rack_misc),
1449             OID_AUTO, "data_after_close", CTLFLAG_RW,
1450             &rack_ignore_data_after_close, 1,
1451             "Do we hold off sending a RST until all pending data is ack'd");
1452         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1453             SYSCTL_CHILDREN(rack_misc),
1454             OID_AUTO, "no_sack_needed", CTLFLAG_RW,
1455             &rack_sack_not_required, 1,
1456             "Do we allow rack to run on connections not supporting SACK");
1457         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1458             SYSCTL_CHILDREN(rack_misc),
1459             OID_AUTO, "prr_sendalot", CTLFLAG_RW,
1460             &rack_send_a_lot_in_prr, 1,
1461             "Send a lot in prr");
1462         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1463             SYSCTL_CHILDREN(rack_misc),
1464             OID_AUTO, "autoscale", CTLFLAG_RW,
1465             &rack_autosndbuf_inc, 20,
1466             "What percentage should rack scale up its snd buffer by?");
1467         /* Sack Attacker detection stuff */
1468         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1469             SYSCTL_CHILDREN(rack_attack),
1470             OID_AUTO, "detect_highsackratio", CTLFLAG_RW,
1471             &rack_highest_sack_thresh_seen, 0,
1472             "Highest sack to ack ratio seen");
1473         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1474             SYSCTL_CHILDREN(rack_attack),
1475             OID_AUTO, "detect_highmoveratio", CTLFLAG_RW,
1476             &rack_highest_move_thresh_seen, 0,
1477             "Highest move to non-move ratio seen");
1478         rack_ack_total = counter_u64_alloc(M_WAITOK);
1479         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1480             SYSCTL_CHILDREN(rack_attack),
1481             OID_AUTO, "acktotal", CTLFLAG_RD,
1482             &rack_ack_total,
1483             "Total number of Ack's");
1484         rack_express_sack = counter_u64_alloc(M_WAITOK);
1485         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1486             SYSCTL_CHILDREN(rack_attack),
1487             OID_AUTO, "exp_sacktotal", CTLFLAG_RD,
1488             &rack_express_sack,
1489             "Total expresss number of Sack's");
1490         rack_sack_total = counter_u64_alloc(M_WAITOK);
1491         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1492             SYSCTL_CHILDREN(rack_attack),
1493             OID_AUTO, "sacktotal", CTLFLAG_RD,
1494             &rack_sack_total,
1495             "Total number of SACKs");
1496         rack_move_none = counter_u64_alloc(M_WAITOK);
1497         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1498             SYSCTL_CHILDREN(rack_attack),
1499             OID_AUTO, "move_none", CTLFLAG_RD,
1500             &rack_move_none,
1501             "Total number of SACK index reuse of postions under threshold");
1502         rack_move_some = counter_u64_alloc(M_WAITOK);
1503         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1504             SYSCTL_CHILDREN(rack_attack),
1505             OID_AUTO, "move_some", CTLFLAG_RD,
1506             &rack_move_some,
1507             "Total number of SACK index reuse of postions over threshold");
1508         rack_sack_attacks_detected = counter_u64_alloc(M_WAITOK);
1509         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1510             SYSCTL_CHILDREN(rack_attack),
1511             OID_AUTO, "attacks", CTLFLAG_RD,
1512             &rack_sack_attacks_detected,
1513             "Total number of SACK attackers that had sack disabled");
1514         rack_sack_attacks_reversed = counter_u64_alloc(M_WAITOK);
1515         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1516             SYSCTL_CHILDREN(rack_attack),
1517             OID_AUTO, "reversed", CTLFLAG_RD,
1518             &rack_sack_attacks_reversed,
1519             "Total number of SACK attackers that were later determined false positive");
1520         rack_sack_used_next_merge = counter_u64_alloc(M_WAITOK);
1521         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1522             SYSCTL_CHILDREN(rack_attack),
1523             OID_AUTO, "nextmerge", CTLFLAG_RD,
1524             &rack_sack_used_next_merge,
1525             "Total number of times we used the next merge");
1526         rack_sack_used_prev_merge = counter_u64_alloc(M_WAITOK);
1527         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1528             SYSCTL_CHILDREN(rack_attack),
1529             OID_AUTO, "prevmerge", CTLFLAG_RD,
1530             &rack_sack_used_prev_merge,
1531             "Total number of times we used the prev merge");
1532         /* Counters */
1533         rack_fto_send = counter_u64_alloc(M_WAITOK);
1534         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1535             SYSCTL_CHILDREN(rack_counters),
1536             OID_AUTO, "fto_send", CTLFLAG_RD,
1537             &rack_fto_send, "Total number of rack_fast_output sends");
1538         rack_fto_rsm_send = counter_u64_alloc(M_WAITOK);
1539         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1540             SYSCTL_CHILDREN(rack_counters),
1541             OID_AUTO, "fto_rsm_send", CTLFLAG_RD,
1542             &rack_fto_rsm_send, "Total number of rack_fast_rsm_output sends");
1543         rack_nfto_resend = counter_u64_alloc(M_WAITOK);
1544         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1545             SYSCTL_CHILDREN(rack_counters),
1546             OID_AUTO, "nfto_resend", CTLFLAG_RD,
1547             &rack_nfto_resend, "Total number of rack_output retransmissions");
1548         rack_non_fto_send = counter_u64_alloc(M_WAITOK);
1549         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1550             SYSCTL_CHILDREN(rack_counters),
1551             OID_AUTO, "nfto_send", CTLFLAG_RD,
1552             &rack_non_fto_send, "Total number of rack_output first sends");
1553         rack_extended_rfo = counter_u64_alloc(M_WAITOK);
1554         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1555             SYSCTL_CHILDREN(rack_counters),
1556             OID_AUTO, "rfo_extended", CTLFLAG_RD,
1557             &rack_extended_rfo, "Total number of times we extended rfo");
1558
1559         rack_hw_pace_init_fail = counter_u64_alloc(M_WAITOK);
1560         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1561             SYSCTL_CHILDREN(rack_counters),
1562             OID_AUTO, "hwpace_init_fail", CTLFLAG_RD,
1563             &rack_hw_pace_init_fail, "Total number of times we failed to initialize hw pacing");
1564         rack_hw_pace_lost = counter_u64_alloc(M_WAITOK);
1565
1566         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1567             SYSCTL_CHILDREN(rack_counters),
1568             OID_AUTO, "hwpace_lost", CTLFLAG_RD,
1569             &rack_hw_pace_lost, "Total number of times we failed to initialize hw pacing");
1570
1571
1572
1573         rack_badfr = counter_u64_alloc(M_WAITOK);
1574         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1575             SYSCTL_CHILDREN(rack_counters),
1576             OID_AUTO, "badfr", CTLFLAG_RD,
1577             &rack_badfr, "Total number of bad FRs");
1578         rack_badfr_bytes = counter_u64_alloc(M_WAITOK);
1579         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1580             SYSCTL_CHILDREN(rack_counters),
1581             OID_AUTO, "badfr_bytes", CTLFLAG_RD,
1582             &rack_badfr_bytes, "Total number of bad FRs");
1583         rack_rtm_prr_retran = counter_u64_alloc(M_WAITOK);
1584         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1585             SYSCTL_CHILDREN(rack_counters),
1586             OID_AUTO, "prrsndret", CTLFLAG_RD,
1587             &rack_rtm_prr_retran,
1588             "Total number of prr based retransmits");
1589         rack_rtm_prr_newdata = counter_u64_alloc(M_WAITOK);
1590         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1591             SYSCTL_CHILDREN(rack_counters),
1592             OID_AUTO, "prrsndnew", CTLFLAG_RD,
1593             &rack_rtm_prr_newdata,
1594             "Total number of prr based new transmits");
1595         rack_timestamp_mismatch = counter_u64_alloc(M_WAITOK);
1596         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1597             SYSCTL_CHILDREN(rack_counters),
1598             OID_AUTO, "tsnf", CTLFLAG_RD,
1599             &rack_timestamp_mismatch,
1600             "Total number of timestamps that we could not find the reported ts");
1601         rack_find_high = counter_u64_alloc(M_WAITOK);
1602         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1603             SYSCTL_CHILDREN(rack_counters),
1604             OID_AUTO, "findhigh", CTLFLAG_RD,
1605             &rack_find_high,
1606             "Total number of FIN causing find-high");
1607         rack_reorder_seen = counter_u64_alloc(M_WAITOK);
1608         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1609             SYSCTL_CHILDREN(rack_counters),
1610             OID_AUTO, "reordering", CTLFLAG_RD,
1611             &rack_reorder_seen,
1612             "Total number of times we added delay due to reordering");
1613         rack_tlp_tot = counter_u64_alloc(M_WAITOK);
1614         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1615             SYSCTL_CHILDREN(rack_counters),
1616             OID_AUTO, "tlp_to_total", CTLFLAG_RD,
1617             &rack_tlp_tot,
1618             "Total number of tail loss probe expirations");
1619         rack_tlp_newdata = counter_u64_alloc(M_WAITOK);
1620         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1621             SYSCTL_CHILDREN(rack_counters),
1622             OID_AUTO, "tlp_new", CTLFLAG_RD,
1623             &rack_tlp_newdata,
1624             "Total number of tail loss probe sending new data");
1625         rack_tlp_retran = counter_u64_alloc(M_WAITOK);
1626         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1627             SYSCTL_CHILDREN(rack_counters),
1628             OID_AUTO, "tlp_retran", CTLFLAG_RD,
1629             &rack_tlp_retran,
1630             "Total number of tail loss probe sending retransmitted data");
1631         rack_tlp_retran_bytes = counter_u64_alloc(M_WAITOK);
1632         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1633             SYSCTL_CHILDREN(rack_counters),
1634             OID_AUTO, "tlp_retran_bytes", CTLFLAG_RD,
1635             &rack_tlp_retran_bytes,
1636             "Total bytes of tail loss probe sending retransmitted data");
1637         rack_tlp_retran_fail = counter_u64_alloc(M_WAITOK);
1638         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1639             SYSCTL_CHILDREN(rack_counters),
1640             OID_AUTO, "tlp_retran_fail", CTLFLAG_RD,
1641             &rack_tlp_retran_fail,
1642             "Total number of tail loss probe sending retransmitted data that failed (wait for t3)");
1643         rack_to_tot = counter_u64_alloc(M_WAITOK);
1644         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1645             SYSCTL_CHILDREN(rack_counters),
1646             OID_AUTO, "rack_to_tot", CTLFLAG_RD,
1647             &rack_to_tot,
1648             "Total number of times the rack to expired");
1649         rack_to_arm_rack = counter_u64_alloc(M_WAITOK);
1650         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1651             SYSCTL_CHILDREN(rack_counters),
1652             OID_AUTO, "arm_rack", CTLFLAG_RD,
1653             &rack_to_arm_rack,
1654             "Total number of times the rack timer armed");
1655         rack_to_arm_tlp = counter_u64_alloc(M_WAITOK);
1656         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1657             SYSCTL_CHILDREN(rack_counters),
1658             OID_AUTO, "arm_tlp", CTLFLAG_RD,
1659             &rack_to_arm_tlp,
1660             "Total number of times the tlp timer armed");
1661         rack_calc_zero = counter_u64_alloc(M_WAITOK);
1662         rack_calc_nonzero = counter_u64_alloc(M_WAITOK);
1663         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1664             SYSCTL_CHILDREN(rack_counters),
1665             OID_AUTO, "calc_zero", CTLFLAG_RD,
1666             &rack_calc_zero,
1667             "Total number of times pacing time worked out to zero");
1668         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1669             SYSCTL_CHILDREN(rack_counters),
1670             OID_AUTO, "calc_nonzero", CTLFLAG_RD,
1671             &rack_calc_nonzero,
1672             "Total number of times pacing time worked out to non-zero");
1673         rack_paced_segments = counter_u64_alloc(M_WAITOK);
1674         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1675             SYSCTL_CHILDREN(rack_counters),
1676             OID_AUTO, "paced", CTLFLAG_RD,
1677             &rack_paced_segments,
1678             "Total number of times a segment send caused hptsi");
1679         rack_unpaced_segments = counter_u64_alloc(M_WAITOK);
1680         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1681             SYSCTL_CHILDREN(rack_counters),
1682             OID_AUTO, "unpaced", CTLFLAG_RD,
1683             &rack_unpaced_segments,
1684             "Total number of times a segment did not cause hptsi");
1685         rack_saw_enobuf = counter_u64_alloc(M_WAITOK);
1686         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1687             SYSCTL_CHILDREN(rack_counters),
1688             OID_AUTO, "saw_enobufs", CTLFLAG_RD,
1689             &rack_saw_enobuf,
1690             "Total number of times a sends returned enobuf for non-hdwr paced connections");
1691         rack_saw_enobuf_hw = counter_u64_alloc(M_WAITOK);
1692         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1693             SYSCTL_CHILDREN(rack_counters),
1694             OID_AUTO, "saw_enobufs_hw", CTLFLAG_RD,
1695             &rack_saw_enobuf_hw,
1696             "Total number of times a send returned enobuf for hdwr paced connections");
1697         rack_saw_enetunreach = counter_u64_alloc(M_WAITOK);
1698         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1699             SYSCTL_CHILDREN(rack_counters),
1700             OID_AUTO, "saw_enetunreach", CTLFLAG_RD,
1701             &rack_saw_enetunreach,
1702             "Total number of times a send received a enetunreachable");
1703         rack_hot_alloc = counter_u64_alloc(M_WAITOK);
1704         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1705             SYSCTL_CHILDREN(rack_counters),
1706             OID_AUTO, "alloc_hot", CTLFLAG_RD,
1707             &rack_hot_alloc,
1708             "Total allocations from the top of our list");
1709         rack_to_alloc = counter_u64_alloc(M_WAITOK);
1710         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1711             SYSCTL_CHILDREN(rack_counters),
1712             OID_AUTO, "allocs", CTLFLAG_RD,
1713             &rack_to_alloc,
1714             "Total allocations of tracking structures");
1715         rack_to_alloc_hard = counter_u64_alloc(M_WAITOK);
1716         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1717             SYSCTL_CHILDREN(rack_counters),
1718             OID_AUTO, "allochard", CTLFLAG_RD,
1719             &rack_to_alloc_hard,
1720             "Total allocations done with sleeping the hard way");
1721         rack_to_alloc_emerg = counter_u64_alloc(M_WAITOK);
1722         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1723             SYSCTL_CHILDREN(rack_counters),
1724             OID_AUTO, "allocemerg", CTLFLAG_RD,
1725             &rack_to_alloc_emerg,
1726             "Total allocations done from emergency cache");
1727         rack_to_alloc_limited = counter_u64_alloc(M_WAITOK);
1728         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1729             SYSCTL_CHILDREN(rack_counters),
1730             OID_AUTO, "alloc_limited", CTLFLAG_RD,
1731             &rack_to_alloc_limited,
1732             "Total allocations dropped due to limit");
1733         rack_alloc_limited_conns = counter_u64_alloc(M_WAITOK);
1734         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1735             SYSCTL_CHILDREN(rack_counters),
1736             OID_AUTO, "alloc_limited_conns", CTLFLAG_RD,
1737             &rack_alloc_limited_conns,
1738             "Connections with allocations dropped due to limit");
1739         rack_split_limited = counter_u64_alloc(M_WAITOK);
1740         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1741             SYSCTL_CHILDREN(rack_counters),
1742             OID_AUTO, "split_limited", CTLFLAG_RD,
1743             &rack_split_limited,
1744             "Split allocations dropped due to limit");
1745
1746         for (i = 0; i < MAX_NUM_OF_CNTS; i++) {
1747                 char name[32];
1748                 sprintf(name, "cmp_ack_cnt_%d", i);
1749                 rack_proc_comp_ack[i] = counter_u64_alloc(M_WAITOK);
1750                 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1751                                        SYSCTL_CHILDREN(rack_counters),
1752                                        OID_AUTO, name, CTLFLAG_RD,
1753                                        &rack_proc_comp_ack[i],
1754                                        "Number of compressed acks we processed");
1755         }
1756         rack_large_ackcmp = counter_u64_alloc(M_WAITOK);
1757         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1758             SYSCTL_CHILDREN(rack_counters),
1759             OID_AUTO, "cmp_large_mbufs", CTLFLAG_RD,
1760             &rack_large_ackcmp,
1761             "Number of TCP connections with large mbuf's for compressed acks");
1762         rack_small_ackcmp = counter_u64_alloc(M_WAITOK);
1763         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1764             SYSCTL_CHILDREN(rack_counters),
1765             OID_AUTO, "cmp_small_mbufs", CTLFLAG_RD,
1766             &rack_small_ackcmp,
1767             "Number of TCP connections with small mbuf's for compressed acks");
1768 #ifdef INVARIANTS
1769         rack_adjust_map_bw = counter_u64_alloc(M_WAITOK);
1770         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1771             SYSCTL_CHILDREN(rack_counters),
1772             OID_AUTO, "map_adjust_req", CTLFLAG_RD,
1773             &rack_adjust_map_bw,
1774             "Number of times we hit the case where the sb went up and down on a sendmap entry");
1775 #endif
1776         rack_multi_single_eq = counter_u64_alloc(M_WAITOK);
1777         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1778             SYSCTL_CHILDREN(rack_counters),
1779             OID_AUTO, "cmp_ack_equiv", CTLFLAG_RD,
1780             &rack_multi_single_eq,
1781             "Number of compressed acks total represented");
1782         rack_proc_non_comp_ack = counter_u64_alloc(M_WAITOK);
1783         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1784             SYSCTL_CHILDREN(rack_counters),
1785             OID_AUTO, "cmp_ack_not", CTLFLAG_RD,
1786             &rack_proc_non_comp_ack,
1787             "Number of non compresseds acks that we processed");
1788
1789
1790         rack_sack_proc_all = counter_u64_alloc(M_WAITOK);
1791         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1792             SYSCTL_CHILDREN(rack_counters),
1793             OID_AUTO, "sack_long", CTLFLAG_RD,
1794             &rack_sack_proc_all,
1795             "Total times we had to walk whole list for sack processing");
1796         rack_sack_proc_restart = counter_u64_alloc(M_WAITOK);
1797         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1798             SYSCTL_CHILDREN(rack_counters),
1799             OID_AUTO, "sack_restart", CTLFLAG_RD,
1800             &rack_sack_proc_restart,
1801             "Total times we had to walk whole list due to a restart");
1802         rack_sack_proc_short = counter_u64_alloc(M_WAITOK);
1803         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1804             SYSCTL_CHILDREN(rack_counters),
1805             OID_AUTO, "sack_short", CTLFLAG_RD,
1806             &rack_sack_proc_short,
1807             "Total times we took shortcut for sack processing");
1808         rack_enter_tlp_calc = counter_u64_alloc(M_WAITOK);
1809         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1810             SYSCTL_CHILDREN(rack_counters),
1811             OID_AUTO, "tlp_calc_entered", CTLFLAG_RD,
1812             &rack_enter_tlp_calc,
1813             "Total times we called calc-tlp");
1814         rack_used_tlpmethod = counter_u64_alloc(M_WAITOK);
1815         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1816             SYSCTL_CHILDREN(rack_counters),
1817             OID_AUTO, "hit_tlp_method", CTLFLAG_RD,
1818             &rack_used_tlpmethod,
1819             "Total number of runt sacks");
1820         rack_used_tlpmethod2 = counter_u64_alloc(M_WAITOK);
1821         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1822             SYSCTL_CHILDREN(rack_counters),
1823             OID_AUTO, "hit_tlp_method2", CTLFLAG_RD,
1824             &rack_used_tlpmethod2,
1825             "Total number of times we hit TLP method 2");
1826         rack_sack_skipped_acked = counter_u64_alloc(M_WAITOK);
1827         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1828             SYSCTL_CHILDREN(rack_attack),
1829             OID_AUTO, "skipacked", CTLFLAG_RD,
1830             &rack_sack_skipped_acked,
1831             "Total number of times we skipped previously sacked");
1832         rack_sack_splits = counter_u64_alloc(M_WAITOK);
1833         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1834             SYSCTL_CHILDREN(rack_attack),
1835             OID_AUTO, "ofsplit", CTLFLAG_RD,
1836             &rack_sack_splits,
1837             "Total number of times we did the old fashion tree split");
1838         rack_progress_drops = counter_u64_alloc(M_WAITOK);
1839         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1840             SYSCTL_CHILDREN(rack_counters),
1841             OID_AUTO, "prog_drops", CTLFLAG_RD,
1842             &rack_progress_drops,
1843             "Total number of progress drops");
1844         rack_input_idle_reduces = counter_u64_alloc(M_WAITOK);
1845         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1846             SYSCTL_CHILDREN(rack_counters),
1847             OID_AUTO, "idle_reduce_oninput", CTLFLAG_RD,
1848             &rack_input_idle_reduces,
1849             "Total number of idle reductions on input");
1850         rack_collapsed_win = counter_u64_alloc(M_WAITOK);
1851         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1852             SYSCTL_CHILDREN(rack_counters),
1853             OID_AUTO, "collapsed_win", CTLFLAG_RD,
1854             &rack_collapsed_win,
1855             "Total number of collapsed windows");
1856         rack_tlp_does_nada = counter_u64_alloc(M_WAITOK);
1857         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1858             SYSCTL_CHILDREN(rack_counters),
1859             OID_AUTO, "tlp_nada", CTLFLAG_RD,
1860             &rack_tlp_does_nada,
1861             "Total number of nada tlp calls");
1862         rack_try_scwnd = counter_u64_alloc(M_WAITOK);
1863         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1864             SYSCTL_CHILDREN(rack_counters),
1865             OID_AUTO, "tried_scwnd", CTLFLAG_RD,
1866             &rack_try_scwnd,
1867             "Total number of scwnd attempts");
1868
1869         rack_per_timer_hole = counter_u64_alloc(M_WAITOK);
1870         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1871             SYSCTL_CHILDREN(rack_counters),
1872             OID_AUTO, "timer_hole", CTLFLAG_RD,
1873             &rack_per_timer_hole,
1874             "Total persists start in timer hole");
1875
1876         rack_sbsndptr_wrong = counter_u64_alloc(M_WAITOK);
1877         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1878             SYSCTL_CHILDREN(rack_counters),
1879             OID_AUTO, "sndptr_wrong", CTLFLAG_RD,
1880             &rack_sbsndptr_wrong, "Total number of times the saved sbsndptr was incorret");
1881         rack_sbsndptr_right = counter_u64_alloc(M_WAITOK);
1882         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1883             SYSCTL_CHILDREN(rack_counters),
1884             OID_AUTO, "sndptr_right", CTLFLAG_RD,
1885             &rack_sbsndptr_right, "Total number of times the saved sbsndptr was corret");
1886
1887         COUNTER_ARRAY_ALLOC(rack_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK);
1888         SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1889             OID_AUTO, "outsize", CTLFLAG_RD,
1890             rack_out_size, TCP_MSS_ACCT_SIZE, "MSS send sizes");
1891         COUNTER_ARRAY_ALLOC(rack_opts_arry, RACK_OPTS_SIZE, M_WAITOK);
1892         SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1893             OID_AUTO, "opts", CTLFLAG_RD,
1894             rack_opts_arry, RACK_OPTS_SIZE, "RACK Option Stats");
1895         SYSCTL_ADD_PROC(&rack_sysctl_ctx,
1896             SYSCTL_CHILDREN(rack_sysctl_root),
1897             OID_AUTO, "clear", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1898             &rack_clear_counter, 0, sysctl_rack_clear, "IU", "Clear counters");
1899 }
1900
1901 static __inline int
1902 rb_map_cmp(struct rack_sendmap *b, struct rack_sendmap *a)
1903 {
1904         if (SEQ_GEQ(b->r_start, a->r_start) &&
1905             SEQ_LT(b->r_start, a->r_end)) {
1906                 /*
1907                  * The entry b is within the
1908                  * block a. i.e.:
1909                  * a --   |-------------|
1910                  * b --   |----|
1911                  * <or>
1912                  * b --       |------|
1913                  * <or>
1914                  * b --       |-----------|
1915                  */
1916                 return (0);
1917         } else if (SEQ_GEQ(b->r_start, a->r_end)) {
1918                 /*
1919                  * b falls as either the next
1920                  * sequence block after a so a
1921                  * is said to be smaller than b.
1922                  * i.e:
1923                  * a --   |------|
1924                  * b --          |--------|
1925                  * or
1926                  * b --              |-----|
1927                  */
1928                 return (1);
1929         }
1930         /*
1931          * Whats left is where a is
1932          * larger than b. i.e:
1933          * a --         |-------|
1934          * b --  |---|
1935          * or even possibly
1936          * b --   |--------------|
1937          */
1938         return (-1);
1939 }
1940
1941 RB_PROTOTYPE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp);
1942 RB_GENERATE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp);
1943
1944 static uint32_t
1945 rc_init_window(struct tcp_rack *rack)
1946 {
1947         uint32_t win;
1948
1949         if (rack->rc_init_win == 0) {
1950                 /*
1951                  * Nothing set by the user, use the system stack
1952                  * default.
1953                  */
1954                 return (tcp_compute_initwnd(tcp_maxseg(rack->rc_tp)));
1955         }
1956         win = ctf_fixed_maxseg(rack->rc_tp) * rack->rc_init_win;
1957         return (win);
1958 }
1959
1960 static uint64_t
1961 rack_get_fixed_pacing_bw(struct tcp_rack *rack)
1962 {
1963         if (IN_FASTRECOVERY(rack->rc_tp->t_flags))
1964                 return (rack->r_ctl.rc_fixed_pacing_rate_rec);
1965         else if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
1966                 return (rack->r_ctl.rc_fixed_pacing_rate_ss);
1967         else
1968                 return (rack->r_ctl.rc_fixed_pacing_rate_ca);
1969 }
1970
1971 static uint64_t
1972 rack_get_bw(struct tcp_rack *rack)
1973 {
1974         if (rack->use_fixed_rate) {
1975                 /* Return the fixed pacing rate */
1976                 return (rack_get_fixed_pacing_bw(rack));
1977         }
1978         if (rack->r_ctl.gp_bw == 0) {
1979                 /*
1980                  * We have yet no b/w measurement,
1981                  * if we have a user set initial bw
1982                  * return it. If we don't have that and
1983                  * we have an srtt, use the tcp IW (10) to
1984                  * calculate a fictional b/w over the SRTT
1985                  * which is more or less a guess. Note
1986                  * we don't use our IW from rack on purpose
1987                  * so if we have like IW=30, we are not
1988                  * calculating a "huge" b/w.
1989                  */
1990                 uint64_t bw, srtt;
1991                 if (rack->r_ctl.init_rate)
1992                         return (rack->r_ctl.init_rate);
1993
1994                 /* Has the user set a max peak rate? */
1995 #ifdef NETFLIX_PEAKRATE
1996                 if (rack->rc_tp->t_maxpeakrate)
1997                         return (rack->rc_tp->t_maxpeakrate);
1998 #endif
1999                 /* Ok lets come up with the IW guess, if we have a srtt */
2000                 if (rack->rc_tp->t_srtt == 0) {
2001                         /*
2002                          * Go with old pacing method
2003                          * i.e. burst mitigation only.
2004                          */
2005                         return (0);
2006                 }
2007                 /* Ok lets get the initial TCP win (not racks) */
2008                 bw = tcp_compute_initwnd(tcp_maxseg(rack->rc_tp));
2009                 srtt = (uint64_t)rack->rc_tp->t_srtt;
2010                 bw *= (uint64_t)USECS_IN_SECOND;
2011                 bw /= srtt;
2012                 if (rack->r_ctl.bw_rate_cap && (bw > rack->r_ctl.bw_rate_cap))
2013                         bw = rack->r_ctl.bw_rate_cap;
2014                 return (bw);
2015         } else {
2016                 uint64_t bw;
2017
2018                 if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
2019                         /* Averaging is done, we can return the value */
2020                         bw = rack->r_ctl.gp_bw;
2021                 } else {
2022                         /* Still doing initial average must calculate */
2023                         bw = rack->r_ctl.gp_bw / rack->r_ctl.num_measurements;
2024                 }
2025 #ifdef NETFLIX_PEAKRATE
2026                 if ((rack->rc_tp->t_maxpeakrate) &&
2027                     (bw > rack->rc_tp->t_maxpeakrate)) {
2028                         /* The user has set a peak rate to pace at
2029                          * don't allow us to pace faster than that.
2030                          */
2031                         return (rack->rc_tp->t_maxpeakrate);
2032                 }
2033 #endif
2034                 if (rack->r_ctl.bw_rate_cap && (bw > rack->r_ctl.bw_rate_cap))
2035                         bw = rack->r_ctl.bw_rate_cap;
2036                 return (bw);
2037         }
2038 }
2039
2040 static uint16_t
2041 rack_get_output_gain(struct tcp_rack *rack, struct rack_sendmap *rsm)
2042 {
2043         if (rack->use_fixed_rate) {
2044                 return (100);
2045         } else if (rack->in_probe_rtt && (rsm == NULL))
2046                 return (rack->r_ctl.rack_per_of_gp_probertt);
2047         else if ((IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
2048                   rack->r_ctl.rack_per_of_gp_rec)) {
2049                 if (rsm) {
2050                         /* a retransmission always use the recovery rate */
2051                         return (rack->r_ctl.rack_per_of_gp_rec);
2052                 } else if (rack->rack_rec_nonrxt_use_cr) {
2053                         /* Directed to use the configured rate */
2054                         goto configured_rate;
2055                 } else if (rack->rack_no_prr &&
2056                            (rack->r_ctl.rack_per_of_gp_rec > 100)) {
2057                         /* No PRR, lets just use the b/w estimate only */
2058                         return (100);
2059                 } else {
2060                         /*
2061                          * Here we may have a non-retransmit but we
2062                          * have no overrides, so just use the recovery
2063                          * rate (prr is in effect).
2064                          */
2065                         return (rack->r_ctl.rack_per_of_gp_rec);
2066                 }
2067         }
2068 configured_rate:
2069         /* For the configured rate we look at our cwnd vs the ssthresh */
2070         if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
2071                 return (rack->r_ctl.rack_per_of_gp_ss);
2072         else
2073                 return (rack->r_ctl.rack_per_of_gp_ca);
2074 }
2075
2076 static void
2077 rack_log_hdwr_pacing(struct tcp_rack *rack,
2078                      uint64_t rate, uint64_t hw_rate, int line,
2079                      int error, uint16_t mod)
2080 {
2081         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2082                 union tcp_log_stackspecific log;
2083                 struct timeval tv;
2084                 const struct ifnet *ifp;
2085
2086                 memset(&log, 0, sizeof(log));
2087                 log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff);
2088                 log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff);
2089                 if (rack->r_ctl.crte) {
2090                         ifp = rack->r_ctl.crte->ptbl->rs_ifp;
2091                 } else if (rack->rc_inp->inp_route.ro_nh &&
2092                            rack->rc_inp->inp_route.ro_nh->nh_ifp) {
2093                         ifp = rack->rc_inp->inp_route.ro_nh->nh_ifp;
2094                 } else
2095                         ifp = NULL;
2096                 if (ifp) {
2097                         log.u_bbr.flex3 = (((uint64_t)ifp  >> 32) & 0x00000000ffffffff);
2098                         log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff);
2099                 }
2100                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2101                 log.u_bbr.bw_inuse = rate;
2102                 log.u_bbr.flex5 = line;
2103                 log.u_bbr.flex6 = error;
2104                 log.u_bbr.flex7 = mod;
2105                 log.u_bbr.applimited = rack->r_ctl.rc_pace_max_segs;
2106                 log.u_bbr.flex8 = rack->use_fixed_rate;
2107                 log.u_bbr.flex8 <<= 1;
2108                 log.u_bbr.flex8 |= rack->rack_hdrw_pacing;
2109                 log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
2110                 log.u_bbr.delRate = rack->r_ctl.crte_prev_rate;
2111                 if (rack->r_ctl.crte)
2112                         log.u_bbr.cur_del_rate = rack->r_ctl.crte->rate;
2113                 else
2114                         log.u_bbr.cur_del_rate = 0;
2115                 log.u_bbr.rttProp = rack->r_ctl.last_hw_bw_req;
2116                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2117                     &rack->rc_inp->inp_socket->so_rcv,
2118                     &rack->rc_inp->inp_socket->so_snd,
2119                     BBR_LOG_HDWR_PACE, 0,
2120                     0, &log, false, &tv);
2121         }
2122 }
2123
2124 static uint64_t
2125 rack_get_output_bw(struct tcp_rack *rack, uint64_t bw, struct rack_sendmap *rsm, int *capped)
2126 {
2127         /*
2128          * We allow rack_per_of_gp_xx to dictate our bw rate we want.
2129          */
2130         uint64_t bw_est, high_rate;
2131         uint64_t gain;
2132
2133         gain = (uint64_t)rack_get_output_gain(rack, rsm);
2134         bw_est = bw * gain;
2135         bw_est /= (uint64_t)100;
2136         /* Never fall below the minimum (def 64kbps) */
2137         if (bw_est < RACK_MIN_BW)
2138                 bw_est = RACK_MIN_BW;
2139         if (rack->r_rack_hw_rate_caps) {
2140                 /* Rate caps are in place */
2141                 if (rack->r_ctl.crte != NULL) {
2142                         /* We have a hdwr rate already */
2143                         high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
2144                         if (bw_est >= high_rate) {
2145                                 /* We are capping bw at the highest rate table entry */
2146                                 rack_log_hdwr_pacing(rack,
2147                                                      bw_est, high_rate, __LINE__,
2148                                                      0, 3);
2149                                 bw_est = high_rate;
2150                                 if (capped)
2151                                         *capped = 1;
2152                         }
2153                 } else if ((rack->rack_hdrw_pacing == 0) &&
2154                            (rack->rack_hdw_pace_ena) &&
2155                            (rack->rack_attempt_hdwr_pace == 0) &&
2156                            (rack->rc_inp->inp_route.ro_nh != NULL) &&
2157                            (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
2158                         /*
2159                          * Special case, we have not yet attempted hardware
2160                          * pacing, and yet we may, when we do, find out if we are
2161                          * above the highest rate. We need to know the maxbw for the interface
2162                          * in question (if it supports ratelimiting). We get back
2163                          * a 0, if the interface is not found in the RL lists.
2164                          */
2165                         high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
2166                         if (high_rate) {
2167                                 /* Yep, we have a rate is it above this rate? */
2168                                 if (bw_est > high_rate) {
2169                                         bw_est = high_rate;
2170                                         if (capped)
2171                                                 *capped = 1;
2172                                 }
2173                         }
2174                 }
2175         }
2176         return (bw_est);
2177 }
2178
2179 static void
2180 rack_log_retran_reason(struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t tsused, uint32_t thresh, int mod)
2181 {
2182         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2183                 union tcp_log_stackspecific log;
2184                 struct timeval tv;
2185
2186                 if ((mod != 1) && (rack_verbose_logging == 0)) {
2187                         /*
2188                          * We get 3 values currently for mod
2189                          * 1 - We are retransmitting and this tells the reason.
2190                          * 2 - We are clearing a dup-ack count.
2191                          * 3 - We are incrementing a dup-ack count.
2192                          *
2193                          * The clear/increment are only logged
2194                          * if you have BBverbose on.
2195                          */
2196                         return;
2197                 }
2198                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2199                 log.u_bbr.flex1 = tsused;
2200                 log.u_bbr.flex2 = thresh;
2201                 log.u_bbr.flex3 = rsm->r_flags;
2202                 log.u_bbr.flex4 = rsm->r_dupack;
2203                 log.u_bbr.flex5 = rsm->r_start;
2204                 log.u_bbr.flex6 = rsm->r_end;
2205                 log.u_bbr.flex8 = mod;
2206                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2207                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2208                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2209                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2210                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2211                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2212                 log.u_bbr.pacing_gain = rack->r_must_retran;
2213                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2214                     &rack->rc_inp->inp_socket->so_rcv,
2215                     &rack->rc_inp->inp_socket->so_snd,
2216                     BBR_LOG_SETTINGS_CHG, 0,
2217                     0, &log, false, &tv);
2218         }
2219 }
2220
2221 static void
2222 rack_log_to_start(struct tcp_rack *rack, uint32_t cts, uint32_t to, int32_t slot, uint8_t which)
2223 {
2224         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2225                 union tcp_log_stackspecific log;
2226                 struct timeval tv;
2227
2228                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2229                 log.u_bbr.flex1 = rack->rc_tp->t_srtt;
2230                 log.u_bbr.flex2 = to;
2231                 log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags;
2232                 log.u_bbr.flex4 = slot;
2233                 log.u_bbr.flex5 = rack->rc_inp->inp_hptsslot;
2234                 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2235                 log.u_bbr.flex7 = rack->rc_in_persist;
2236                 log.u_bbr.flex8 = which;
2237                 if (rack->rack_no_prr)
2238                         log.u_bbr.pkts_out = 0;
2239                 else
2240                         log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
2241                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2242                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2243                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2244                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2245                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2246                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2247                 log.u_bbr.pacing_gain = rack->r_must_retran;
2248                 log.u_bbr.lt_epoch = rack->rc_tp->t_rxtshift;
2249                 log.u_bbr.lost = rack_rto_min;
2250                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2251                     &rack->rc_inp->inp_socket->so_rcv,
2252                     &rack->rc_inp->inp_socket->so_snd,
2253                     BBR_LOG_TIMERSTAR, 0,
2254                     0, &log, false, &tv);
2255         }
2256 }
2257
2258 static void
2259 rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm)
2260 {
2261         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2262                 union tcp_log_stackspecific log;
2263                 struct timeval tv;
2264
2265                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2266                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2267                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2268                 log.u_bbr.flex8 = to_num;
2269                 log.u_bbr.flex1 = rack->r_ctl.rc_rack_min_rtt;
2270                 log.u_bbr.flex2 = rack->rc_rack_rtt;
2271                 if (rsm == NULL)
2272                         log.u_bbr.flex3 = 0;
2273                 else
2274                         log.u_bbr.flex3 = rsm->r_end - rsm->r_start;
2275                 if (rack->rack_no_prr)
2276                         log.u_bbr.flex5 = 0;
2277                 else
2278                         log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2279                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2280                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2281                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2282                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2283                 log.u_bbr.pacing_gain = rack->r_must_retran;
2284                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2285                     &rack->rc_inp->inp_socket->so_rcv,
2286                     &rack->rc_inp->inp_socket->so_snd,
2287                     BBR_LOG_RTO, 0,
2288                     0, &log, false, &tv);
2289         }
2290 }
2291
2292 static void
2293 rack_log_map_chg(struct tcpcb *tp, struct tcp_rack *rack,
2294                  struct rack_sendmap *prev,
2295                  struct rack_sendmap *rsm,
2296                  struct rack_sendmap *next,
2297                  int flag, uint32_t th_ack, int line)
2298 {
2299         if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) {
2300                 union tcp_log_stackspecific log;
2301                 struct timeval tv;
2302
2303                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2304                 log.u_bbr.flex8 = flag;
2305                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2306                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2307                 log.u_bbr.cur_del_rate = (uint64_t)prev;
2308                 log.u_bbr.delRate = (uint64_t)rsm;
2309                 log.u_bbr.rttProp = (uint64_t)next;
2310                 log.u_bbr.flex7 = 0;
2311                 if (prev) {
2312                         log.u_bbr.flex1 = prev->r_start;
2313                         log.u_bbr.flex2 = prev->r_end;
2314                         log.u_bbr.flex7 |= 0x4;
2315                 }
2316                 if (rsm) {
2317                         log.u_bbr.flex3 = rsm->r_start;
2318                         log.u_bbr.flex4 = rsm->r_end;
2319                         log.u_bbr.flex7 |= 0x2;
2320                 }
2321                 if (next) {
2322                         log.u_bbr.flex5 = next->r_start;
2323                         log.u_bbr.flex6 = next->r_end;
2324                         log.u_bbr.flex7 |= 0x1;
2325                 }
2326                 log.u_bbr.applimited = line;
2327                 log.u_bbr.pkts_out = th_ack;
2328                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2329                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2330                 if (rack->rack_no_prr)
2331                         log.u_bbr.lost = 0;
2332                 else
2333                         log.u_bbr.lost = rack->r_ctl.rc_prr_sndcnt;
2334                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2335                     &rack->rc_inp->inp_socket->so_rcv,
2336                     &rack->rc_inp->inp_socket->so_snd,
2337                     TCP_LOG_MAPCHG, 0,
2338                     0, &log, false, &tv);
2339         }
2340 }
2341
2342 static void
2343 rack_log_rtt_upd(struct tcpcb *tp, struct tcp_rack *rack, uint32_t t, uint32_t len,
2344                  struct rack_sendmap *rsm, int conf)
2345 {
2346         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
2347                 union tcp_log_stackspecific log;
2348                 struct timeval tv;
2349                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2350                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2351                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2352                 log.u_bbr.flex1 = t;
2353                 log.u_bbr.flex2 = len;
2354                 log.u_bbr.flex3 = rack->r_ctl.rc_rack_min_rtt;
2355                 log.u_bbr.flex4 = rack->r_ctl.rack_rs.rs_rtt_lowest;
2356                 log.u_bbr.flex5 = rack->r_ctl.rack_rs.rs_rtt_highest;
2357                 log.u_bbr.flex6 = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2358                 log.u_bbr.flex7 = conf;
2359                 log.u_bbr.rttProp = (uint64_t)rack->r_ctl.rack_rs.rs_rtt_tot;
2360                 log.u_bbr.flex8 = rack->r_ctl.rc_rate_sample_method;
2361                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2362                 log.u_bbr.delivered = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2363                 log.u_bbr.pkts_out = rack->r_ctl.rack_rs.rs_flags;
2364                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2365                 if (rsm) {
2366                         log.u_bbr.pkt_epoch = rsm->r_start;
2367                         log.u_bbr.lost = rsm->r_end;
2368                         log.u_bbr.cwnd_gain = rsm->r_rtr_cnt;
2369                         log.u_bbr.pacing_gain = rsm->r_flags;
2370                 } else {
2371                         /* Its a SYN */
2372                         log.u_bbr.pkt_epoch = rack->rc_tp->iss;
2373                         log.u_bbr.lost = 0;
2374                         log.u_bbr.cwnd_gain = 0;
2375                         log.u_bbr.pacing_gain = 0;
2376                 }
2377                 /* Write out general bits of interest rrs here */
2378                 log.u_bbr.use_lt_bw = rack->rc_highly_buffered;
2379                 log.u_bbr.use_lt_bw <<= 1;
2380                 log.u_bbr.use_lt_bw |= rack->forced_ack;
2381                 log.u_bbr.use_lt_bw <<= 1;
2382                 log.u_bbr.use_lt_bw |= rack->rc_gp_dyn_mul;
2383                 log.u_bbr.use_lt_bw <<= 1;
2384                 log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
2385                 log.u_bbr.use_lt_bw <<= 1;
2386                 log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
2387                 log.u_bbr.use_lt_bw <<= 1;
2388                 log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
2389                 log.u_bbr.use_lt_bw <<= 1;
2390                 log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
2391                 log.u_bbr.use_lt_bw <<= 1;
2392                 log.u_bbr.use_lt_bw |= rack->rc_dragged_bottom;
2393                 log.u_bbr.applimited = rack->r_ctl.rc_target_probertt_flight;
2394                 log.u_bbr.epoch = rack->r_ctl.rc_time_probertt_starts;
2395                 log.u_bbr.lt_epoch = rack->r_ctl.rc_time_probertt_entered;
2396                 log.u_bbr.cur_del_rate = rack->r_ctl.rc_lower_rtt_us_cts;
2397                 log.u_bbr.delRate = rack->r_ctl.rc_gp_srtt;
2398                 log.u_bbr.bw_inuse = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
2399                 log.u_bbr.bw_inuse <<= 32;
2400                 if (rsm)
2401                         log.u_bbr.bw_inuse |= ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]);
2402                 TCP_LOG_EVENTP(tp, NULL,
2403                     &rack->rc_inp->inp_socket->so_rcv,
2404                     &rack->rc_inp->inp_socket->so_snd,
2405                     BBR_LOG_BBRRTT, 0,
2406                     0, &log, false, &tv);
2407
2408
2409         }
2410 }
2411
2412 static void
2413 rack_log_rtt_sample(struct tcp_rack *rack, uint32_t rtt)
2414 {
2415         /*
2416          * Log the rtt sample we are
2417          * applying to the srtt algorithm in
2418          * useconds.
2419          */
2420         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2421                 union tcp_log_stackspecific log;
2422                 struct timeval tv;
2423
2424                 /* Convert our ms to a microsecond */
2425                 memset(&log, 0, sizeof(log));
2426                 log.u_bbr.flex1 = rtt;
2427                 log.u_bbr.flex2 = rack->r_ctl.ack_count;
2428                 log.u_bbr.flex3 = rack->r_ctl.sack_count;
2429                 log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
2430                 log.u_bbr.flex5 = rack->r_ctl.sack_moved_extra;
2431                 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2432                 log.u_bbr.flex7 = 1;
2433                 log.u_bbr.flex8 = rack->sack_attack_disable;
2434                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2435                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2436                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2437                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2438                 log.u_bbr.pacing_gain = rack->r_must_retran;
2439                 /*
2440                  * We capture in delRate the upper 32 bits as
2441                  * the confidence level we had declared, and the
2442                  * lower 32 bits as the actual RTT using the arrival
2443                  * timestamp.
2444                  */
2445                 log.u_bbr.delRate = rack->r_ctl.rack_rs.confidence;
2446                 log.u_bbr.delRate <<= 32;
2447                 log.u_bbr.delRate |= rack->r_ctl.rack_rs.rs_us_rtt;
2448                 /* Lets capture all the things that make up t_rtxcur */
2449                 log.u_bbr.applimited = rack_rto_min;
2450                 log.u_bbr.epoch = rack_rto_max;
2451                 log.u_bbr.lt_epoch = rtt;
2452                 log.u_bbr.lost = rack_rto_min;
2453                 log.u_bbr.pkt_epoch = TICKS_2_USEC(tcp_rexmit_slop);
2454                 log.u_bbr.rttProp = RACK_REXMTVAL(rack->rc_tp);
2455                 log.u_bbr.bw_inuse = rack->r_ctl.act_rcv_time.tv_sec;
2456                 log.u_bbr.bw_inuse *= HPTS_USEC_IN_SEC;
2457                 log.u_bbr.bw_inuse += rack->r_ctl.act_rcv_time.tv_usec;
2458                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2459                     &rack->rc_inp->inp_socket->so_rcv,
2460                     &rack->rc_inp->inp_socket->so_snd,
2461                     TCP_LOG_RTT, 0,
2462                     0, &log, false, &tv);
2463         }
2464 }
2465
2466 static void
2467 rack_log_rtt_sample_calc(struct tcp_rack *rack, uint32_t rtt, uint32_t send_time, uint32_t ack_time, int where)
2468 {
2469         if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2470                 union tcp_log_stackspecific log;
2471                 struct timeval tv;
2472
2473                 /* Convert our ms to a microsecond */
2474                 memset(&log, 0, sizeof(log));
2475                 log.u_bbr.flex1 = rtt;
2476                 log.u_bbr.flex2 = send_time;
2477                 log.u_bbr.flex3 = ack_time;
2478                 log.u_bbr.flex4 = where;
2479                 log.u_bbr.flex7 = 2;
2480                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2481                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2482                     &rack->rc_inp->inp_socket->so_rcv,
2483                     &rack->rc_inp->inp_socket->so_snd,
2484                     TCP_LOG_RTT, 0,
2485                     0, &log, false, &tv);
2486         }
2487 }
2488
2489
2490
2491 static inline void
2492 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line)
2493 {
2494         if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) {
2495                 union tcp_log_stackspecific log;
2496                 struct timeval tv;
2497
2498                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2499                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2500                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2501                 log.u_bbr.flex1 = line;
2502                 log.u_bbr.flex2 = tick;
2503                 log.u_bbr.flex3 = tp->t_maxunacktime;
2504                 log.u_bbr.flex4 = tp->t_acktime;
2505                 log.u_bbr.flex8 = event;
2506                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2507                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2508                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2509                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2510                 log.u_bbr.pacing_gain = rack->r_must_retran;
2511                 TCP_LOG_EVENTP(tp, NULL,
2512                     &rack->rc_inp->inp_socket->so_rcv,
2513                     &rack->rc_inp->inp_socket->so_snd,
2514                     BBR_LOG_PROGRESS, 0,
2515                     0, &log, false, &tv);
2516         }
2517 }
2518
2519 static void
2520 rack_log_type_bbrsnd(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint32_t cts, struct timeval *tv)
2521 {
2522         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2523                 union tcp_log_stackspecific log;
2524
2525                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2526                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2527                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2528                 log.u_bbr.flex1 = slot;
2529                 if (rack->rack_no_prr)
2530                         log.u_bbr.flex2 = 0;
2531                 else
2532                         log.u_bbr.flex2 = rack->r_ctl.rc_prr_sndcnt;
2533                 log.u_bbr.flex7 = (0x0000ffff & rack->r_ctl.rc_hpts_flags);
2534                 log.u_bbr.flex8 = rack->rc_in_persist;
2535                 log.u_bbr.timeStamp = cts;
2536                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2537                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2538                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2539                 log.u_bbr.pacing_gain = rack->r_must_retran;
2540                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2541                     &rack->rc_inp->inp_socket->so_rcv,
2542                     &rack->rc_inp->inp_socket->so_snd,
2543                     BBR_LOG_BBRSND, 0,
2544                     0, &log, false, tv);
2545         }
2546 }
2547
2548 static void
2549 rack_log_doseg_done(struct tcp_rack *rack, uint32_t cts, int32_t nxt_pkt, int32_t did_out, int way_out, int nsegs)
2550 {
2551         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2552                 union tcp_log_stackspecific log;
2553                 struct timeval tv;
2554
2555                 memset(&log, 0, sizeof(log));
2556                 log.u_bbr.flex1 = did_out;
2557                 log.u_bbr.flex2 = nxt_pkt;
2558                 log.u_bbr.flex3 = way_out;
2559                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
2560                 if (rack->rack_no_prr)
2561                         log.u_bbr.flex5 = 0;
2562                 else
2563                         log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2564                 log.u_bbr.flex6 = nsegs;
2565                 log.u_bbr.applimited = rack->r_ctl.rc_pace_min_segs;
2566                 log.u_bbr.flex7 = rack->rc_ack_can_sendout_data;        /* Do we have ack-can-send set */
2567                 log.u_bbr.flex7 <<= 1;
2568                 log.u_bbr.flex7 |= rack->r_fast_output; /* is fast output primed */
2569                 log.u_bbr.flex7 <<= 1;
2570                 log.u_bbr.flex7 |= rack->r_wanted_output;       /* Do we want output */
2571                 log.u_bbr.flex8 = rack->rc_in_persist;
2572                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2573                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2574                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2575                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
2576                 log.u_bbr.use_lt_bw <<= 1;
2577                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
2578                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2579                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2580                 log.u_bbr.pacing_gain = rack->r_must_retran;
2581                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2582                     &rack->rc_inp->inp_socket->so_rcv,
2583                     &rack->rc_inp->inp_socket->so_snd,
2584                     BBR_LOG_DOSEG_DONE, 0,
2585                     0, &log, false, &tv);
2586         }
2587 }
2588
2589 static void
2590 rack_log_type_pacing_sizes(struct tcpcb *tp, struct tcp_rack *rack, uint32_t arg1, uint32_t arg2, uint32_t arg3, uint8_t frm)
2591 {
2592         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
2593                 union tcp_log_stackspecific log;
2594                 struct timeval tv;
2595                 uint32_t cts;
2596
2597                 memset(&log, 0, sizeof(log));
2598                 cts = tcp_get_usecs(&tv);
2599                 log.u_bbr.flex1 = rack->r_ctl.rc_pace_min_segs;
2600                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
2601                 log.u_bbr.flex4 = arg1;
2602                 log.u_bbr.flex5 = arg2;
2603                 log.u_bbr.flex6 = arg3;
2604                 log.u_bbr.flex8 = frm;
2605                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2606                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2607                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2608                 log.u_bbr.applimited = rack->r_ctl.rc_sacked;
2609                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2610                 log.u_bbr.pacing_gain = rack->r_must_retran;
2611                 TCP_LOG_EVENTP(tp, NULL,
2612                     &tp->t_inpcb->inp_socket->so_rcv,
2613                     &tp->t_inpcb->inp_socket->so_snd,
2614                     TCP_HDWR_PACE_SIZE, 0,
2615                     0, &log, false, &tv);
2616         }
2617 }
2618
2619 static void
2620 rack_log_type_just_return(struct tcp_rack *rack, uint32_t cts, uint32_t tlen, uint32_t slot,
2621                           uint8_t hpts_calling, int reason, uint32_t cwnd_to_use)
2622 {
2623         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2624                 union tcp_log_stackspecific log;
2625                 struct timeval tv;
2626
2627                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2628                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2629                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2630                 log.u_bbr.flex1 = slot;
2631                 log.u_bbr.flex2 = rack->r_ctl.rc_hpts_flags;
2632                 log.u_bbr.flex4 = reason;
2633                 if (rack->rack_no_prr)
2634                         log.u_bbr.flex5 = 0;
2635                 else
2636                         log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2637                 log.u_bbr.flex7 = hpts_calling;
2638                 log.u_bbr.flex8 = rack->rc_in_persist;
2639                 log.u_bbr.lt_epoch = cwnd_to_use;
2640                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2641                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2642                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2643                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2644                 log.u_bbr.pacing_gain = rack->r_must_retran;
2645                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2646                     &rack->rc_inp->inp_socket->so_rcv,
2647                     &rack->rc_inp->inp_socket->so_snd,
2648                     BBR_LOG_JUSTRET, 0,
2649                     tlen, &log, false, &tv);
2650         }
2651 }
2652
2653 static void
2654 rack_log_to_cancel(struct tcp_rack *rack, int32_t hpts_removed, int line, uint32_t us_cts,
2655                    struct timeval *tv, uint32_t flags_on_entry)
2656 {
2657         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2658                 union tcp_log_stackspecific log;
2659
2660                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2661                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2662                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2663                 log.u_bbr.flex1 = line;
2664                 log.u_bbr.flex2 = rack->r_ctl.rc_last_output_to;
2665                 log.u_bbr.flex3 = flags_on_entry;
2666                 log.u_bbr.flex4 = us_cts;
2667                 if (rack->rack_no_prr)
2668                         log.u_bbr.flex5 = 0;
2669                 else
2670                         log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2671                 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2672                 log.u_bbr.flex7 = hpts_removed;
2673                 log.u_bbr.flex8 = 1;
2674                 log.u_bbr.applimited = rack->r_ctl.rc_hpts_flags;
2675                 log.u_bbr.timeStamp = us_cts;
2676                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2677                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2678                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2679                 log.u_bbr.pacing_gain = rack->r_must_retran;
2680                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2681                     &rack->rc_inp->inp_socket->so_rcv,
2682                     &rack->rc_inp->inp_socket->so_snd,
2683                     BBR_LOG_TIMERCANC, 0,
2684                     0, &log, false, tv);
2685         }
2686 }
2687
2688 static void
2689 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
2690                           uint32_t flex1, uint32_t flex2,
2691                           uint32_t flex3, uint32_t flex4,
2692                           uint32_t flex5, uint32_t flex6,
2693                           uint16_t flex7, uint8_t mod)
2694 {
2695         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2696                 union tcp_log_stackspecific log;
2697                 struct timeval tv;
2698
2699                 if (mod == 1) {
2700                         /* No you can't use 1, its for the real to cancel */
2701                         return;
2702                 }
2703                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2704                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2705                 log.u_bbr.flex1 = flex1;
2706                 log.u_bbr.flex2 = flex2;
2707                 log.u_bbr.flex3 = flex3;
2708                 log.u_bbr.flex4 = flex4;
2709                 log.u_bbr.flex5 = flex5;
2710                 log.u_bbr.flex6 = flex6;
2711                 log.u_bbr.flex7 = flex7;
2712                 log.u_bbr.flex8 = mod;
2713                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2714                     &rack->rc_inp->inp_socket->so_rcv,
2715                     &rack->rc_inp->inp_socket->so_snd,
2716                     BBR_LOG_TIMERCANC, 0,
2717                     0, &log, false, &tv);
2718         }
2719 }
2720
2721 static void
2722 rack_log_to_processing(struct tcp_rack *rack, uint32_t cts, int32_t ret, int32_t timers)
2723 {
2724         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2725                 union tcp_log_stackspecific log;
2726                 struct timeval tv;
2727
2728                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2729                 log.u_bbr.flex1 = timers;
2730                 log.u_bbr.flex2 = ret;
2731                 log.u_bbr.flex3 = rack->r_ctl.rc_timer_exp;
2732                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
2733                 log.u_bbr.flex5 = cts;
2734                 if (rack->rack_no_prr)
2735                         log.u_bbr.flex6 = 0;
2736                 else
2737                         log.u_bbr.flex6 = rack->r_ctl.rc_prr_sndcnt;
2738                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2739                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2740                 log.u_bbr.pacing_gain = rack->r_must_retran;
2741                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2742                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2743                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2744                     &rack->rc_inp->inp_socket->so_rcv,
2745                     &rack->rc_inp->inp_socket->so_snd,
2746                     BBR_LOG_TO_PROCESS, 0,
2747                     0, &log, false, &tv);
2748         }
2749 }
2750
2751 static void
2752 rack_log_to_prr(struct tcp_rack *rack, int frm, int orig_cwnd)
2753 {
2754         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2755                 union tcp_log_stackspecific log;
2756                 struct timeval tv;
2757
2758                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2759                 log.u_bbr.flex1 = rack->r_ctl.rc_prr_out;
2760                 log.u_bbr.flex2 = rack->r_ctl.rc_prr_recovery_fs;
2761                 if (rack->rack_no_prr)
2762                         log.u_bbr.flex3 = 0;
2763                 else
2764                         log.u_bbr.flex3 = rack->r_ctl.rc_prr_sndcnt;
2765                 log.u_bbr.flex4 = rack->r_ctl.rc_prr_delivered;
2766                 log.u_bbr.flex5 = rack->r_ctl.rc_sacked;
2767                 log.u_bbr.flex6 = rack->r_ctl.rc_holes_rxt;
2768                 log.u_bbr.flex8 = frm;
2769                 log.u_bbr.pkts_out = orig_cwnd;
2770                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2771                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2772                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
2773                 log.u_bbr.use_lt_bw <<= 1;
2774                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
2775                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2776                     &rack->rc_inp->inp_socket->so_rcv,
2777                     &rack->rc_inp->inp_socket->so_snd,
2778                     BBR_LOG_BBRUPD, 0,
2779                     0, &log, false, &tv);
2780         }
2781 }
2782
2783 #ifdef NETFLIX_EXP_DETECTION
2784 static void
2785 rack_log_sad(struct tcp_rack *rack, int event)
2786 {
2787         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2788                 union tcp_log_stackspecific log;
2789                 struct timeval tv;
2790
2791                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2792                 log.u_bbr.flex1 = rack->r_ctl.sack_count;
2793                 log.u_bbr.flex2 = rack->r_ctl.ack_count;
2794                 log.u_bbr.flex3 = rack->r_ctl.sack_moved_extra;
2795                 log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
2796                 log.u_bbr.flex5 = rack->r_ctl.rc_num_maps_alloced;
2797                 log.u_bbr.flex6 = tcp_sack_to_ack_thresh;
2798                 log.u_bbr.pkts_out = tcp_sack_to_move_thresh;
2799                 log.u_bbr.lt_epoch = (tcp_force_detection << 8);
2800                 log.u_bbr.lt_epoch |= rack->do_detection;
2801                 log.u_bbr.applimited = tcp_map_minimum;
2802                 log.u_bbr.flex7 = rack->sack_attack_disable;
2803                 log.u_bbr.flex8 = event;
2804                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2805                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2806                 log.u_bbr.delivered = tcp_sad_decay_val;
2807                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2808                     &rack->rc_inp->inp_socket->so_rcv,
2809                     &rack->rc_inp->inp_socket->so_snd,
2810                     TCP_SAD_DETECTION, 0,
2811                     0, &log, false, &tv);
2812         }
2813 }
2814 #endif
2815
2816 static void
2817 rack_counter_destroy(void)
2818 {
2819         int i;
2820
2821         counter_u64_free(rack_fto_send);
2822         counter_u64_free(rack_fto_rsm_send);
2823         counter_u64_free(rack_nfto_resend);
2824         counter_u64_free(rack_hw_pace_init_fail);
2825         counter_u64_free(rack_hw_pace_lost);
2826         counter_u64_free(rack_non_fto_send);
2827         counter_u64_free(rack_extended_rfo);
2828         counter_u64_free(rack_ack_total);
2829         counter_u64_free(rack_express_sack);
2830         counter_u64_free(rack_sack_total);
2831         counter_u64_free(rack_move_none);
2832         counter_u64_free(rack_move_some);
2833         counter_u64_free(rack_sack_attacks_detected);
2834         counter_u64_free(rack_sack_attacks_reversed);
2835         counter_u64_free(rack_sack_used_next_merge);
2836         counter_u64_free(rack_sack_used_prev_merge);
2837         counter_u64_free(rack_badfr);
2838         counter_u64_free(rack_badfr_bytes);
2839         counter_u64_free(rack_rtm_prr_retran);
2840         counter_u64_free(rack_rtm_prr_newdata);
2841         counter_u64_free(rack_timestamp_mismatch);
2842         counter_u64_free(rack_find_high);
2843         counter_u64_free(rack_reorder_seen);
2844         counter_u64_free(rack_tlp_tot);
2845         counter_u64_free(rack_tlp_newdata);
2846         counter_u64_free(rack_tlp_retran);
2847         counter_u64_free(rack_tlp_retran_bytes);
2848         counter_u64_free(rack_tlp_retran_fail);
2849         counter_u64_free(rack_to_tot);
2850         counter_u64_free(rack_to_arm_rack);
2851         counter_u64_free(rack_to_arm_tlp);
2852         counter_u64_free(rack_calc_zero);
2853         counter_u64_free(rack_calc_nonzero);
2854         counter_u64_free(rack_paced_segments);
2855         counter_u64_free(rack_unpaced_segments);
2856         counter_u64_free(rack_saw_enobuf);
2857         counter_u64_free(rack_saw_enobuf_hw);
2858         counter_u64_free(rack_saw_enetunreach);
2859         counter_u64_free(rack_hot_alloc);
2860         counter_u64_free(rack_to_alloc);
2861         counter_u64_free(rack_to_alloc_hard);
2862         counter_u64_free(rack_to_alloc_emerg);
2863         counter_u64_free(rack_to_alloc_limited);
2864         counter_u64_free(rack_alloc_limited_conns);
2865         counter_u64_free(rack_split_limited);
2866         for (i = 0; i < MAX_NUM_OF_CNTS; i++) {
2867                 counter_u64_free(rack_proc_comp_ack[i]);
2868         }
2869         counter_u64_free(rack_multi_single_eq);
2870         counter_u64_free(rack_proc_non_comp_ack);
2871         counter_u64_free(rack_sack_proc_all);
2872         counter_u64_free(rack_sack_proc_restart);
2873         counter_u64_free(rack_sack_proc_short);
2874         counter_u64_free(rack_enter_tlp_calc);
2875         counter_u64_free(rack_used_tlpmethod);
2876         counter_u64_free(rack_used_tlpmethod2);
2877         counter_u64_free(rack_sack_skipped_acked);
2878         counter_u64_free(rack_sack_splits);
2879         counter_u64_free(rack_progress_drops);
2880         counter_u64_free(rack_input_idle_reduces);
2881         counter_u64_free(rack_collapsed_win);
2882         counter_u64_free(rack_tlp_does_nada);
2883         counter_u64_free(rack_try_scwnd);
2884         counter_u64_free(rack_per_timer_hole);
2885         counter_u64_free(rack_large_ackcmp);
2886         counter_u64_free(rack_small_ackcmp);
2887 #ifdef INVARIANTS
2888         counter_u64_free(rack_adjust_map_bw);
2889 #endif
2890         COUNTER_ARRAY_FREE(rack_out_size, TCP_MSS_ACCT_SIZE);
2891         COUNTER_ARRAY_FREE(rack_opts_arry, RACK_OPTS_SIZE);
2892 }
2893
2894 static struct rack_sendmap *
2895 rack_alloc(struct tcp_rack *rack)
2896 {
2897         struct rack_sendmap *rsm;
2898
2899         /*
2900          * First get the top of the list it in
2901          * theory is the "hottest" rsm we have,
2902          * possibly just freed by ack processing.
2903          */
2904         if (rack->rc_free_cnt > rack_free_cache) {
2905                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
2906                 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2907                 counter_u64_add(rack_hot_alloc, 1);
2908                 rack->rc_free_cnt--;
2909                 return (rsm);
2910         }
2911         /*
2912          * Once we get under our free cache we probably
2913          * no longer have a "hot" one available. Lets
2914          * get one from UMA.
2915          */
2916         rsm = uma_zalloc(rack_zone, M_NOWAIT);
2917         if (rsm) {
2918                 rack->r_ctl.rc_num_maps_alloced++;
2919                 counter_u64_add(rack_to_alloc, 1);
2920                 return (rsm);
2921         }
2922         /*
2923          * Dig in to our aux rsm's (the last two) since
2924          * UMA failed to get us one.
2925          */
2926         if (rack->rc_free_cnt) {
2927                 counter_u64_add(rack_to_alloc_emerg, 1);
2928                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
2929                 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2930                 rack->rc_free_cnt--;
2931                 return (rsm);
2932         }
2933         return (NULL);
2934 }
2935
2936 static struct rack_sendmap *
2937 rack_alloc_full_limit(struct tcp_rack *rack)
2938 {
2939         if ((V_tcp_map_entries_limit > 0) &&
2940             (rack->do_detection == 0) &&
2941             (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
2942                 counter_u64_add(rack_to_alloc_limited, 1);
2943                 if (!rack->alloc_limit_reported) {
2944                         rack->alloc_limit_reported = 1;
2945                         counter_u64_add(rack_alloc_limited_conns, 1);
2946                 }
2947                 return (NULL);
2948         }
2949         return (rack_alloc(rack));
2950 }
2951
2952 /* wrapper to allocate a sendmap entry, subject to a specific limit */
2953 static struct rack_sendmap *
2954 rack_alloc_limit(struct tcp_rack *rack, uint8_t limit_type)
2955 {
2956         struct rack_sendmap *rsm;
2957
2958         if (limit_type) {
2959                 /* currently there is only one limit type */
2960                 if (V_tcp_map_split_limit > 0 &&
2961                     (rack->do_detection == 0) &&
2962                     rack->r_ctl.rc_num_split_allocs >= V_tcp_map_split_limit) {
2963                         counter_u64_add(rack_split_limited, 1);
2964                         if (!rack->alloc_limit_reported) {
2965                                 rack->alloc_limit_reported = 1;
2966                                 counter_u64_add(rack_alloc_limited_conns, 1);
2967                         }
2968                         return (NULL);
2969                 }
2970         }
2971
2972         /* allocate and mark in the limit type, if set */
2973         rsm = rack_alloc(rack);
2974         if (rsm != NULL && limit_type) {
2975                 rsm->r_limit_type = limit_type;
2976                 rack->r_ctl.rc_num_split_allocs++;
2977         }
2978         return (rsm);
2979 }
2980
2981 static void
2982 rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm)
2983 {
2984         if (rsm->r_flags & RACK_APP_LIMITED) {
2985                 if (rack->r_ctl.rc_app_limited_cnt > 0) {
2986                         rack->r_ctl.rc_app_limited_cnt--;
2987                 }
2988         }
2989         if (rsm->r_limit_type) {
2990                 /* currently there is only one limit type */
2991                 rack->r_ctl.rc_num_split_allocs--;
2992         }
2993         if (rsm == rack->r_ctl.rc_first_appl) {
2994                 if (rack->r_ctl.rc_app_limited_cnt == 0)
2995                         rack->r_ctl.rc_first_appl = NULL;
2996                 else {
2997                         /* Follow the next one out */
2998                         struct rack_sendmap fe;
2999
3000                         fe.r_start = rsm->r_nseq_appl;
3001                         rack->r_ctl.rc_first_appl = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
3002                 }
3003         }
3004         if (rsm == rack->r_ctl.rc_resend)
3005                 rack->r_ctl.rc_resend = NULL;
3006         if (rsm == rack->r_ctl.rc_rsm_at_retran)
3007                 rack->r_ctl.rc_rsm_at_retran = NULL;
3008         if (rsm == rack->r_ctl.rc_end_appl)
3009                 rack->r_ctl.rc_end_appl = NULL;
3010         if (rack->r_ctl.rc_tlpsend == rsm)
3011                 rack->r_ctl.rc_tlpsend = NULL;
3012         if (rack->r_ctl.rc_sacklast == rsm)
3013                 rack->r_ctl.rc_sacklast = NULL;
3014         memset(rsm, 0, sizeof(struct rack_sendmap));
3015         TAILQ_INSERT_HEAD(&rack->r_ctl.rc_free, rsm, r_tnext);
3016         rack->rc_free_cnt++;
3017 }
3018
3019 static void
3020 rack_free_trim(struct tcp_rack *rack)
3021 {
3022         struct rack_sendmap *rsm;
3023
3024         /*
3025          * Free up all the tail entries until
3026          * we get our list down to the limit.
3027          */
3028         while (rack->rc_free_cnt > rack_free_cache) {
3029                 rsm = TAILQ_LAST(&rack->r_ctl.rc_free, rack_head);
3030                 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3031                 rack->rc_free_cnt--;
3032                 uma_zfree(rack_zone, rsm);
3033         }
3034 }
3035
3036
3037 static uint32_t
3038 rack_get_measure_window(struct tcpcb *tp, struct tcp_rack *rack)
3039 {
3040         uint64_t srtt, bw, len, tim;
3041         uint32_t segsiz, def_len, minl;
3042
3043         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3044         def_len = rack_def_data_window * segsiz;
3045         if (rack->rc_gp_filled == 0) {
3046                 /*
3047                  * We have no measurement (IW is in flight?) so
3048                  * we can only guess using our data_window sysctl
3049                  * value (usually 100MSS).
3050                  */
3051                 return (def_len);
3052         }
3053         /*
3054          * Now we have a number of factors to consider.
3055          *
3056          * 1) We have a desired BDP which is usually
3057          *    at least 2.
3058          * 2) We have a minimum number of rtt's usually 1 SRTT
3059          *    but we allow it too to be more.
3060          * 3) We want to make sure a measurement last N useconds (if
3061          *    we have set rack_min_measure_usec.
3062          *
3063          * We handle the first concern here by trying to create a data
3064          * window of max(rack_def_data_window, DesiredBDP). The
3065          * second concern we handle in not letting the measurement
3066          * window end normally until at least the required SRTT's
3067          * have gone by which is done further below in
3068          * rack_enough_for_measurement(). Finally the third concern
3069          * we also handle here by calculating how long that time
3070          * would take at the current BW and then return the
3071          * max of our first calculation and that length. Note
3072          * that if rack_min_measure_usec is 0, we don't deal
3073          * with concern 3. Also for both Concern 1 and 3 an
3074          * application limited period could end the measurement
3075          * earlier.
3076          *
3077          * So lets calculate the BDP with the "known" b/w using
3078          * the SRTT has our rtt and then multiply it by the
3079          * goal.
3080          */
3081         bw = rack_get_bw(rack);
3082         srtt = (uint64_t)tp->t_srtt;
3083         len = bw * srtt;
3084         len /= (uint64_t)HPTS_USEC_IN_SEC;
3085         len *= max(1, rack_goal_bdp);
3086         /* Now we need to round up to the nearest MSS */
3087         len = roundup(len, segsiz);
3088         if (rack_min_measure_usec) {
3089                 /* Now calculate our min length for this b/w */
3090                 tim = rack_min_measure_usec;
3091                 minl = (tim * bw) / (uint64_t)HPTS_USEC_IN_SEC;
3092                 if (minl == 0)
3093                         minl = 1;
3094                 minl = roundup(minl, segsiz);
3095                 if (len < minl)
3096                         len = minl;
3097         }
3098         /*
3099          * Now if we have a very small window we want
3100          * to attempt to get the window that is
3101          * as small as possible. This happens on
3102          * low b/w connections and we don't want to
3103          * span huge numbers of rtt's between measurements.
3104          *
3105          * We basically include 2 over our "MIN window" so
3106          * that the measurement can be shortened (possibly) by
3107          * an ack'ed packet.
3108          */
3109         if (len < def_len)
3110                 return (max((uint32_t)len, ((MIN_GP_WIN+2) * segsiz)));
3111         else
3112                 return (max((uint32_t)len, def_len));
3113
3114 }
3115
3116 static int
3117 rack_enough_for_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq th_ack)
3118 {
3119         uint32_t tim, srtts, segsiz;
3120
3121         /*
3122          * Has enough time passed for the GP measurement to be valid?
3123          */
3124         if ((tp->snd_max == tp->snd_una) ||
3125             (th_ack == tp->snd_max)){
3126                 /* All is acked */
3127                 return (1);
3128         }
3129         if (SEQ_LT(th_ack, tp->gput_seq)) {
3130                 /* Not enough bytes yet */
3131                 return (0);
3132         }
3133         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3134         if (SEQ_LT(th_ack, tp->gput_ack) &&
3135             ((th_ack - tp->gput_seq) < max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
3136                 /* Not enough bytes yet */
3137                 return (0);
3138         }
3139         if (rack->r_ctl.rc_first_appl &&
3140             (rack->r_ctl.rc_first_appl->r_start == th_ack)) {
3141                 /*
3142                  * We are up to the app limited point
3143                  * we have to measure irrespective of the time..
3144                  */
3145                 return (1);
3146         }
3147         /* Now what about time? */
3148         srtts = (rack->r_ctl.rc_gp_srtt * rack_min_srtts);
3149         tim = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - tp->gput_ts;
3150         if (tim >= srtts) {
3151                 return (1);
3152         }
3153         /* Nope not even a full SRTT has passed */
3154         return (0);
3155 }
3156
3157 static void
3158 rack_log_timely(struct tcp_rack *rack,
3159                 uint32_t logged, uint64_t cur_bw, uint64_t low_bnd,
3160                 uint64_t up_bnd, int line, uint8_t method)
3161 {
3162         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
3163                 union tcp_log_stackspecific log;
3164                 struct timeval tv;
3165
3166                 memset(&log, 0, sizeof(log));
3167                 log.u_bbr.flex1 = logged;
3168                 log.u_bbr.flex2 = rack->rc_gp_timely_inc_cnt;
3169                 log.u_bbr.flex2 <<= 4;
3170                 log.u_bbr.flex2 |= rack->rc_gp_timely_dec_cnt;
3171                 log.u_bbr.flex2 <<= 4;
3172                 log.u_bbr.flex2 |= rack->rc_gp_incr;
3173                 log.u_bbr.flex2 <<= 4;
3174                 log.u_bbr.flex2 |= rack->rc_gp_bwred;
3175                 log.u_bbr.flex3 = rack->rc_gp_incr;
3176                 log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3177                 log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ca;
3178                 log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_rec;
3179                 log.u_bbr.flex7 = rack->rc_gp_bwred;
3180                 log.u_bbr.flex8 = method;
3181                 log.u_bbr.cur_del_rate = cur_bw;
3182                 log.u_bbr.delRate = low_bnd;
3183                 log.u_bbr.bw_inuse = up_bnd;
3184                 log.u_bbr.rttProp = rack_get_bw(rack);
3185                 log.u_bbr.pkt_epoch = line;
3186                 log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3187                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3188                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3189                 log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3190                 log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3191                 log.u_bbr.cwnd_gain = rack->rc_dragged_bottom;
3192                 log.u_bbr.cwnd_gain <<= 1;
3193                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_rec;
3194                 log.u_bbr.cwnd_gain <<= 1;
3195                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
3196                 log.u_bbr.cwnd_gain <<= 1;
3197                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
3198                 log.u_bbr.lost = rack->r_ctl.rc_loss_count;
3199                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3200                     &rack->rc_inp->inp_socket->so_rcv,
3201                     &rack->rc_inp->inp_socket->so_snd,
3202                     TCP_TIMELY_WORK, 0,
3203                     0, &log, false, &tv);
3204         }
3205 }
3206
3207 static int
3208 rack_bw_can_be_raised(struct tcp_rack *rack, uint64_t cur_bw, uint64_t last_bw_est, uint16_t mult)
3209 {
3210         /*
3211          * Before we increase we need to know if
3212          * the estimate just made was less than
3213          * our pacing goal (i.e. (cur_bw * mult) > last_bw_est)
3214          *
3215          * If we already are pacing at a fast enough
3216          * rate to push us faster there is no sense of
3217          * increasing.
3218          *
3219          * We first caculate our actual pacing rate (ss or ca multipler
3220          * times our cur_bw).
3221          *
3222          * Then we take the last measured rate and multipy by our
3223          * maximum pacing overage to give us a max allowable rate.
3224          *
3225          * If our act_rate is smaller than our max_allowable rate
3226          * then we should increase. Else we should hold steady.
3227          *
3228          */
3229         uint64_t act_rate, max_allow_rate;
3230
3231         if (rack_timely_no_stopping)
3232                 return (1);
3233
3234         if ((cur_bw == 0) || (last_bw_est == 0)) {
3235                 /*
3236                  * Initial startup case or
3237                  * everything is acked case.
3238                  */
3239                 rack_log_timely(rack,  mult, cur_bw, 0, 0,
3240                                 __LINE__, 9);
3241                 return (1);
3242         }
3243         if (mult <= 100) {
3244                 /*
3245                  * We can always pace at or slightly above our rate.
3246                  */
3247                 rack_log_timely(rack,  mult, cur_bw, 0, 0,
3248                                 __LINE__, 9);
3249                 return (1);
3250         }
3251         act_rate = cur_bw * (uint64_t)mult;
3252         act_rate /= 100;
3253         max_allow_rate = last_bw_est * ((uint64_t)rack_max_per_above + (uint64_t)100);
3254         max_allow_rate /= 100;
3255         if (act_rate < max_allow_rate) {
3256                 /*
3257                  * Here the rate we are actually pacing at
3258                  * is smaller than 10% above our last measurement.
3259                  * This means we are pacing below what we would
3260                  * like to try to achieve (plus some wiggle room).
3261                  */
3262                 rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3263                                 __LINE__, 9);
3264                 return (1);
3265         } else {
3266                 /*
3267                  * Here we are already pacing at least rack_max_per_above(10%)
3268                  * what we are getting back. This indicates most likely
3269                  * that we are being limited (cwnd/rwnd/app) and can't
3270                  * get any more b/w. There is no sense of trying to
3271                  * raise up the pacing rate its not speeding us up
3272                  * and we already are pacing faster than we are getting.
3273                  */
3274                 rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3275                                 __LINE__, 8);
3276                 return (0);
3277         }
3278 }
3279
3280 static void
3281 rack_validate_multipliers_at_or_above100(struct tcp_rack *rack)
3282 {
3283         /*
3284          * When we drag bottom, we want to assure
3285          * that no multiplier is below 1.0, if so
3286          * we want to restore it to at least that.
3287          */
3288         if (rack->r_ctl.rack_per_of_gp_rec  < 100) {
3289                 /* This is unlikely we usually do not touch recovery */
3290                 rack->r_ctl.rack_per_of_gp_rec = 100;
3291         }
3292         if (rack->r_ctl.rack_per_of_gp_ca < 100) {
3293                 rack->r_ctl.rack_per_of_gp_ca = 100;
3294         }
3295         if (rack->r_ctl.rack_per_of_gp_ss < 100) {
3296                 rack->r_ctl.rack_per_of_gp_ss = 100;
3297         }
3298 }
3299
3300 static void
3301 rack_validate_multipliers_at_or_below_100(struct tcp_rack *rack)
3302 {
3303         if (rack->r_ctl.rack_per_of_gp_ca > 100) {
3304                 rack->r_ctl.rack_per_of_gp_ca = 100;
3305         }
3306         if (rack->r_ctl.rack_per_of_gp_ss > 100) {
3307                 rack->r_ctl.rack_per_of_gp_ss = 100;
3308         }
3309 }
3310
3311 static void
3312 rack_increase_bw_mul(struct tcp_rack *rack, int timely_says, uint64_t cur_bw, uint64_t last_bw_est, int override)
3313 {
3314         int32_t  calc, logged, plus;
3315
3316         logged = 0;
3317
3318         if (override) {
3319                 /*
3320                  * override is passed when we are
3321                  * loosing b/w and making one last
3322                  * gasp at trying to not loose out
3323                  * to a new-reno flow.
3324                  */
3325                 goto extra_boost;
3326         }
3327         /* In classic timely we boost by 5x if we have 5 increases in a row, lets not */
3328         if (rack->rc_gp_incr &&
3329             ((rack->rc_gp_timely_inc_cnt + 1) >= RACK_TIMELY_CNT_BOOST)) {
3330                 /*
3331                  * Reset and get 5 strokes more before the boost. Note
3332                  * that the count is 0 based so we have to add one.
3333                  */
3334 extra_boost:
3335                 plus = (uint32_t)rack_gp_increase_per * RACK_TIMELY_CNT_BOOST;
3336                 rack->rc_gp_timely_inc_cnt = 0;
3337         } else
3338                 plus = (uint32_t)rack_gp_increase_per;
3339         /* Must be at least 1% increase for true timely increases */
3340         if ((plus < 1) &&
3341             ((rack->r_ctl.rc_rtt_diff <= 0) || (timely_says <= 0)))
3342                 plus = 1;
3343         if (rack->rc_gp_saw_rec &&
3344             (rack->rc_gp_no_rec_chg == 0) &&
3345             rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3346                                   rack->r_ctl.rack_per_of_gp_rec)) {
3347                 /* We have been in recovery ding it too */
3348                 calc = rack->r_ctl.rack_per_of_gp_rec + plus;
3349                 if (calc > 0xffff)
3350                         calc = 0xffff;
3351                 logged |= 1;
3352                 rack->r_ctl.rack_per_of_gp_rec = (uint16_t)calc;
3353                 if (rack_per_upper_bound_ss &&
3354                     (rack->rc_dragged_bottom == 0) &&
3355                     (rack->r_ctl.rack_per_of_gp_rec > rack_per_upper_bound_ss))
3356                         rack->r_ctl.rack_per_of_gp_rec = rack_per_upper_bound_ss;
3357         }
3358         if (rack->rc_gp_saw_ca &&
3359             (rack->rc_gp_saw_ss == 0) &&
3360             rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3361                                   rack->r_ctl.rack_per_of_gp_ca)) {
3362                 /* In CA */
3363                 calc = rack->r_ctl.rack_per_of_gp_ca + plus;
3364                 if (calc > 0xffff)
3365                         calc = 0xffff;
3366                 logged |= 2;
3367                 rack->r_ctl.rack_per_of_gp_ca = (uint16_t)calc;
3368                 if (rack_per_upper_bound_ca &&
3369                     (rack->rc_dragged_bottom == 0) &&
3370                     (rack->r_ctl.rack_per_of_gp_ca > rack_per_upper_bound_ca))
3371                         rack->r_ctl.rack_per_of_gp_ca = rack_per_upper_bound_ca;
3372         }
3373         if (rack->rc_gp_saw_ss &&
3374             rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3375                                   rack->r_ctl.rack_per_of_gp_ss)) {
3376                 /* In SS */
3377                 calc = rack->r_ctl.rack_per_of_gp_ss + plus;
3378                 if (calc > 0xffff)
3379                         calc = 0xffff;
3380                 rack->r_ctl.rack_per_of_gp_ss = (uint16_t)calc;
3381                 if (rack_per_upper_bound_ss &&
3382                     (rack->rc_dragged_bottom == 0) &&
3383                     (rack->r_ctl.rack_per_of_gp_ss > rack_per_upper_bound_ss))
3384                         rack->r_ctl.rack_per_of_gp_ss = rack_per_upper_bound_ss;
3385                 logged |= 4;
3386         }
3387         if (logged &&
3388             (rack->rc_gp_incr == 0)){
3389                 /* Go into increment mode */
3390                 rack->rc_gp_incr = 1;
3391                 rack->rc_gp_timely_inc_cnt = 0;
3392         }
3393         if (rack->rc_gp_incr &&
3394             logged &&
3395             (rack->rc_gp_timely_inc_cnt < RACK_TIMELY_CNT_BOOST)) {
3396                 rack->rc_gp_timely_inc_cnt++;
3397         }
3398         rack_log_timely(rack,  logged, plus, 0, 0,
3399                         __LINE__, 1);
3400 }
3401
3402 static uint32_t
3403 rack_get_decrease(struct tcp_rack *rack, uint32_t curper, int32_t rtt_diff)
3404 {
3405         /*
3406          * norm_grad = rtt_diff / minrtt;
3407          * new_per = curper * (1 - B * norm_grad)
3408          *
3409          * B = rack_gp_decrease_per (default 10%)
3410          * rtt_dif = input var current rtt-diff
3411          * curper = input var current percentage
3412          * minrtt = from rack filter
3413          *
3414          */
3415         uint64_t perf;
3416
3417         perf = (((uint64_t)curper * ((uint64_t)1000000 -
3418                     ((uint64_t)rack_gp_decrease_per * (uint64_t)10000 *
3419                      (((uint64_t)rtt_diff * (uint64_t)1000000)/
3420                       (uint64_t)get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt)))/
3421                      (uint64_t)1000000)) /
3422                 (uint64_t)1000000);
3423         if (perf > curper) {
3424                 /* TSNH */
3425                 perf = curper - 1;
3426         }
3427         return ((uint32_t)perf);
3428 }
3429
3430 static uint32_t
3431 rack_decrease_highrtt(struct tcp_rack *rack, uint32_t curper, uint32_t rtt)
3432 {
3433         /*
3434          *                                   highrttthresh
3435          * result = curper * (1 - (B * ( 1 -  ------          ))
3436          *                                     gp_srtt
3437          *
3438          * B = rack_gp_decrease_per (default 10%)
3439          * highrttthresh = filter_min * rack_gp_rtt_maxmul
3440          */
3441         uint64_t perf;
3442         uint32_t highrttthresh;
3443
3444         highrttthresh = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
3445
3446         perf = (((uint64_t)curper * ((uint64_t)1000000 -
3447                                      ((uint64_t)rack_gp_decrease_per * ((uint64_t)1000000 -
3448                                         ((uint64_t)highrttthresh * (uint64_t)1000000) /
3449                                                     (uint64_t)rtt)) / 100)) /(uint64_t)1000000);
3450         return (perf);
3451 }
3452
3453 static void
3454 rack_decrease_bw_mul(struct tcp_rack *rack, int timely_says, uint32_t rtt, int32_t rtt_diff)
3455 {
3456         uint64_t logvar, logvar2, logvar3;
3457         uint32_t logged, new_per, ss_red, ca_red, rec_red, alt, val;
3458
3459         if (rack->rc_gp_incr) {
3460                 /* Turn off increment counting */
3461                 rack->rc_gp_incr = 0;
3462                 rack->rc_gp_timely_inc_cnt = 0;
3463         }
3464         ss_red = ca_red = rec_red = 0;
3465         logged = 0;
3466         /* Calculate the reduction value */
3467         if (rtt_diff < 0) {
3468                 rtt_diff *= -1;
3469         }
3470         /* Must be at least 1% reduction */
3471         if (rack->rc_gp_saw_rec && (rack->rc_gp_no_rec_chg == 0)) {
3472                 /* We have been in recovery ding it too */
3473                 if (timely_says == 2) {
3474                         new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_rec, rtt);
3475                         alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3476                         if (alt < new_per)
3477                                 val = alt;
3478                         else
3479                                 val = new_per;
3480                 } else
3481                          val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3482                 if (rack->r_ctl.rack_per_of_gp_rec > val) {
3483                         rec_red = (rack->r_ctl.rack_per_of_gp_rec - val);
3484                         rack->r_ctl.rack_per_of_gp_rec = (uint16_t)val;
3485                 } else {
3486                         rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
3487                         rec_red = 0;
3488                 }
3489                 if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_rec)
3490                         rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
3491                 logged |= 1;
3492         }
3493         if (rack->rc_gp_saw_ss) {
3494                 /* Sent in SS */
3495                 if (timely_says == 2) {
3496                         new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ss, rtt);
3497                         alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3498                         if (alt < new_per)
3499                                 val = alt;
3500                         else
3501                                 val = new_per;
3502                 } else
3503                         val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff);
3504                 if (rack->r_ctl.rack_per_of_gp_ss > new_per) {
3505                         ss_red = rack->r_ctl.rack_per_of_gp_ss - val;
3506                         rack->r_ctl.rack_per_of_gp_ss = (uint16_t)val;
3507                 } else {
3508                         ss_red = new_per;
3509                         rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
3510                         logvar = new_per;
3511                         logvar <<= 32;
3512                         logvar |= alt;
3513                         logvar2 = (uint32_t)rtt;
3514                         logvar2 <<= 32;
3515                         logvar2 |= (uint32_t)rtt_diff;
3516                         logvar3 = rack_gp_rtt_maxmul;
3517                         logvar3 <<= 32;
3518                         logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3519                         rack_log_timely(rack, timely_says,
3520                                         logvar2, logvar3,
3521                                         logvar, __LINE__, 10);
3522                 }
3523                 if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ss)
3524                         rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
3525                 logged |= 4;
3526         } else if (rack->rc_gp_saw_ca) {
3527                 /* Sent in CA */
3528                 if (timely_says == 2) {
3529                         new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ca, rtt);
3530                         alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3531                         if (alt < new_per)
3532                                 val = alt;
3533                         else
3534                                 val = new_per;
3535                 } else
3536                         val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff);
3537                 if (rack->r_ctl.rack_per_of_gp_ca > val) {
3538                         ca_red = rack->r_ctl.rack_per_of_gp_ca - val;
3539                         rack->r_ctl.rack_per_of_gp_ca = (uint16_t)val;
3540                 } else {
3541                         rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
3542                         ca_red = 0;
3543                         logvar = new_per;
3544                         logvar <<= 32;
3545                         logvar |= alt;
3546                         logvar2 = (uint32_t)rtt;
3547                         logvar2 <<= 32;
3548                         logvar2 |= (uint32_t)rtt_diff;
3549                         logvar3 = rack_gp_rtt_maxmul;
3550                         logvar3 <<= 32;
3551                         logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3552                         rack_log_timely(rack, timely_says,
3553                                         logvar2, logvar3,
3554                                         logvar, __LINE__, 10);
3555                 }
3556                 if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ca)
3557                         rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
3558                 logged |= 2;
3559         }
3560         if (rack->rc_gp_timely_dec_cnt < 0x7) {
3561                 rack->rc_gp_timely_dec_cnt++;
3562                 if (rack_timely_dec_clear &&
3563                     (rack->rc_gp_timely_dec_cnt == rack_timely_dec_clear))
3564                         rack->rc_gp_timely_dec_cnt = 0;
3565         }
3566         logvar = ss_red;
3567         logvar <<= 32;
3568         logvar |= ca_red;
3569         rack_log_timely(rack,  logged, rec_red, rack_per_lower_bound, logvar,
3570                         __LINE__, 2);
3571 }
3572
3573 static void
3574 rack_log_rtt_shrinks(struct tcp_rack *rack, uint32_t us_cts,
3575                      uint32_t rtt, uint32_t line, uint8_t reas)
3576 {
3577         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
3578                 union tcp_log_stackspecific log;
3579                 struct timeval tv;
3580
3581                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3582                 log.u_bbr.flex1 = line;
3583                 log.u_bbr.flex2 = rack->r_ctl.rc_time_probertt_starts;
3584                 log.u_bbr.flex3 = rack->r_ctl.rc_lower_rtt_us_cts;
3585                 log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3586                 log.u_bbr.flex5 = rtt;
3587                 log.u_bbr.flex6 = rack->rc_highly_buffered;
3588                 log.u_bbr.flex6 <<= 1;
3589                 log.u_bbr.flex6 |= rack->forced_ack;
3590                 log.u_bbr.flex6 <<= 1;
3591                 log.u_bbr.flex6 |= rack->rc_gp_dyn_mul;
3592                 log.u_bbr.flex6 <<= 1;
3593                 log.u_bbr.flex6 |= rack->in_probe_rtt;
3594                 log.u_bbr.flex6 <<= 1;
3595                 log.u_bbr.flex6 |= rack->measure_saw_probe_rtt;
3596                 log.u_bbr.flex7 = rack->r_ctl.rack_per_of_gp_probertt;
3597                 log.u_bbr.pacing_gain = rack->r_ctl.rack_per_of_gp_ca;
3598                 log.u_bbr.cwnd_gain = rack->r_ctl.rack_per_of_gp_rec;
3599                 log.u_bbr.flex8 = reas;
3600                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3601                 log.u_bbr.delRate = rack_get_bw(rack);
3602                 log.u_bbr.cur_del_rate = rack->r_ctl.rc_highest_us_rtt;
3603                 log.u_bbr.cur_del_rate <<= 32;
3604                 log.u_bbr.cur_del_rate |= rack->r_ctl.rc_lowest_us_rtt;
3605                 log.u_bbr.applimited = rack->r_ctl.rc_time_probertt_entered;
3606                 log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3607                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3608                 log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3609                 log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3610                 log.u_bbr.pkt_epoch = rack->r_ctl.rc_lower_rtt_us_cts;
3611                 log.u_bbr.delivered = rack->r_ctl.rc_target_probertt_flight;
3612                 log.u_bbr.lost = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3613                 log.u_bbr.rttProp = us_cts;
3614                 log.u_bbr.rttProp <<= 32;
3615                 log.u_bbr.rttProp |= rack->r_ctl.rc_entry_gp_rtt;
3616                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3617                     &rack->rc_inp->inp_socket->so_rcv,
3618                     &rack->rc_inp->inp_socket->so_snd,
3619                     BBR_LOG_RTT_SHRINKS, 0,
3620                     0, &log, false, &rack->r_ctl.act_rcv_time);
3621         }
3622 }
3623
3624 static void
3625 rack_set_prtt_target(struct tcp_rack *rack, uint32_t segsiz, uint32_t rtt)
3626 {
3627         uint64_t bwdp;
3628
3629         bwdp = rack_get_bw(rack);
3630         bwdp *= (uint64_t)rtt;
3631         bwdp /= (uint64_t)HPTS_USEC_IN_SEC;
3632         rack->r_ctl.rc_target_probertt_flight = roundup((uint32_t)bwdp, segsiz);
3633         if (rack->r_ctl.rc_target_probertt_flight < (segsiz * rack_timely_min_segs)) {
3634                 /*
3635                  * A window protocol must be able to have 4 packets
3636                  * outstanding as the floor in order to function
3637                  * (especially considering delayed ack :D).
3638                  */
3639                 rack->r_ctl.rc_target_probertt_flight = (segsiz * rack_timely_min_segs);
3640         }
3641 }
3642
3643 static void
3644 rack_enter_probertt(struct tcp_rack *rack, uint32_t us_cts)
3645 {
3646         /**
3647          * ProbeRTT is a bit different in rack_pacing than in
3648          * BBR. It is like BBR in that it uses the lowering of
3649          * the RTT as a signal that we saw something new and
3650          * counts from there for how long between. But it is
3651          * different in that its quite simple. It does not
3652          * play with the cwnd and wait until we get down
3653          * to N segments outstanding and hold that for
3654          * 200ms. Instead it just sets the pacing reduction
3655          * rate to a set percentage (70 by default) and hold
3656          * that for a number of recent GP Srtt's.
3657          */
3658         uint32_t segsiz;
3659
3660         if (rack->rc_gp_dyn_mul == 0)
3661                 return;
3662
3663         if (rack->rc_tp->snd_max == rack->rc_tp->snd_una) {
3664                 /* We are idle */
3665                 return;
3666         }
3667         if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
3668             SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
3669                 /*
3670                  * Stop the goodput now, the idea here is
3671                  * that future measurements with in_probe_rtt
3672                  * won't register if they are not greater so
3673                  * we want to get what info (if any) is available
3674                  * now.
3675                  */
3676                 rack_do_goodput_measurement(rack->rc_tp, rack,
3677                                             rack->rc_tp->snd_una, __LINE__);
3678         }
3679         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
3680         rack->r_ctl.rc_time_probertt_entered = us_cts;
3681         segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
3682                      rack->r_ctl.rc_pace_min_segs);
3683         rack->in_probe_rtt = 1;
3684         rack->measure_saw_probe_rtt = 1;
3685         rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
3686         rack->r_ctl.rc_time_probertt_starts = 0;
3687         rack->r_ctl.rc_entry_gp_rtt = rack->r_ctl.rc_gp_srtt;
3688         if (rack_probertt_use_min_rtt_entry)
3689                 rack_set_prtt_target(rack, segsiz, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
3690         else
3691                 rack_set_prtt_target(rack, segsiz, rack->r_ctl.rc_gp_srtt);
3692         rack_log_rtt_shrinks(rack,  us_cts,  get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3693                              __LINE__, RACK_RTTS_ENTERPROBE);
3694 }
3695
3696 static void
3697 rack_exit_probertt(struct tcp_rack *rack, uint32_t us_cts)
3698 {
3699         struct rack_sendmap *rsm;
3700         uint32_t segsiz;
3701
3702         segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
3703                      rack->r_ctl.rc_pace_min_segs);
3704         rack->in_probe_rtt = 0;
3705         if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
3706             SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
3707                 /*
3708                  * Stop the goodput now, the idea here is
3709                  * that future measurements with in_probe_rtt
3710                  * won't register if they are not greater so
3711                  * we want to get what info (if any) is available
3712                  * now.
3713                  */
3714                 rack_do_goodput_measurement(rack->rc_tp, rack,
3715                                             rack->rc_tp->snd_una, __LINE__);
3716         } else if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
3717                 /*
3718                  * We don't have enough data to make a measurement.
3719                  * So lets just stop and start here after exiting
3720                  * probe-rtt. We probably are not interested in
3721                  * the results anyway.
3722                  */
3723                 rack->rc_tp->t_flags &= ~TF_GPUTINPROG;
3724         }
3725         /*
3726          * Measurements through the current snd_max are going
3727          * to be limited by the slower pacing rate.
3728          *
3729          * We need to mark these as app-limited so we
3730          * don't collapse the b/w.
3731          */
3732         rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
3733         if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
3734                 if (rack->r_ctl.rc_app_limited_cnt == 0)
3735                         rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
3736                 else {
3737                         /*
3738                          * Go out to the end app limited and mark
3739                          * this new one as next and move the end_appl up
3740                          * to this guy.
3741                          */
3742                         if (rack->r_ctl.rc_end_appl)
3743                                 rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
3744                         rack->r_ctl.rc_end_appl = rsm;
3745                 }
3746                 rsm->r_flags |= RACK_APP_LIMITED;
3747                 rack->r_ctl.rc_app_limited_cnt++;
3748         }
3749         /*
3750          * Now, we need to examine our pacing rate multipliers.
3751          * If its under 100%, we need to kick it back up to
3752          * 100%. We also don't let it be over our "max" above
3753          * the actual rate i.e. 100% + rack_clamp_atexit_prtt.
3754          * Note setting clamp_atexit_prtt to 0 has the effect
3755          * of setting CA/SS to 100% always at exit (which is
3756          * the default behavior).
3757          */
3758         if (rack_probertt_clear_is) {
3759                 rack->rc_gp_incr = 0;
3760                 rack->rc_gp_bwred = 0;
3761                 rack->rc_gp_timely_inc_cnt = 0;
3762                 rack->rc_gp_timely_dec_cnt = 0;
3763         }
3764         /* Do we do any clamping at exit? */
3765         if (rack->rc_highly_buffered && rack_atexit_prtt_hbp) {
3766                 rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt_hbp;
3767                 rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt_hbp;
3768         }
3769         if ((rack->rc_highly_buffered == 0) && rack_atexit_prtt) {
3770                 rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt;
3771                 rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt;
3772         }
3773         /*
3774          * Lets set rtt_diff to 0, so that we will get a "boost"
3775          * after exiting.
3776          */
3777         rack->r_ctl.rc_rtt_diff = 0;
3778
3779         /* Clear all flags so we start fresh */
3780         rack->rc_tp->t_bytes_acked = 0;
3781         rack->rc_tp->ccv->flags &= ~CCF_ABC_SENTAWND;
3782         /*
3783          * If configured to, set the cwnd and ssthresh to
3784          * our targets.
3785          */
3786         if (rack_probe_rtt_sets_cwnd) {
3787                 uint64_t ebdp;
3788                 uint32_t setto;
3789
3790                 /* Set ssthresh so we get into CA once we hit our target */
3791                 if (rack_probertt_use_min_rtt_exit == 1) {
3792                         /* Set to min rtt */
3793                         rack_set_prtt_target(rack, segsiz,
3794                                              get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
3795                 } else if (rack_probertt_use_min_rtt_exit == 2) {
3796                         /* Set to current gp rtt */
3797                         rack_set_prtt_target(rack, segsiz,
3798                                              rack->r_ctl.rc_gp_srtt);
3799                 } else if (rack_probertt_use_min_rtt_exit == 3) {
3800                         /* Set to entry gp rtt */
3801                         rack_set_prtt_target(rack, segsiz,
3802                                              rack->r_ctl.rc_entry_gp_rtt);
3803                 } else {
3804                         uint64_t sum;
3805                         uint32_t setval;
3806
3807                         sum = rack->r_ctl.rc_entry_gp_rtt;
3808                         sum *= 10;
3809                         sum /= (uint64_t)(max(1, rack->r_ctl.rc_gp_srtt));
3810                         if (sum >= 20) {
3811                                 /*
3812                                  * A highly buffered path needs
3813                                  * cwnd space for timely to work.
3814                                  * Lets set things up as if
3815                                  * we are heading back here again.
3816                                  */
3817                                 setval = rack->r_ctl.rc_entry_gp_rtt;
3818                         } else if (sum >= 15) {
3819                                 /*
3820                                  * Lets take the smaller of the
3821                                  * two since we are just somewhat
3822                                  * buffered.
3823                                  */
3824                                 setval = rack->r_ctl.rc_gp_srtt;
3825                                 if (setval > rack->r_ctl.rc_entry_gp_rtt)
3826                                         setval = rack->r_ctl.rc_entry_gp_rtt;
3827                         } else {
3828                                 /*
3829                                  * Here we are not highly buffered
3830                                  * and should pick the min we can to
3831                                  * keep from causing loss.
3832                                  */
3833                                 setval = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3834                         }
3835                         rack_set_prtt_target(rack, segsiz,
3836                                              setval);
3837                 }
3838                 if (rack_probe_rtt_sets_cwnd > 1) {
3839                         /* There is a percentage here to boost */
3840                         ebdp = rack->r_ctl.rc_target_probertt_flight;
3841                         ebdp *= rack_probe_rtt_sets_cwnd;
3842                         ebdp /= 100;
3843                         setto = rack->r_ctl.rc_target_probertt_flight + ebdp;
3844                 } else
3845                         setto = rack->r_ctl.rc_target_probertt_flight;
3846                 rack->rc_tp->snd_cwnd = roundup(setto, segsiz);
3847                 if (rack->rc_tp->snd_cwnd < (segsiz * rack_timely_min_segs)) {
3848                         /* Enforce a min */
3849                         rack->rc_tp->snd_cwnd = segsiz * rack_timely_min_segs;
3850                 }
3851                 /* If we set in the cwnd also set the ssthresh point so we are in CA */
3852                 rack->rc_tp->snd_ssthresh = (rack->rc_tp->snd_cwnd - 1);
3853         }
3854         rack_log_rtt_shrinks(rack,  us_cts,
3855                              get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3856                              __LINE__, RACK_RTTS_EXITPROBE);
3857         /* Clear times last so log has all the info */
3858         rack->r_ctl.rc_probertt_sndmax_atexit = rack->rc_tp->snd_max;
3859         rack->r_ctl.rc_time_probertt_entered = us_cts;
3860         rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
3861         rack->r_ctl.rc_time_of_last_probertt = us_cts;
3862 }
3863
3864 static void
3865 rack_check_probe_rtt(struct tcp_rack *rack, uint32_t us_cts)
3866 {
3867         /* Check in on probe-rtt */
3868         if (rack->rc_gp_filled == 0) {
3869                 /* We do not do p-rtt unless we have gp measurements */
3870                 return;
3871         }
3872         if (rack->in_probe_rtt) {
3873                 uint64_t no_overflow;
3874                 uint32_t endtime, must_stay;
3875
3876                 if (rack->r_ctl.rc_went_idle_time &&
3877                     ((us_cts - rack->r_ctl.rc_went_idle_time) > rack_min_probertt_hold)) {
3878                         /*
3879                          * We went idle during prtt, just exit now.
3880                          */
3881                         rack_exit_probertt(rack, us_cts);
3882                 } else if (rack_probe_rtt_safety_val &&
3883                     TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered) &&
3884                     ((us_cts - rack->r_ctl.rc_time_probertt_entered) > rack_probe_rtt_safety_val)) {
3885                         /*
3886                          * Probe RTT safety value triggered!
3887                          */
3888                         rack_log_rtt_shrinks(rack,  us_cts,
3889                                              get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3890                                              __LINE__, RACK_RTTS_SAFETY);
3891                         rack_exit_probertt(rack, us_cts);
3892                 }
3893                 /* Calculate the max we will wait */
3894                 endtime = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_max_drain_wait);
3895                 if (rack->rc_highly_buffered)
3896                         endtime += (rack->r_ctl.rc_gp_srtt * rack_max_drain_hbp);
3897                 /* Calculate the min we must wait */
3898                 must_stay = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_must_drain);
3899                 if ((ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.rc_target_probertt_flight) &&
3900                     TSTMP_LT(us_cts, endtime)) {
3901                         uint32_t calc;
3902                         /* Do we lower more? */
3903 no_exit:
3904                         if (TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered))
3905                                 calc = us_cts - rack->r_ctl.rc_time_probertt_entered;
3906                         else
3907                                 calc = 0;
3908                         calc /= max(rack->r_ctl.rc_gp_srtt, 1);
3909                         if (calc) {
3910                                 /* Maybe */
3911                                 calc *= rack_per_of_gp_probertt_reduce;
3912                                 rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt - calc;
3913                                 /* Limit it too */
3914                                 if (rack->r_ctl.rack_per_of_gp_probertt < rack_per_of_gp_lowthresh)
3915                                         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_lowthresh;
3916                         }
3917                         /* We must reach target or the time set */
3918                         return;
3919                 }
3920                 if (rack->r_ctl.rc_time_probertt_starts == 0) {
3921                         if ((TSTMP_LT(us_cts, must_stay) &&
3922                              rack->rc_highly_buffered) ||
3923                              (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) >
3924                               rack->r_ctl.rc_target_probertt_flight)) {
3925                                 /* We are not past the must_stay time */
3926                                 goto no_exit;
3927                         }
3928                         rack_log_rtt_shrinks(rack,  us_cts,
3929                                              get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3930                                              __LINE__, RACK_RTTS_REACHTARGET);
3931                         rack->r_ctl.rc_time_probertt_starts = us_cts;
3932                         if (rack->r_ctl.rc_time_probertt_starts == 0)
3933                                 rack->r_ctl.rc_time_probertt_starts = 1;
3934                         /* Restore back to our rate we want to pace at in prtt */
3935                         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
3936                 }
3937                 /*
3938                  * Setup our end time, some number of gp_srtts plus 200ms.
3939                  */
3940                 no_overflow = ((uint64_t)rack->r_ctl.rc_gp_srtt *
3941                                (uint64_t)rack_probertt_gpsrtt_cnt_mul);
3942                 if (rack_probertt_gpsrtt_cnt_div)
3943                         endtime = (uint32_t)(no_overflow / (uint64_t)rack_probertt_gpsrtt_cnt_div);
3944                 else
3945                         endtime = 0;
3946                 endtime += rack_min_probertt_hold;
3947                 endtime += rack->r_ctl.rc_time_probertt_starts;
3948                 if (TSTMP_GEQ(us_cts,  endtime)) {
3949                         /* yes, exit probertt */
3950                         rack_exit_probertt(rack, us_cts);
3951                 }
3952
3953         } else if ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= rack_time_between_probertt) {
3954                 /* Go into probertt, its been too long since we went lower */
3955                 rack_enter_probertt(rack, us_cts);
3956         }
3957 }
3958
3959 static void
3960 rack_update_multiplier(struct tcp_rack *rack, int32_t timely_says, uint64_t last_bw_est,
3961                        uint32_t rtt, int32_t rtt_diff)
3962 {
3963         uint64_t cur_bw, up_bnd, low_bnd, subfr;
3964         uint32_t losses;
3965
3966         if ((rack->rc_gp_dyn_mul == 0) ||
3967             (rack->use_fixed_rate) ||
3968             (rack->in_probe_rtt) ||
3969             (rack->rc_always_pace == 0)) {
3970                 /* No dynamic GP multipler in play */
3971                 return;
3972         }
3973         losses = rack->r_ctl.rc_loss_count - rack->r_ctl.rc_loss_at_start;
3974         cur_bw = rack_get_bw(rack);
3975         /* Calculate our up and down range */
3976         up_bnd = rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_up;
3977         up_bnd /= 100;
3978         up_bnd += rack->r_ctl.last_gp_comp_bw;
3979
3980         subfr = (uint64_t)rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_down;
3981         subfr /= 100;
3982         low_bnd = rack->r_ctl.last_gp_comp_bw - subfr;
3983         if ((timely_says == 2) && (rack->r_ctl.rc_no_push_at_mrtt)) {
3984                 /*
3985                  * This is the case where our RTT is above
3986                  * the max target and we have been configured
3987                  * to just do timely no bonus up stuff in that case.
3988                  *
3989                  * There are two configurations, set to 1, and we
3990                  * just do timely if we are over our max. If its
3991                  * set above 1 then we slam the multipliers down
3992                  * to 100 and then decrement per timely.
3993                  */
3994                 rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3995                                 __LINE__, 3);
3996                 if (rack->r_ctl.rc_no_push_at_mrtt > 1)
3997                         rack_validate_multipliers_at_or_below_100(rack);
3998                 rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
3999         } else if ((last_bw_est < low_bnd) && !losses) {
4000                 /*
4001                  * We are decreasing this is a bit complicated this
4002                  * means we are loosing ground. This could be
4003                  * because another flow entered and we are competing
4004                  * for b/w with it. This will push the RTT up which
4005                  * makes timely unusable unless we want to get shoved
4006                  * into a corner and just be backed off (the age
4007                  * old problem with delay based CC).
4008                  *
4009                  * On the other hand if it was a route change we
4010                  * would like to stay somewhat contained and not
4011                  * blow out the buffers.
4012                  */
4013                 rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4014                                 __LINE__, 3);
4015                 rack->r_ctl.last_gp_comp_bw = cur_bw;
4016                 if (rack->rc_gp_bwred == 0) {
4017                         /* Go into reduction counting */
4018                         rack->rc_gp_bwred = 1;
4019                         rack->rc_gp_timely_dec_cnt = 0;
4020                 }
4021                 if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) ||
4022                     (timely_says == 0)) {
4023                         /*
4024                          * Push another time with a faster pacing
4025                          * to try to gain back (we include override to
4026                          * get a full raise factor).
4027                          */
4028                         if ((rack->rc_gp_saw_ca && rack->r_ctl.rack_per_of_gp_ca <= rack_down_raise_thresh) ||
4029                             (rack->rc_gp_saw_ss && rack->r_ctl.rack_per_of_gp_ss <= rack_down_raise_thresh) ||
4030                             (timely_says == 0) ||
4031                             (rack_down_raise_thresh == 0)) {
4032                                 /*
4033                                  * Do an override up in b/w if we were
4034                                  * below the threshold or if the threshold
4035                                  * is zero we always do the raise.
4036                                  */
4037                                 rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 1);
4038                         } else {
4039                                 /* Log it stays the same */
4040                                 rack_log_timely(rack,  0, last_bw_est, low_bnd, 0,
4041                                                 __LINE__, 11);
4042                         }
4043                         rack->rc_gp_timely_dec_cnt++;
4044                         /* We are not incrementing really no-count */
4045                         rack->rc_gp_incr = 0;
4046                         rack->rc_gp_timely_inc_cnt = 0;
4047                 } else {
4048                         /*
4049                          * Lets just use the RTT
4050                          * information and give up
4051                          * pushing.
4052                          */
4053                         goto use_timely;
4054                 }
4055         } else if ((timely_says != 2) &&
4056                     !losses &&
4057                     (last_bw_est > up_bnd)) {
4058                 /*
4059                  * We are increasing b/w lets keep going, updating
4060                  * our b/w and ignoring any timely input, unless
4061                  * of course we are at our max raise (if there is one).
4062                  */
4063
4064                 rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4065                                 __LINE__, 3);
4066                 rack->r_ctl.last_gp_comp_bw = cur_bw;
4067                 if (rack->rc_gp_saw_ss &&
4068                     rack_per_upper_bound_ss &&
4069                      (rack->r_ctl.rack_per_of_gp_ss == rack_per_upper_bound_ss)) {
4070                             /*
4071                              * In cases where we can't go higher
4072                              * we should just use timely.
4073                              */
4074                             goto use_timely;
4075                 }
4076                 if (rack->rc_gp_saw_ca &&
4077                     rack_per_upper_bound_ca &&
4078                     (rack->r_ctl.rack_per_of_gp_ca == rack_per_upper_bound_ca)) {
4079                             /*
4080                              * In cases where we can't go higher
4081                              * we should just use timely.
4082                              */
4083                             goto use_timely;
4084                 }
4085                 rack->rc_gp_bwred = 0;
4086                 rack->rc_gp_timely_dec_cnt = 0;
4087                 /* You get a set number of pushes if timely is trying to reduce */
4088                 if ((rack->rc_gp_incr < rack_timely_max_push_rise) || (timely_says == 0)) {
4089                         rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4090                 } else {
4091                         /* Log it stays the same */
4092                         rack_log_timely(rack,  0, last_bw_est, up_bnd, 0,
4093                             __LINE__, 12);
4094                 }
4095                 return;
4096         } else {
4097                 /*
4098                  * We are staying between the lower and upper range bounds
4099                  * so use timely to decide.
4100                  */
4101                 rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4102                                 __LINE__, 3);
4103 use_timely:
4104                 if (timely_says) {
4105                         rack->rc_gp_incr = 0;
4106                         rack->rc_gp_timely_inc_cnt = 0;
4107                         if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) &&
4108                             !losses &&
4109                             (last_bw_est < low_bnd)) {
4110                                 /* We are loosing ground */
4111                                 rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4112                                 rack->rc_gp_timely_dec_cnt++;
4113                                 /* We are not incrementing really no-count */
4114                                 rack->rc_gp_incr = 0;
4115                                 rack->rc_gp_timely_inc_cnt = 0;
4116                         } else
4117                                 rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
4118                 } else {
4119                         rack->rc_gp_bwred = 0;
4120                         rack->rc_gp_timely_dec_cnt = 0;
4121                         rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4122                 }
4123         }
4124 }
4125
4126 static int32_t
4127 rack_make_timely_judgement(struct tcp_rack *rack, uint32_t rtt, int32_t rtt_diff, uint32_t prev_rtt)
4128 {
4129         int32_t timely_says;
4130         uint64_t log_mult, log_rtt_a_diff;
4131
4132         log_rtt_a_diff = rtt;
4133         log_rtt_a_diff <<= 32;
4134         log_rtt_a_diff |= (uint32_t)rtt_diff;
4135         if (rtt >= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) *
4136                     rack_gp_rtt_maxmul)) {
4137                 /* Reduce the b/w multipler */
4138                 timely_says = 2;
4139                 log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
4140                 log_mult <<= 32;
4141                 log_mult |= prev_rtt;
4142                 rack_log_timely(rack,  timely_says, log_mult,
4143                                 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4144                                 log_rtt_a_diff, __LINE__, 4);
4145         } else if (rtt <= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4146                            ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4147                             max(rack_gp_rtt_mindiv , 1)))) {
4148                 /* Increase the b/w multipler */
4149                 log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4150                         ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4151                          max(rack_gp_rtt_mindiv , 1));
4152                 log_mult <<= 32;
4153                 log_mult |= prev_rtt;
4154                 timely_says = 0;
4155                 rack_log_timely(rack,  timely_says, log_mult ,
4156                                 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4157                                 log_rtt_a_diff, __LINE__, 5);
4158         } else {
4159                 /*
4160                  * Use a gradient to find it the timely gradient
4161                  * is:
4162                  * grad = rc_rtt_diff / min_rtt;
4163                  *
4164                  * anything below or equal to 0 will be
4165                  * a increase indication. Anything above
4166                  * zero is a decrease. Note we take care
4167                  * of the actual gradient calculation
4168                  * in the reduction (its not needed for
4169                  * increase).
4170                  */
4171                 log_mult = prev_rtt;
4172                 if (rtt_diff <= 0) {
4173                         /*
4174                          * Rttdiff is less than zero, increase the
4175                          * b/w multipler (its 0 or negative)
4176                          */
4177                         timely_says = 0;
4178                         rack_log_timely(rack,  timely_says, log_mult,
4179                                         get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 6);
4180                 } else {
4181                         /* Reduce the b/w multipler */
4182                         timely_says = 1;
4183                         rack_log_timely(rack,  timely_says, log_mult,
4184                                         get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 7);
4185                 }
4186         }
4187         return (timely_says);
4188 }
4189
4190 static void
4191 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
4192                             tcp_seq th_ack, int line)
4193 {
4194         uint64_t tim, bytes_ps, ltim, stim, utim;
4195         uint32_t segsiz, bytes, reqbytes, us_cts;
4196         int32_t gput, new_rtt_diff, timely_says;
4197         uint64_t  resid_bw, subpart = 0, addpart = 0, srtt;
4198         int did_add = 0;
4199
4200         us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4201         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
4202         if (TSTMP_GEQ(us_cts, tp->gput_ts))
4203                 tim = us_cts - tp->gput_ts;
4204         else
4205                 tim = 0;
4206
4207         if (rack->r_ctl.rc_gp_cumack_ts > rack->r_ctl.rc_gp_output_ts)
4208                 stim = rack->r_ctl.rc_gp_cumack_ts - rack->r_ctl.rc_gp_output_ts;
4209         else
4210                 stim = 0;
4211         /*
4212          * Use the larger of the send time or ack time. This prevents us
4213          * from being influenced by ack artifacts to come up with too
4214          * high of measurement. Note that since we are spanning over many more
4215          * bytes in most of our measurements hopefully that is less likely to
4216          * occur.
4217          */
4218         if (tim > stim)
4219                 utim = max(tim, 1);
4220         else
4221                 utim = max(stim, 1);
4222         /* Lets get a msec time ltim too for the old stuff */
4223         ltim = max(1, (utim / HPTS_USEC_IN_MSEC));
4224         gput = (((uint64_t) (th_ack - tp->gput_seq)) << 3) / ltim;
4225         reqbytes = min(rc_init_window(rack), (MIN_GP_WIN * segsiz));
4226         if ((tim == 0) && (stim == 0)) {
4227                 /*
4228                  * Invalid measurement time, maybe
4229                  * all on one ack/one send?
4230                  */
4231                 bytes = 0;
4232                 bytes_ps = 0;
4233                 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4234                                            0, 0, 0, 10, __LINE__, NULL);
4235                 goto skip_measurement;
4236         }
4237         if (rack->r_ctl.rc_gp_lowrtt == 0xffffffff) {
4238                 /* We never made a us_rtt measurement? */
4239                 bytes = 0;
4240                 bytes_ps = 0;
4241                 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4242                                            0, 0, 0, 10, __LINE__, NULL);
4243                 goto skip_measurement;
4244         }
4245         /*
4246          * Calculate the maximum possible b/w this connection
4247          * could have. We base our calculation on the lowest
4248          * rtt we have seen during the measurement and the
4249          * largest rwnd the client has given us in that time. This
4250          * forms a BDP that is the maximum that we could ever
4251          * get to the client. Anything larger is not valid.
4252          *
4253          * I originally had code here that rejected measurements
4254          * where the time was less than 1/2 the latest us_rtt.
4255          * But after thinking on that I realized its wrong since
4256          * say you had a 150Mbps or even 1Gbps link, and you
4257          * were a long way away.. example I am in Europe (100ms rtt)
4258          * talking to my 1Gbps link in S.C. Now measuring say 150,000
4259          * bytes my time would be 1.2ms, and yet my rtt would say
4260          * the measurement was invalid the time was < 50ms. The
4261          * same thing is true for 150Mb (8ms of time).
4262          *
4263          * A better way I realized is to look at what the maximum
4264          * the connection could possibly do. This is gated on
4265          * the lowest RTT we have seen and the highest rwnd.
4266          * We should in theory never exceed that, if we are
4267          * then something on the path is storing up packets
4268          * and then feeding them all at once to our endpoint
4269          * messing up our measurement.
4270          */
4271         rack->r_ctl.last_max_bw = rack->r_ctl.rc_gp_high_rwnd;
4272         rack->r_ctl.last_max_bw *= HPTS_USEC_IN_SEC;
4273         rack->r_ctl.last_max_bw /= rack->r_ctl.rc_gp_lowrtt;
4274         if (SEQ_LT(th_ack, tp->gput_seq)) {
4275                 /* No measurement can be made */
4276                 bytes = 0;
4277                 bytes_ps = 0;
4278                 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4279                                            0, 0, 0, 10, __LINE__, NULL);
4280                 goto skip_measurement;
4281         } else
4282                 bytes = (th_ack - tp->gput_seq);
4283         bytes_ps = (uint64_t)bytes;
4284         /*
4285          * Don't measure a b/w for pacing unless we have gotten at least
4286          * an initial windows worth of data in this measurement interval.
4287          *
4288          * Small numbers of bytes get badly influenced by delayed ack and
4289          * other artifacts. Note we take the initial window or our
4290          * defined minimum GP (defaulting to 10 which hopefully is the
4291          * IW).
4292          */
4293         if (rack->rc_gp_filled == 0) {
4294                 /*
4295                  * The initial estimate is special. We
4296                  * have blasted out an IW worth of packets
4297                  * without a real valid ack ts results. We
4298                  * then setup the app_limited_needs_set flag,
4299                  * this should get the first ack in (probably 2
4300                  * MSS worth) to be recorded as the timestamp.
4301                  * We thus allow a smaller number of bytes i.e.
4302                  * IW - 2MSS.
4303                  */
4304                 reqbytes -= (2 * segsiz);
4305                 /* Also lets fill previous for our first measurement to be neutral */
4306                 rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
4307         }
4308         if ((bytes_ps < reqbytes) || rack->app_limited_needs_set) {
4309                 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4310                                            rack->r_ctl.rc_app_limited_cnt,
4311                                            0, 0, 10, __LINE__, NULL);
4312                 goto skip_measurement;
4313         }
4314         /*
4315          * We now need to calculate the Timely like status so
4316          * we can update (possibly) the b/w multipliers.
4317          */
4318         new_rtt_diff = (int32_t)rack->r_ctl.rc_gp_srtt - (int32_t)rack->r_ctl.rc_prev_gp_srtt;
4319         if (rack->rc_gp_filled == 0) {
4320                 /* No previous reading */
4321                 rack->r_ctl.rc_rtt_diff = new_rtt_diff;
4322         } else {
4323                 if (rack->measure_saw_probe_rtt == 0) {
4324                         /*
4325                          * We don't want a probertt to be counted
4326                          * since it will be negative incorrectly. We
4327                          * expect to be reducing the RTT when we
4328                          * pace at a slower rate.
4329                          */
4330                         rack->r_ctl.rc_rtt_diff -= (rack->r_ctl.rc_rtt_diff / 8);
4331                         rack->r_ctl.rc_rtt_diff += (new_rtt_diff / 8);
4332                 }
4333         }
4334         timely_says = rack_make_timely_judgement(rack,
4335                 rack->r_ctl.rc_gp_srtt,
4336                 rack->r_ctl.rc_rtt_diff,
4337                 rack->r_ctl.rc_prev_gp_srtt
4338                 );
4339         bytes_ps *= HPTS_USEC_IN_SEC;
4340         bytes_ps /= utim;
4341         if (bytes_ps > rack->r_ctl.last_max_bw) {
4342                 /*
4343                  * Something is on path playing
4344                  * since this b/w is not possible based
4345                  * on our BDP (highest rwnd and lowest rtt
4346                  * we saw in the measurement window).
4347                  *
4348                  * Another option here would be to
4349                  * instead skip the measurement.
4350                  */
4351                 rack_log_pacing_delay_calc(rack, bytes, reqbytes,
4352                                            bytes_ps, rack->r_ctl.last_max_bw, 0,
4353                                            11, __LINE__, NULL);
4354                 bytes_ps = rack->r_ctl.last_max_bw;
4355         }
4356         /* We store gp for b/w in bytes per second */
4357         if (rack->rc_gp_filled == 0) {
4358                 /* Initial measurment */
4359                 if (bytes_ps) {
4360                         rack->r_ctl.gp_bw = bytes_ps;
4361                         rack->rc_gp_filled = 1;
4362                         rack->r_ctl.num_measurements = 1;
4363                         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
4364                 } else {
4365                         rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4366                                                    rack->r_ctl.rc_app_limited_cnt,
4367                                                    0, 0, 10, __LINE__, NULL);
4368                 }
4369                 if (rack->rc_inp->inp_in_hpts &&
4370                     (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
4371                         /*
4372                          * Ok we can't trust the pacer in this case
4373                          * where we transition from un-paced to paced.
4374                          * Or for that matter when the burst mitigation
4375                          * was making a wild guess and got it wrong.
4376                          * Stop the pacer and clear up all the aggregate
4377                          * delays etc.
4378                          */
4379                         tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
4380                         rack->r_ctl.rc_hpts_flags = 0;
4381                         rack->r_ctl.rc_last_output_to = 0;
4382                 }
4383                 did_add = 2;
4384         } else if (rack->r_ctl.num_measurements < RACK_REQ_AVG) {
4385                 /* Still a small number run an average */
4386                 rack->r_ctl.gp_bw += bytes_ps;
4387                 addpart = rack->r_ctl.num_measurements;
4388                 rack->r_ctl.num_measurements++;
4389                 if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
4390                         /* We have collected enought to move forward */
4391                         rack->r_ctl.gp_bw /= (uint64_t)rack->r_ctl.num_measurements;
4392                 }
4393                 did_add = 3;
4394         } else {
4395                 /*
4396                  * We want to take 1/wma of the goodput and add in to 7/8th
4397                  * of the old value weighted by the srtt. So if your measurement
4398                  * period is say 2 SRTT's long you would get 1/4 as the
4399                  * value, if it was like 1/2 SRTT then you would get 1/16th.
4400                  *
4401                  * But we must be careful not to take too much i.e. if the
4402                  * srtt is say 20ms and the measurement is taken over
4403                  * 400ms our weight would be 400/20 i.e. 20. On the
4404                  * other hand if we get a measurement over 1ms with a
4405                  * 10ms rtt we only want to take a much smaller portion.
4406                  */
4407                 if (rack->r_ctl.num_measurements < 0xff) {
4408                         rack->r_ctl.num_measurements++;
4409                 }
4410                 srtt = (uint64_t)tp->t_srtt;
4411                 if (srtt == 0) {
4412                         /*
4413                          * Strange why did t_srtt go back to zero?
4414                          */
4415                         if (rack->r_ctl.rc_rack_min_rtt)
4416                                 srtt = rack->r_ctl.rc_rack_min_rtt;
4417                         else
4418                                 srtt = HPTS_USEC_IN_MSEC;
4419                 }
4420                 /*
4421                  * XXXrrs: Note for reviewers, in playing with
4422                  * dynamic pacing I discovered this GP calculation
4423                  * as done originally leads to some undesired results.
4424                  * Basically you can get longer measurements contributing
4425                  * too much to the WMA. Thus I changed it if you are doing
4426                  * dynamic adjustments to only do the aportioned adjustment
4427                  * if we have a very small (time wise) measurement. Longer
4428                  * measurements just get there weight (defaulting to 1/8)
4429                  * add to the WMA. We may want to think about changing
4430                  * this to always do that for both sides i.e. dynamic
4431                  * and non-dynamic... but considering lots of folks
4432                  * were playing with this I did not want to change the
4433                  * calculation per.se. without your thoughts.. Lawerence?
4434                  * Peter??
4435                  */
4436                 if (rack->rc_gp_dyn_mul == 0) {
4437                         subpart = rack->r_ctl.gp_bw * utim;
4438                         subpart /= (srtt * 8);
4439                         if (subpart < (rack->r_ctl.gp_bw / 2)) {
4440                                 /*
4441                                  * The b/w update takes no more
4442                                  * away then 1/2 our running total
4443                                  * so factor it in.
4444                                  */
4445                                 addpart = bytes_ps * utim;
4446                                 addpart /= (srtt * 8);
4447                         } else {
4448                                 /*
4449                                  * Don't allow a single measurement
4450                                  * to account for more than 1/2 of the
4451                                  * WMA. This could happen on a retransmission
4452                                  * where utim becomes huge compared to
4453                                  * srtt (multiple retransmissions when using
4454                                  * the sending rate which factors in all the
4455                                  * transmissions from the first one).
4456                                  */
4457                                 subpart = rack->r_ctl.gp_bw / 2;
4458                                 addpart = bytes_ps / 2;
4459                         }
4460                         resid_bw = rack->r_ctl.gp_bw - subpart;
4461                         rack->r_ctl.gp_bw = resid_bw + addpart;
4462                         did_add = 1;
4463                 } else {
4464                         if ((utim / srtt) <= 1) {
4465                                 /*
4466                                  * The b/w update was over a small period
4467                                  * of time. The idea here is to prevent a small
4468                                  * measurement time period from counting
4469                                  * too much. So we scale it based on the
4470                                  * time so it attributes less than 1/rack_wma_divisor
4471                                  * of its measurement.
4472                                  */
4473                                 subpart = rack->r_ctl.gp_bw * utim;
4474                                 subpart /= (srtt * rack_wma_divisor);
4475                                 addpart = bytes_ps * utim;
4476                                 addpart /= (srtt * rack_wma_divisor);
4477                         } else {
4478                                 /*
4479                                  * The scaled measurement was long
4480                                  * enough so lets just add in the
4481                                  * portion of the measurment i.e. 1/rack_wma_divisor
4482                                  */
4483                                 subpart = rack->r_ctl.gp_bw / rack_wma_divisor;
4484                                 addpart = bytes_ps / rack_wma_divisor;
4485                         }
4486                         if ((rack->measure_saw_probe_rtt == 0) ||
4487                             (bytes_ps > rack->r_ctl.gp_bw)) {
4488                                 /*
4489                                  * For probe-rtt we only add it in
4490                                  * if its larger, all others we just
4491                                  * add in.
4492                                  */
4493                                 did_add = 1;
4494                                 resid_bw = rack->r_ctl.gp_bw - subpart;
4495                                 rack->r_ctl.gp_bw = resid_bw + addpart;
4496                         }
4497                 }
4498         }
4499         if ((rack->gp_ready == 0) &&
4500             (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
4501                 /* We have enough measurements now */
4502                 rack->gp_ready = 1;
4503                 rack_set_cc_pacing(rack);
4504                 if (rack->defer_options)
4505                         rack_apply_deferred_options(rack);
4506         }
4507         rack_log_pacing_delay_calc(rack, subpart, addpart, bytes_ps, stim,
4508                                    rack_get_bw(rack), 22, did_add, NULL);
4509         /* We do not update any multipliers if we are in or have seen a probe-rtt */
4510         if ((rack->measure_saw_probe_rtt == 0) && rack->rc_gp_rtt_set)
4511                 rack_update_multiplier(rack, timely_says, bytes_ps,
4512                                        rack->r_ctl.rc_gp_srtt,
4513                                        rack->r_ctl.rc_rtt_diff);
4514         rack_log_pacing_delay_calc(rack, bytes, tim, bytes_ps, stim,
4515                                    rack_get_bw(rack), 3, line, NULL);
4516         /* reset the gp srtt and setup the new prev */
4517         rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
4518         /* Record the lost count for the next measurement */
4519         rack->r_ctl.rc_loss_at_start = rack->r_ctl.rc_loss_count;
4520         /*
4521          * We restart our diffs based on the gpsrtt in the
4522          * measurement window.
4523          */
4524         rack->rc_gp_rtt_set = 0;
4525         rack->rc_gp_saw_rec = 0;
4526         rack->rc_gp_saw_ca = 0;
4527         rack->rc_gp_saw_ss = 0;
4528         rack->rc_dragged_bottom = 0;
4529 skip_measurement:
4530
4531 #ifdef STATS
4532         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT,
4533                                  gput);
4534         /*
4535          * XXXLAS: This is a temporary hack, and should be
4536          * chained off VOI_TCP_GPUT when stats(9) grows an
4537          * API to deal with chained VOIs.
4538          */
4539         if (tp->t_stats_gput_prev > 0)
4540                 stats_voi_update_abs_s32(tp->t_stats,
4541                                          VOI_TCP_GPUT_ND,
4542                                          ((gput - tp->t_stats_gput_prev) * 100) /
4543                                          tp->t_stats_gput_prev);
4544 #endif
4545         tp->t_flags &= ~TF_GPUTINPROG;
4546         tp->t_stats_gput_prev = gput;
4547         /*
4548          * Now are we app limited now and there is space from where we
4549          * were to where we want to go?
4550          *
4551          * We don't do the other case i.e. non-applimited here since
4552          * the next send will trigger us picking up the missing data.
4553          */
4554         if (rack->r_ctl.rc_first_appl &&
4555             TCPS_HAVEESTABLISHED(tp->t_state) &&
4556             rack->r_ctl.rc_app_limited_cnt &&
4557             (SEQ_GT(rack->r_ctl.rc_first_appl->r_start, th_ack)) &&
4558             ((rack->r_ctl.rc_first_appl->r_start - th_ack) >
4559              max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
4560                 /*
4561                  * Yep there is enough outstanding to make a measurement here.
4562                  */
4563                 struct rack_sendmap *rsm, fe;
4564
4565                 tp->t_flags |= TF_GPUTINPROG;
4566                 rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
4567                 rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
4568                 tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4569                 rack->app_limited_needs_set = 0;
4570                 tp->gput_seq = th_ack;
4571                 if (rack->in_probe_rtt)
4572                         rack->measure_saw_probe_rtt = 1;
4573                 else if ((rack->measure_saw_probe_rtt) &&
4574                          (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
4575                         rack->measure_saw_probe_rtt = 0;
4576                 if ((rack->r_ctl.rc_first_appl->r_start - th_ack) >= rack_get_measure_window(tp, rack)) {
4577                         /* There is a full window to gain info from */
4578                         tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
4579                 } else {
4580                         /* We can only measure up to the applimited point */
4581                         tp->gput_ack = tp->gput_seq + (rack->r_ctl.rc_first_appl->r_start - th_ack);
4582                 }
4583                 /*
4584                  * Now we need to find the timestamp of the send at tp->gput_seq
4585                  * for the send based measurement.
4586                  */
4587                 fe.r_start = tp->gput_seq;
4588                 rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
4589                 if (rsm) {
4590                         /* Ok send-based limit is set */
4591                         if (SEQ_LT(rsm->r_start, tp->gput_seq)) {
4592                                 /*
4593                                  * Move back to include the earlier part
4594                                  * so our ack time lines up right (this may
4595                                  * make an overlapping measurement but thats
4596                                  * ok).
4597                                  */
4598                                 tp->gput_seq = rsm->r_start;
4599                         }
4600                         if (rsm->r_flags & RACK_ACKED)
4601                                 tp->gput_ts = (uint32_t)rsm->r_ack_arrival;
4602                         else
4603                                 rack->app_limited_needs_set = 1;
4604                         rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
4605                 } else {
4606                         /*
4607                          * If we don't find the rsm due to some
4608                          * send-limit set the current time, which
4609                          * basically disables the send-limit.
4610                          */
4611                         struct timeval tv;
4612
4613                         microuptime(&tv);
4614                         rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
4615                 }
4616                 rack_log_pacing_delay_calc(rack,
4617                                            tp->gput_seq,
4618                                            tp->gput_ack,
4619                                            (uint64_t)rsm,
4620                                            tp->gput_ts,
4621                                            rack->r_ctl.rc_app_limited_cnt,
4622                                            9,
4623                                            __LINE__, NULL);
4624         }
4625 }
4626
4627 /*
4628  * CC wrapper hook functions
4629  */
4630 static void
4631 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, uint32_t th_ack, uint16_t nsegs,
4632     uint16_t type, int32_t recovery)
4633 {
4634         uint32_t prior_cwnd, acked;
4635         struct tcp_log_buffer *lgb = NULL;
4636         uint8_t labc_to_use;
4637
4638         INP_WLOCK_ASSERT(tp->t_inpcb);
4639         tp->ccv->nsegs = nsegs;
4640         acked = tp->ccv->bytes_this_ack = (th_ack - tp->snd_una);
4641         if ((recovery) && (rack->r_ctl.rc_early_recovery_segs)) {
4642                 uint32_t max;
4643
4644                 max = rack->r_ctl.rc_early_recovery_segs * ctf_fixed_maxseg(tp);
4645                 if (tp->ccv->bytes_this_ack > max) {
4646                         tp->ccv->bytes_this_ack = max;
4647                 }
4648         }
4649 #ifdef STATS
4650         stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF,
4651             ((int32_t)rack->r_ctl.cwnd_to_use) - tp->snd_wnd);
4652 #endif
4653         if ((tp->t_flags & TF_GPUTINPROG) &&
4654             rack_enough_for_measurement(tp, rack, th_ack)) {
4655                 /* Measure the Goodput */
4656                 rack_do_goodput_measurement(tp, rack, th_ack, __LINE__);
4657 #ifdef NETFLIX_PEAKRATE
4658                 if ((type == CC_ACK) &&
4659                     (tp->t_maxpeakrate)) {
4660                         /*
4661                          * We update t_peakrate_thr. This gives us roughly
4662                          * one update per round trip time. Note
4663                          * it will only be used if pace_always is off i.e
4664                          * we don't do this for paced flows.
4665                          */
4666                         rack_update_peakrate_thr(tp);
4667                 }
4668 #endif
4669         }
4670         /* Which way our we limited, if not cwnd limited no advance in CA */
4671         if (tp->snd_cwnd <= tp->snd_wnd)
4672                 tp->ccv->flags |= CCF_CWND_LIMITED;
4673         else
4674                 tp->ccv->flags &= ~CCF_CWND_LIMITED;
4675         if (tp->snd_cwnd > tp->snd_ssthresh) {
4676                 tp->t_bytes_acked += min(tp->ccv->bytes_this_ack,
4677                          nsegs * V_tcp_abc_l_var * ctf_fixed_maxseg(tp));
4678                 /* For the setting of a window past use the actual scwnd we are using */
4679                 if (tp->t_bytes_acked >= rack->r_ctl.cwnd_to_use) {
4680                         tp->t_bytes_acked -= rack->r_ctl.cwnd_to_use;
4681                         tp->ccv->flags |= CCF_ABC_SENTAWND;
4682                 }
4683         } else {
4684                 tp->ccv->flags &= ~CCF_ABC_SENTAWND;
4685                 tp->t_bytes_acked = 0;
4686         }
4687         prior_cwnd = tp->snd_cwnd;
4688         if ((recovery == 0) || (rack_max_abc_post_recovery == 0) || rack->r_use_labc_for_rec ||
4689             (rack_client_low_buf && (rack->client_bufferlvl < rack_client_low_buf)))
4690                 labc_to_use = rack->rc_labc;
4691         else
4692                 labc_to_use = rack_max_abc_post_recovery;
4693         if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
4694                 union tcp_log_stackspecific log;
4695                 struct timeval tv;
4696
4697                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4698                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4699                 log.u_bbr.flex1 = th_ack;
4700                 log.u_bbr.flex2 = tp->ccv->flags;
4701                 log.u_bbr.flex3 = tp->ccv->bytes_this_ack;
4702                 log.u_bbr.flex4 = tp->ccv->nsegs;
4703                 log.u_bbr.flex5 = labc_to_use;
4704                 log.u_bbr.flex6 = prior_cwnd;
4705                 log.u_bbr.flex7 = V_tcp_do_newsack;
4706                 log.u_bbr.flex8 = 1;
4707                 lgb = tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
4708                                      0, &log, false, NULL, NULL, 0, &tv);
4709         }
4710         if (CC_ALGO(tp)->ack_received != NULL) {
4711                 /* XXXLAS: Find a way to live without this */
4712                 tp->ccv->curack = th_ack;
4713                 tp->ccv->labc = labc_to_use;
4714                 tp->ccv->flags |= CCF_USE_LOCAL_ABC;
4715                 CC_ALGO(tp)->ack_received(tp->ccv, type);
4716         }
4717         if (lgb) {
4718                 lgb->tlb_stackinfo.u_bbr.flex6 = tp->snd_cwnd;
4719         }
4720         if (rack->r_must_retran) {
4721                 if (SEQ_GEQ(th_ack, rack->r_ctl.rc_snd_max_at_rto)) {
4722                         /*
4723                          * We now are beyond the rxt point so lets disable
4724                          * the flag.
4725                          */
4726                         rack->r_ctl.rc_out_at_rto = 0;
4727                         rack->r_must_retran = 0;
4728                 } else if ((prior_cwnd + ctf_fixed_maxseg(tp)) <= tp->snd_cwnd) {
4729                         /*
4730                          * Only decrement the rc_out_at_rto if the cwnd advances
4731                          * at least a whole segment. Otherwise next time the peer
4732                          * acks, we won't be able to send this generaly happens
4733                          * when we are in Congestion Avoidance.
4734                          */
4735                         if (acked <= rack->r_ctl.rc_out_at_rto){
4736                                 rack->r_ctl.rc_out_at_rto -= acked;
4737                         } else {
4738                                 rack->r_ctl.rc_out_at_rto = 0;
4739                         }
4740                 }
4741         }
4742 #ifdef STATS
4743         stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, rack->r_ctl.cwnd_to_use);
4744 #endif
4745         if (rack->r_ctl.rc_rack_largest_cwnd < rack->r_ctl.cwnd_to_use) {
4746                 rack->r_ctl.rc_rack_largest_cwnd = rack->r_ctl.cwnd_to_use;
4747         }
4748 #ifdef NETFLIX_PEAKRATE
4749         /* we enforce max peak rate if it is set and we are not pacing */
4750         if ((rack->rc_always_pace == 0) &&
4751             tp->t_peakrate_thr &&
4752             (tp->snd_cwnd > tp->t_peakrate_thr)) {
4753                 tp->snd_cwnd = tp->t_peakrate_thr;
4754         }
4755 #endif
4756 }
4757
4758 static void
4759 tcp_rack_partialack(struct tcpcb *tp)
4760 {
4761         struct tcp_rack *rack;
4762
4763         rack = (struct tcp_rack *)tp->t_fb_ptr;
4764         INP_WLOCK_ASSERT(tp->t_inpcb);
4765         /*
4766          * If we are doing PRR and have enough
4767          * room to send <or> we are pacing and prr
4768          * is disabled we will want to see if we
4769          * can send data (by setting r_wanted_output to
4770          * true).
4771          */
4772         if ((rack->r_ctl.rc_prr_sndcnt > 0) ||
4773             rack->rack_no_prr)
4774                 rack->r_wanted_output = 1;
4775 }
4776
4777 static void
4778 rack_post_recovery(struct tcpcb *tp, uint32_t th_ack)
4779 {
4780         struct tcp_rack *rack;
4781         uint32_t orig_cwnd;
4782
4783         orig_cwnd = tp->snd_cwnd;
4784         INP_WLOCK_ASSERT(tp->t_inpcb);
4785         rack = (struct tcp_rack *)tp->t_fb_ptr;
4786         /* only alert CC if we alerted when we entered */
4787         if (CC_ALGO(tp)->post_recovery != NULL) {
4788                 tp->ccv->curack = th_ack;
4789                 CC_ALGO(tp)->post_recovery(tp->ccv);
4790                 if (tp->snd_cwnd < tp->snd_ssthresh) {
4791                         /*
4792                          * Rack has burst control and pacing
4793                          * so lets not set this any lower than
4794                          * snd_ssthresh per RFC-6582 (option 2).
4795                          */
4796                         tp->snd_cwnd = tp->snd_ssthresh;
4797                 }
4798         }
4799         if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
4800                 union tcp_log_stackspecific log;
4801                 struct timeval tv;
4802
4803                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4804                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4805                 log.u_bbr.flex1 = th_ack;
4806                 log.u_bbr.flex2 = tp->ccv->flags;
4807                 log.u_bbr.flex3 = tp->ccv->bytes_this_ack;
4808                 log.u_bbr.flex4 = tp->ccv->nsegs;
4809                 log.u_bbr.flex5 = V_tcp_abc_l_var;
4810                 log.u_bbr.flex6 = orig_cwnd;
4811                 log.u_bbr.flex7 = V_tcp_do_newsack;
4812                 log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
4813                 log.u_bbr.flex8 = 2;
4814                 tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
4815                                0, &log, false, NULL, NULL, 0, &tv);
4816         }
4817         if ((rack->rack_no_prr == 0) &&
4818             (rack->no_prr_addback == 0) &&
4819             (rack->r_ctl.rc_prr_sndcnt > 0)) {
4820                 /*
4821                  * Suck the next prr cnt back into cwnd, but
4822                  * only do that if we are not application limited.
4823                  */
4824                 if (ctf_outstanding(tp) <= sbavail(&(tp->t_inpcb->inp_socket->so_snd))) {
4825                         /*
4826                          * We are allowed to add back to the cwnd the amount we did
4827                          * not get out if:
4828                          * a) no_prr_addback is off.
4829                          * b) we are not app limited
4830                          * c) we are doing prr
4831                          * <and>
4832                          * d) it is bounded by rack_prr_addbackmax (if addback is 0, then none).
4833                          */
4834                         tp->snd_cwnd += min((ctf_fixed_maxseg(tp) * rack_prr_addbackmax),
4835                                             rack->r_ctl.rc_prr_sndcnt);
4836                 }
4837                 rack->r_ctl.rc_prr_sndcnt = 0;
4838                 rack_log_to_prr(rack, 1, 0);
4839         }
4840         rack_log_to_prr(rack, 14, orig_cwnd);
4841         tp->snd_recover = tp->snd_una;
4842         EXIT_RECOVERY(tp->t_flags);
4843 }
4844
4845 static void
4846 rack_cong_signal(struct tcpcb *tp, uint32_t type, uint32_t ack)
4847 {
4848         struct tcp_rack *rack;
4849         uint32_t ssthresh_enter, cwnd_enter, in_rec_at_entry, orig_cwnd;
4850
4851         INP_WLOCK_ASSERT(tp->t_inpcb);
4852 #ifdef STATS
4853         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type);
4854 #endif
4855         if (IN_RECOVERY(tp->t_flags) == 0) {
4856                 in_rec_at_entry = 0;
4857                 ssthresh_enter = tp->snd_ssthresh;
4858                 cwnd_enter = tp->snd_cwnd;
4859         } else
4860                 in_rec_at_entry = 1;
4861         rack = (struct tcp_rack *)tp->t_fb_ptr;
4862         switch (type) {
4863         case CC_NDUPACK:
4864                 tp->t_flags &= ~TF_WASFRECOVERY;
4865                 tp->t_flags &= ~TF_WASCRECOVERY;
4866                 if (!IN_FASTRECOVERY(tp->t_flags)) {
4867                         rack->r_ctl.rc_prr_delivered = 0;
4868                         rack->r_ctl.rc_prr_out = 0;
4869                         if (rack->rack_no_prr == 0) {
4870                                 rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
4871                                 rack_log_to_prr(rack, 2, in_rec_at_entry);
4872                         }
4873                         rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una;
4874                         tp->snd_recover = tp->snd_max;
4875                         if (tp->t_flags2 & TF2_ECN_PERMIT)
4876                                 tp->t_flags2 |= TF2_ECN_SND_CWR;
4877                 }
4878                 break;
4879         case CC_ECN:
4880                 if (!IN_CONGRECOVERY(tp->t_flags) ||
4881                     /*
4882                      * Allow ECN reaction on ACK to CWR, if
4883                      * that data segment was also CE marked.
4884                      */
4885                     SEQ_GEQ(ack, tp->snd_recover)) {
4886                         EXIT_CONGRECOVERY(tp->t_flags);
4887                         KMOD_TCPSTAT_INC(tcps_ecn_rcwnd);
4888                         tp->snd_recover = tp->snd_max + 1;
4889                         if (tp->t_flags2 & TF2_ECN_PERMIT)
4890                                 tp->t_flags2 |= TF2_ECN_SND_CWR;
4891                 }
4892                 break;
4893         case CC_RTO:
4894                 tp->t_dupacks = 0;
4895                 tp->t_bytes_acked = 0;
4896                 EXIT_RECOVERY(tp->t_flags);
4897                 tp->snd_ssthresh = max(2, min(tp->snd_wnd, rack->r_ctl.cwnd_to_use) / 2 /
4898                     ctf_fixed_maxseg(tp)) * ctf_fixed_maxseg(tp);
4899                 orig_cwnd = tp->snd_cwnd;
4900                 tp->snd_cwnd = ctf_fixed_maxseg(tp);
4901                 rack_log_to_prr(rack, 16, orig_cwnd);
4902                 if (tp->t_flags2 & TF2_ECN_PERMIT)
4903                         tp->t_flags2 |= TF2_ECN_SND_CWR;
4904                 break;
4905         case CC_RTO_ERR:
4906                 KMOD_TCPSTAT_INC(tcps_sndrexmitbad);
4907                 /* RTO was unnecessary, so reset everything. */
4908                 tp->snd_cwnd = tp->snd_cwnd_prev;
4909                 tp->snd_ssthresh = tp->snd_ssthresh_prev;
4910                 tp->snd_recover = tp->snd_recover_prev;
4911                 if (tp->t_flags & TF_WASFRECOVERY) {
4912                         ENTER_FASTRECOVERY(tp->t_flags);
4913                         tp->t_flags &= ~TF_WASFRECOVERY;
4914                 }
4915                 if (tp->t_flags & TF_WASCRECOVERY) {
4916                         ENTER_CONGRECOVERY(tp->t_flags);
4917                         tp->t_flags &= ~TF_WASCRECOVERY;
4918                 }
4919                 tp->snd_nxt = tp->snd_max;
4920                 tp->t_badrxtwin = 0;
4921                 break;
4922         }
4923         if ((CC_ALGO(tp)->cong_signal != NULL)  &&
4924             (type != CC_RTO)){
4925                 tp->ccv->curack = ack;
4926                 CC_ALGO(tp)->cong_signal(tp->ccv, type);
4927         }
4928         if ((in_rec_at_entry == 0) && IN_RECOVERY(tp->t_flags)) {
4929                 rack_log_to_prr(rack, 15, cwnd_enter);
4930                 rack->r_ctl.dsack_byte_cnt = 0;
4931                 rack->r_ctl.retran_during_recovery = 0;
4932                 rack->r_ctl.rc_cwnd_at_erec = cwnd_enter;
4933                 rack->r_ctl.rc_ssthresh_at_erec = ssthresh_enter;
4934                 rack->r_ent_rec_ns = 1;
4935         }
4936 }
4937
4938 static inline void
4939 rack_cc_after_idle(struct tcp_rack *rack, struct tcpcb *tp)
4940 {
4941         uint32_t i_cwnd;
4942
4943         INP_WLOCK_ASSERT(tp->t_inpcb);
4944
4945 #ifdef NETFLIX_STATS
4946         KMOD_TCPSTAT_INC(tcps_idle_restarts);
4947         if (tp->t_state == TCPS_ESTABLISHED)
4948                 KMOD_TCPSTAT_INC(tcps_idle_estrestarts);
4949 #endif
4950         if (CC_ALGO(tp)->after_idle != NULL)
4951                 CC_ALGO(tp)->after_idle(tp->ccv);
4952
4953         if (tp->snd_cwnd == 1)
4954                 i_cwnd = tp->t_maxseg;          /* SYN(-ACK) lost */
4955         else
4956                 i_cwnd = rc_init_window(rack);
4957
4958         /*
4959          * Being idle is no differnt than the initial window. If the cc
4960          * clamps it down below the initial window raise it to the initial
4961          * window.
4962          */
4963         if (tp->snd_cwnd < i_cwnd) {
4964                 tp->snd_cwnd = i_cwnd;
4965         }
4966 }
4967
4968 /*
4969  * Indicate whether this ack should be delayed.  We can delay the ack if
4970  * following conditions are met:
4971  *      - There is no delayed ack timer in progress.
4972  *      - Our last ack wasn't a 0-sized window. We never want to delay
4973  *        the ack that opens up a 0-sized window.
4974  *      - LRO wasn't used for this segment. We make sure by checking that the
4975  *        segment size is not larger than the MSS.
4976  *      - Delayed acks are enabled or this is a half-synchronized T/TCP
4977  *        connection.
4978  */
4979 #define DELAY_ACK(tp, tlen)                      \
4980         (((tp->t_flags & TF_RXWIN0SENT) == 0) && \
4981         ((tp->t_flags & TF_DELACK) == 0) &&      \
4982         (tlen <= tp->t_maxseg) &&                \
4983         (tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN)))
4984
4985 static struct rack_sendmap *
4986 rack_find_lowest_rsm(struct tcp_rack *rack)
4987 {
4988         struct rack_sendmap *rsm;
4989
4990         /*
4991          * Walk the time-order transmitted list looking for an rsm that is
4992          * not acked. This will be the one that was sent the longest time
4993          * ago that is still outstanding.
4994          */
4995         TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
4996                 if (rsm->r_flags & RACK_ACKED) {
4997                         continue;
4998                 }
4999                 goto finish;
5000         }
5001 finish:
5002         return (rsm);
5003 }
5004
5005 static struct rack_sendmap *
5006 rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm)
5007 {
5008         struct rack_sendmap *prsm;
5009
5010         /*
5011          * Walk the sequence order list backward until we hit and arrive at
5012          * the highest seq not acked. In theory when this is called it
5013          * should be the last segment (which it was not).
5014          */
5015         counter_u64_add(rack_find_high, 1);
5016         prsm = rsm;
5017         RB_FOREACH_REVERSE_FROM(prsm, rack_rb_tree_head, rsm) {
5018                 if (prsm->r_flags & (RACK_ACKED | RACK_HAS_FIN)) {
5019                         continue;
5020                 }
5021                 return (prsm);
5022         }
5023         return (NULL);
5024 }
5025
5026 static uint32_t
5027 rack_calc_thresh_rack(struct tcp_rack *rack, uint32_t srtt, uint32_t cts)
5028 {
5029         int32_t lro;
5030         uint32_t thresh;
5031
5032         /*
5033          * lro is the flag we use to determine if we have seen reordering.
5034          * If it gets set we have seen reordering. The reorder logic either
5035          * works in one of two ways:
5036          *
5037          * If reorder-fade is configured, then we track the last time we saw
5038          * re-ordering occur. If we reach the point where enough time as
5039          * passed we no longer consider reordering has occuring.
5040          *
5041          * Or if reorder-face is 0, then once we see reordering we consider
5042          * the connection to alway be subject to reordering and just set lro
5043          * to 1.
5044          *
5045          * In the end if lro is non-zero we add the extra time for
5046          * reordering in.
5047          */
5048         if (srtt == 0)
5049                 srtt = 1;
5050         if (rack->r_ctl.rc_reorder_ts) {
5051                 if (rack->r_ctl.rc_reorder_fade) {
5052                         if (SEQ_GEQ(cts, rack->r_ctl.rc_reorder_ts)) {
5053                                 lro = cts - rack->r_ctl.rc_reorder_ts;
5054                                 if (lro == 0) {
5055                                         /*
5056                                          * No time as passed since the last
5057                                          * reorder, mark it as reordering.
5058                                          */
5059                                         lro = 1;
5060                                 }
5061                         } else {
5062                                 /* Negative time? */
5063                                 lro = 0;
5064                         }
5065                         if (lro > rack->r_ctl.rc_reorder_fade) {
5066                                 /* Turn off reordering seen too */
5067                                 rack->r_ctl.rc_reorder_ts = 0;
5068                                 lro = 0;
5069                         }
5070                 } else {
5071                         /* Reodering does not fade */
5072                         lro = 1;
5073                 }
5074         } else {
5075                 lro = 0;
5076         }
5077         thresh = srtt + rack->r_ctl.rc_pkt_delay;
5078         if (lro) {
5079                 /* It must be set, if not you get 1/4 rtt */
5080                 if (rack->r_ctl.rc_reorder_shift)
5081                         thresh += (srtt >> rack->r_ctl.rc_reorder_shift);
5082                 else
5083                         thresh += (srtt >> 2);
5084         } else {
5085                 thresh += 1;
5086         }
5087         /* We don't let the rack timeout be above a RTO */
5088         if (thresh > rack->rc_tp->t_rxtcur) {
5089                 thresh = rack->rc_tp->t_rxtcur;
5090         }
5091         /* And we don't want it above the RTO max either */
5092         if (thresh > rack_rto_max) {
5093                 thresh = rack_rto_max;
5094         }
5095         return (thresh);
5096 }
5097
5098 static uint32_t
5099 rack_calc_thresh_tlp(struct tcpcb *tp, struct tcp_rack *rack,
5100                      struct rack_sendmap *rsm, uint32_t srtt)
5101 {
5102         struct rack_sendmap *prsm;
5103         uint32_t thresh, len;
5104         int segsiz;
5105
5106         if (srtt == 0)
5107                 srtt = 1;
5108         if (rack->r_ctl.rc_tlp_threshold)
5109                 thresh = srtt + (srtt / rack->r_ctl.rc_tlp_threshold);
5110         else
5111                 thresh = (srtt * 2);
5112
5113         /* Get the previous sent packet, if any */
5114         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
5115         counter_u64_add(rack_enter_tlp_calc, 1);
5116         len = rsm->r_end - rsm->r_start;
5117         if (rack->rack_tlp_threshold_use == TLP_USE_ID) {
5118                 /* Exactly like the ID */
5119                 if (((tp->snd_max - tp->snd_una) - rack->r_ctl.rc_sacked + rack->r_ctl.rc_holes_rxt) <= segsiz) {
5120                         uint32_t alt_thresh;
5121                         /*
5122                          * Compensate for delayed-ack with the d-ack time.
5123                          */
5124                         counter_u64_add(rack_used_tlpmethod, 1);
5125                         alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5126                         if (alt_thresh > thresh)
5127                                 thresh = alt_thresh;
5128                 }
5129         } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_ONE) {
5130                 /* 2.1 behavior */
5131                 prsm = TAILQ_PREV(rsm, rack_head, r_tnext);
5132                 if (prsm && (len <= segsiz)) {
5133                         /*
5134                          * Two packets outstanding, thresh should be (2*srtt) +
5135                          * possible inter-packet delay (if any).
5136                          */
5137                         uint32_t inter_gap = 0;
5138                         int idx, nidx;
5139
5140                         counter_u64_add(rack_used_tlpmethod, 1);
5141                         idx = rsm->r_rtr_cnt - 1;
5142                         nidx = prsm->r_rtr_cnt - 1;
5143                         if (rsm->r_tim_lastsent[nidx] >= prsm->r_tim_lastsent[idx]) {
5144                                 /* Yes it was sent later (or at the same time) */
5145                                 inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx];
5146                         }
5147                         thresh += inter_gap;
5148                 } else if (len <= segsiz) {
5149                         /*
5150                          * Possibly compensate for delayed-ack.
5151                          */
5152                         uint32_t alt_thresh;
5153
5154                         counter_u64_add(rack_used_tlpmethod2, 1);
5155                         alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5156                         if (alt_thresh > thresh)
5157                                 thresh = alt_thresh;
5158                 }
5159         } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_TWO) {
5160                 /* 2.2 behavior */
5161                 if (len <= segsiz) {
5162                         uint32_t alt_thresh;
5163                         /*
5164                          * Compensate for delayed-ack with the d-ack time.
5165                          */
5166                         counter_u64_add(rack_used_tlpmethod, 1);
5167                         alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5168                         if (alt_thresh > thresh)
5169                                 thresh = alt_thresh;
5170                 }
5171         }
5172         /* Not above an RTO */
5173         if (thresh > tp->t_rxtcur) {
5174                 thresh = tp->t_rxtcur;
5175         }
5176         /* Not above a RTO max */
5177         if (thresh > rack_rto_max) {
5178                 thresh = rack_rto_max;
5179         }
5180         /* Apply user supplied min TLP */
5181         if (thresh < rack_tlp_min) {
5182                 thresh = rack_tlp_min;
5183         }
5184         return (thresh);
5185 }
5186
5187 static uint32_t
5188 rack_grab_rtt(struct tcpcb *tp, struct tcp_rack *rack)
5189 {
5190         /*
5191          * We want the rack_rtt which is the
5192          * last rtt we measured. However if that
5193          * does not exist we fallback to the srtt (which
5194          * we probably will never do) and then as a last
5195          * resort we use RACK_INITIAL_RTO if no srtt is
5196          * yet set.
5197          */
5198         if (rack->rc_rack_rtt)
5199                 return (rack->rc_rack_rtt);
5200         else if (tp->t_srtt == 0)
5201                 return (RACK_INITIAL_RTO);
5202         return (tp->t_srtt);
5203 }
5204
5205 static struct rack_sendmap *
5206 rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused)
5207 {
5208         /*
5209          * Check to see that we don't need to fall into recovery. We will
5210          * need to do so if our oldest transmit is past the time we should
5211          * have had an ack.
5212          */
5213         struct tcp_rack *rack;
5214         struct rack_sendmap *rsm;
5215         int32_t idx;
5216         uint32_t srtt, thresh;
5217
5218         rack = (struct tcp_rack *)tp->t_fb_ptr;
5219         if (RB_EMPTY(&rack->r_ctl.rc_mtree)) {
5220                 return (NULL);
5221         }
5222         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5223         if (rsm == NULL)
5224                 return (NULL);
5225
5226         if (rsm->r_flags & RACK_ACKED) {
5227                 rsm = rack_find_lowest_rsm(rack);
5228                 if (rsm == NULL)
5229                         return (NULL);
5230         }
5231         idx = rsm->r_rtr_cnt - 1;
5232         srtt = rack_grab_rtt(tp, rack);
5233         thresh = rack_calc_thresh_rack(rack, srtt, tsused);
5234         if (TSTMP_LT(tsused, ((uint32_t)rsm->r_tim_lastsent[idx]))) {
5235                 return (NULL);
5236         }
5237         if ((tsused - ((uint32_t)rsm->r_tim_lastsent[idx])) < thresh) {
5238                 return (NULL);
5239         }
5240         /* Ok if we reach here we are over-due and this guy can be sent */
5241         if (IN_RECOVERY(tp->t_flags) == 0) {
5242                 /*
5243                  * For the one that enters us into recovery record undo
5244                  * info.
5245                  */
5246                 rack->r_ctl.rc_rsm_start = rsm->r_start;
5247                 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
5248                 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
5249         }
5250         rack_cong_signal(tp, CC_NDUPACK, tp->snd_una);
5251         return (rsm);
5252 }
5253
5254 static uint32_t
5255 rack_get_persists_timer_val(struct tcpcb *tp, struct tcp_rack *rack)
5256 {
5257         int32_t t;
5258         int32_t tt;
5259         uint32_t ret_val;
5260
5261         t = (tp->t_srtt + (tp->t_rttvar << 2));
5262         RACK_TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
5263             rack_persist_min, rack_persist_max);
5264         if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
5265                 tp->t_rxtshift++;
5266         rack->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT;
5267         ret_val = (uint32_t)tt;
5268         return (ret_val);
5269 }
5270
5271 static uint32_t
5272 rack_timer_start(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int sup_rack)
5273 {
5274         /*
5275          * Start the FR timer, we do this based on getting the first one in
5276          * the rc_tmap. Note that if its NULL we must stop the timer. in all
5277          * events we need to stop the running timer (if its running) before
5278          * starting the new one.
5279          */
5280         uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse;
5281         uint32_t srtt_cur;
5282         int32_t idx;
5283         int32_t is_tlp_timer = 0;
5284         struct rack_sendmap *rsm;
5285
5286         if (rack->t_timers_stopped) {
5287                 /* All timers have been stopped none are to run */
5288                 return (0);
5289         }
5290         if (rack->rc_in_persist) {
5291                 /* We can't start any timer in persists */
5292                 return (rack_get_persists_timer_val(tp, rack));
5293         }
5294         rack->rc_on_min_to = 0;
5295         if ((tp->t_state < TCPS_ESTABLISHED) ||
5296             ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
5297                 goto activate_rxt;
5298         }
5299         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5300         if ((rsm == NULL) || sup_rack) {
5301                 /* Nothing on the send map or no rack */
5302 activate_rxt:
5303                 time_since_sent = 0;
5304                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5305                 if (rsm) {
5306                         /*
5307                          * Should we discount the RTX timer any?
5308                          *
5309                          * We want to discount it the smallest amount.
5310                          * If a timer (Rack/TLP or RXT) has gone off more
5311                          * recently thats the discount we want to use (now - timer time).
5312                          * If the retransmit of the oldest packet was more recent then
5313                          * we want to use that (now - oldest-packet-last_transmit_time).
5314                          *
5315                          */
5316                         idx = rsm->r_rtr_cnt - 1;
5317                         if (TSTMP_GEQ(rack->r_ctl.rc_tlp_rxt_last_time, ((uint32_t)rsm->r_tim_lastsent[idx])))
5318                                 tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
5319                         else
5320                                 tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
5321                         if (TSTMP_GT(cts, tstmp_touse))
5322                             time_since_sent = cts - tstmp_touse;
5323                 }
5324                 if (SEQ_LT(tp->snd_una, tp->snd_max) || sbavail(&(tp->t_inpcb->inp_socket->so_snd))) {
5325                         rack->r_ctl.rc_hpts_flags |= PACE_TMR_RXT;
5326                         to = tp->t_rxtcur;
5327                         if (to > time_since_sent)
5328                                 to -= time_since_sent;
5329                         else
5330                                 to = rack->r_ctl.rc_min_to;
5331                         if (to == 0)
5332                                 to = 1;
5333                         /* Special case for KEEPINIT */
5334                         if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
5335                             (TP_KEEPINIT(tp) != 0) &&
5336                             rsm) {
5337                                 /*
5338                                  * We have to put a ceiling on the rxt timer
5339                                  * of the keep-init timeout.
5340                                  */
5341                                 uint32_t max_time, red;
5342
5343                                 max_time = TICKS_2_USEC(TP_KEEPINIT(tp));
5344                                 if (TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) {
5345                                         red = (cts - (uint32_t)rsm->r_tim_lastsent[0]);
5346                                         if (red < max_time)
5347                                                 max_time -= red;
5348                                         else
5349                                                 max_time = 1;
5350                                 }
5351                                 /* Reduce timeout to the keep value if needed */
5352                                 if (max_time < to)
5353                                         to = max_time;
5354                         }
5355                         return (to);
5356                 }
5357                 return (0);
5358         }
5359         if (rsm->r_flags & RACK_ACKED) {
5360                 rsm = rack_find_lowest_rsm(rack);
5361                 if (rsm == NULL) {
5362                         /* No lowest? */
5363                         goto activate_rxt;
5364                 }
5365         }
5366         if (rack->sack_attack_disable) {
5367                 /*
5368                  * We don't want to do
5369                  * any TLP's if you are an attacker.
5370                  * Though if you are doing what
5371                  * is expected you may still have
5372                  * SACK-PASSED marks.
5373                  */
5374                 goto activate_rxt;
5375         }
5376         /* Convert from ms to usecs */
5377         if ((rsm->r_flags & RACK_SACK_PASSED) || (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
5378                 if ((tp->t_flags & TF_SENTFIN) &&
5379                     ((tp->snd_max - tp->snd_una) == 1) &&
5380                     (rsm->r_flags & RACK_HAS_FIN)) {
5381                         /*
5382                          * We don't start a rack timer if all we have is a
5383                          * FIN outstanding.
5384                          */
5385                         goto activate_rxt;
5386                 }
5387                 if ((rack->use_rack_rr == 0) &&
5388                     (IN_FASTRECOVERY(tp->t_flags)) &&
5389                     (rack->rack_no_prr == 0) &&
5390                      (rack->r_ctl.rc_prr_sndcnt  < ctf_fixed_maxseg(tp))) {
5391                         /*
5392                          * We are not cheating, in recovery  and
5393                          * not enough ack's to yet get our next
5394                          * retransmission out.
5395                          *
5396                          * Note that classified attackers do not
5397                          * get to use the rack-cheat.
5398                          */
5399                         goto activate_tlp;
5400                 }
5401                 srtt = rack_grab_rtt(tp, rack);
5402                 thresh = rack_calc_thresh_rack(rack, srtt, cts);
5403                 idx = rsm->r_rtr_cnt - 1;
5404                 exp = ((uint32_t)rsm->r_tim_lastsent[idx]) + thresh;
5405                 if (SEQ_GEQ(exp, cts)) {
5406                         to = exp - cts;
5407                         if (to < rack->r_ctl.rc_min_to) {
5408                                 to = rack->r_ctl.rc_min_to;
5409                                 if (rack->r_rr_config == 3)
5410                                         rack->rc_on_min_to = 1;
5411                         }
5412                 } else {
5413                         to = rack->r_ctl.rc_min_to;
5414                         if (rack->r_rr_config == 3)
5415                                 rack->rc_on_min_to = 1;
5416                 }
5417         } else {
5418                 /* Ok we need to do a TLP not RACK */
5419 activate_tlp:
5420                 if ((rack->rc_tlp_in_progress != 0) &&
5421                     (rack->r_ctl.rc_tlp_cnt_out >= rack_tlp_limit)) {
5422                         /*
5423                          * The previous send was a TLP and we have sent
5424                          * N TLP's without sending new data.
5425                          */
5426                         goto activate_rxt;
5427                 }
5428                 rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
5429                 if (rsm == NULL) {
5430                         /* We found no rsm to TLP with. */
5431                         goto activate_rxt;
5432                 }
5433                 if (rsm->r_flags & RACK_HAS_FIN) {
5434                         /* If its a FIN we dont do TLP */
5435                         rsm = NULL;
5436                         goto activate_rxt;
5437                 }
5438                 idx = rsm->r_rtr_cnt - 1;
5439                 time_since_sent = 0;
5440                 if (TSTMP_GEQ(((uint32_t)rsm->r_tim_lastsent[idx]), rack->r_ctl.rc_tlp_rxt_last_time))
5441                         tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
5442                 else
5443                         tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
5444                 if (TSTMP_GT(cts, tstmp_touse))
5445                     time_since_sent = cts - tstmp_touse;
5446                 is_tlp_timer = 1;
5447                 if (tp->t_srtt) {
5448                         if ((rack->rc_srtt_measure_made == 0) &&
5449                             (tp->t_srtt == 1)) {
5450                                 /*
5451                                  * If another stack as run and set srtt to 1,
5452                                  * then the srtt was 0, so lets use the initial.
5453                                  */
5454                                 srtt = RACK_INITIAL_RTO;
5455                         } else {
5456                                 srtt_cur = tp->t_srtt;
5457                                 srtt = srtt_cur;
5458                         }
5459                 } else
5460                         srtt = RACK_INITIAL_RTO;
5461                 /*
5462                  * If the SRTT is not keeping up and the
5463                  * rack RTT has spiked we want to use
5464                  * the last RTT not the smoothed one.
5465                  */
5466                 if (rack_tlp_use_greater &&
5467                     tp->t_srtt &&
5468                     (srtt < rack_grab_rtt(tp, rack))) {
5469                         srtt = rack_grab_rtt(tp, rack);
5470                 }
5471                 thresh = rack_calc_thresh_tlp(tp, rack, rsm, srtt);
5472                 if (thresh > time_since_sent) {
5473                         to = thresh - time_since_sent;
5474                 } else {
5475                         to = rack->r_ctl.rc_min_to;
5476                         rack_log_alt_to_to_cancel(rack,
5477                                                   thresh,               /* flex1 */
5478                                                   time_since_sent,      /* flex2 */
5479                                                   tstmp_touse,          /* flex3 */
5480                                                   rack->r_ctl.rc_tlp_rxt_last_time, /* flex4 */
5481                                                   (uint32_t)rsm->r_tim_lastsent[idx],
5482                                                   srtt,
5483                                                   idx, 99);
5484                 }
5485                 if (to < rack_tlp_min) {
5486                         to = rack_tlp_min;
5487                 }
5488                 if (to > TICKS_2_USEC(TCPTV_REXMTMAX)) {
5489                         /*
5490                          * If the TLP time works out to larger than the max
5491                          * RTO lets not do TLP.. just RTO.
5492                          */
5493                         goto activate_rxt;
5494                 }
5495         }
5496         if (is_tlp_timer == 0) {
5497                 rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK;
5498         } else {
5499                 rack->r_ctl.rc_hpts_flags |= PACE_TMR_TLP;
5500         }
5501         if (to == 0)
5502                 to = 1;
5503         return (to);
5504 }
5505
5506 static void
5507 rack_enter_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5508 {
5509         if (rack->rc_in_persist == 0) {
5510                 if (tp->t_flags & TF_GPUTINPROG) {
5511                         /*
5512                          * Stop the goodput now, the calling of the
5513                          * measurement function clears the flag.
5514                          */
5515                         rack_do_goodput_measurement(tp, rack, tp->snd_una, __LINE__);
5516                 }
5517 #ifdef NETFLIX_SHARED_CWND
5518                 if (rack->r_ctl.rc_scw) {
5519                         tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
5520                         rack->rack_scwnd_is_idle = 1;
5521                 }
5522 #endif
5523                 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
5524                 if (rack->r_ctl.rc_went_idle_time == 0)
5525                         rack->r_ctl.rc_went_idle_time = 1;
5526                 rack_timer_cancel(tp, rack, cts, __LINE__);
5527                 tp->t_rxtshift = 0;
5528                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
5529                               rack_rto_min, rack_rto_max);
5530                 rack->rc_in_persist = 1;
5531         }
5532 }
5533
5534 static void
5535 rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5536 {
5537         if (rack->rc_inp->inp_in_hpts) {
5538                 tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
5539                 rack->r_ctl.rc_hpts_flags = 0;
5540         }
5541 #ifdef NETFLIX_SHARED_CWND
5542         if (rack->r_ctl.rc_scw) {
5543                 tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
5544                 rack->rack_scwnd_is_idle = 0;
5545         }
5546 #endif
5547         if (rack->rc_gp_dyn_mul &&
5548             (rack->use_fixed_rate == 0) &&
5549             (rack->rc_always_pace)) {
5550                 /*
5551                  * Do we count this as if a probe-rtt just
5552                  * finished?
5553                  */
5554                 uint32_t time_idle, idle_min;
5555
5556                 time_idle = tcp_get_usecs(NULL) - rack->r_ctl.rc_went_idle_time;
5557                 idle_min = rack_min_probertt_hold;
5558                 if (rack_probertt_gpsrtt_cnt_div) {
5559                         uint64_t extra;
5560                         extra = (uint64_t)rack->r_ctl.rc_gp_srtt *
5561                                 (uint64_t)rack_probertt_gpsrtt_cnt_mul;
5562                         extra /= (uint64_t)rack_probertt_gpsrtt_cnt_div;
5563                         idle_min += (uint32_t)extra;
5564                 }
5565                 if (time_idle >= idle_min) {
5566                         /* Yes, we count it as a probe-rtt. */
5567                         uint32_t us_cts;
5568
5569                         us_cts = tcp_get_usecs(NULL);
5570                         if (rack->in_probe_rtt == 0) {
5571                                 rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
5572                                 rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
5573                                 rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
5574                                 rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
5575                         } else {
5576                                 rack_exit_probertt(rack, us_cts);
5577                         }
5578                 }
5579         }
5580         rack->rc_in_persist = 0;
5581         rack->r_ctl.rc_went_idle_time = 0;
5582         tp->t_rxtshift = 0;
5583         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
5584            rack_rto_min, rack_rto_max);
5585         rack->r_ctl.rc_agg_delayed = 0;
5586         rack->r_early = 0;
5587         rack->r_late = 0;
5588         rack->r_ctl.rc_agg_early = 0;
5589 }
5590
5591 static void
5592 rack_log_hpts_diag(struct tcp_rack *rack, uint32_t cts,
5593                    struct hpts_diag *diag, struct timeval *tv)
5594 {
5595         if (rack_verbose_logging && rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
5596                 union tcp_log_stackspecific log;
5597
5598                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5599                 log.u_bbr.flex1 = diag->p_nxt_slot;
5600                 log.u_bbr.flex2 = diag->p_cur_slot;
5601                 log.u_bbr.flex3 = diag->slot_req;
5602                 log.u_bbr.flex4 = diag->inp_hptsslot;
5603                 log.u_bbr.flex5 = diag->slot_remaining;
5604                 log.u_bbr.flex6 = diag->need_new_to;
5605                 log.u_bbr.flex7 = diag->p_hpts_active;
5606                 log.u_bbr.flex8 = diag->p_on_min_sleep;
5607                 /* Hijack other fields as needed */
5608                 log.u_bbr.epoch = diag->have_slept;
5609                 log.u_bbr.lt_epoch = diag->yet_to_sleep;
5610                 log.u_bbr.pkts_out = diag->co_ret;
5611                 log.u_bbr.applimited = diag->hpts_sleep_time;
5612                 log.u_bbr.delivered = diag->p_prev_slot;
5613                 log.u_bbr.inflight = diag->p_runningtick;
5614                 log.u_bbr.bw_inuse = diag->wheel_tick;
5615                 log.u_bbr.rttProp = diag->wheel_cts;
5616                 log.u_bbr.timeStamp = cts;
5617                 log.u_bbr.delRate = diag->maxticks;
5618                 log.u_bbr.cur_del_rate = diag->p_curtick;
5619                 log.u_bbr.cur_del_rate <<= 32;
5620                 log.u_bbr.cur_del_rate |= diag->p_lasttick;
5621                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
5622                     &rack->rc_inp->inp_socket->so_rcv,
5623                     &rack->rc_inp->inp_socket->so_snd,
5624                     BBR_LOG_HPTSDIAG, 0,
5625                     0, &log, false, tv);
5626         }
5627
5628 }
5629
5630 static void
5631 rack_log_wakeup(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb, uint32_t len, int type)
5632 {
5633         if (rack_verbose_logging && rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
5634                 union tcp_log_stackspecific log;
5635                 struct timeval tv;
5636
5637                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5638                 log.u_bbr.flex1 = sb->sb_flags;
5639                 log.u_bbr.flex2 = len;
5640                 log.u_bbr.flex3 = sb->sb_state;
5641                 log.u_bbr.flex8 = type;
5642                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5643                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
5644                     &rack->rc_inp->inp_socket->so_rcv,
5645                     &rack->rc_inp->inp_socket->so_snd,
5646                     TCP_LOG_SB_WAKE, 0,
5647                     len, &log, false, &tv);
5648         }
5649 }
5650
5651 static void
5652 rack_start_hpts_timer(struct tcp_rack *rack, struct tcpcb *tp, uint32_t cts,
5653       int32_t slot, uint32_t tot_len_this_send, int sup_rack)
5654 {
5655         struct hpts_diag diag;
5656         struct inpcb *inp;
5657         struct timeval tv;
5658         uint32_t delayed_ack = 0;
5659         uint32_t hpts_timeout;
5660         uint32_t entry_slot = slot;
5661         uint8_t stopped;
5662         uint32_t left = 0;
5663         uint32_t us_cts;
5664
5665         inp = tp->t_inpcb;
5666         if ((tp->t_state == TCPS_CLOSED) ||
5667             (tp->t_state == TCPS_LISTEN)) {
5668                 return;
5669         }
5670         if (inp->inp_in_hpts) {
5671                 /* Already on the pacer */
5672                 return;
5673         }
5674         stopped = rack->rc_tmr_stopped;
5675         if (stopped && TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
5676                 left = rack->r_ctl.rc_timer_exp - cts;
5677         }
5678         rack->r_ctl.rc_timer_exp = 0;
5679         rack->r_ctl.rc_hpts_flags = 0;
5680         us_cts = tcp_get_usecs(&tv);
5681         /* Now early/late accounting */
5682         rack_log_pacing_delay_calc(rack, entry_slot, slot, 0, 0, 0, 26, __LINE__, NULL);
5683         if (rack->r_early && (rack->rc_ack_can_sendout_data == 0)) {
5684                 /*
5685                  * We have a early carry over set,
5686                  * we can always add more time so we
5687                  * can always make this compensation.
5688                  *
5689                  * Note if ack's are allowed to wake us do not
5690                  * penalize the next timer for being awoke
5691                  * by an ack aka the rc_agg_early (non-paced mode).
5692                  */
5693                 slot += rack->r_ctl.rc_agg_early;
5694                 rack->r_early = 0;
5695                 rack->r_ctl.rc_agg_early = 0;
5696         }
5697         if (rack->r_late) {
5698                 /*
5699                  * This is harder, we can
5700                  * compensate some but it
5701                  * really depends on what
5702                  * the current pacing time is.
5703                  */
5704                 if (rack->r_ctl.rc_agg_delayed >= slot) {
5705                         /*
5706                          * We can't compensate for it all.
5707                          * And we have to have some time
5708                          * on the clock. We always have a min
5709                          * 10 slots (10 x 10 i.e. 100 usecs).
5710                          */
5711                         if (slot <= HPTS_TICKS_PER_USEC) {
5712                                 /* We gain delay */
5713                                 rack->r_ctl.rc_agg_delayed += (HPTS_TICKS_PER_USEC - slot);
5714                                 slot = HPTS_TICKS_PER_USEC;
5715                         } else {
5716                                 /* We take off some */
5717                                 rack->r_ctl.rc_agg_delayed -= (slot - HPTS_TICKS_PER_USEC);
5718                                 slot = HPTS_TICKS_PER_USEC;
5719                         }
5720                 } else {
5721                         slot -= rack->r_ctl.rc_agg_delayed;
5722                         rack->r_ctl.rc_agg_delayed = 0;
5723                         /* Make sure we have 100 useconds at minimum */
5724                         if (slot < HPTS_TICKS_PER_USEC) {
5725                                 rack->r_ctl.rc_agg_delayed = HPTS_TICKS_PER_USEC - slot;
5726                                 slot = HPTS_TICKS_PER_USEC;
5727                         }
5728                         if (rack->r_ctl.rc_agg_delayed == 0)
5729                                 rack->r_late = 0;
5730                 }
5731         }
5732         if (slot) {
5733                 /* We are pacing too */
5734                 rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT;
5735         }
5736         hpts_timeout = rack_timer_start(tp, rack, cts, sup_rack);
5737 #ifdef NETFLIX_EXP_DETECTION
5738         if (rack->sack_attack_disable &&
5739             (slot < tcp_sad_pacing_interval)) {
5740                 /*
5741                  * We have a potential attacker on
5742                  * the line. We have possibly some
5743                  * (or now) pacing time set. We want to
5744                  * slow down the processing of sacks by some
5745                  * amount (if it is an attacker). Set the default
5746                  * slot for attackers in place (unless the orginal
5747                  * interval is longer). Its stored in
5748                  * micro-seconds, so lets convert to msecs.
5749                  */
5750                 slot = tcp_sad_pacing_interval;
5751         }
5752 #endif
5753         if (tp->t_flags & TF_DELACK) {
5754                 delayed_ack = TICKS_2_USEC(tcp_delacktime);
5755                 rack->r_ctl.rc_hpts_flags |= PACE_TMR_DELACK;
5756         }
5757         if (delayed_ack && ((hpts_timeout == 0) ||
5758                             (delayed_ack < hpts_timeout)))
5759                 hpts_timeout = delayed_ack;
5760         else
5761                 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
5762         /*
5763          * If no timers are going to run and we will fall off the hptsi
5764          * wheel, we resort to a keep-alive timer if its configured.
5765          */
5766         if ((hpts_timeout == 0) &&
5767             (slot == 0)) {
5768                 if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
5769                     (tp->t_state <= TCPS_CLOSING)) {
5770                         /*
5771                          * Ok we have no timer (persists, rack, tlp, rxt  or
5772                          * del-ack), we don't have segments being paced. So
5773                          * all that is left is the keepalive timer.
5774                          */
5775                         if (TCPS_HAVEESTABLISHED(tp->t_state)) {
5776                                 /* Get the established keep-alive time */
5777                                 hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp));
5778                         } else {
5779                                 /*
5780                                  * Get the initial setup keep-alive time,
5781                                  * note that this is probably not going to
5782                                  * happen, since rack will be running a rxt timer
5783                                  * if a SYN of some sort is outstanding. It is
5784                                  * actually handled in rack_timeout_rxt().
5785                                  */
5786                                 hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp));
5787                         }
5788                         rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP;
5789                         if (rack->in_probe_rtt) {
5790                                 /*
5791                                  * We want to instead not wake up a long time from
5792                                  * now but to wake up about the time we would
5793                                  * exit probe-rtt and initiate a keep-alive ack.
5794                                  * This will get us out of probe-rtt and update
5795                                  * our min-rtt.
5796                                  */
5797                                 hpts_timeout = rack_min_probertt_hold;
5798                         }
5799                 }
5800         }
5801         if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) ==
5802             (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) {
5803                 /*
5804                  * RACK, TLP, persists and RXT timers all are restartable
5805                  * based on actions input .. i.e we received a packet (ack
5806                  * or sack) and that changes things (rw, or snd_una etc).
5807                  * Thus we can restart them with a new value. For
5808                  * keep-alive, delayed_ack we keep track of what was left
5809                  * and restart the timer with a smaller value.
5810                  */
5811                 if (left < hpts_timeout)
5812                         hpts_timeout = left;
5813         }
5814         if (hpts_timeout) {
5815                 /*
5816                  * Hack alert for now we can't time-out over 2,147,483
5817                  * seconds (a bit more than 596 hours), which is probably ok
5818                  * :).
5819                  */
5820                 if (hpts_timeout > 0x7ffffffe)
5821                         hpts_timeout = 0x7ffffffe;
5822                 rack->r_ctl.rc_timer_exp = cts + hpts_timeout;
5823         }
5824         rack_log_pacing_delay_calc(rack, entry_slot, slot, hpts_timeout, 0, 0, 27, __LINE__, NULL);
5825         if ((rack->gp_ready == 0) &&
5826             (rack->use_fixed_rate == 0) &&
5827             (hpts_timeout < slot) &&
5828             (rack->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) {
5829                 /*
5830                  * We have no good estimate yet for the
5831                  * old clunky burst mitigation or the
5832                  * real pacing. And the tlp or rxt is smaller
5833                  * than the pacing calculation. Lets not
5834                  * pace that long since we know the calculation
5835                  * so far is not accurate.
5836                  */
5837                 slot = hpts_timeout;
5838         }
5839         rack->r_ctl.last_pacing_time = slot;
5840         /**
5841          * Turn off all the flags for queuing by default. The
5842          * flags have important meanings to what happens when
5843          * LRO interacts with the transport. Most likely (by default now)
5844          * mbuf_queueing and ack compression are on. So the transport
5845          * has a couple of flags that control what happens (if those
5846          * are not on then these flags won't have any effect since it
5847          * won't go through the queuing LRO path).
5848          *
5849          * INP_MBUF_QUEUE_READY - This flags says that I am busy
5850          *                        pacing output, so don't disturb. But
5851          *                        it also means LRO can wake me if there
5852          *                        is a SACK arrival.
5853          *
5854          * INP_DONT_SACK_QUEUE - This flag is used in conjunction
5855          *                       with the above flag (QUEUE_READY) and
5856          *                       when present it says don't even wake me
5857          *                       if a SACK arrives.
5858          *
5859          * The idea behind these flags is that if we are pacing we
5860          * set the MBUF_QUEUE_READY and only get woken up if
5861          * a SACK arrives (which could change things) or if
5862          * our pacing timer expires. If, however, we have a rack
5863          * timer running, then we don't even want a sack to wake
5864          * us since the rack timer has to expire before we can send.
5865          *
5866          * Other cases should usually have none of the flags set
5867          * so LRO can call into us.
5868          */
5869         inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY);
5870         if (slot) {
5871                 rack->r_ctl.rc_last_output_to = us_cts + slot;
5872                 /*
5873                  * A pacing timer (slot) is being set, in
5874                  * such a case we cannot send (we are blocked by
5875                  * the timer). So lets tell LRO that it should not
5876                  * wake us unless there is a SACK. Note this only
5877                  * will be effective if mbuf queueing is on or
5878                  * compressed acks are being processed.
5879                  */
5880                 inp->inp_flags2 |= INP_MBUF_QUEUE_READY;
5881                 /*
5882                  * But wait if we have a Rack timer running
5883                  * even a SACK should not disturb us (with
5884                  * the exception of r_rr_config 3).
5885                  */
5886                 if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK) &&
5887                     (rack->r_rr_config != 3))
5888                         inp->inp_flags2 |= INP_DONT_SACK_QUEUE;
5889                 if (rack->rc_ack_can_sendout_data) {
5890                         /*
5891                          * Ahh but wait, this is that special case
5892                          * where the pacing timer can be disturbed
5893                          * backout the changes (used for non-paced
5894                          * burst limiting).
5895                          */
5896                         inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY);
5897                 }
5898                 if ((rack->use_rack_rr) &&
5899                     (rack->r_rr_config < 2) &&
5900                     ((hpts_timeout) && (hpts_timeout < slot))) {
5901                         /*
5902                          * Arrange for the hpts to kick back in after the
5903                          * t-o if the t-o does not cause a send.
5904                          */
5905                         (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(hpts_timeout),
5906                                                    __LINE__, &diag);
5907                         rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5908                         rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
5909                 } else {
5910                         (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(slot),
5911                                                    __LINE__, &diag);
5912                         rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5913                         rack_log_to_start(rack, cts, hpts_timeout, slot, 1);
5914                 }
5915         } else if (hpts_timeout) {
5916                 /*
5917                  * With respect to inp_flags2 here, lets let any new acks wake
5918                  * us up here. Since we are not pacing (no pacing timer), output
5919                  * can happen so we should let it. If its a Rack timer, then any inbound
5920                  * packet probably won't change the sending (we will be blocked)
5921                  * but it may change the prr stats so letting it in (the set defaults
5922                  * at the start of this block) are good enough.
5923                  */
5924                 (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(hpts_timeout),
5925                                            __LINE__, &diag);
5926                 rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5927                 rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
5928         } else {
5929                 /* No timer starting */
5930 #ifdef INVARIANTS
5931                 if (SEQ_GT(tp->snd_max, tp->snd_una)) {
5932                         panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?",
5933                             tp, rack, tot_len_this_send, cts, slot, hpts_timeout);
5934                 }
5935 #endif
5936         }
5937         rack->rc_tmr_stopped = 0;
5938         if (slot)
5939                 rack_log_type_bbrsnd(rack, tot_len_this_send, slot, us_cts, &tv);
5940 }
5941
5942 /*
5943  * RACK Timer, here we simply do logging and house keeping.
5944  * the normal rack_output() function will call the
5945  * appropriate thing to check if we need to do a RACK retransmit.
5946  * We return 1, saying don't proceed with rack_output only
5947  * when all timers have been stopped (destroyed PCB?).
5948  */
5949 static int
5950 rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5951 {
5952         /*
5953          * This timer simply provides an internal trigger to send out data.
5954          * The check_recovery_mode call will see if there are needed
5955          * retransmissions, if so we will enter fast-recovery. The output
5956          * call may or may not do the same thing depending on sysctl
5957          * settings.
5958          */
5959         struct rack_sendmap *rsm;
5960
5961         if (tp->t_timers->tt_flags & TT_STOPPED) {
5962                 return (1);
5963         }
5964         counter_u64_add(rack_to_tot, 1);
5965         if (rack->r_state && (rack->r_state != tp->t_state))
5966                 rack_set_state(tp, rack);
5967         rack->rc_on_min_to = 0;
5968         rsm = rack_check_recovery_mode(tp, cts);
5969         rack_log_to_event(rack, RACK_TO_FRM_RACK, rsm);
5970         if (rsm) {
5971                 rack->r_ctl.rc_resend = rsm;
5972                 rack->r_timer_override = 1;
5973                 if (rack->use_rack_rr) {
5974                         /*
5975                          * Don't accumulate extra pacing delay
5976                          * we are allowing the rack timer to
5977                          * over-ride pacing i.e. rrr takes precedence
5978                          * if the pacing interval is longer than the rrr
5979                          * time (in other words we get the min pacing
5980                          * time versus rrr pacing time).
5981                          */
5982                         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
5983                 }
5984         }
5985         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK;
5986         if (rsm == NULL) {
5987                 /* restart a timer and return 1 */
5988                 rack_start_hpts_timer(rack, tp, cts,
5989                                       0, 0, 0);
5990                 return (1);
5991         }
5992         return (0);
5993 }
5994
5995 static void
5996 rack_adjust_orig_mlen(struct rack_sendmap *rsm)
5997 {
5998         if (rsm->m->m_len > rsm->orig_m_len) {
5999                 /*
6000                  * Mbuf grew, caused by sbcompress, our offset does
6001                  * not change.
6002                  */
6003                 rsm->orig_m_len = rsm->m->m_len;
6004         } else if (rsm->m->m_len < rsm->orig_m_len) {
6005                 /*
6006                  * Mbuf shrank, trimmed off the top by an ack, our
6007                  * offset changes.
6008                  */
6009                 rsm->soff -= (rsm->orig_m_len - rsm->m->m_len);
6010                 rsm->orig_m_len = rsm->m->m_len;
6011         }
6012 }
6013
6014 static void
6015 rack_setup_offset_for_rsm(struct rack_sendmap *src_rsm, struct rack_sendmap *rsm)
6016 {
6017         struct mbuf *m;
6018         uint32_t soff;
6019
6020         if (src_rsm->orig_m_len != src_rsm->m->m_len) {
6021                 /* Fix up the orig_m_len and possibly the mbuf offset */
6022                 rack_adjust_orig_mlen(src_rsm);
6023         }
6024         m = src_rsm->m;
6025         soff = src_rsm->soff + (src_rsm->r_end - src_rsm->r_start);
6026         while (soff >= m->m_len) {
6027                 /* Move out past this mbuf */
6028                 soff -= m->m_len;
6029                 m = m->m_next;
6030                 KASSERT((m != NULL),
6031                         ("rsm:%p nrsm:%p hit at soff:%u null m",
6032                          src_rsm, rsm, soff));
6033         }
6034         rsm->m = m;
6035         rsm->soff = soff;
6036         rsm->orig_m_len = m->m_len;
6037 }
6038
6039 static __inline void
6040 rack_clone_rsm(struct tcp_rack *rack, struct rack_sendmap *nrsm,
6041                struct rack_sendmap *rsm, uint32_t start)
6042 {
6043         int idx;
6044
6045         nrsm->r_start = start;
6046         nrsm->r_end = rsm->r_end;
6047         nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
6048         nrsm->r_flags = rsm->r_flags;
6049         nrsm->r_dupack = rsm->r_dupack;
6050         nrsm->r_no_rtt_allowed = rsm->r_no_rtt_allowed;
6051         nrsm->r_rtr_bytes = 0;
6052         rsm->r_end = nrsm->r_start;
6053         nrsm->r_just_ret = rsm->r_just_ret;
6054         for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
6055                 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
6056         }
6057         /*
6058          * Now we need to find nrsm's new location in the mbuf chain
6059          * we basically calculate a new offset, which is soff +
6060          * how much is left in original rsm. Then we walk out the mbuf
6061          * chain to find the righ postion, it may be the same mbuf
6062          * or maybe not.
6063          */
6064         KASSERT((rsm->m != NULL),
6065                 ("rsm:%p nrsm:%p rack:%p -- rsm->m is NULL?", rsm, nrsm, rack));
6066         rack_setup_offset_for_rsm(rsm, nrsm);
6067 }
6068
6069 static struct rack_sendmap *
6070 rack_merge_rsm(struct tcp_rack *rack,
6071                struct rack_sendmap *l_rsm,
6072                struct rack_sendmap *r_rsm)
6073 {
6074         /*
6075          * We are merging two ack'd RSM's,
6076          * the l_rsm is on the left (lower seq
6077          * values) and the r_rsm is on the right
6078          * (higher seq value). The simplest way
6079          * to merge these is to move the right
6080          * one into the left. I don't think there
6081          * is any reason we need to try to find
6082          * the oldest (or last oldest retransmitted).
6083          */
6084         struct rack_sendmap *rm;
6085
6086         rack_log_map_chg(rack->rc_tp, rack, NULL,
6087                          l_rsm, r_rsm, MAP_MERGE, r_rsm->r_end, __LINE__);
6088         l_rsm->r_end = r_rsm->r_end;
6089         if (l_rsm->r_dupack < r_rsm->r_dupack)
6090                 l_rsm->r_dupack = r_rsm->r_dupack;
6091         if (r_rsm->r_rtr_bytes)
6092                 l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes;
6093         if (r_rsm->r_in_tmap) {
6094                 /* This really should not happen */
6095                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, r_rsm, r_tnext);
6096                 r_rsm->r_in_tmap = 0;
6097         }
6098
6099         /* Now the flags */
6100         if (r_rsm->r_flags & RACK_HAS_FIN)
6101                 l_rsm->r_flags |= RACK_HAS_FIN;
6102         if (r_rsm->r_flags & RACK_TLP)
6103                 l_rsm->r_flags |= RACK_TLP;
6104         if (r_rsm->r_flags & RACK_RWND_COLLAPSED)
6105                 l_rsm->r_flags |= RACK_RWND_COLLAPSED;
6106         if ((r_rsm->r_flags & RACK_APP_LIMITED)  &&
6107             ((l_rsm->r_flags & RACK_APP_LIMITED) == 0)) {
6108                 /*
6109                  * If both are app-limited then let the
6110                  * free lower the count. If right is app
6111                  * limited and left is not, transfer.
6112                  */
6113                 l_rsm->r_flags |= RACK_APP_LIMITED;
6114                 r_rsm->r_flags &= ~RACK_APP_LIMITED;
6115                 if (r_rsm == rack->r_ctl.rc_first_appl)
6116                         rack->r_ctl.rc_first_appl = l_rsm;
6117         }
6118         rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, r_rsm);
6119 #ifdef INVARIANTS
6120         if (rm != r_rsm) {
6121                 panic("removing head in rack:%p rsm:%p rm:%p",
6122                       rack, r_rsm, rm);
6123         }
6124 #endif
6125         if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) {
6126                 /* Transfer the split limit to the map we free */
6127                 r_rsm->r_limit_type = l_rsm->r_limit_type;
6128                 l_rsm->r_limit_type = 0;
6129         }
6130         rack_free(rack, r_rsm);
6131         return (l_rsm);
6132 }
6133
6134 /*
6135  * TLP Timer, here we simply setup what segment we want to
6136  * have the TLP expire on, the normal rack_output() will then
6137  * send it out.
6138  *
6139  * We return 1, saying don't proceed with rack_output only
6140  * when all timers have been stopped (destroyed PCB?).
6141  */
6142 static int
6143 rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6144 {
6145         /*
6146          * Tail Loss Probe.
6147          */
6148         struct rack_sendmap *rsm = NULL;
6149         struct rack_sendmap *insret;
6150         struct socket *so;
6151         uint32_t amm;
6152         uint32_t out, avail;
6153         int collapsed_win = 0;
6154
6155         if (tp->t_timers->tt_flags & TT_STOPPED) {
6156                 return (1);
6157         }
6158         if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
6159                 /* Its not time yet */
6160                 return (0);
6161         }
6162         if (ctf_progress_timeout_check(tp, true)) {
6163                 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6164                 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
6165                 return (1);
6166         }
6167         /*
6168          * A TLP timer has expired. We have been idle for 2 rtts. So we now
6169          * need to figure out how to force a full MSS segment out.
6170          */
6171         rack_log_to_event(rack, RACK_TO_FRM_TLP, NULL);
6172         rack->r_ctl.retran_during_recovery = 0;
6173         rack->r_ctl.dsack_byte_cnt = 0;
6174         counter_u64_add(rack_tlp_tot, 1);
6175         if (rack->r_state && (rack->r_state != tp->t_state))
6176                 rack_set_state(tp, rack);
6177         so = tp->t_inpcb->inp_socket;
6178         avail = sbavail(&so->so_snd);
6179         out = tp->snd_max - tp->snd_una;
6180         if (out > tp->snd_wnd) {
6181                 /* special case, we need a retransmission */
6182                 collapsed_win = 1;
6183                 goto need_retran;
6184         }
6185         /*
6186          * Check our send oldest always settings, and if
6187          * there is an oldest to send jump to the need_retran.
6188          */
6189         if (rack_always_send_oldest && (TAILQ_EMPTY(&rack->r_ctl.rc_tmap) == 0))
6190                 goto need_retran;
6191
6192         if (avail > out) {
6193                 /* New data is available */
6194                 amm = avail - out;
6195                 if (amm > ctf_fixed_maxseg(tp)) {
6196                         amm = ctf_fixed_maxseg(tp);
6197                         if ((amm + out) > tp->snd_wnd) {
6198                                 /* We are rwnd limited */
6199                                 goto need_retran;
6200                         }
6201                 } else if (amm < ctf_fixed_maxseg(tp)) {
6202                         /* not enough to fill a MTU */
6203                         goto need_retran;
6204                 }
6205                 if (IN_FASTRECOVERY(tp->t_flags)) {
6206                         /* Unlikely */
6207                         if (rack->rack_no_prr == 0) {
6208                                 if (out + amm <= tp->snd_wnd) {
6209                                         rack->r_ctl.rc_prr_sndcnt = amm;
6210                                         rack_log_to_prr(rack, 4, 0);
6211                                 }
6212                         } else
6213                                 goto need_retran;
6214                 } else {
6215                         /* Set the send-new override */
6216                         if (out + amm <= tp->snd_wnd)
6217                                 rack->r_ctl.rc_tlp_new_data = amm;
6218                         else
6219                                 goto need_retran;
6220                 }
6221                 rack->r_ctl.rc_tlpsend = NULL;
6222                 counter_u64_add(rack_tlp_newdata, 1);
6223                 goto send;
6224         }
6225 need_retran:
6226         /*
6227          * Ok we need to arrange the last un-acked segment to be re-sent, or
6228          * optionally the first un-acked segment.
6229          */
6230         if (collapsed_win == 0) {
6231                 if (rack_always_send_oldest)
6232                         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6233                 else {
6234                         rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6235                         if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) {
6236                                 rsm = rack_find_high_nonack(rack, rsm);
6237                         }
6238                 }
6239                 if (rsm == NULL) {
6240                         counter_u64_add(rack_tlp_does_nada, 1);
6241 #ifdef TCP_BLACKBOX
6242                         tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
6243 #endif
6244                         goto out;
6245                 }
6246         } else {
6247                 /*
6248                  * We must find the last segment
6249                  * that was acceptable by the client.
6250                  */
6251                 RB_FOREACH_REVERSE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
6252                         if ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0) {
6253                                 /* Found one */
6254                                 break;
6255                         }
6256                 }
6257                 if (rsm == NULL) {
6258                         /* None? if so send the first */
6259                         rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6260                         if (rsm == NULL) {
6261                                 counter_u64_add(rack_tlp_does_nada, 1);
6262 #ifdef TCP_BLACKBOX
6263                                 tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
6264 #endif
6265                                 goto out;
6266                         }
6267                 }
6268         }
6269         if ((rsm->r_end - rsm->r_start) > ctf_fixed_maxseg(tp)) {
6270                 /*
6271                  * We need to split this the last segment in two.
6272                  */
6273                 struct rack_sendmap *nrsm;
6274
6275                 nrsm = rack_alloc_full_limit(rack);
6276                 if (nrsm == NULL) {
6277                         /*
6278                          * No memory to split, we will just exit and punt
6279                          * off to the RXT timer.
6280                          */
6281                         counter_u64_add(rack_tlp_does_nada, 1);
6282                         goto out;
6283                 }
6284                 rack_clone_rsm(rack, nrsm, rsm,
6285                                (rsm->r_end - ctf_fixed_maxseg(tp)));
6286                 rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
6287                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
6288 #ifdef INVARIANTS
6289                 if (insret != NULL) {
6290                         panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
6291                               nrsm, insret, rack, rsm);
6292                 }
6293 #endif
6294                 if (rsm->r_in_tmap) {
6295                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
6296                         nrsm->r_in_tmap = 1;
6297                 }
6298                 rsm->r_flags &= (~RACK_HAS_FIN);
6299                 rsm = nrsm;
6300         }
6301         rack->r_ctl.rc_tlpsend = rsm;
6302 send:
6303         rack->r_timer_override = 1;
6304         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
6305         return (0);
6306 out:
6307         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
6308         return (0);
6309 }
6310
6311 /*
6312  * Delayed ack Timer, here we simply need to setup the
6313  * ACK_NOW flag and remove the DELACK flag. From there
6314  * the output routine will send the ack out.
6315  *
6316  * We only return 1, saying don't proceed, if all timers
6317  * are stopped (destroyed PCB?).
6318  */
6319 static int
6320 rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6321 {
6322         if (tp->t_timers->tt_flags & TT_STOPPED) {
6323                 return (1);
6324         }
6325         rack_log_to_event(rack, RACK_TO_FRM_DELACK, NULL);
6326         tp->t_flags &= ~TF_DELACK;
6327         tp->t_flags |= TF_ACKNOW;
6328         KMOD_TCPSTAT_INC(tcps_delack);
6329         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
6330         return (0);
6331 }
6332
6333 /*
6334  * Persists timer, here we simply send the
6335  * same thing as a keepalive will.
6336  * the one byte send.
6337  *
6338  * We only return 1, saying don't proceed, if all timers
6339  * are stopped (destroyed PCB?).
6340  */
6341 static int
6342 rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6343 {
6344         struct tcptemp *t_template;
6345         struct inpcb *inp;
6346         int32_t retval = 1;
6347
6348         inp = tp->t_inpcb;
6349
6350         if (tp->t_timers->tt_flags & TT_STOPPED) {
6351                 return (1);
6352         }
6353         if (rack->rc_in_persist == 0)
6354                 return (0);
6355         if (ctf_progress_timeout_check(tp, false)) {
6356                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6357                 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6358                 tcp_set_inp_to_drop(inp, ETIMEDOUT);
6359                 return (1);
6360         }
6361         KASSERT(inp != NULL, ("%s: tp %p tp->t_inpcb == NULL", __func__, tp));
6362         /*
6363          * Persistence timer into zero window. Force a byte to be output, if
6364          * possible.
6365          */
6366         KMOD_TCPSTAT_INC(tcps_persisttimeo);
6367         /*
6368          * Hack: if the peer is dead/unreachable, we do not time out if the
6369          * window is closed.  After a full backoff, drop the connection if
6370          * the idle time (no responses to probes) reaches the maximum
6371          * backoff that we would use if retransmitting.
6372          */
6373         if (tp->t_rxtshift == TCP_MAXRXTSHIFT &&
6374             (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
6375              TICKS_2_USEC(ticks - tp->t_rcvtime) >= RACK_REXMTVAL(tp) * tcp_totbackoff)) {
6376                 KMOD_TCPSTAT_INC(tcps_persistdrop);
6377                 retval = 1;
6378                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6379                 tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT);
6380                 goto out;
6381         }
6382         if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) &&
6383             tp->snd_una == tp->snd_max)
6384                 rack_exit_persist(tp, rack, cts);
6385         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT;
6386         /*
6387          * If the user has closed the socket then drop a persisting
6388          * connection after a much reduced timeout.
6389          */
6390         if (tp->t_state > TCPS_CLOSE_WAIT &&
6391             (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
6392                 retval = 1;
6393                 KMOD_TCPSTAT_INC(tcps_persistdrop);
6394                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6395                 tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT);
6396                 goto out;
6397         }
6398         t_template = tcpip_maketemplate(rack->rc_inp);
6399         if (t_template) {
6400                 /* only set it if we were answered */
6401                 if (rack->forced_ack == 0) {
6402                         rack->forced_ack = 1;
6403                         rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
6404                 }
6405                 tcp_respond(tp, t_template->tt_ipgen,
6406                             &t_template->tt_t, (struct mbuf *)NULL,
6407                             tp->rcv_nxt, tp->snd_una - 1, 0);
6408                 /* This sends an ack */
6409                 if (tp->t_flags & TF_DELACK)
6410                         tp->t_flags &= ~TF_DELACK;
6411                 free(t_template, M_TEMP);
6412         }
6413         if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
6414                 tp->t_rxtshift++;
6415 out:
6416         rack_log_to_event(rack, RACK_TO_FRM_PERSIST, NULL);
6417         rack_start_hpts_timer(rack, tp, cts,
6418                               0, 0, 0);
6419         return (retval);
6420 }
6421
6422 /*
6423  * If a keepalive goes off, we had no other timers
6424  * happening. We always return 1 here since this
6425  * routine either drops the connection or sends
6426  * out a segment with respond.
6427  */
6428 static int
6429 rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6430 {
6431         struct tcptemp *t_template;
6432         struct inpcb *inp;
6433
6434         if (tp->t_timers->tt_flags & TT_STOPPED) {
6435                 return (1);
6436         }
6437         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP;
6438         inp = tp->t_inpcb;
6439         rack_log_to_event(rack, RACK_TO_FRM_KEEP, NULL);
6440         /*
6441          * Keep-alive timer went off; send something or drop connection if
6442          * idle for too long.
6443          */
6444         KMOD_TCPSTAT_INC(tcps_keeptimeo);
6445         if (tp->t_state < TCPS_ESTABLISHED)
6446                 goto dropit;
6447         if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
6448             tp->t_state <= TCPS_CLOSING) {
6449                 if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
6450                         goto dropit;
6451                 /*
6452                  * Send a packet designed to force a response if the peer is
6453                  * up and reachable: either an ACK if the connection is
6454                  * still alive, or an RST if the peer has closed the
6455                  * connection due to timeout or reboot. Using sequence
6456                  * number tp->snd_una-1 causes the transmitted zero-length
6457                  * segment to lie outside the receive window; by the
6458                  * protocol spec, this requires the correspondent TCP to
6459                  * respond.
6460                  */
6461                 KMOD_TCPSTAT_INC(tcps_keepprobe);
6462                 t_template = tcpip_maketemplate(inp);
6463                 if (t_template) {
6464                         if (rack->forced_ack == 0) {
6465                                 rack->forced_ack = 1;
6466                                 rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
6467                         }
6468                         tcp_respond(tp, t_template->tt_ipgen,
6469                             &t_template->tt_t, (struct mbuf *)NULL,
6470                             tp->rcv_nxt, tp->snd_una - 1, 0);
6471                         free(t_template, M_TEMP);
6472                 }
6473         }
6474         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
6475         return (1);
6476 dropit:
6477         KMOD_TCPSTAT_INC(tcps_keepdrops);
6478         tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
6479         tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT);
6480         return (1);
6481 }
6482
6483 /*
6484  * Retransmit helper function, clear up all the ack
6485  * flags and take care of important book keeping.
6486  */
6487 static void
6488 rack_remxt_tmr(struct tcpcb *tp)
6489 {
6490         /*
6491          * The retransmit timer went off, all sack'd blocks must be
6492          * un-acked.
6493          */
6494         struct rack_sendmap *rsm, *trsm = NULL;
6495         struct tcp_rack *rack;
6496
6497         rack = (struct tcp_rack *)tp->t_fb_ptr;
6498         rack_timer_cancel(tp, rack, tcp_get_usecs(NULL), __LINE__);
6499         rack_log_to_event(rack, RACK_TO_FRM_TMR, NULL);
6500         if (rack->r_state && (rack->r_state != tp->t_state))
6501                 rack_set_state(tp, rack);
6502         /*
6503          * Ideally we would like to be able to
6504          * mark SACK-PASS on anything not acked here.
6505          *
6506          * However, if we do that we would burst out
6507          * all that data 1ms apart. This would be unwise,
6508          * so for now we will just let the normal rxt timer
6509          * and tlp timer take care of it.
6510          *
6511          * Also we really need to stick them back in sequence
6512          * order. This way we send in the proper order and any
6513          * sacks that come floating in will "re-ack" the data.
6514          * To do this we zap the tmap with an INIT and then
6515          * walk through and place every rsm in the RB tree
6516          * back in its seq ordered place.
6517          */
6518         TAILQ_INIT(&rack->r_ctl.rc_tmap);
6519         RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
6520                 rsm->r_dupack = 0;
6521                 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
6522                 /* We must re-add it back to the tlist */
6523                 if (trsm == NULL) {
6524                         TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
6525                 } else {
6526                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext);
6527                 }
6528                 rsm->r_in_tmap = 1;
6529                 trsm = rsm;
6530                 if (rsm->r_flags & RACK_ACKED)
6531                         rsm->r_flags |= RACK_WAS_ACKED;
6532                 rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS);
6533         }
6534         /* Clear the count (we just un-acked them) */
6535         rack->r_ctl.rc_last_timeout_snduna = tp->snd_una;
6536         rack->r_ctl.rc_sacked = 0;
6537         rack->r_ctl.rc_sacklast = NULL;
6538         rack->r_ctl.rc_agg_delayed = 0;
6539         rack->r_early = 0;
6540         rack->r_ctl.rc_agg_early = 0;
6541         rack->r_late = 0;
6542         /* Clear the tlp rtx mark */
6543         rack->r_ctl.rc_resend = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6544         if (rack->r_ctl.rc_resend != NULL)
6545                 rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
6546         rack->r_ctl.rc_prr_sndcnt = 0;
6547         rack_log_to_prr(rack, 6, 0);
6548         rack->r_timer_override = 1;
6549         if ((((tp->t_flags & TF_SACK_PERMIT) == 0)
6550 #ifdef NETFLIX_EXP_DETECTION
6551             || (rack->sack_attack_disable != 0)
6552 #endif
6553                     ) && ((tp->t_flags & TF_SENTFIN) == 0)) {
6554                 /*
6555                  * For non-sack customers new data
6556                  * needs to go out as retransmits until
6557                  * we retransmit up to snd_max.
6558                  */
6559                 rack->r_must_retran = 1;
6560                 rack->r_ctl.rc_out_at_rto = ctf_flight_size(rack->rc_tp,
6561                                                 rack->r_ctl.rc_sacked);
6562         }
6563         rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
6564 }
6565
6566 static void
6567 rack_convert_rtts(struct tcpcb *tp)
6568 {
6569         if (tp->t_srtt > 1) {
6570                 uint32_t val, frac;
6571
6572                 val = tp->t_srtt >> TCP_RTT_SHIFT;
6573                 frac = tp->t_srtt & 0x1f;
6574                 tp->t_srtt = TICKS_2_USEC(val);
6575                 /*
6576                  * frac is the fractional part of the srtt (if any)
6577                  * but its in ticks and every bit represents
6578                  * 1/32nd of a hz.
6579                  */
6580                 if (frac) {
6581                         if (hz == 1000) {
6582                                 frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_MSEC) / (uint64_t)TCP_RTT_SCALE);
6583                         } else {
6584                                 frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_SEC) / ((uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE));
6585                         }
6586                         tp->t_srtt += frac;
6587                 }
6588         }
6589         if (tp->t_rttvar) {
6590                 uint32_t val, frac;
6591
6592                 val = tp->t_rttvar >> TCP_RTTVAR_SHIFT;
6593                 frac = tp->t_rttvar & 0x1f;
6594                 tp->t_rttvar = TICKS_2_USEC(val);
6595                 /*
6596                  * frac is the fractional part of the srtt (if any)
6597                  * but its in ticks and every bit represents
6598                  * 1/32nd of a hz.
6599                  */
6600                 if (frac) {
6601                         if (hz == 1000) {
6602                                 frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_MSEC) / (uint64_t)TCP_RTT_SCALE);
6603                         } else {
6604                                 frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_SEC) / ((uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE));
6605                         }
6606                         tp->t_rttvar += frac;
6607                 }
6608         }
6609         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
6610                            rack_rto_min, rack_rto_max);
6611 }
6612
6613 static void
6614 rack_cc_conn_init(struct tcpcb *tp)
6615 {
6616         struct tcp_rack *rack;
6617         uint32_t srtt;
6618
6619         rack = (struct tcp_rack *)tp->t_fb_ptr;
6620         srtt = tp->t_srtt;
6621         cc_conn_init(tp);
6622         /*
6623          * Now convert to rack's internal format,
6624          * if required.
6625          */
6626         if ((srtt == 0) && (tp->t_srtt != 0))
6627                 rack_convert_rtts(tp);
6628         /*
6629          * We want a chance to stay in slowstart as
6630          * we create a connection. TCP spec says that
6631          * initially ssthresh is infinite. For our
6632          * purposes that is the snd_wnd.
6633          */
6634         if (tp->snd_ssthresh < tp->snd_wnd) {
6635                 tp->snd_ssthresh = tp->snd_wnd;
6636         }
6637         /*
6638          * We also want to assure a IW worth of
6639          * data can get inflight.
6640          */
6641         if (rc_init_window(rack) < tp->snd_cwnd)
6642                 tp->snd_cwnd = rc_init_window(rack);
6643 }
6644
6645 /*
6646  * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
6647  * we will setup to retransmit the lowest seq number outstanding.
6648  */
6649 static int
6650 rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6651 {
6652         int32_t rexmt;
6653         struct inpcb *inp;
6654         int32_t retval = 0;
6655         bool isipv6;
6656
6657         inp = tp->t_inpcb;
6658         if (tp->t_timers->tt_flags & TT_STOPPED) {
6659                 return (1);
6660         }
6661         if (ctf_progress_timeout_check(tp, false)) {
6662                 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
6663                 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6664                 tcp_set_inp_to_drop(inp, ETIMEDOUT);
6665                 return (1);
6666         }
6667         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
6668         rack->r_ctl.retran_during_recovery = 0;
6669         rack->r_ctl.dsack_byte_cnt = 0;
6670         if (IN_FASTRECOVERY(tp->t_flags))
6671                 tp->t_flags |= TF_WASFRECOVERY;
6672         else
6673                 tp->t_flags &= ~TF_WASFRECOVERY;
6674         if (IN_CONGRECOVERY(tp->t_flags))
6675                 tp->t_flags |= TF_WASCRECOVERY;
6676         else
6677                 tp->t_flags &= ~TF_WASCRECOVERY;
6678         if (TCPS_HAVEESTABLISHED(tp->t_state) &&
6679             (tp->snd_una == tp->snd_max)) {
6680                 /* Nothing outstanding .. nothing to do */
6681                 return (0);
6682         }
6683         /*
6684          * Rack can only run one timer  at a time, so we cannot
6685          * run a KEEPINIT (gating SYN sending) and a retransmit
6686          * timer for the SYN. So if we are in a front state and
6687          * have a KEEPINIT timer we need to check the first transmit
6688          * against now to see if we have exceeded the KEEPINIT time
6689          * (if one is set).
6690          */
6691         if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
6692             (TP_KEEPINIT(tp) != 0)) {
6693                 struct rack_sendmap *rsm;
6694
6695                 rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6696                 if (rsm) {
6697                         /* Ok we have something outstanding to test keepinit with */
6698                         if ((TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) &&
6699                             ((cts - (uint32_t)rsm->r_tim_lastsent[0]) >= TICKS_2_USEC(TP_KEEPINIT(tp)))) {
6700                                 /* We have exceeded the KEEPINIT time */
6701                                 tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
6702                                 goto drop_it;
6703                         }
6704                 }
6705         }
6706         /*
6707          * Retransmission timer went off.  Message has not been acked within
6708          * retransmit interval.  Back off to a longer retransmit interval
6709          * and retransmit one segment.
6710          */
6711         rack_remxt_tmr(tp);
6712         if ((rack->r_ctl.rc_resend == NULL) ||
6713             ((rack->r_ctl.rc_resend->r_flags & RACK_RWND_COLLAPSED) == 0)) {
6714                 /*
6715                  * If the rwnd collapsed on
6716                  * the one we are retransmitting
6717                  * it does not count against the
6718                  * rxt count.
6719                  */
6720                 tp->t_rxtshift++;
6721         }
6722         if (tp->t_rxtshift > TCP_MAXRXTSHIFT) {
6723                 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
6724 drop_it:
6725                 tp->t_rxtshift = TCP_MAXRXTSHIFT;
6726                 KMOD_TCPSTAT_INC(tcps_timeoutdrop);
6727                 retval = 1;
6728                 tcp_set_inp_to_drop(rack->rc_inp,
6729                     (tp->t_softerror ? (uint16_t) tp->t_softerror : ETIMEDOUT));
6730                 goto out;
6731         }
6732         if (tp->t_state == TCPS_SYN_SENT) {
6733                 /*
6734                  * If the SYN was retransmitted, indicate CWND to be limited
6735                  * to 1 segment in cc_conn_init().
6736                  */
6737                 tp->snd_cwnd = 1;
6738         } else if (tp->t_rxtshift == 1) {
6739                 /*
6740                  * first retransmit; record ssthresh and cwnd so they can be
6741                  * recovered if this turns out to be a "bad" retransmit. A
6742                  * retransmit is considered "bad" if an ACK for this segment
6743                  * is received within RTT/2 interval; the assumption here is
6744                  * that the ACK was already in flight.  See "On Estimating
6745                  * End-to-End Network Path Properties" by Allman and Paxson
6746                  * for more details.
6747                  */
6748                 tp->snd_cwnd_prev = tp->snd_cwnd;
6749                 tp->snd_ssthresh_prev = tp->snd_ssthresh;
6750                 tp->snd_recover_prev = tp->snd_recover;
6751                 tp->t_badrxtwin = ticks + (USEC_2_TICKS(tp->t_srtt)/2);
6752                 tp->t_flags |= TF_PREVVALID;
6753         } else if ((tp->t_flags & TF_RCVD_TSTMP) == 0)
6754                 tp->t_flags &= ~TF_PREVVALID;
6755         KMOD_TCPSTAT_INC(tcps_rexmttimeo);
6756         if ((tp->t_state == TCPS_SYN_SENT) ||
6757             (tp->t_state == TCPS_SYN_RECEIVED))
6758                 rexmt = RACK_INITIAL_RTO * tcp_backoff[tp->t_rxtshift];
6759         else
6760                 rexmt = max(rack_rto_min, (tp->t_srtt + (tp->t_rttvar << 2))) * tcp_backoff[tp->t_rxtshift];
6761
6762         RACK_TCPT_RANGESET(tp->t_rxtcur, rexmt,
6763            max(rack_rto_min, rexmt), rack_rto_max);
6764         /*
6765          * We enter the path for PLMTUD if connection is established or, if
6766          * connection is FIN_WAIT_1 status, reason for the last is that if
6767          * amount of data we send is very small, we could send it in couple
6768          * of packets and process straight to FIN. In that case we won't
6769          * catch ESTABLISHED state.
6770          */
6771 #ifdef INET6
6772         isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) ? true : false;
6773 #else
6774         isipv6 = false;
6775 #endif
6776         if (((V_tcp_pmtud_blackhole_detect == 1) ||
6777             (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
6778             (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
6779             ((tp->t_state == TCPS_ESTABLISHED) ||
6780             (tp->t_state == TCPS_FIN_WAIT_1))) {
6781                 /*
6782                  * Idea here is that at each stage of mtu probe (usually,
6783                  * 1448 -> 1188 -> 524) should be given 2 chances to recover
6784                  * before further clamping down. 'tp->t_rxtshift % 2 == 0'
6785                  * should take care of that.
6786                  */
6787                 if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
6788                     (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
6789                     (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
6790                     tp->t_rxtshift % 2 == 0)) {
6791                         /*
6792                          * Enter Path MTU Black-hole Detection mechanism: -
6793                          * Disable Path MTU Discovery (IP "DF" bit). -
6794                          * Reduce MTU to lower value than what we negotiated
6795                          * with peer.
6796                          */
6797                         if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
6798                                 /* Record that we may have found a black hole. */
6799                                 tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
6800                                 /* Keep track of previous MSS. */
6801                                 tp->t_pmtud_saved_maxseg = tp->t_maxseg;
6802                         }
6803
6804                         /*
6805                          * Reduce the MSS to blackhole value or to the
6806                          * default in an attempt to retransmit.
6807                          */
6808 #ifdef INET6
6809                         if (isipv6 &&
6810                             tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
6811                                 /* Use the sysctl tuneable blackhole MSS. */
6812                                 tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
6813                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
6814                         } else if (isipv6) {
6815                                 /* Use the default MSS. */
6816                                 tp->t_maxseg = V_tcp_v6mssdflt;
6817                                 /*
6818                                  * Disable Path MTU Discovery when we switch
6819                                  * to minmss.
6820                                  */
6821                                 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
6822                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
6823                         }
6824 #endif
6825 #if defined(INET6) && defined(INET)
6826                         else
6827 #endif
6828 #ifdef INET
6829                         if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
6830                                 /* Use the sysctl tuneable blackhole MSS. */
6831                                 tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
6832                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
6833                         } else {
6834                                 /* Use the default MSS. */
6835                                 tp->t_maxseg = V_tcp_mssdflt;
6836                                 /*
6837                                  * Disable Path MTU Discovery when we switch
6838                                  * to minmss.
6839                                  */
6840                                 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
6841                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
6842                         }
6843 #endif
6844                 } else {
6845                         /*
6846                          * If further retransmissions are still unsuccessful
6847                          * with a lowered MTU, maybe this isn't a blackhole
6848                          * and we restore the previous MSS and blackhole
6849                          * detection flags. The limit '6' is determined by
6850                          * giving each probe stage (1448, 1188, 524) 2
6851                          * chances to recover.
6852                          */
6853                         if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
6854                             (tp->t_rxtshift >= 6)) {
6855                                 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
6856                                 tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
6857                                 tp->t_maxseg = tp->t_pmtud_saved_maxseg;
6858                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed);
6859                         }
6860                 }
6861         }
6862         /*
6863          * Disable RFC1323 and SACK if we haven't got any response to
6864          * our third SYN to work-around some broken terminal servers
6865          * (most of which have hopefully been retired) that have bad VJ
6866          * header compression code which trashes TCP segments containing
6867          * unknown-to-them TCP options.
6868          */
6869         if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
6870             (tp->t_rxtshift == 3))
6871                 tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP|TF_SACK_PERMIT);
6872         /*
6873          * If we backed off this far, our srtt estimate is probably bogus.
6874          * Clobber it so we'll take the next rtt measurement as our srtt;
6875          * move the current srtt into rttvar to keep the current retransmit
6876          * times until then.
6877          */
6878         if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
6879 #ifdef INET6
6880                 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0)
6881                         in6_losing(tp->t_inpcb);
6882                 else
6883 #endif
6884                         in_losing(tp->t_inpcb);
6885                 tp->t_rttvar += tp->t_srtt;
6886                 tp->t_srtt = 0;
6887         }
6888         sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
6889         tp->snd_recover = tp->snd_max;
6890         tp->t_flags |= TF_ACKNOW;
6891         tp->t_rtttime = 0;
6892         rack_cong_signal(tp, CC_RTO, tp->snd_una);
6893 out:
6894         return (retval);
6895 }
6896
6897 static int
6898 rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling)
6899 {
6900         int32_t ret = 0;
6901         int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
6902
6903         if (timers == 0) {
6904                 return (0);
6905         }
6906         if (tp->t_state == TCPS_LISTEN) {
6907                 /* no timers on listen sockets */
6908                 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
6909                         return (0);
6910                 return (1);
6911         }
6912         if ((timers & PACE_TMR_RACK) &&
6913             rack->rc_on_min_to) {
6914                 /*
6915                  * For the rack timer when we
6916                  * are on a min-timeout (which means rrr_conf = 3)
6917                  * we don't want to check the timer. It may
6918                  * be going off for a pace and thats ok we
6919                  * want to send the retransmit (if its ready).
6920                  *
6921                  * If its on a normal rack timer (non-min) then
6922                  * we will check if its expired.
6923                  */
6924                 goto skip_time_check;
6925         }
6926         if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
6927                 uint32_t left;
6928
6929                 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
6930                         ret = -1;
6931                         rack_log_to_processing(rack, cts, ret, 0);
6932                         return (0);
6933                 }
6934                 if (hpts_calling == 0) {
6935                         /*
6936                          * A user send or queued mbuf (sack) has called us? We
6937                          * return 0 and let the pacing guards
6938                          * deal with it if they should or
6939                          * should not cause a send.
6940                          */
6941                         ret = -2;
6942                         rack_log_to_processing(rack, cts, ret, 0);
6943                         return (0);
6944                 }
6945                 /*
6946                  * Ok our timer went off early and we are not paced false
6947                  * alarm, go back to sleep.
6948                  */
6949                 ret = -3;
6950                 left = rack->r_ctl.rc_timer_exp - cts;
6951                 tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(left));
6952                 rack_log_to_processing(rack, cts, ret, left);
6953                 return (1);
6954         }
6955 skip_time_check:
6956         rack->rc_tmr_stopped = 0;
6957         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
6958         if (timers & PACE_TMR_DELACK) {
6959                 ret = rack_timeout_delack(tp, rack, cts);
6960         } else if (timers & PACE_TMR_RACK) {
6961                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
6962                 rack->r_fast_output = 0;
6963                 ret = rack_timeout_rack(tp, rack, cts);
6964         } else if (timers & PACE_TMR_TLP) {
6965                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
6966                 ret = rack_timeout_tlp(tp, rack, cts);
6967         } else if (timers & PACE_TMR_RXT) {
6968                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
6969                 rack->r_fast_output = 0;
6970                 ret = rack_timeout_rxt(tp, rack, cts);
6971         } else if (timers & PACE_TMR_PERSIT) {
6972                 ret = rack_timeout_persist(tp, rack, cts);
6973         } else if (timers & PACE_TMR_KEEP) {
6974                 ret = rack_timeout_keepalive(tp, rack, cts);
6975         }
6976         rack_log_to_processing(rack, cts, ret, timers);
6977         return (ret);
6978 }
6979
6980 static void
6981 rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line)
6982 {
6983         struct timeval tv;
6984         uint32_t us_cts, flags_on_entry;
6985         uint8_t hpts_removed = 0;
6986
6987         flags_on_entry = rack->r_ctl.rc_hpts_flags;
6988         us_cts = tcp_get_usecs(&tv);
6989         if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
6990             ((TSTMP_GEQ(us_cts, rack->r_ctl.rc_last_output_to)) ||
6991              ((tp->snd_max - tp->snd_una) == 0))) {
6992                 tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
6993                 hpts_removed = 1;
6994                 /* If we were not delayed cancel out the flag. */
6995                 if ((tp->snd_max - tp->snd_una) == 0)
6996                         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
6997                 rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
6998         }
6999         if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
7000                 rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
7001                 if (rack->rc_inp->inp_in_hpts &&
7002                     ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) {
7003                         /*
7004                          * Canceling timer's when we have no output being
7005                          * paced. We also must remove ourselves from the
7006                          * hpts.
7007                          */
7008                         tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
7009                         hpts_removed = 1;
7010                 }
7011                 rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
7012         }
7013         if (hpts_removed == 0)
7014                 rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
7015 }
7016
7017 static void
7018 rack_timer_stop(struct tcpcb *tp, uint32_t timer_type)
7019 {
7020         return;
7021 }
7022
7023 static int
7024 rack_stopall(struct tcpcb *tp)
7025 {
7026         struct tcp_rack *rack;
7027         rack = (struct tcp_rack *)tp->t_fb_ptr;
7028         rack->t_timers_stopped = 1;
7029         return (0);
7030 }
7031
7032 static void
7033 rack_timer_activate(struct tcpcb *tp, uint32_t timer_type, uint32_t delta)
7034 {
7035         return;
7036 }
7037
7038 static int
7039 rack_timer_active(struct tcpcb *tp, uint32_t timer_type)
7040 {
7041         return (0);
7042 }
7043
7044 static void
7045 rack_stop_all_timers(struct tcpcb *tp)
7046 {
7047         struct tcp_rack *rack;
7048
7049         /*
7050          * Assure no timers are running.
7051          */
7052         if (tcp_timer_active(tp, TT_PERSIST)) {
7053                 /* We enter in persists, set the flag appropriately */
7054                 rack = (struct tcp_rack *)tp->t_fb_ptr;
7055                 rack->rc_in_persist = 1;
7056         }
7057         tcp_timer_suspend(tp, TT_PERSIST);
7058         tcp_timer_suspend(tp, TT_REXMT);
7059         tcp_timer_suspend(tp, TT_KEEP);
7060         tcp_timer_suspend(tp, TT_DELACK);
7061 }
7062
7063 static void
7064 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
7065     struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag)
7066 {
7067         int32_t idx;
7068         uint16_t stripped_flags;
7069
7070         rsm->r_rtr_cnt++;
7071         rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7072         rsm->r_dupack = 0;
7073         if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) {
7074                 rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS;
7075                 rsm->r_flags |= RACK_OVERMAX;
7076         }
7077         if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & RACK_TLP) == 0)) {
7078                 rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
7079                 rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
7080         }
7081         idx = rsm->r_rtr_cnt - 1;
7082         rsm->r_tim_lastsent[idx] = ts;
7083         stripped_flags = rsm->r_flags & ~(RACK_SENT_SP|RACK_SENT_FP);
7084         if (rsm->r_flags & RACK_ACKED) {
7085                 /* Problably MTU discovery messing with us */
7086                 rsm->r_flags &= ~RACK_ACKED;
7087                 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
7088         }
7089         if (rsm->r_in_tmap) {
7090                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7091                 rsm->r_in_tmap = 0;
7092         }
7093         TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7094         rsm->r_in_tmap = 1;
7095         if (rsm->r_flags & RACK_SACK_PASSED) {
7096                 /* We have retransmitted due to the SACK pass */
7097                 rsm->r_flags &= ~RACK_SACK_PASSED;
7098                 rsm->r_flags |= RACK_WAS_SACKPASS;
7099         }
7100 }
7101
7102 static uint32_t
7103 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
7104     struct rack_sendmap *rsm, uint64_t ts, int32_t *lenp, uint16_t add_flag)
7105 {
7106         /*
7107          * We (re-)transmitted starting at rsm->r_start for some length
7108          * (possibly less than r_end.
7109          */
7110         struct rack_sendmap *nrsm, *insret;
7111         uint32_t c_end;
7112         int32_t len;
7113
7114         len = *lenp;
7115         c_end = rsm->r_start + len;
7116         if (SEQ_GEQ(c_end, rsm->r_end)) {
7117                 /*
7118                  * We retransmitted the whole piece or more than the whole
7119                  * slopping into the next rsm.
7120                  */
7121                 rack_update_rsm(tp, rack, rsm, ts, add_flag);
7122                 if (c_end == rsm->r_end) {
7123                         *lenp = 0;
7124                         return (0);
7125                 } else {
7126                         int32_t act_len;
7127
7128                         /* Hangs over the end return whats left */
7129                         act_len = rsm->r_end - rsm->r_start;
7130                         *lenp = (len - act_len);
7131                         return (rsm->r_end);
7132                 }
7133                 /* We don't get out of this block. */
7134         }
7135         /*
7136          * Here we retransmitted less than the whole thing which means we
7137          * have to split this into what was transmitted and what was not.
7138          */
7139         nrsm = rack_alloc_full_limit(rack);
7140         if (nrsm == NULL) {
7141                 /*
7142                  * We can't get memory, so lets not proceed.
7143                  */
7144                 *lenp = 0;
7145                 return (0);
7146         }
7147         /*
7148          * So here we are going to take the original rsm and make it what we
7149          * retransmitted. nrsm will be the tail portion we did not
7150          * retransmit. For example say the chunk was 1, 11 (10 bytes). And
7151          * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
7152          * 1, 6 and the new piece will be 6, 11.
7153          */
7154         rack_clone_rsm(rack, nrsm, rsm, c_end);
7155         nrsm->r_dupack = 0;
7156         rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
7157         insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7158 #ifdef INVARIANTS
7159         if (insret != NULL) {
7160                 panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7161                       nrsm, insret, rack, rsm);
7162         }
7163 #endif
7164         if (rsm->r_in_tmap) {
7165                 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7166                 nrsm->r_in_tmap = 1;
7167         }
7168         rsm->r_flags &= (~RACK_HAS_FIN);
7169         rack_update_rsm(tp, rack, rsm, ts, add_flag);
7170         /* Log a split of rsm into rsm and nrsm */
7171         rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7172         *lenp = 0;
7173         return (0);
7174 }
7175
7176 static void
7177 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
7178                 uint32_t seq_out, uint8_t th_flags, int32_t err, uint64_t cts,
7179                 struct rack_sendmap *hintrsm, uint16_t add_flag, struct mbuf *s_mb, uint32_t s_moff)
7180 {
7181         struct tcp_rack *rack;
7182         struct rack_sendmap *rsm, *nrsm, *insret, fe;
7183         register uint32_t snd_max, snd_una;
7184
7185         /*
7186          * Add to the RACK log of packets in flight or retransmitted. If
7187          * there is a TS option we will use the TS echoed, if not we will
7188          * grab a TS.
7189          *
7190          * Retransmissions will increment the count and move the ts to its
7191          * proper place. Note that if options do not include TS's then we
7192          * won't be able to effectively use the ACK for an RTT on a retran.
7193          *
7194          * Notes about r_start and r_end. Lets consider a send starting at
7195          * sequence 1 for 10 bytes. In such an example the r_start would be
7196          * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
7197          * This means that r_end is actually the first sequence for the next
7198          * slot (11).
7199          *
7200          */
7201         /*
7202          * If err is set what do we do XXXrrs? should we not add the thing?
7203          * -- i.e. return if err != 0 or should we pretend we sent it? --
7204          * i.e. proceed with add ** do this for now.
7205          */
7206         INP_WLOCK_ASSERT(tp->t_inpcb);
7207         if (err)
7208                 /*
7209                  * We don't log errors -- we could but snd_max does not
7210                  * advance in this case either.
7211                  */
7212                 return;
7213
7214         if (th_flags & TH_RST) {
7215                 /*
7216                  * We don't log resets and we return immediately from
7217                  * sending
7218                  */
7219                 return;
7220         }
7221         rack = (struct tcp_rack *)tp->t_fb_ptr;
7222         snd_una = tp->snd_una;
7223         snd_max = tp->snd_max;
7224         if (th_flags & (TH_SYN | TH_FIN)) {
7225                 /*
7226                  * The call to rack_log_output is made before bumping
7227                  * snd_max. This means we can record one extra byte on a SYN
7228                  * or FIN if seq_out is adding more on and a FIN is present
7229                  * (and we are not resending).
7230                  */
7231                 if ((th_flags & TH_SYN) && (seq_out == tp->iss))
7232                         len++;
7233                 if (th_flags & TH_FIN)
7234                         len++;
7235                 if (SEQ_LT(snd_max, tp->snd_nxt)) {
7236                         /*
7237                          * The add/update as not been done for the FIN/SYN
7238                          * yet.
7239                          */
7240                         snd_max = tp->snd_nxt;
7241                 }
7242         }
7243         if (SEQ_LEQ((seq_out + len), snd_una)) {
7244                 /* Are sending an old segment to induce an ack (keep-alive)? */
7245                 return;
7246         }
7247         if (SEQ_LT(seq_out, snd_una)) {
7248                 /* huh? should we panic? */
7249                 uint32_t end;
7250
7251                 end = seq_out + len;
7252                 seq_out = snd_una;
7253                 if (SEQ_GEQ(end, seq_out))
7254                         len = end - seq_out;
7255                 else
7256                         len = 0;
7257         }
7258         if (len == 0) {
7259                 /* We don't log zero window probes */
7260                 return;
7261         }
7262         rack->r_ctl.rc_time_last_sent = cts;
7263         if (IN_FASTRECOVERY(tp->t_flags)) {
7264                 rack->r_ctl.rc_prr_out += len;
7265         }
7266         /* First question is it a retransmission or new? */
7267         if (seq_out == snd_max) {
7268                 /* Its new */
7269 again:
7270                 rsm = rack_alloc(rack);
7271                 if (rsm == NULL) {
7272                         /*
7273                          * Hmm out of memory and the tcb got destroyed while
7274                          * we tried to wait.
7275                          */
7276                         return;
7277                 }
7278                 if (th_flags & TH_FIN) {
7279                         rsm->r_flags = RACK_HAS_FIN|add_flag;
7280                 } else {
7281                         rsm->r_flags = add_flag;
7282                 }
7283                 rsm->r_tim_lastsent[0] = cts;
7284                 rsm->r_rtr_cnt = 1;
7285                 rsm->r_rtr_bytes = 0;
7286                 if (th_flags & TH_SYN) {
7287                         /* The data space is one beyond snd_una */
7288                         rsm->r_flags |= RACK_HAS_SYN;
7289                 }
7290                 rsm->r_start = seq_out;
7291                 rsm->r_end = rsm->r_start + len;
7292                 rsm->r_dupack = 0;
7293                 /*
7294                  * save off the mbuf location that
7295                  * sndmbuf_noadv returned (which is
7296                  * where we started copying from)..
7297                  */
7298                 rsm->m = s_mb;
7299                 rsm->soff = s_moff;
7300                 /* rsm->m will be NULL if RACK_HAS_SYN or RACK_HAS_FIN is set */
7301                 if (rsm->m) {
7302                         if (rsm->m->m_len <= rsm->soff) {
7303                                 /*
7304                                  * XXXrrs Question, will this happen?
7305                                  *
7306                                  * If sbsndptr is set at the correct place
7307                                  * then s_moff should always be somewhere
7308                                  * within rsm->m. But if the sbsndptr was
7309                                  * off then that won't be true. If it occurs
7310                                  * we need to walkout to the correct location.
7311                                  */
7312                                 struct mbuf *lm;
7313
7314                                 lm = rsm->m;
7315                                 while (lm->m_len <= rsm->soff) {
7316                                         rsm->soff -= lm->m_len;
7317                                         lm = lm->m_next;
7318                                         KASSERT(lm != NULL, ("%s rack:%p lm goes null orig_off:%u origmb:%p rsm->soff:%u",
7319                                                              __func__, rack, s_moff, s_mb, rsm->soff));
7320                                 }
7321                                 rsm->m = lm;
7322                                 counter_u64_add(rack_sbsndptr_wrong, 1);
7323                         } else
7324                                 counter_u64_add(rack_sbsndptr_right, 1);
7325                         rsm->orig_m_len = rsm->m->m_len;
7326                 } else
7327                         rsm->orig_m_len = 0;
7328                 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7329                 /* Log a new rsm */
7330                 rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_NEW, 0, __LINE__);
7331                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7332 #ifdef INVARIANTS
7333                 if (insret != NULL) {
7334                         panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7335                               nrsm, insret, rack, rsm);
7336                 }
7337 #endif
7338                 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7339                 rsm->r_in_tmap = 1;
7340                 /*
7341                  * Special case detection, is there just a single
7342                  * packet outstanding when we are not in recovery?
7343                  *
7344                  * If this is true mark it so.
7345                  */
7346                 if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
7347                     (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) == ctf_fixed_maxseg(tp))) {
7348                         struct rack_sendmap *prsm;
7349
7350                         prsm = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7351                         if (prsm)
7352                                 prsm->r_one_out_nr = 1;
7353                 }
7354                 return;
7355         }
7356         /*
7357          * If we reach here its a retransmission and we need to find it.
7358          */
7359         memset(&fe, 0, sizeof(fe));
7360 more:
7361         if (hintrsm && (hintrsm->r_start == seq_out)) {
7362                 rsm = hintrsm;
7363                 hintrsm = NULL;
7364         } else {
7365                 /* No hints sorry */
7366                 rsm = NULL;
7367         }
7368         if ((rsm) && (rsm->r_start == seq_out)) {
7369                 seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag);
7370                 if (len == 0) {
7371                         return;
7372                 } else {
7373                         goto more;
7374                 }
7375         }
7376         /* Ok it was not the last pointer go through it the hard way. */
7377 refind:
7378         fe.r_start = seq_out;
7379         rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
7380         if (rsm) {
7381                 if (rsm->r_start == seq_out) {
7382                         seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag);
7383                         if (len == 0) {
7384                                 return;
7385                         } else {
7386                                 goto refind;
7387                         }
7388                 }
7389                 if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
7390                         /* Transmitted within this piece */
7391                         /*
7392                          * Ok we must split off the front and then let the
7393                          * update do the rest
7394                          */
7395                         nrsm = rack_alloc_full_limit(rack);
7396                         if (nrsm == NULL) {
7397                                 rack_update_rsm(tp, rack, rsm, cts, add_flag);
7398                                 return;
7399                         }
7400                         /*
7401                          * copy rsm to nrsm and then trim the front of rsm
7402                          * to not include this part.
7403                          */
7404                         rack_clone_rsm(rack, nrsm, rsm, seq_out);
7405                         insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7406                         rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7407 #ifdef INVARIANTS
7408                         if (insret != NULL) {
7409                                 panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7410                                       nrsm, insret, rack, rsm);
7411                         }
7412 #endif
7413                         if (rsm->r_in_tmap) {
7414                                 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7415                                 nrsm->r_in_tmap = 1;
7416                         }
7417                         rsm->r_flags &= (~RACK_HAS_FIN);
7418                         seq_out = rack_update_entry(tp, rack, nrsm, cts, &len, add_flag);
7419                         if (len == 0) {
7420                                 return;
7421                         } else if (len > 0)
7422                                 goto refind;
7423                 }
7424         }
7425         /*
7426          * Hmm not found in map did they retransmit both old and on into the
7427          * new?
7428          */
7429         if (seq_out == tp->snd_max) {
7430                 goto again;
7431         } else if (SEQ_LT(seq_out, tp->snd_max)) {
7432 #ifdef INVARIANTS
7433                 printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
7434                        seq_out, len, tp->snd_una, tp->snd_max);
7435                 printf("Starting Dump of all rack entries\n");
7436                 RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
7437                         printf("rsm:%p start:%u end:%u\n",
7438                                rsm, rsm->r_start, rsm->r_end);
7439                 }
7440                 printf("Dump complete\n");
7441                 panic("seq_out not found rack:%p tp:%p",
7442                       rack, tp);
7443 #endif
7444         } else {
7445 #ifdef INVARIANTS
7446                 /*
7447                  * Hmm beyond sndmax? (only if we are using the new rtt-pack
7448                  * flag)
7449                  */
7450                 panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
7451                       seq_out, len, tp->snd_max, tp);
7452 #endif
7453         }
7454 }
7455
7456 /*
7457  * Record one of the RTT updates from an ack into
7458  * our sample structure.
7459  */
7460
7461 static void
7462 tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt, uint32_t len, uint32_t us_rtt,
7463                     int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt)
7464 {
7465         if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7466             (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) {
7467                 rack->r_ctl.rack_rs.rs_rtt_lowest = rtt;
7468         }
7469         if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7470             (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) {
7471                 rack->r_ctl.rack_rs.rs_rtt_highest = rtt;
7472         }
7473         if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
7474             if (us_rtt < rack->r_ctl.rc_gp_lowrtt)
7475                 rack->r_ctl.rc_gp_lowrtt = us_rtt;
7476             if (rack->rc_tp->snd_wnd > rack->r_ctl.rc_gp_high_rwnd)
7477                     rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
7478         }
7479         if ((confidence == 1) &&
7480             ((rsm == NULL) ||
7481              (rsm->r_just_ret) ||
7482              (rsm->r_one_out_nr &&
7483               len < (ctf_fixed_maxseg(rack->rc_tp) * 2)))) {
7484                 /*
7485                  * If the rsm had a just return
7486                  * hit it then we can't trust the
7487                  * rtt measurement for buffer deterimination
7488                  * Note that a confidence of 2, indicates
7489                  * SACK'd which overrides the r_just_ret or
7490                  * the r_one_out_nr. If it was a CUM-ACK and
7491                  * we had only two outstanding, but get an
7492                  * ack for only 1. Then that also lowers our
7493                  * confidence.
7494                  */
7495                 confidence = 0;
7496         }
7497         if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7498             (rack->r_ctl.rack_rs.rs_us_rtt > us_rtt)) {
7499                 if (rack->r_ctl.rack_rs.confidence == 0) {
7500                         /*
7501                          * We take anything with no current confidence
7502                          * saved.
7503                          */
7504                         rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
7505                         rack->r_ctl.rack_rs.confidence = confidence;
7506                         rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
7507                 } else if (confidence || rack->r_ctl.rack_rs.confidence) {
7508                         /*
7509                          * Once we have a confident number,
7510                          * we can update it with a smaller
7511                          * value since this confident number
7512                          * may include the DSACK time until
7513                          * the next segment (the second one) arrived.
7514                          */
7515                         rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
7516                         rack->r_ctl.rack_rs.confidence = confidence;
7517                         rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
7518                 }
7519         }
7520         rack_log_rtt_upd(rack->rc_tp, rack, us_rtt, len, rsm, confidence);
7521         rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID;
7522         rack->r_ctl.rack_rs.rs_rtt_tot += rtt;
7523         rack->r_ctl.rack_rs.rs_rtt_cnt++;
7524 }
7525
7526 /*
7527  * Collect new round-trip time estimate
7528  * and update averages and current timeout.
7529  */
7530 static void
7531 tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp)
7532 {
7533         int32_t delta;
7534         uint32_t o_srtt, o_var;
7535         int32_t hrtt_up = 0;
7536         int32_t rtt;
7537
7538         if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY)
7539                 /* No valid sample */
7540                 return;
7541         if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) {
7542                 /* We are to use the lowest RTT seen in a single ack */
7543                 rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
7544         } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) {
7545                 /* We are to use the highest RTT seen in a single ack */
7546                 rtt = rack->r_ctl.rack_rs.rs_rtt_highest;
7547         } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) {
7548                 /* We are to use the average RTT seen in a single ack */
7549                 rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot /
7550                                 (uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt);
7551         } else {
7552 #ifdef INVARIANTS
7553                 panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method);
7554 #endif
7555                 return;
7556         }
7557         if (rtt == 0)
7558                 rtt = 1;
7559         if (rack->rc_gp_rtt_set == 0) {
7560                 /*
7561                  * With no RTT we have to accept
7562                  * even one we are not confident of.
7563                  */
7564                 rack->r_ctl.rc_gp_srtt = rack->r_ctl.rack_rs.rs_us_rtt;
7565                 rack->rc_gp_rtt_set = 1;
7566         } else if (rack->r_ctl.rack_rs.confidence) {
7567                 /* update the running gp srtt */
7568                 rack->r_ctl.rc_gp_srtt -= (rack->r_ctl.rc_gp_srtt/8);
7569                 rack->r_ctl.rc_gp_srtt += rack->r_ctl.rack_rs.rs_us_rtt / 8;
7570         }
7571         if (rack->r_ctl.rack_rs.confidence) {
7572                 /*
7573                  * record the low and high for highly buffered path computation,
7574                  * we only do this if we are confident (not a retransmission).
7575                  */
7576                 if (rack->r_ctl.rc_highest_us_rtt < rack->r_ctl.rack_rs.rs_us_rtt) {
7577                         rack->r_ctl.rc_highest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7578                         hrtt_up = 1;
7579                 }
7580                 if (rack->rc_highly_buffered == 0) {
7581                         /*
7582                          * Currently once we declare a path has
7583                          * highly buffered there is no going
7584                          * back, which may be a problem...
7585                          */
7586                         if ((rack->r_ctl.rc_highest_us_rtt / rack->r_ctl.rc_lowest_us_rtt) > rack_hbp_thresh) {
7587                                 rack_log_rtt_shrinks(rack, rack->r_ctl.rack_rs.rs_us_rtt,
7588                                                      rack->r_ctl.rc_highest_us_rtt,
7589                                                      rack->r_ctl.rc_lowest_us_rtt,
7590                                                      RACK_RTTS_SEEHBP);
7591                                 rack->rc_highly_buffered = 1;
7592                         }
7593                 }
7594         }
7595         if ((rack->r_ctl.rack_rs.confidence) ||
7596             (rack->r_ctl.rack_rs.rs_us_rtrcnt == 1)) {
7597                 /*
7598                  * If we are highly confident of it <or> it was
7599                  * never retransmitted we accept it as the last us_rtt.
7600                  */
7601                 rack->r_ctl.rc_last_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7602                 /* The lowest rtt can be set if its was not retransmited */
7603                 if (rack->r_ctl.rc_lowest_us_rtt > rack->r_ctl.rack_rs.rs_us_rtt) {
7604                         rack->r_ctl.rc_lowest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7605                         if (rack->r_ctl.rc_lowest_us_rtt == 0)
7606                                 rack->r_ctl.rc_lowest_us_rtt = 1;
7607                 }
7608         }
7609         o_srtt = tp->t_srtt;
7610         o_var = tp->t_rttvar;
7611         rack = (struct tcp_rack *)tp->t_fb_ptr;
7612         if (tp->t_srtt != 0) {
7613                 /*
7614                  * We keep a simple srtt in microseconds, like our rtt
7615                  * measurement. We don't need to do any tricks with shifting
7616                  * etc. Instead we just add in 1/8th of the new measurement
7617                  * and subtract out 1/8 of the old srtt. We do the same with
7618                  * the variance after finding the absolute value of the
7619                  * difference between this sample and the current srtt.
7620                  */
7621                 delta = tp->t_srtt - rtt;
7622                 /* Take off 1/8th of the current sRTT */
7623                 tp->t_srtt -= (tp->t_srtt >> 3);
7624                 /* Add in 1/8th of the new RTT just measured */
7625                 tp->t_srtt += (rtt >> 3);
7626                 if (tp->t_srtt <= 0)
7627                         tp->t_srtt = 1;
7628                 /* Now lets make the absolute value of the variance */
7629                 if (delta < 0)
7630                         delta = -delta;
7631                 /* Subtract out 1/8th */
7632                 tp->t_rttvar -= (tp->t_rttvar >> 3);
7633                 /* Add in 1/8th of the new variance we just saw */
7634                 tp->t_rttvar += (delta >> 3);
7635                 if (tp->t_rttvar <= 0)
7636                         tp->t_rttvar = 1;
7637                 if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar)
7638                         tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
7639         } else {
7640                 /*
7641                  * No rtt measurement yet - use the unsmoothed rtt. Set the
7642                  * variance to half the rtt (so our first retransmit happens
7643                  * at 3*rtt).
7644                  */
7645                 tp->t_srtt = rtt;
7646                 tp->t_rttvar = rtt >> 1;
7647                 tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
7648         }
7649         rack->rc_srtt_measure_made = 1;
7650         KMOD_TCPSTAT_INC(tcps_rttupdated);
7651         tp->t_rttupdated++;
7652 #ifdef STATS
7653         if (rack_stats_gets_ms_rtt == 0) {
7654                 /* Send in the microsecond rtt used for rxt timeout purposes */
7655                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt));
7656         } else if (rack_stats_gets_ms_rtt == 1) {
7657                 /* Send in the millisecond rtt used for rxt timeout purposes */
7658                 int32_t ms_rtt;
7659
7660                 /* Round up */
7661                 ms_rtt = (rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
7662                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
7663         } else if (rack_stats_gets_ms_rtt == 2) {
7664                 /* Send in the millisecond rtt has close to the path RTT as we can get  */
7665                 int32_t ms_rtt;
7666
7667                 /* Round up */
7668                 ms_rtt = (rack->r_ctl.rack_rs.rs_us_rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
7669                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
7670         }  else {
7671                 /* Send in the microsecond rtt has close to the path RTT as we can get  */
7672                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
7673         }
7674
7675 #endif
7676         /*
7677          * the retransmit should happen at rtt + 4 * rttvar. Because of the
7678          * way we do the smoothing, srtt and rttvar will each average +1/2
7679          * tick of bias.  When we compute the retransmit timer, we want 1/2
7680          * tick of rounding and 1 extra tick because of +-1/2 tick
7681          * uncertainty in the firing of the timer.  The bias will give us
7682          * exactly the 1.5 tick we need.  But, because the bias is
7683          * statistical, we have to test that we don't drop below the minimum
7684          * feasible timer (which is 2 ticks).
7685          */
7686         tp->t_rxtshift = 0;
7687         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
7688                       max(rack_rto_min, rtt + 2), rack_rto_max);
7689         rack_log_rtt_sample(rack, rtt);
7690         tp->t_softerror = 0;
7691 }
7692
7693
7694 static void
7695 rack_apply_updated_usrtt(struct tcp_rack *rack, uint32_t us_rtt, uint32_t us_cts)
7696 {
7697         /*
7698          * Apply to filter the inbound us-rtt at us_cts.
7699          */
7700         uint32_t old_rtt;
7701
7702         old_rtt = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
7703         apply_filter_min_small(&rack->r_ctl.rc_gp_min_rtt,
7704                                us_rtt, us_cts);
7705         if (rack->r_ctl.last_pacing_time &&
7706             rack->rc_gp_dyn_mul &&
7707             (rack->r_ctl.last_pacing_time > us_rtt))
7708                 rack->pacing_longer_than_rtt = 1;
7709         else
7710                 rack->pacing_longer_than_rtt = 0;
7711         if (old_rtt > us_rtt) {
7712                 /* We just hit a new lower rtt time */
7713                 rack_log_rtt_shrinks(rack,  us_cts,  old_rtt,
7714                                      __LINE__, RACK_RTTS_NEWRTT);
7715                 /*
7716                  * Only count it if its lower than what we saw within our
7717                  * calculated range.
7718                  */
7719                 if ((old_rtt - us_rtt) > rack_min_rtt_movement) {
7720                         if (rack_probertt_lower_within &&
7721                             rack->rc_gp_dyn_mul &&
7722                             (rack->use_fixed_rate == 0) &&
7723                             (rack->rc_always_pace)) {
7724                                 /*
7725                                  * We are seeing a new lower rtt very close
7726                                  * to the time that we would have entered probe-rtt.
7727                                  * This is probably due to the fact that a peer flow
7728                                  * has entered probe-rtt. Lets go in now too.
7729                                  */
7730                                 uint32_t val;
7731
7732                                 val = rack_probertt_lower_within * rack_time_between_probertt;
7733                                 val /= 100;
7734                                 if ((rack->in_probe_rtt == 0)  &&
7735                                     ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= (rack_time_between_probertt - val))) {
7736                                         rack_enter_probertt(rack, us_cts);
7737                                 }
7738                         }
7739                         rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
7740                 }
7741         }
7742 }
7743
7744 static int
7745 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
7746     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack)
7747 {
7748         int32_t i, all;
7749         uint32_t t, len_acked;
7750
7751         if ((rsm->r_flags & RACK_ACKED) ||
7752             (rsm->r_flags & RACK_WAS_ACKED))
7753                 /* Already done */
7754                 return (0);
7755         if (rsm->r_no_rtt_allowed) {
7756                 /* Not allowed */
7757                 return (0);
7758         }
7759         if (ack_type == CUM_ACKED) {
7760                 if (SEQ_GT(th_ack, rsm->r_end)) {
7761                         len_acked = rsm->r_end - rsm->r_start;
7762                         all = 1;
7763                 } else {
7764                         len_acked = th_ack - rsm->r_start;
7765                         all = 0;
7766                 }
7767         } else {
7768                 len_acked = rsm->r_end - rsm->r_start;
7769                 all = 0;
7770         }
7771         if (rsm->r_rtr_cnt == 1) {
7772                 uint32_t us_rtt;
7773
7774                 t = cts - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7775                 if ((int)t <= 0)
7776                         t = 1;
7777                 if (!tp->t_rttlow || tp->t_rttlow > t)
7778                         tp->t_rttlow = t;
7779                 if (!rack->r_ctl.rc_rack_min_rtt ||
7780                     SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7781                         rack->r_ctl.rc_rack_min_rtt = t;
7782                         if (rack->r_ctl.rc_rack_min_rtt == 0) {
7783                                 rack->r_ctl.rc_rack_min_rtt = 1;
7784                         }
7785                 }
7786                 if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]))
7787                         us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
7788                 else
7789                         us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
7790                 if (us_rtt == 0)
7791                         us_rtt = 1;
7792                 rack_apply_updated_usrtt(rack, us_rtt, tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
7793                 if (ack_type == SACKED) {
7794                         rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 1);
7795                         tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt, 2 , rsm, rsm->r_rtr_cnt);
7796                 } else {
7797                         /*
7798                          * We need to setup what our confidence
7799                          * is in this ack.
7800                          *
7801                          * If the rsm was app limited and it is
7802                          * less than a mss in length (the end
7803                          * of the send) then we have a gap. If we
7804                          * were app limited but say we were sending
7805                          * multiple MSS's then we are more confident
7806                          * int it.
7807                          *
7808                          * When we are not app-limited then we see if
7809                          * the rsm is being included in the current
7810                          * measurement, we tell this by the app_limited_needs_set
7811                          * flag.
7812                          *
7813                          * Note that being cwnd blocked is not applimited
7814                          * as well as the pacing delay between packets which
7815                          * are sending only 1 or 2 MSS's also will show up
7816                          * in the RTT. We probably need to examine this algorithm
7817                          * a bit more and enhance it to account for the delay
7818                          * between rsm's. We could do that by saving off the
7819                          * pacing delay of each rsm (in an rsm) and then
7820                          * factoring that in somehow though for now I am
7821                          * not sure how :)
7822                          */
7823                         int calc_conf = 0;
7824
7825                         if (rsm->r_flags & RACK_APP_LIMITED) {
7826                                 if (all && (len_acked <= ctf_fixed_maxseg(tp)))
7827                                         calc_conf = 0;
7828                                 else
7829                                         calc_conf = 1;
7830                         } else if (rack->app_limited_needs_set == 0) {
7831                                 calc_conf = 1;
7832                         } else {
7833                                 calc_conf = 0;
7834                         }
7835                         rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 2);
7836                         tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt,
7837                                             calc_conf, rsm, rsm->r_rtr_cnt);
7838                 }
7839                 if ((rsm->r_flags & RACK_TLP) &&
7840                     (!IN_FASTRECOVERY(tp->t_flags))) {
7841                         /* Segment was a TLP and our retrans matched */
7842                         if (rack->r_ctl.rc_tlp_cwnd_reduce) {
7843                                 rack->r_ctl.rc_rsm_start = tp->snd_max;
7844                                 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
7845                                 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
7846                                 rack_cong_signal(tp, CC_NDUPACK, tp->snd_una);
7847                         }
7848                 }
7849                 if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
7850                         /* New more recent rack_tmit_time */
7851                         rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7852                         rack->rc_rack_rtt = t;
7853                 }
7854                 return (1);
7855         }
7856         /*
7857          * We clear the soft/rxtshift since we got an ack.
7858          * There is no assurance we will call the commit() function
7859          * so we need to clear these to avoid incorrect handling.
7860          */
7861         tp->t_rxtshift = 0;
7862         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
7863                       rack_rto_min, rack_rto_max);
7864         tp->t_softerror = 0;
7865         if (to && (to->to_flags & TOF_TS) &&
7866             (ack_type == CUM_ACKED) &&
7867             (to->to_tsecr) &&
7868             ((rsm->r_flags & RACK_OVERMAX) == 0)) {
7869                 /*
7870                  * Now which timestamp does it match? In this block the ACK
7871                  * must be coming from a previous transmission.
7872                  */
7873                 for (i = 0; i < rsm->r_rtr_cnt; i++) {
7874                         if (rack_ts_to_msec(rsm->r_tim_lastsent[i]) == to->to_tsecr) {
7875                                 t = cts - (uint32_t)rsm->r_tim_lastsent[i];
7876                                 if ((int)t <= 0)
7877                                         t = 1;
7878                                 if ((i + 1) < rsm->r_rtr_cnt) {
7879                                         /*
7880                                          * The peer ack'd from our previous
7881                                          * transmission. We have a spurious
7882                                          * retransmission and thus we dont
7883                                          * want to update our rack_rtt.
7884                                          */
7885                                         return (0);
7886                                 }
7887                                 if (!tp->t_rttlow || tp->t_rttlow > t)
7888                                         tp->t_rttlow = t;
7889                                 if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7890                                         rack->r_ctl.rc_rack_min_rtt = t;
7891                                         if (rack->r_ctl.rc_rack_min_rtt == 0) {
7892                                                 rack->r_ctl.rc_rack_min_rtt = 1;
7893                                         }
7894                                 }
7895                                 if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
7896                                            (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
7897                                         /* New more recent rack_tmit_time */
7898                                         rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7899                                         rack->rc_rack_rtt = t;
7900                                 }
7901                                 rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[i], cts, 3);
7902                                 tcp_rack_xmit_timer(rack, t + 1, len_acked, t, 0, rsm,
7903                                                     rsm->r_rtr_cnt);
7904                                 return (1);
7905                         }
7906                 }
7907                 goto ts_not_found;
7908         } else {
7909                 /*
7910                  * Ok its a SACK block that we retransmitted. or a windows
7911                  * machine without timestamps. We can tell nothing from the
7912                  * time-stamp since its not there or the time the peer last
7913                  * recieved a segment that moved forward its cum-ack point.
7914                  */
7915 ts_not_found:
7916                 i = rsm->r_rtr_cnt - 1;
7917                 t = cts - (uint32_t)rsm->r_tim_lastsent[i];
7918                 if ((int)t <= 0)
7919                         t = 1;
7920                 if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7921                         /*
7922                          * We retransmitted and the ack came back in less
7923                          * than the smallest rtt we have observed. We most
7924                          * likely did an improper retransmit as outlined in
7925                          * 6.2 Step 2 point 2 in the rack-draft so we
7926                          * don't want to update our rack_rtt. We in
7927                          * theory (in future) might want to think about reverting our
7928                          * cwnd state but we won't for now.
7929                          */
7930                         return (0);
7931                 } else if (rack->r_ctl.rc_rack_min_rtt) {
7932                         /*
7933                          * We retransmitted it and the retransmit did the
7934                          * job.
7935                          */
7936                         if (!rack->r_ctl.rc_rack_min_rtt ||
7937                             SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7938                                 rack->r_ctl.rc_rack_min_rtt = t;
7939                                 if (rack->r_ctl.rc_rack_min_rtt == 0) {
7940                                         rack->r_ctl.rc_rack_min_rtt = 1;
7941                                 }
7942                         }
7943                         if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[i])) {
7944                                 /* New more recent rack_tmit_time */
7945                                 rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[i];
7946                                 rack->rc_rack_rtt = t;
7947                         }
7948                         return (1);
7949                 }
7950         }
7951         return (0);
7952 }
7953
7954 /*
7955  * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
7956  */
7957 static void
7958 rack_log_sack_passed(struct tcpcb *tp,
7959     struct tcp_rack *rack, struct rack_sendmap *rsm)
7960 {
7961         struct rack_sendmap *nrsm;
7962
7963         nrsm = rsm;
7964         TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap,
7965             rack_head, r_tnext) {
7966                 if (nrsm == rsm) {
7967                         /* Skip orginal segment he is acked */
7968                         continue;
7969                 }
7970                 if (nrsm->r_flags & RACK_ACKED) {
7971                         /*
7972                          * Skip ack'd segments, though we
7973                          * should not see these, since tmap
7974                          * should not have ack'd segments.
7975                          */
7976                         continue;
7977                 }
7978                 if (nrsm->r_flags & RACK_SACK_PASSED) {
7979                         /*
7980                          * We found one that is already marked
7981                          * passed, we have been here before and
7982                          * so all others below this are marked.
7983                          */
7984                         break;
7985                 }
7986                 nrsm->r_flags |= RACK_SACK_PASSED;
7987                 nrsm->r_flags &= ~RACK_WAS_SACKPASS;
7988         }
7989 }
7990
7991 static void
7992 rack_need_set_test(struct tcpcb *tp,
7993                    struct tcp_rack *rack,
7994                    struct rack_sendmap *rsm,
7995                    tcp_seq th_ack,
7996                    int line,
7997                    int use_which)
7998 {
7999
8000         if ((tp->t_flags & TF_GPUTINPROG) &&
8001             SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
8002                 /*
8003                  * We were app limited, and this ack
8004                  * butts up or goes beyond the point where we want
8005                  * to start our next measurement. We need
8006                  * to record the new gput_ts as here and
8007                  * possibly update the start sequence.
8008                  */
8009                 uint32_t seq, ts;
8010
8011                 if (rsm->r_rtr_cnt > 1) {
8012                         /*
8013                          * This is a retransmit, can we
8014                          * really make any assessment at this
8015                          * point?  We are not really sure of
8016                          * the timestamp, is it this or the
8017                          * previous transmission?
8018                          *
8019                          * Lets wait for something better that
8020                          * is not retransmitted.
8021                          */
8022                         return;
8023                 }
8024                 seq = tp->gput_seq;
8025                 ts = tp->gput_ts;
8026                 rack->app_limited_needs_set = 0;
8027                 tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
8028                 /* Do we start at a new end? */
8029                 if ((use_which == RACK_USE_BEG) &&
8030                     SEQ_GEQ(rsm->r_start, tp->gput_seq)) {
8031                         /*
8032                          * When we get an ACK that just eats
8033                          * up some of the rsm, we set RACK_USE_BEG
8034                          * since whats at r_start (i.e. th_ack)
8035                          * is left unacked and thats where the
8036                          * measurement not starts.
8037                          */
8038                         tp->gput_seq = rsm->r_start;
8039                         rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8040                 }
8041                 if ((use_which == RACK_USE_END) &&
8042                     SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
8043                             /*
8044                              * We use the end when the cumack
8045                              * is moving forward and completely
8046                              * deleting the rsm passed so basically
8047                              * r_end holds th_ack.
8048                              *
8049                              * For SACK's we also want to use the end
8050                              * since this piece just got sacked and
8051                              * we want to target anything after that
8052                              * in our measurement.
8053                              */
8054                             tp->gput_seq = rsm->r_end;
8055                             rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8056                 }
8057                 if (use_which == RACK_USE_END_OR_THACK) {
8058                         /*
8059                          * special case for ack moving forward,
8060                          * not a sack, we need to move all the
8061                          * way up to where this ack cum-ack moves
8062                          * to.
8063                          */
8064                         if (SEQ_GT(th_ack, rsm->r_end))
8065                                 tp->gput_seq = th_ack;
8066                         else
8067                                 tp->gput_seq = rsm->r_end;
8068                         rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8069                 }
8070                 if (SEQ_GT(tp->gput_seq, tp->gput_ack)) {
8071                         /*
8072                          * We moved beyond this guy's range, re-calculate
8073                          * the new end point.
8074                          */
8075                         if (rack->rc_gp_filled == 0) {
8076                                 tp->gput_ack = tp->gput_seq + max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
8077                         } else {
8078                                 tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
8079                         }
8080                 }
8081                 /*
8082                  * We are moving the goal post, we may be able to clear the
8083                  * measure_saw_probe_rtt flag.
8084                  */
8085                 if ((rack->in_probe_rtt == 0) &&
8086                     (rack->measure_saw_probe_rtt) &&
8087                     (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
8088                         rack->measure_saw_probe_rtt = 0;
8089                 rack_log_pacing_delay_calc(rack, ts, tp->gput_ts,
8090                                            seq, tp->gput_seq, 0, 5, line, NULL);
8091                 if (rack->rc_gp_filled &&
8092                     ((tp->gput_ack - tp->gput_seq) <
8093                      max(rc_init_window(rack), (MIN_GP_WIN *
8094                                                 ctf_fixed_maxseg(tp))))) {
8095                         uint32_t ideal_amount;
8096
8097                         ideal_amount = rack_get_measure_window(tp, rack);
8098                         if (ideal_amount > sbavail(&tp->t_inpcb->inp_socket->so_snd)) {
8099                                 /*
8100                                  * There is no sense of continuing this measurement
8101                                  * because its too small to gain us anything we
8102                                  * trust. Skip it and that way we can start a new
8103                                  * measurement quicker.
8104                                  */
8105                                 tp->t_flags &= ~TF_GPUTINPROG;
8106                                 rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
8107                                                            0, 0, 0, 6, __LINE__, NULL);
8108                         } else {
8109                                 /*
8110                                  * Reset the window further out.
8111                                  */
8112                                 tp->gput_ack = tp->gput_seq + ideal_amount;
8113                         }
8114                 }
8115         }
8116 }
8117
8118 static uint32_t
8119 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack,
8120                    struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts, int *moved_two)
8121 {
8122         uint32_t start, end, changed = 0;
8123         struct rack_sendmap stack_map;
8124         struct rack_sendmap *rsm, *nrsm, fe, *insret, *prev, *next;
8125         int32_t used_ref = 1;
8126         int moved = 0;
8127
8128         start = sack->start;
8129         end = sack->end;
8130         rsm = *prsm;
8131         memset(&fe, 0, sizeof(fe));
8132 do_rest_ofb:
8133         if ((rsm == NULL) ||
8134             (SEQ_LT(end, rsm->r_start)) ||
8135             (SEQ_GEQ(start, rsm->r_end)) ||
8136             (SEQ_LT(start, rsm->r_start))) {
8137                 /*
8138                  * We are not in the right spot,
8139                  * find the correct spot in the tree.
8140                  */
8141                 used_ref = 0;
8142                 fe.r_start = start;
8143                 rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
8144                 moved++;
8145         }
8146         if (rsm == NULL) {
8147                 /* TSNH */
8148                 goto out;
8149         }
8150         /* Ok we have an ACK for some piece of this rsm */
8151         if (rsm->r_start != start) {
8152                 if ((rsm->r_flags & RACK_ACKED) == 0) {
8153                         /**
8154                          * Need to split this in two pieces the before and after,
8155                          * the before remains in the map, the after must be
8156                          * added. In other words we have:
8157                          * rsm        |--------------|
8158                          * sackblk        |------->
8159                          * rsm will become
8160                          *     rsm    |---|
8161                          * and nrsm will be  the sacked piece
8162                          *     nrsm       |----------|
8163                          *
8164                          * But before we start down that path lets
8165                          * see if the sack spans over on top of
8166                          * the next guy and it is already sacked.
8167                          */
8168                         next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8169                         if (next && (next->r_flags & RACK_ACKED) &&
8170                             SEQ_GEQ(end, next->r_start)) {
8171                                 /**
8172                                  * So the next one is already acked, and
8173                                  * we can thus by hookery use our stack_map
8174                                  * to reflect the piece being sacked and
8175                                  * then adjust the two tree entries moving
8176                                  * the start and ends around. So we start like:
8177                                  *  rsm     |------------|             (not-acked)
8178                                  *  next                 |-----------| (acked)
8179                                  *  sackblk        |-------->
8180                                  *  We want to end like so:
8181                                  *  rsm     |------|                   (not-acked)
8182                                  *  next           |-----------------| (acked)
8183                                  *  nrsm           |-----|
8184                                  * Where nrsm is a temporary stack piece we
8185                                  * use to update all the gizmos.
8186                                  */
8187                                 /* Copy up our fudge block */
8188                                 nrsm = &stack_map;
8189                                 memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
8190                                 /* Now adjust our tree blocks */
8191                                 rsm->r_end = start;
8192                                 next->r_start = start;
8193                                 /* Now we must adjust back where next->m is */
8194                                 rack_setup_offset_for_rsm(rsm, next);
8195
8196                                 /* We don't need to adjust rsm, it did not change */
8197                                 /* Clear out the dup ack count of the remainder */
8198                                 rsm->r_dupack = 0;
8199                                 rsm->r_just_ret = 0;
8200                                 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8201                                 /* Now lets make sure our fudge block is right */
8202                                 nrsm->r_start = start;
8203                                 /* Now lets update all the stats and such */
8204                                 rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
8205                                 if (rack->app_limited_needs_set)
8206                                         rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
8207                                 changed += (nrsm->r_end - nrsm->r_start);
8208                                 rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
8209                                 if (nrsm->r_flags & RACK_SACK_PASSED) {
8210                                         counter_u64_add(rack_reorder_seen, 1);
8211                                         rack->r_ctl.rc_reorder_ts = cts;
8212                                 }
8213                                 /*
8214                                  * Now we want to go up from rsm (the
8215                                  * one left un-acked) to the next one
8216                                  * in the tmap. We do this so when
8217                                  * we walk backwards we include marking
8218                                  * sack-passed on rsm (The one passed in
8219                                  * is skipped since it is generally called
8220                                  * on something sacked before removing it
8221                                  * from the tmap).
8222                                  */
8223                                 if (rsm->r_in_tmap) {
8224                                         nrsm = TAILQ_NEXT(rsm, r_tnext);
8225                                         /*
8226                                          * Now that we have the next
8227                                          * one walk backwards from there.
8228                                          */
8229                                         if (nrsm && nrsm->r_in_tmap)
8230                                                 rack_log_sack_passed(tp, rack, nrsm);
8231                                 }
8232                                 /* Now are we done? */
8233                                 if (SEQ_LT(end, next->r_end) ||
8234                                     (end == next->r_end)) {
8235                                         /* Done with block */
8236                                         goto out;
8237                                 }
8238                                 rack_log_map_chg(tp, rack, &stack_map, rsm, next, MAP_SACK_M1, end, __LINE__);
8239                                 counter_u64_add(rack_sack_used_next_merge, 1);
8240                                 /* Postion for the next block */
8241                                 start = next->r_end;
8242                                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, next);
8243                                 if (rsm == NULL)
8244                                         goto out;
8245                         } else {
8246                                 /**
8247                                  * We can't use any hookery here, so we
8248                                  * need to split the map. We enter like
8249                                  * so:
8250                                  *  rsm      |--------|
8251                                  *  sackblk       |----->
8252                                  * We will add the new block nrsm and
8253                                  * that will be the new portion, and then
8254                                  * fall through after reseting rsm. So we
8255                                  * split and look like this:
8256                                  *  rsm      |----|
8257                                  *  sackblk       |----->
8258                                  *  nrsm          |---|
8259                                  * We then fall through reseting
8260                                  * rsm to nrsm, so the next block
8261                                  * picks it up.
8262                                  */
8263                                 nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
8264                                 if (nrsm == NULL) {
8265                                         /*
8266                                          * failed XXXrrs what can we do but loose the sack
8267                                          * info?
8268                                          */
8269                                         goto out;
8270                                 }
8271                                 counter_u64_add(rack_sack_splits, 1);
8272                                 rack_clone_rsm(rack, nrsm, rsm, start);
8273                                 rsm->r_just_ret = 0;
8274                                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8275 #ifdef INVARIANTS
8276                                 if (insret != NULL) {
8277                                         panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
8278                                               nrsm, insret, rack, rsm);
8279                                 }
8280 #endif
8281                                 if (rsm->r_in_tmap) {
8282                                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8283                                         nrsm->r_in_tmap = 1;
8284                                 }
8285                                 rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M2, end, __LINE__);
8286                                 rsm->r_flags &= (~RACK_HAS_FIN);
8287                                 /* Position us to point to the new nrsm that starts the sack blk */
8288                                 rsm = nrsm;
8289                         }
8290                 } else {
8291                         /* Already sacked this piece */
8292                         counter_u64_add(rack_sack_skipped_acked, 1);
8293                         moved++;
8294                         if (end == rsm->r_end) {
8295                                 /* Done with block */
8296                                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8297                                 goto out;
8298                         } else if (SEQ_LT(end, rsm->r_end)) {
8299                                 /* A partial sack to a already sacked block */
8300                                 moved++;
8301                                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8302                                 goto out;
8303                         } else {
8304                                 /*
8305                                  * The end goes beyond this guy
8306                                  * repostion the start to the
8307                                  * next block.
8308                                  */
8309                                 start = rsm->r_end;
8310                                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8311                                 if (rsm == NULL)
8312                                         goto out;
8313                         }
8314                 }
8315         }
8316         if (SEQ_GEQ(end, rsm->r_end)) {
8317                 /**
8318                  * The end of this block is either beyond this guy or right
8319                  * at this guy. I.e.:
8320                  *  rsm ---                 |-----|
8321                  *  end                     |-----|
8322                  *  <or>
8323                  *  end                     |---------|
8324                  */
8325                 if ((rsm->r_flags & RACK_ACKED) == 0) {
8326                         rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
8327                         changed += (rsm->r_end - rsm->r_start);
8328                         rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
8329                         if (rsm->r_in_tmap) /* should be true */
8330                                 rack_log_sack_passed(tp, rack, rsm);
8331                         /* Is Reordering occuring? */
8332                         if (rsm->r_flags & RACK_SACK_PASSED) {
8333                                 rsm->r_flags &= ~RACK_SACK_PASSED;
8334                                 counter_u64_add(rack_reorder_seen, 1);
8335                                 rack->r_ctl.rc_reorder_ts = cts;
8336                         }
8337                         if (rack->app_limited_needs_set)
8338                                 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
8339                         rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8340                         rsm->r_flags |= RACK_ACKED;
8341                         rsm->r_flags &= ~RACK_TLP;
8342                         if (rsm->r_in_tmap) {
8343                                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8344                                 rsm->r_in_tmap = 0;
8345                         }
8346                         rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_SACK_M3, end, __LINE__);
8347                 } else {
8348                         counter_u64_add(rack_sack_skipped_acked, 1);
8349                         moved++;
8350                 }
8351                 if (end == rsm->r_end) {
8352                         /* This block only - done, setup for next */
8353                         goto out;
8354                 }
8355                 /*
8356                  * There is more not coverend by this rsm move on
8357                  * to the next block in the RB tree.
8358                  */
8359                 nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8360                 start = rsm->r_end;
8361                 rsm = nrsm;
8362                 if (rsm == NULL)
8363                         goto out;
8364                 goto do_rest_ofb;
8365         }
8366         /**
8367          * The end of this sack block is smaller than
8368          * our rsm i.e.:
8369          *  rsm ---                 |-----|
8370          *  end                     |--|
8371          */
8372         if ((rsm->r_flags & RACK_ACKED) == 0) {
8373                 prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8374                 if (prev && (prev->r_flags & RACK_ACKED)) {
8375                         /**
8376                          * Goal, we want the right remainder of rsm to shrink
8377                          * in place and span from (rsm->r_start = end) to rsm->r_end.
8378                          * We want to expand prev to go all the way
8379                          * to prev->r_end <- end.
8380                          * so in the tree we have before:
8381                          *   prev     |--------|         (acked)
8382                          *   rsm               |-------| (non-acked)
8383                          *   sackblk           |-|
8384                          * We churn it so we end up with
8385                          *   prev     |----------|       (acked)
8386                          *   rsm                 |-----| (non-acked)
8387                          *   nrsm              |-| (temporary)
8388                          */
8389                         nrsm = &stack_map;
8390                         memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
8391                         prev->r_end = end;
8392                         rsm->r_start = end;
8393                         /* Now adjust nrsm (stack copy) to be
8394                          * the one that is the small
8395                          * piece that was "sacked".
8396                          */
8397                         nrsm->r_end = end;
8398                         rsm->r_dupack = 0;
8399                         rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8400                         /*
8401                          * Now that the rsm has had its start moved forward
8402                          * lets go ahead and get its new place in the world.
8403                          */
8404                         rack_setup_offset_for_rsm(prev, rsm);
8405                         /*
8406                          * Now nrsm is our new little piece
8407                          * that is acked (which was merged
8408                          * to prev). Update the rtt and changed
8409                          * based on that. Also check for reordering.
8410                          */
8411                         rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
8412                         if (rack->app_limited_needs_set)
8413                                 rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
8414                         changed += (nrsm->r_end - nrsm->r_start);
8415                         rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
8416                         if (nrsm->r_flags & RACK_SACK_PASSED) {
8417                                 counter_u64_add(rack_reorder_seen, 1);
8418                                 rack->r_ctl.rc_reorder_ts = cts;
8419                         }
8420                         rack_log_map_chg(tp, rack, prev, &stack_map, rsm, MAP_SACK_M4, end, __LINE__);
8421                         rsm = prev;
8422                         counter_u64_add(rack_sack_used_prev_merge, 1);
8423                 } else {
8424                         /**
8425                          * This is the case where our previous
8426                          * block is not acked either, so we must
8427                          * split the block in two.
8428                          */
8429                         nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
8430                         if (nrsm == NULL) {
8431                                 /* failed rrs what can we do but loose the sack info? */
8432                                 goto out;
8433                         }
8434                         /**
8435                          * In this case nrsm becomes
8436                          * nrsm->r_start = end;
8437                          * nrsm->r_end = rsm->r_end;
8438                          * which is un-acked.
8439                          * <and>
8440                          * rsm->r_end = nrsm->r_start;
8441                          * i.e. the remaining un-acked
8442                          * piece is left on the left
8443                          * hand side.
8444                          *
8445                          * So we start like this
8446                          * rsm      |----------| (not acked)
8447                          * sackblk  |---|
8448                          * build it so we have
8449                          * rsm      |---|         (acked)
8450                          * nrsm         |------|  (not acked)
8451                          */
8452                         counter_u64_add(rack_sack_splits, 1);
8453                         rack_clone_rsm(rack, nrsm, rsm, end);
8454                         rsm->r_flags &= (~RACK_HAS_FIN);
8455                         rsm->r_just_ret = 0;
8456                         insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8457 #ifdef INVARIANTS
8458                         if (insret != NULL) {
8459                                 panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
8460                                       nrsm, insret, rack, rsm);
8461                         }
8462 #endif
8463                         if (rsm->r_in_tmap) {
8464                                 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8465                                 nrsm->r_in_tmap = 1;
8466                         }
8467                         nrsm->r_dupack = 0;
8468                         rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
8469                         rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
8470                         changed += (rsm->r_end - rsm->r_start);
8471                         rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
8472                         if (rsm->r_in_tmap) /* should be true */
8473                                 rack_log_sack_passed(tp, rack, rsm);
8474                         /* Is Reordering occuring? */
8475                         if (rsm->r_flags & RACK_SACK_PASSED) {
8476                                 rsm->r_flags &= ~RACK_SACK_PASSED;
8477                                 counter_u64_add(rack_reorder_seen, 1);
8478                                 rack->r_ctl.rc_reorder_ts = cts;
8479                         }
8480                         if (rack->app_limited_needs_set)
8481                                 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
8482                         rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8483                         rsm->r_flags |= RACK_ACKED;
8484                         rsm->r_flags &= ~RACK_TLP;
8485                         rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M5, end, __LINE__);
8486                         if (rsm->r_in_tmap) {
8487                                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8488                                 rsm->r_in_tmap = 0;
8489                         }
8490                 }
8491         } else if (start != end){
8492                 /*
8493                  * The block was already acked.
8494                  */
8495                 counter_u64_add(rack_sack_skipped_acked, 1);
8496                 moved++;
8497         }
8498 out:
8499         if (rsm && (rsm->r_flags & RACK_ACKED)) {
8500                 /*
8501                  * Now can we merge where we worked
8502                  * with either the previous or
8503                  * next block?
8504                  */
8505                 next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8506                 while (next) {
8507                     if (next->r_flags & RACK_ACKED) {
8508                         /* yep this and next can be merged */
8509                         rsm = rack_merge_rsm(rack, rsm, next);
8510                         next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8511                     } else
8512                             break;
8513                 }
8514                 /* Now what about the previous? */
8515                 prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8516                 while (prev) {
8517                     if (prev->r_flags & RACK_ACKED) {
8518                         /* yep the previous and this can be merged */
8519                         rsm = rack_merge_rsm(rack, prev, rsm);
8520                         prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8521                     } else
8522                             break;
8523                 }
8524         }
8525         if (used_ref == 0) {
8526                 counter_u64_add(rack_sack_proc_all, 1);
8527         } else {
8528                 counter_u64_add(rack_sack_proc_short, 1);
8529         }
8530         /* Save off the next one for quick reference. */
8531         if (rsm)
8532                 nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8533         else
8534                 nrsm = NULL;
8535         *prsm = rack->r_ctl.rc_sacklast = nrsm;
8536         /* Pass back the moved. */
8537         *moved_two = moved;
8538         return (changed);
8539 }
8540
8541 static void inline
8542 rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack)
8543 {
8544         struct rack_sendmap *tmap;
8545
8546         tmap = NULL;
8547         while (rsm && (rsm->r_flags & RACK_ACKED)) {
8548                 /* Its no longer sacked, mark it so */
8549                 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
8550 #ifdef INVARIANTS
8551                 if (rsm->r_in_tmap) {
8552                         panic("rack:%p rsm:%p flags:0x%x in tmap?",
8553                               rack, rsm, rsm->r_flags);
8554                 }
8555 #endif
8556                 rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS);
8557                 /* Rebuild it into our tmap */
8558                 if (tmap == NULL) {
8559                         TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8560                         tmap = rsm;
8561                 } else {
8562                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext);
8563                         tmap = rsm;
8564                 }
8565                 tmap->r_in_tmap = 1;
8566                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8567         }
8568         /*
8569          * Now lets possibly clear the sack filter so we start
8570          * recognizing sacks that cover this area.
8571          */
8572         sack_filter_clear(&rack->r_ctl.rack_sf, th_ack);
8573
8574 }
8575
8576 static void
8577 rack_do_decay(struct tcp_rack *rack)
8578 {
8579         struct timeval res;
8580
8581 #define timersub(tvp, uvp, vvp)                                         \
8582         do {                                                            \
8583                 (vvp)->tv_sec = (tvp)->tv_sec - (uvp)->tv_sec;          \
8584                 (vvp)->tv_usec = (tvp)->tv_usec - (uvp)->tv_usec;       \
8585                 if ((vvp)->tv_usec < 0) {                               \
8586                         (vvp)->tv_sec--;                                \
8587                         (vvp)->tv_usec += 1000000;                      \
8588                 }                                                       \
8589         } while (0)
8590
8591         timersub(&rack->r_ctl.act_rcv_time, &rack->r_ctl.rc_last_time_decay, &res);
8592 #undef timersub
8593
8594         rack->r_ctl.input_pkt++;
8595         if ((rack->rc_in_persist) ||
8596             (res.tv_sec >= 1) ||
8597             (rack->rc_tp->snd_max == rack->rc_tp->snd_una)) {
8598                 /*
8599                  * Check for decay of non-SAD,
8600                  * we want all SAD detection metrics to
8601                  * decay 1/4 per second (or more) passed.
8602                  */
8603                 uint32_t pkt_delta;
8604
8605                 pkt_delta = rack->r_ctl.input_pkt - rack->r_ctl.saved_input_pkt;
8606                 /* Update our saved tracking values */
8607                 rack->r_ctl.saved_input_pkt = rack->r_ctl.input_pkt;
8608                 rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
8609                 /* Now do we escape without decay? */
8610 #ifdef NETFLIX_EXP_DETECTION
8611                 if (rack->rc_in_persist ||
8612                     (rack->rc_tp->snd_max == rack->rc_tp->snd_una) ||
8613                     (pkt_delta < tcp_sad_low_pps)){
8614                         /*
8615                          * We don't decay idle connections
8616                          * or ones that have a low input pps.
8617                          */
8618                         return;
8619                 }
8620                 /* Decay the counters */
8621                 rack->r_ctl.ack_count = ctf_decay_count(rack->r_ctl.ack_count,
8622                                                         tcp_sad_decay_val);
8623                 rack->r_ctl.sack_count = ctf_decay_count(rack->r_ctl.sack_count,
8624                                                          tcp_sad_decay_val);
8625                 rack->r_ctl.sack_moved_extra = ctf_decay_count(rack->r_ctl.sack_moved_extra,
8626                                                                tcp_sad_decay_val);
8627                 rack->r_ctl.sack_noextra_move = ctf_decay_count(rack->r_ctl.sack_noextra_move,
8628                                                                 tcp_sad_decay_val);
8629 #endif
8630         }
8631 }
8632
8633 static void
8634 rack_process_to_cumack(struct tcpcb *tp, struct tcp_rack *rack, register uint32_t th_ack, uint32_t cts, struct tcpopt *to)
8635 {
8636         struct rack_sendmap *rsm, *rm;
8637
8638         /*
8639          * The ACK point is advancing to th_ack, we must drop off
8640          * the packets in the rack log and calculate any eligble
8641          * RTT's.
8642          */
8643         rack->r_wanted_output = 1;
8644 more:
8645         rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
8646         if (rsm == NULL) {
8647                 if ((th_ack - 1) == tp->iss) {
8648                         /*
8649                          * For the SYN incoming case we will not
8650                          * have called tcp_output for the sending of
8651                          * the SYN, so there will be no map. All
8652                          * other cases should probably be a panic.
8653                          */
8654                         return;
8655                 }
8656                 if (tp->t_flags & TF_SENTFIN) {
8657                         /* if we sent a FIN we often will not have map */
8658                         return;
8659                 }
8660 #ifdef INVARIANTS
8661                 panic("No rack map tp:%p for state:%d ack:%u rack:%p snd_una:%u snd_max:%u snd_nxt:%u\n",
8662                       tp,
8663                       tp->t_state, th_ack, rack,
8664                       tp->snd_una, tp->snd_max, tp->snd_nxt);
8665 #endif
8666                 return;
8667         }
8668         if (SEQ_LT(th_ack, rsm->r_start)) {
8669                 /* Huh map is missing this */
8670 #ifdef INVARIANTS
8671                 printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n",
8672                        rsm->r_start,
8673                        th_ack, tp->t_state, rack->r_state);
8674 #endif
8675                 return;
8676         }
8677         rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED, th_ack);
8678         /* Now do we consume the whole thing? */
8679         if (SEQ_GEQ(th_ack, rsm->r_end)) {
8680                 /* Its all consumed. */
8681                 uint32_t left;
8682                 uint8_t newly_acked;
8683
8684                 rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_FREE, rsm->r_end, __LINE__);
8685                 rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
8686                 rsm->r_rtr_bytes = 0;
8687                 /* Record the time of highest cumack sent */
8688                 rack->r_ctl.rc_gp_cumack_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8689                 rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8690 #ifdef INVARIANTS
8691                 if (rm != rsm) {
8692                         panic("removing head in rack:%p rsm:%p rm:%p",
8693                               rack, rsm, rm);
8694                 }
8695 #endif
8696                 if (rsm->r_in_tmap) {
8697                         TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8698                         rsm->r_in_tmap = 0;
8699                 }
8700                 newly_acked = 1;
8701                 if (rsm->r_flags & RACK_ACKED) {
8702                         /*
8703                          * It was acked on the scoreboard -- remove
8704                          * it from total
8705                          */
8706                         rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
8707                         newly_acked = 0;
8708                 } else if (rsm->r_flags & RACK_SACK_PASSED) {
8709                         /*
8710                          * There are segments ACKED on the
8711                          * scoreboard further up. We are seeing
8712                          * reordering.
8713                          */
8714                         rsm->r_flags &= ~RACK_SACK_PASSED;
8715                         counter_u64_add(rack_reorder_seen, 1);
8716                         rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8717                         rsm->r_flags |= RACK_ACKED;
8718                         rack->r_ctl.rc_reorder_ts = cts;
8719                         if (rack->r_ent_rec_ns) {
8720                                 /*
8721                                  * We have sent no more, and we saw an sack
8722                                  * then ack arrive.
8723                                  */
8724                                 rack->r_might_revert = 1;
8725                         }
8726                 }
8727                 if ((rsm->r_flags & RACK_TO_REXT) &&
8728                     (tp->t_flags & TF_RCVD_TSTMP) &&
8729                     (to->to_flags & TOF_TS) &&
8730                     (tp->t_flags & TF_PREVVALID)) {
8731                         /*
8732                          * We can use the timestamp to see
8733                          * if this retransmission was from the
8734                          * first transmit. If so we made a mistake.
8735                          */
8736                         tp->t_flags &= ~TF_PREVVALID;
8737                         if (to->to_tsecr == rack_ts_to_msec(rsm->r_tim_lastsent[0])) {
8738                                 /* The first transmit is what this ack is for */
8739                                 rack_cong_signal(tp, CC_RTO_ERR, th_ack);
8740                         }
8741                 }
8742                 left = th_ack - rsm->r_end;
8743                 if (rack->app_limited_needs_set && newly_acked)
8744                         rack_need_set_test(tp, rack, rsm, th_ack, __LINE__, RACK_USE_END_OR_THACK);
8745                 /* Free back to zone */
8746                 rack_free(rack, rsm);
8747                 if (left) {
8748                         goto more;
8749                 }
8750                 /* Check for reneging */
8751                 rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
8752                 if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) {
8753                         /*
8754                          * The peer has moved snd_una up to
8755                          * the edge of this send, i.e. one
8756                          * that it had previously acked. The only
8757                          * way that can be true if the peer threw
8758                          * away data (space issues) that it had
8759                          * previously sacked (else it would have
8760                          * given us snd_una up to (rsm->r_end).
8761                          * We need to undo the acked markings here.
8762                          *
8763                          * Note we have to look to make sure th_ack is
8764                          * our rsm->r_start in case we get an old ack
8765                          * where th_ack is behind snd_una.
8766                          */
8767                         rack_peer_reneges(rack, rsm, th_ack);
8768                 }
8769                 return;
8770         }
8771         if (rsm->r_flags & RACK_ACKED) {
8772                 /*
8773                  * It was acked on the scoreboard -- remove it from
8774                  * total for the part being cum-acked.
8775                  */
8776                 rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
8777         }
8778         /*
8779          * Clear the dup ack count for
8780          * the piece that remains.
8781          */
8782         rsm->r_dupack = 0;
8783         rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8784         if (rsm->r_rtr_bytes) {
8785                 /*
8786                  * It was retransmitted adjust the
8787                  * sack holes for what was acked.
8788                  */
8789                 int ack_am;
8790
8791                 ack_am = (th_ack - rsm->r_start);
8792                 if (ack_am >= rsm->r_rtr_bytes) {
8793                         rack->r_ctl.rc_holes_rxt -= ack_am;
8794                         rsm->r_rtr_bytes -= ack_am;
8795                 }
8796         }
8797         /*
8798          * Update where the piece starts and record
8799          * the time of send of highest cumack sent.
8800          */
8801         rack->r_ctl.rc_gp_cumack_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8802         rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_TRIM_HEAD, th_ack, __LINE__);
8803         /* Now we need to move our offset forward too */
8804         if (rsm->orig_m_len != rsm->m->m_len) {
8805                 /* Fix up the orig_m_len and possibly the mbuf offset */
8806                 rack_adjust_orig_mlen(rsm);
8807         }
8808         rsm->soff += (th_ack - rsm->r_start);
8809         rsm->r_start = th_ack;
8810         /* Now do we need to move the mbuf fwd too? */
8811         while (rsm->soff >= rsm->m->m_len) {
8812                 rsm->soff -= rsm->m->m_len;
8813                 rsm->m = rsm->m->m_next;
8814                 KASSERT((rsm->m != NULL),
8815                         (" nrsm:%p hit at soff:%u null m",
8816                          rsm, rsm->soff));
8817         }
8818         rsm->orig_m_len = rsm->m->m_len;
8819         if (rack->app_limited_needs_set)
8820                 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_BEG);
8821 }
8822
8823 static void
8824 rack_handle_might_revert(struct tcpcb *tp, struct tcp_rack *rack)
8825 {
8826         struct rack_sendmap *rsm;
8827         int sack_pass_fnd = 0;
8828
8829         if (rack->r_might_revert) {
8830                 /*
8831                  * Ok we have reordering, have not sent anything, we
8832                  * might want to revert the congestion state if nothing
8833                  * further has SACK_PASSED on it. Lets check.
8834                  *
8835                  * We also get here when we have DSACKs come in for
8836                  * all the data that we FR'd. Note that a rxt or tlp
8837                  * timer clears this from happening.
8838                  */
8839
8840                 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
8841                         if (rsm->r_flags & RACK_SACK_PASSED) {
8842                                 sack_pass_fnd = 1;
8843                                 break;
8844                         }
8845                 }
8846                 if (sack_pass_fnd == 0) {
8847                         /*
8848                          * We went into recovery
8849                          * incorrectly due to reordering!
8850                          */
8851                         int orig_cwnd;
8852
8853                         rack->r_ent_rec_ns = 0;
8854                         orig_cwnd = tp->snd_cwnd;
8855                         tp->snd_cwnd = rack->r_ctl.rc_cwnd_at_erec;
8856                         tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at_erec;
8857                         tp->snd_recover = tp->snd_una;
8858                         rack_log_to_prr(rack, 14, orig_cwnd);
8859                         EXIT_RECOVERY(tp->t_flags);
8860                 }
8861                 rack->r_might_revert = 0;
8862         }
8863 }
8864
8865 #ifdef NETFLIX_EXP_DETECTION
8866 static void
8867 rack_do_detection(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t bytes_this_ack, uint32_t segsiz)
8868 {
8869         if ((rack->do_detection || tcp_force_detection) &&
8870             tcp_sack_to_ack_thresh &&
8871             tcp_sack_to_move_thresh &&
8872             ((rack->r_ctl.rc_num_maps_alloced > tcp_map_minimum) || rack->sack_attack_disable)) {
8873                 /*
8874                  * We have thresholds set to find
8875                  * possible attackers and disable sack.
8876                  * Check them.
8877                  */
8878                 uint64_t ackratio, moveratio, movetotal;
8879
8880                 /* Log detecting */
8881                 rack_log_sad(rack, 1);
8882                 ackratio = (uint64_t)(rack->r_ctl.sack_count);
8883                 ackratio *= (uint64_t)(1000);
8884                 if (rack->r_ctl.ack_count)
8885                         ackratio /= (uint64_t)(rack->r_ctl.ack_count);
8886                 else {
8887                         /* We really should not hit here */
8888                         ackratio = 1000;
8889                 }
8890                 if ((rack->sack_attack_disable == 0) &&
8891                     (ackratio > rack_highest_sack_thresh_seen))
8892                         rack_highest_sack_thresh_seen = (uint32_t)ackratio;
8893                 movetotal = rack->r_ctl.sack_moved_extra;
8894                 movetotal += rack->r_ctl.sack_noextra_move;
8895                 moveratio = rack->r_ctl.sack_moved_extra;
8896                 moveratio *= (uint64_t)1000;
8897                 if (movetotal)
8898                         moveratio /= movetotal;
8899                 else {
8900                         /* No moves, thats pretty good */
8901                         moveratio = 0;
8902                 }
8903                 if ((rack->sack_attack_disable == 0) &&
8904                     (moveratio > rack_highest_move_thresh_seen))
8905                         rack_highest_move_thresh_seen = (uint32_t)moveratio;
8906                 if (rack->sack_attack_disable == 0) {
8907                         if ((ackratio > tcp_sack_to_ack_thresh) &&
8908                             (moveratio > tcp_sack_to_move_thresh)) {
8909                                 /* Disable sack processing */
8910                                 rack->sack_attack_disable = 1;
8911                                 if (rack->r_rep_attack == 0) {
8912                                         rack->r_rep_attack = 1;
8913                                         counter_u64_add(rack_sack_attacks_detected, 1);
8914                                 }
8915                                 if (tcp_attack_on_turns_on_logging) {
8916                                         /*
8917                                          * Turn on logging, used for debugging
8918                                          * false positives.
8919                                          */
8920                                         rack->rc_tp->t_logstate = tcp_attack_on_turns_on_logging;
8921                                 }
8922                                 /* Clamp the cwnd at flight size */
8923                                 rack->r_ctl.rc_saved_cwnd = rack->rc_tp->snd_cwnd;
8924                                 rack->rc_tp->snd_cwnd = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
8925                                 rack_log_sad(rack, 2);
8926                         }
8927                 } else {
8928                         /* We are sack-disabled check for false positives */
8929                         if ((ackratio <= tcp_restoral_thresh) ||
8930                             (rack->r_ctl.rc_num_maps_alloced  < tcp_map_minimum)) {
8931                                 rack->sack_attack_disable = 0;
8932                                 rack_log_sad(rack, 3);
8933                                 /* Restart counting */
8934                                 rack->r_ctl.sack_count = 0;
8935                                 rack->r_ctl.sack_moved_extra = 0;
8936                                 rack->r_ctl.sack_noextra_move = 1;
8937                                 rack->r_ctl.ack_count = max(1,
8938                                       (bytes_this_ack / segsiz));
8939
8940                                 if (rack->r_rep_reverse == 0) {
8941                                         rack->r_rep_reverse = 1;
8942                                         counter_u64_add(rack_sack_attacks_reversed, 1);
8943                                 }
8944                                 /* Restore the cwnd */
8945                                 if (rack->r_ctl.rc_saved_cwnd > rack->rc_tp->snd_cwnd)
8946                                         rack->rc_tp->snd_cwnd = rack->r_ctl.rc_saved_cwnd;
8947                         }
8948                 }
8949         }
8950 }
8951 #endif
8952
8953 static void
8954 rack_note_dsack(struct tcp_rack *rack, tcp_seq start, tcp_seq end)
8955 {
8956
8957         uint32_t am;
8958
8959         if (SEQ_GT(end, start))
8960                 am = end - start;
8961         else
8962                 am = 0;
8963         /*
8964          * We keep track of how many DSACK blocks we get
8965          * after a recovery incident.
8966          */
8967         rack->r_ctl.dsack_byte_cnt += am;
8968         if (!IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
8969             rack->r_ctl.retran_during_recovery &&
8970             (rack->r_ctl.dsack_byte_cnt >= rack->r_ctl.retran_during_recovery)) {
8971                 /*
8972                  * False recovery most likely culprit is reordering. If
8973                  * nothing else is missing we need to revert.
8974                  */
8975                 rack->r_might_revert = 1;
8976                 rack_handle_might_revert(rack->rc_tp, rack);
8977                 rack->r_might_revert = 0;
8978                 rack->r_ctl.retran_during_recovery = 0;
8979                 rack->r_ctl.dsack_byte_cnt = 0;
8980         }
8981 }
8982
8983 static void
8984 rack_update_prr(struct tcpcb *tp, struct tcp_rack *rack, uint32_t changed, tcp_seq th_ack)
8985 {
8986         /* Deal with changed and PRR here (in recovery only) */
8987         uint32_t pipe, snd_una;
8988
8989         rack->r_ctl.rc_prr_delivered += changed;
8990
8991         if (sbavail(&rack->rc_inp->inp_socket->so_snd) <= (tp->snd_max - tp->snd_una)) {
8992                 /*
8993                  * It is all outstanding, we are application limited
8994                  * and thus we don't need more room to send anything.
8995                  * Note we use tp->snd_una here and not th_ack because
8996                  * the data as yet not been cut from the sb.
8997                  */
8998                 rack->r_ctl.rc_prr_sndcnt = 0;
8999                 return;
9000         }
9001         /* Compute prr_sndcnt */
9002         if (SEQ_GT(tp->snd_una, th_ack)) {
9003                 snd_una = tp->snd_una;
9004         } else {
9005                 snd_una = th_ack;
9006         }
9007         pipe = ((tp->snd_max - snd_una) - rack->r_ctl.rc_sacked) + rack->r_ctl.rc_holes_rxt;
9008         if (pipe > tp->snd_ssthresh) {
9009                 long sndcnt;
9010
9011                 sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh;
9012                 if (rack->r_ctl.rc_prr_recovery_fs > 0)
9013                         sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs;
9014                 else {
9015                         rack->r_ctl.rc_prr_sndcnt = 0;
9016                         rack_log_to_prr(rack, 9, 0);
9017                         sndcnt = 0;
9018                 }
9019                 sndcnt++;
9020                 if (sndcnt > (long)rack->r_ctl.rc_prr_out)
9021                         sndcnt -= rack->r_ctl.rc_prr_out;
9022                 else
9023                         sndcnt = 0;
9024                 rack->r_ctl.rc_prr_sndcnt = sndcnt;
9025                 rack_log_to_prr(rack, 10, 0);
9026         } else {
9027                 uint32_t limit;
9028
9029                 if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out)
9030                         limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out);
9031                 else
9032                         limit = 0;
9033                 if (changed > limit)
9034                         limit = changed;
9035                 limit += ctf_fixed_maxseg(tp);
9036                 if (tp->snd_ssthresh > pipe) {
9037                         rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit);
9038                         rack_log_to_prr(rack, 11, 0);
9039                 } else {
9040                         rack->r_ctl.rc_prr_sndcnt = min(0, limit);
9041                         rack_log_to_prr(rack, 12, 0);
9042                 }
9043         }
9044 }
9045
9046 static void
9047 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, int entered_recovery, int dup_ack_struck)
9048 {
9049         uint32_t changed;
9050         struct tcp_rack *rack;
9051         struct rack_sendmap *rsm;
9052         struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
9053         register uint32_t th_ack;
9054         int32_t i, j, k, num_sack_blks = 0;
9055         uint32_t cts, acked, ack_point, sack_changed = 0;
9056         int loop_start = 0, moved_two = 0;
9057         uint32_t tsused;
9058
9059
9060         INP_WLOCK_ASSERT(tp->t_inpcb);
9061         if (th->th_flags & TH_RST) {
9062                 /* We don't log resets */
9063                 return;
9064         }
9065         rack = (struct tcp_rack *)tp->t_fb_ptr;
9066         cts = tcp_get_usecs(NULL);
9067         rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9068         changed = 0;
9069         th_ack = th->th_ack;
9070         if (rack->sack_attack_disable == 0)
9071                 rack_do_decay(rack);
9072         if (BYTES_THIS_ACK(tp, th) >= ctf_fixed_maxseg(rack->rc_tp)) {
9073                 /*
9074                  * You only get credit for
9075                  * MSS and greater (and you get extra
9076                  * credit for larger cum-ack moves).
9077                  */
9078                 int ac;
9079
9080                 ac = BYTES_THIS_ACK(tp, th) / ctf_fixed_maxseg(rack->rc_tp);
9081                 rack->r_ctl.ack_count += ac;
9082                 counter_u64_add(rack_ack_total, ac);
9083         }
9084         if (rack->r_ctl.ack_count > 0xfff00000) {
9085                 /*
9086                  * reduce the number to keep us under
9087                  * a uint32_t.
9088                  */
9089                 rack->r_ctl.ack_count /= 2;
9090                 rack->r_ctl.sack_count /= 2;
9091         }
9092         if (SEQ_GT(th_ack, tp->snd_una)) {
9093                 rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__);
9094                 tp->t_acktime = ticks;
9095         }
9096         if (rsm && SEQ_GT(th_ack, rsm->r_start))
9097                 changed = th_ack - rsm->r_start;
9098         if (changed) {
9099                 rack_process_to_cumack(tp, rack, th_ack, cts, to);
9100         }
9101         if ((to->to_flags & TOF_SACK) == 0) {
9102                 /* We are done nothing left and no sack. */
9103                 rack_handle_might_revert(tp, rack);
9104                 /*
9105                  * For cases where we struck a dup-ack
9106                  * with no SACK, add to the changes so
9107                  * PRR will work right.
9108                  */
9109                 if (dup_ack_struck && (changed == 0)) {
9110                         changed += ctf_fixed_maxseg(rack->rc_tp);
9111                 }
9112                 goto out;
9113         }
9114         /* Sack block processing */
9115         if (SEQ_GT(th_ack, tp->snd_una))
9116                 ack_point = th_ack;
9117         else
9118                 ack_point = tp->snd_una;
9119         for (i = 0; i < to->to_nsacks; i++) {
9120                 bcopy((to->to_sacks + i * TCPOLEN_SACK),
9121                       &sack, sizeof(sack));
9122                 sack.start = ntohl(sack.start);
9123                 sack.end = ntohl(sack.end);
9124                 if (SEQ_GT(sack.end, sack.start) &&
9125                     SEQ_GT(sack.start, ack_point) &&
9126                     SEQ_LT(sack.start, tp->snd_max) &&
9127                     SEQ_GT(sack.end, ack_point) &&
9128                     SEQ_LEQ(sack.end, tp->snd_max)) {
9129                         sack_blocks[num_sack_blks] = sack;
9130                         num_sack_blks++;
9131 #ifdef NETFLIX_STATS
9132                 } else if (SEQ_LEQ(sack.start, th_ack) &&
9133                            SEQ_LEQ(sack.end, th_ack)) {
9134                         /*
9135                          * Its a D-SACK block.
9136                          */
9137                         tcp_record_dsack(sack.start, sack.end);
9138 #endif
9139                         rack_note_dsack(rack, sack.start, sack.end);
9140                 }
9141         }
9142         /*
9143          * Sort the SACK blocks so we can update the rack scoreboard with
9144          * just one pass.
9145          */
9146         num_sack_blks = sack_filter_blks(&rack->r_ctl.rack_sf, sack_blocks,
9147                                          num_sack_blks, th->th_ack);
9148         ctf_log_sack_filter(rack->rc_tp, num_sack_blks, sack_blocks);
9149         if (num_sack_blks == 0) {
9150                 /* Nothing to sack (DSACKs?) */
9151                 goto out_with_totals;
9152         }
9153         if (num_sack_blks < 2) {
9154                 /* Only one, we don't need to sort */
9155                 goto do_sack_work;
9156         }
9157         /* Sort the sacks */
9158         for (i = 0; i < num_sack_blks; i++) {
9159                 for (j = i + 1; j < num_sack_blks; j++) {
9160                         if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
9161                                 sack = sack_blocks[i];
9162                                 sack_blocks[i] = sack_blocks[j];
9163                                 sack_blocks[j] = sack;
9164                         }
9165                 }
9166         }
9167         /*
9168          * Now are any of the sack block ends the same (yes some
9169          * implementations send these)?
9170          */
9171 again:
9172         if (num_sack_blks == 0)
9173                 goto out_with_totals;
9174         if (num_sack_blks > 1) {
9175                 for (i = 0; i < num_sack_blks; i++) {
9176                         for (j = i + 1; j < num_sack_blks; j++) {
9177                                 if (sack_blocks[i].end == sack_blocks[j].end) {
9178                                         /*
9179                                          * Ok these two have the same end we
9180                                          * want the smallest end and then
9181                                          * throw away the larger and start
9182                                          * again.
9183                                          */
9184                                         if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
9185                                                 /*
9186                                                  * The second block covers
9187                                                  * more area use that
9188                                                  */
9189                                                 sack_blocks[i].start = sack_blocks[j].start;
9190                                         }
9191                                         /*
9192                                          * Now collapse out the dup-sack and
9193                                          * lower the count
9194                                          */
9195                                         for (k = (j + 1); k < num_sack_blks; k++) {
9196                                                 sack_blocks[j].start = sack_blocks[k].start;
9197                                                 sack_blocks[j].end = sack_blocks[k].end;
9198                                                 j++;
9199                                         }
9200                                         num_sack_blks--;
9201                                         goto again;
9202                                 }
9203                         }
9204                 }
9205         }
9206 do_sack_work:
9207         /*
9208          * First lets look to see if
9209          * we have retransmitted and
9210          * can use the transmit next?
9211          */
9212         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
9213         if (rsm &&
9214             SEQ_GT(sack_blocks[0].end, rsm->r_start) &&
9215             SEQ_LT(sack_blocks[0].start, rsm->r_end)) {
9216                 /*
9217                  * We probably did the FR and the next
9218                  * SACK in continues as we would expect.
9219                  */
9220                 acked = rack_proc_sack_blk(tp, rack, &sack_blocks[0], to, &rsm, cts, &moved_two);
9221                 if (acked) {
9222                         rack->r_wanted_output = 1;
9223                         changed += acked;
9224                         sack_changed += acked;
9225                 }
9226                 if (num_sack_blks == 1) {
9227                         /*
9228                          * This is what we would expect from
9229                          * a normal implementation to happen
9230                          * after we have retransmitted the FR,
9231                          * i.e the sack-filter pushes down
9232                          * to 1 block and the next to be retransmitted
9233                          * is the sequence in the sack block (has more
9234                          * are acked). Count this as ACK'd data to boost
9235                          * up the chances of recovering any false positives.
9236                          */
9237                         rack->r_ctl.ack_count += (acked / ctf_fixed_maxseg(rack->rc_tp));
9238                         counter_u64_add(rack_ack_total, (acked / ctf_fixed_maxseg(rack->rc_tp)));
9239                         counter_u64_add(rack_express_sack, 1);
9240                         if (rack->r_ctl.ack_count > 0xfff00000) {
9241                                 /*
9242                                  * reduce the number to keep us under
9243                                  * a uint32_t.
9244                                  */
9245                                 rack->r_ctl.ack_count /= 2;
9246                                 rack->r_ctl.sack_count /= 2;
9247                         }
9248                         goto out_with_totals;
9249                 } else {
9250                         /*
9251                          * Start the loop through the
9252                          * rest of blocks, past the first block.
9253                          */
9254                         moved_two = 0;
9255                         loop_start = 1;
9256                 }
9257         }
9258         /* Its a sack of some sort */
9259         rack->r_ctl.sack_count++;
9260         if (rack->r_ctl.sack_count > 0xfff00000) {
9261                 /*
9262                  * reduce the number to keep us under
9263                  * a uint32_t.
9264                  */
9265                 rack->r_ctl.ack_count /= 2;
9266                 rack->r_ctl.sack_count /= 2;
9267         }
9268         counter_u64_add(rack_sack_total, 1);
9269         if (rack->sack_attack_disable) {
9270                 /* An attacker disablement is in place */
9271                 if (num_sack_blks > 1) {
9272                         rack->r_ctl.sack_count += (num_sack_blks - 1);
9273                         rack->r_ctl.sack_moved_extra++;
9274                         counter_u64_add(rack_move_some, 1);
9275                         if (rack->r_ctl.sack_moved_extra > 0xfff00000) {
9276                                 rack->r_ctl.sack_moved_extra /= 2;
9277                                 rack->r_ctl.sack_noextra_move /= 2;
9278                         }
9279                 }
9280                 goto out;
9281         }
9282         rsm = rack->r_ctl.rc_sacklast;
9283         for (i = loop_start; i < num_sack_blks; i++) {
9284                 acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts, &moved_two);
9285                 if (acked) {
9286                         rack->r_wanted_output = 1;
9287                         changed += acked;
9288                         sack_changed += acked;
9289                 }
9290                 if (moved_two) {
9291                         /*
9292                          * If we did not get a SACK for at least a MSS and
9293                          * had to move at all, or if we moved more than our
9294                          * threshold, it counts against the "extra" move.
9295                          */
9296                         rack->r_ctl.sack_moved_extra += moved_two;
9297                         counter_u64_add(rack_move_some, 1);
9298                 } else {
9299                         /*
9300                          * else we did not have to move
9301                          * any more than we would expect.
9302                          */
9303                         rack->r_ctl.sack_noextra_move++;
9304                         counter_u64_add(rack_move_none, 1);
9305                 }
9306                 if (moved_two && (acked < ctf_fixed_maxseg(rack->rc_tp))) {
9307                         /*
9308                          * If the SACK was not a full MSS then
9309                          * we add to sack_count the number of
9310                          * MSS's (or possibly more than
9311                          * a MSS if its a TSO send) we had to skip by.
9312                          */
9313                         rack->r_ctl.sack_count += moved_two;
9314                         counter_u64_add(rack_sack_total, moved_two);
9315                 }
9316                 /*
9317                  * Now we need to setup for the next
9318                  * round. First we make sure we won't
9319                  * exceed the size of our uint32_t on
9320                  * the various counts, and then clear out
9321                  * moved_two.
9322                  */
9323                 if ((rack->r_ctl.sack_moved_extra > 0xfff00000) ||
9324                     (rack->r_ctl.sack_noextra_move > 0xfff00000)) {
9325                         rack->r_ctl.sack_moved_extra /= 2;
9326                         rack->r_ctl.sack_noextra_move /= 2;
9327                 }
9328                 if (rack->r_ctl.sack_count > 0xfff00000) {
9329                         rack->r_ctl.ack_count /= 2;
9330                         rack->r_ctl.sack_count /= 2;
9331                 }
9332                 moved_two = 0;
9333         }
9334 out_with_totals:
9335         if (num_sack_blks > 1) {
9336                 /*
9337                  * You get an extra stroke if
9338                  * you have more than one sack-blk, this
9339                  * could be where we are skipping forward
9340                  * and the sack-filter is still working, or
9341                  * it could be an attacker constantly
9342                  * moving us.
9343                  */
9344                 rack->r_ctl.sack_moved_extra++;
9345                 counter_u64_add(rack_move_some, 1);
9346         }
9347 out:
9348 #ifdef NETFLIX_EXP_DETECTION
9349         rack_do_detection(tp, rack, BYTES_THIS_ACK(tp, th), ctf_fixed_maxseg(rack->rc_tp));
9350 #endif
9351         if (changed) {
9352                 /* Something changed cancel the rack timer */
9353                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
9354         }
9355         tsused = tcp_get_usecs(NULL);
9356         rsm = tcp_rack_output(tp, rack, tsused);
9357         if ((!IN_FASTRECOVERY(tp->t_flags)) &&
9358             rsm) {
9359                 /* Enter recovery */
9360                 rack->r_ctl.rc_rsm_start = rsm->r_start;
9361                 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
9362                 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
9363                 entered_recovery = 1;
9364                 rack_cong_signal(tp, CC_NDUPACK, tp->snd_una);
9365                 /*
9366                  * When we enter recovery we need to assure we send
9367                  * one packet.
9368                  */
9369                 if (rack->rack_no_prr == 0) {
9370                         rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
9371                         rack_log_to_prr(rack, 8, 0);
9372                 }
9373                 rack->r_timer_override = 1;
9374                 rack->r_early = 0;
9375                 rack->r_ctl.rc_agg_early = 0;
9376         } else if (IN_FASTRECOVERY(tp->t_flags) &&
9377                    rsm &&
9378                    (rack->r_rr_config == 3)) {
9379                 /*
9380                  * Assure we can output and we get no
9381                  * remembered pace time except the retransmit.
9382                  */
9383                 rack->r_timer_override = 1;
9384                 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
9385                 rack->r_ctl.rc_resend = rsm;
9386         }
9387         if (IN_FASTRECOVERY(tp->t_flags) &&
9388             (rack->rack_no_prr == 0) &&
9389             (entered_recovery == 0)) {
9390                 rack_update_prr(tp, rack, changed, th_ack);
9391                 if ((rsm && (rack->r_ctl.rc_prr_sndcnt >= ctf_fixed_maxseg(tp)) &&
9392                      ((rack->rc_inp->inp_in_hpts == 0) &&
9393                       ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)))) {
9394                         /*
9395                          * If you are pacing output you don't want
9396                          * to override.
9397                          */
9398                         rack->r_early = 0;
9399                         rack->r_ctl.rc_agg_early = 0;
9400                         rack->r_timer_override = 1;
9401                 }
9402         }
9403 }
9404
9405 static void
9406 rack_strike_dupack(struct tcp_rack *rack)
9407 {
9408         struct rack_sendmap *rsm;
9409
9410         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
9411         while (rsm && (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
9412                 rsm = TAILQ_NEXT(rsm, r_tnext);
9413         }
9414         if (rsm && (rsm->r_dupack < 0xff)) {
9415                 rsm->r_dupack++;
9416                 if (rsm->r_dupack >= DUP_ACK_THRESHOLD) {
9417                         struct timeval tv;
9418                         uint32_t cts;
9419                         /*
9420                          * Here we see if we need to retransmit. For
9421                          * a SACK type connection if enough time has passed
9422                          * we will get a return of the rsm. For a non-sack
9423                          * connection we will get the rsm returned if the
9424                          * dupack value is 3 or more.
9425                          */
9426                         cts = tcp_get_usecs(&tv);
9427                         rack->r_ctl.rc_resend = tcp_rack_output(rack->rc_tp, rack, cts);
9428                         if (rack->r_ctl.rc_resend != NULL) {
9429                                 if (!IN_FASTRECOVERY(rack->rc_tp->t_flags)) {
9430                                         rack_cong_signal(rack->rc_tp, CC_NDUPACK,
9431                                                          rack->rc_tp->snd_una);
9432                                 }
9433                                 rack->r_wanted_output = 1;
9434                                 rack->r_timer_override = 1;
9435                                 rack_log_retran_reason(rack, rsm, __LINE__, 1, 3);
9436                         }
9437                 } else {
9438                         rack_log_retran_reason(rack, rsm, __LINE__, 0, 3);
9439                 }
9440         }
9441 }
9442
9443 static void
9444 rack_check_bottom_drag(struct tcpcb *tp,
9445                        struct tcp_rack *rack,
9446                        struct socket *so, int32_t acked)
9447 {
9448         uint32_t segsiz, minseg;
9449
9450         segsiz = ctf_fixed_maxseg(tp);
9451         minseg = segsiz;
9452
9453         if (tp->snd_max == tp->snd_una) {
9454                 /*
9455                  * We are doing dynamic pacing and we are way
9456                  * under. Basically everything got acked while
9457                  * we were still waiting on the pacer to expire.
9458                  *
9459                  * This means we need to boost the b/w in
9460                  * addition to any earlier boosting of
9461                  * the multipler.
9462                  */
9463                 rack->rc_dragged_bottom = 1;
9464                 rack_validate_multipliers_at_or_above100(rack);
9465                 /*
9466                  * Lets use the segment bytes acked plus
9467                  * the lowest RTT seen as the basis to
9468                  * form a b/w estimate. This will be off
9469                  * due to the fact that the true estimate
9470                  * should be around 1/2 the time of the RTT
9471                  * but we can settle for that.
9472                  */
9473                 if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_VALID) &&
9474                     acked) {
9475                         uint64_t bw, calc_bw, rtt;
9476
9477                         rtt = rack->r_ctl.rack_rs.rs_us_rtt;
9478                         if (rtt == 0) {
9479                                 /* no us sample is there a ms one? */
9480                                 if (rack->r_ctl.rack_rs.rs_rtt_lowest) {
9481                                         rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
9482                                 } else {
9483                                         goto no_measurement;
9484                                 }
9485                         }
9486                         bw = acked;
9487                         calc_bw = bw * 1000000;
9488                         calc_bw /= rtt;
9489                         if (rack->r_ctl.last_max_bw &&
9490                             (rack->r_ctl.last_max_bw < calc_bw)) {
9491                                 /*
9492                                  * If we have a last calculated max bw
9493                                  * enforce it.
9494                                  */
9495                                 calc_bw = rack->r_ctl.last_max_bw;
9496                         }
9497                         /* now plop it in */
9498                         if (rack->rc_gp_filled == 0) {
9499                                 if (calc_bw > ONE_POINT_TWO_MEG) {
9500                                         /*
9501                                          * If we have no measurement
9502                                          * don't let us set in more than
9503                                          * 1.2Mbps. If we are still too
9504                                          * low after pacing with this we
9505                                          * will hopefully have a max b/w
9506                                          * available to sanity check things.
9507                                          */
9508                                         calc_bw = ONE_POINT_TWO_MEG;
9509                                 }
9510                                 rack->r_ctl.rc_rtt_diff = 0;
9511                                 rack->r_ctl.gp_bw = calc_bw;
9512                                 rack->rc_gp_filled = 1;
9513                                 if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
9514                                         rack->r_ctl.num_measurements = RACK_REQ_AVG;
9515                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
9516                         } else if (calc_bw > rack->r_ctl.gp_bw) {
9517                                 rack->r_ctl.rc_rtt_diff = 0;
9518                                 if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
9519                                         rack->r_ctl.num_measurements = RACK_REQ_AVG;
9520                                 rack->r_ctl.gp_bw = calc_bw;
9521                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
9522                         } else
9523                                 rack_increase_bw_mul(rack, -1, 0, 0, 1);
9524                         if ((rack->gp_ready == 0) &&
9525                             (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
9526                                 /* We have enough measurements now */
9527                                 rack->gp_ready = 1;
9528                                 rack_set_cc_pacing(rack);
9529                                 if (rack->defer_options)
9530                                         rack_apply_deferred_options(rack);
9531                         }
9532                         /*
9533                          * For acks over 1mss we do a extra boost to simulate
9534                          * where we would get 2 acks (we want 110 for the mul).
9535                          */
9536                         if (acked > segsiz)
9537                                 rack_increase_bw_mul(rack, -1, 0, 0, 1);
9538                 } else {
9539                         /*
9540                          * zero rtt possibly?, settle for just an old increase.
9541                          */
9542 no_measurement:
9543                         rack_increase_bw_mul(rack, -1, 0, 0, 1);
9544                 }
9545         } else if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
9546                    (sbavail(&so->so_snd) > max((segsiz * (4 + rack_req_segs)),
9547                                                minseg)) &&
9548                    (rack->r_ctl.cwnd_to_use > max((segsiz * (rack_req_segs + 2)), minseg)) &&
9549                    (tp->snd_wnd > max((segsiz * (rack_req_segs + 2)), minseg)) &&
9550                    (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) <=
9551                     (segsiz * rack_req_segs))) {
9552                 /*
9553                  * We are doing dynamic GP pacing and
9554                  * we have everything except 1MSS or less
9555                  * bytes left out. We are still pacing away.
9556                  * And there is data that could be sent, This
9557                  * means we are inserting delayed ack time in
9558                  * our measurements because we are pacing too slow.
9559                  */
9560                 rack_validate_multipliers_at_or_above100(rack);
9561                 rack->rc_dragged_bottom = 1;
9562                 rack_increase_bw_mul(rack, -1, 0, 0, 1);
9563         }
9564 }
9565
9566
9567
9568 static void
9569 rack_gain_for_fastoutput(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t acked_amount)
9570 {
9571         /*
9572          * The fast output path is enabled and we
9573          * have moved the cumack forward. Lets see if
9574          * we can expand forward the fast path length by
9575          * that amount. What we would ideally like to
9576          * do is increase the number of bytes in the
9577          * fast path block (left_to_send) by the
9578          * acked amount. However we have to gate that
9579          * by two factors:
9580          * 1) The amount outstanding and the rwnd of the peer
9581          *    (i.e. we don't want to exceed the rwnd of the peer).
9582          *    <and>
9583          * 2) The amount of data left in the socket buffer (i.e.
9584          *    we can't send beyond what is in the buffer).
9585          *
9586          * Note that this does not take into account any increase
9587          * in the cwnd. We will only extend the fast path by
9588          * what was acked.
9589          */
9590         uint32_t new_total, gating_val;
9591
9592         new_total = acked_amount + rack->r_ctl.fsb.left_to_send;
9593         gating_val = min((sbavail(&so->so_snd) - (tp->snd_max - tp->snd_una)),
9594                          (tp->snd_wnd - (tp->snd_max - tp->snd_una)));
9595         if (new_total <= gating_val) {
9596                 /* We can increase left_to_send by the acked amount */
9597                 counter_u64_add(rack_extended_rfo, 1);
9598                 rack->r_ctl.fsb.left_to_send = new_total;
9599                 KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(&rack->rc_inp->inp_socket->so_snd) - (tp->snd_max - tp->snd_una))),
9600                         ("rack:%p left_to_send:%u sbavail:%u out:%u",
9601                          rack, rack->r_ctl.fsb.left_to_send,
9602                          sbavail(&rack->rc_inp->inp_socket->so_snd),
9603                          (tp->snd_max - tp->snd_una)));
9604
9605         }
9606 }
9607
9608 static void
9609 rack_adjust_sendmap(struct tcp_rack *rack, struct sockbuf *sb, tcp_seq snd_una)
9610 {
9611         /*
9612          * Here any sendmap entry that points to the
9613          * beginning mbuf must be adjusted to the correct
9614          * offset. This must be called with:
9615          * 1) The socket buffer locked
9616          * 2) snd_una adjusted to its new postion.
9617          *
9618          * Note that (2) implies rack_ack_received has also
9619          * been called.
9620          *
9621          * We grab the first mbuf in the socket buffer and
9622          * then go through the front of the sendmap, recalculating
9623          * the stored offset for any sendmap entry that has
9624          * that mbuf. We must use the sb functions to do this
9625          * since its possible an add was done has well as
9626          * the subtraction we may have just completed. This should
9627          * not be a penalty though, since we just referenced the sb
9628          * to go in and trim off the mbufs that we freed (of course
9629          * there will be a penalty for the sendmap references though).
9630          */
9631         struct mbuf *m;
9632         struct rack_sendmap *rsm;
9633
9634         SOCKBUF_LOCK_ASSERT(sb);
9635         m = sb->sb_mb;
9636         rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9637         if ((rsm == NULL) || (m == NULL)) {
9638                 /* Nothing outstanding */
9639                 return;
9640         }
9641         while (rsm->m == m) {
9642                 /* one to adjust */
9643 #ifdef INVARIANTS
9644                 struct mbuf *tm;
9645                 uint32_t soff;
9646
9647                 tm = sbsndmbuf(sb, (rsm->r_start - snd_una), &soff);
9648                 if (rsm->orig_m_len != m->m_len) {
9649                         rack_adjust_orig_mlen(rsm);
9650                 }
9651                 if (rsm->soff != soff) {
9652                         /*
9653                          * This is not a fatal error, we anticipate it
9654                          * might happen (the else code), so we count it here
9655                          * so that under invariant we can see that it really
9656                          * does happen.
9657                          */
9658                         counter_u64_add(rack_adjust_map_bw, 1);
9659                 }
9660                 rsm->m = tm;
9661                 rsm->soff = soff;
9662                 rsm->orig_m_len = rsm->m->m_len;
9663 #else
9664                 rsm->m = sbsndmbuf(sb, (rsm->r_start - snd_una), &rsm->soff);
9665                 rsm->orig_m_len = rsm->m->m_len;
9666 #endif
9667                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
9668                               rsm);
9669                 if (rsm == NULL)
9670                         break;
9671         }
9672 }
9673
9674 /*
9675  * Return value of 1, we do not need to call rack_process_data().
9676  * return value of 0, rack_process_data can be called.
9677  * For ret_val if its 0 the TCP is locked, if its non-zero
9678  * its unlocked and probably unsafe to touch the TCB.
9679  */
9680 static int
9681 rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
9682     struct tcpcb *tp, struct tcpopt *to,
9683     uint32_t tiwin, int32_t tlen,
9684     int32_t * ofia, int32_t thflags, int32_t *ret_val)
9685 {
9686         int32_t ourfinisacked = 0;
9687         int32_t nsegs, acked_amount;
9688         int32_t acked;
9689         struct mbuf *mfree;
9690         struct tcp_rack *rack;
9691         int32_t under_pacing = 0;
9692         int32_t recovery = 0;
9693
9694         rack = (struct tcp_rack *)tp->t_fb_ptr;
9695         if (SEQ_GT(th->th_ack, tp->snd_max)) {
9696                 __ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val,
9697                                       &rack->r_ctl.challenge_ack_ts,
9698                                       &rack->r_ctl.challenge_ack_cnt);
9699                 rack->r_wanted_output = 1;
9700                 return (1);
9701         }
9702         if (rack->gp_ready &&
9703             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
9704                 under_pacing = 1;
9705         }
9706         if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
9707                 int in_rec, dup_ack_struck = 0;
9708
9709                 in_rec = IN_FASTRECOVERY(tp->t_flags);
9710                 if (rack->rc_in_persist) {
9711                         tp->t_rxtshift = 0;
9712                         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
9713                                       rack_rto_min, rack_rto_max);
9714                 }
9715                 if ((th->th_ack == tp->snd_una) && (tiwin == tp->snd_wnd)) {
9716                         rack_strike_dupack(rack);
9717                         dup_ack_struck = 1;
9718                 }
9719                 rack_log_ack(tp, to, th, ((in_rec == 0) && IN_FASTRECOVERY(tp->t_flags)), dup_ack_struck);
9720         }
9721         if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
9722                 /*
9723                  * Old ack, behind (or duplicate to) the last one rcv'd
9724                  * Note: We mark reordering is occuring if its
9725                  * less than and we have not closed our window.
9726                  */
9727                 if (SEQ_LT(th->th_ack, tp->snd_una) && (sbspace(&so->so_rcv) > ctf_fixed_maxseg(tp))) {
9728                         counter_u64_add(rack_reorder_seen, 1);
9729                         rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
9730                 }
9731                 return (0);
9732         }
9733         /*
9734          * If we reach this point, ACK is not a duplicate, i.e., it ACKs
9735          * something we sent.
9736          */
9737         if (tp->t_flags & TF_NEEDSYN) {
9738                 /*
9739                  * T/TCP: Connection was half-synchronized, and our SYN has
9740                  * been ACK'd (so connection is now fully synchronized).  Go
9741                  * to non-starred state, increment snd_una for ACK of SYN,
9742                  * and check if we can do window scaling.
9743                  */
9744                 tp->t_flags &= ~TF_NEEDSYN;
9745                 tp->snd_una++;
9746                 /* Do window scaling? */
9747                 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
9748                     (TF_RCVD_SCALE | TF_REQ_SCALE)) {
9749                         tp->rcv_scale = tp->request_r_scale;
9750                         /* Send window already scaled. */
9751                 }
9752         }
9753         nsegs = max(1, m->m_pkthdr.lro_nsegs);
9754         INP_WLOCK_ASSERT(tp->t_inpcb);
9755
9756         acked = BYTES_THIS_ACK(tp, th);
9757         KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
9758         KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
9759         /*
9760          * If we just performed our first retransmit, and the ACK arrives
9761          * within our recovery window, then it was a mistake to do the
9762          * retransmit in the first place.  Recover our original cwnd and
9763          * ssthresh, and proceed to transmit where we left off.
9764          */
9765         if ((tp->t_flags & TF_PREVVALID) &&
9766             ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
9767                 tp->t_flags &= ~TF_PREVVALID;
9768                 if (tp->t_rxtshift == 1 &&
9769                     (int)(ticks - tp->t_badrxtwin) < 0)
9770                         rack_cong_signal(tp, CC_RTO_ERR, th->th_ack);
9771         }
9772         if (acked) {
9773                 /* assure we are not backed off */
9774                 tp->t_rxtshift = 0;
9775                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
9776                               rack_rto_min, rack_rto_max);
9777                 rack->rc_tlp_in_progress = 0;
9778                 rack->r_ctl.rc_tlp_cnt_out = 0;
9779                 /*
9780                  * If it is the RXT timer we want to
9781                  * stop it, so we can restart a TLP.
9782                  */
9783                 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
9784                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
9785 #ifdef NETFLIX_HTTP_LOGGING
9786                 tcp_http_check_for_comp(rack->rc_tp, th->th_ack);
9787 #endif
9788         }
9789         /*
9790          * If we have a timestamp reply, update smoothed round trip time. If
9791          * no timestamp is present but transmit timer is running and timed
9792          * sequence number was acked, update smoothed round trip time. Since
9793          * we now have an rtt measurement, cancel the timer backoff (cf.,
9794          * Phil Karn's retransmit alg.). Recompute the initial retransmit
9795          * timer.
9796          *
9797          * Some boxes send broken timestamp replies during the SYN+ACK
9798          * phase, ignore timestamps of 0 or we could calculate a huge RTT
9799          * and blow up the retransmit timer.
9800          */
9801         /*
9802          * If all outstanding data is acked, stop retransmit timer and
9803          * remember to restart (more output or persist). If there is more
9804          * data to be acked, restart retransmit timer, using current
9805          * (possibly backed-off) value.
9806          */
9807         if (acked == 0) {
9808                 if (ofia)
9809                         *ofia = ourfinisacked;
9810                 return (0);
9811         }
9812         if (IN_RECOVERY(tp->t_flags)) {
9813                 if (SEQ_LT(th->th_ack, tp->snd_recover) &&
9814                     (SEQ_LT(th->th_ack, tp->snd_max))) {
9815                         tcp_rack_partialack(tp);
9816                 } else {
9817                         rack_post_recovery(tp, th->th_ack);
9818                         recovery = 1;
9819                 }
9820         }
9821         /*
9822          * Let the congestion control algorithm update congestion control
9823          * related information. This typically means increasing the
9824          * congestion window.
9825          */
9826         rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, recovery);
9827         SOCKBUF_LOCK(&so->so_snd);
9828         acked_amount = min(acked, (int)sbavail(&so->so_snd));
9829         tp->snd_wnd -= acked_amount;
9830         mfree = sbcut_locked(&so->so_snd, acked_amount);
9831         if ((sbused(&so->so_snd) == 0) &&
9832             (acked > acked_amount) &&
9833             (tp->t_state >= TCPS_FIN_WAIT_1) &&
9834             (tp->t_flags & TF_SENTFIN)) {
9835                 /*
9836                  * We must be sure our fin
9837                  * was sent and acked (we can be
9838                  * in FIN_WAIT_1 without having
9839                  * sent the fin).
9840                  */
9841                 ourfinisacked = 1;
9842         }
9843         tp->snd_una = th->th_ack;
9844         if (acked_amount && sbavail(&so->so_snd))
9845                 rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
9846         rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
9847         SOCKBUF_UNLOCK(&so->so_snd);
9848         tp->t_flags |= TF_WAKESOW;
9849         m_freem(mfree);
9850         if (SEQ_GT(tp->snd_una, tp->snd_recover))
9851                 tp->snd_recover = tp->snd_una;
9852
9853         if (SEQ_LT(tp->snd_nxt, tp->snd_una)) {
9854                 tp->snd_nxt = tp->snd_una;
9855         }
9856         if (under_pacing &&
9857             (rack->use_fixed_rate == 0) &&
9858             (rack->in_probe_rtt == 0) &&
9859             rack->rc_gp_dyn_mul &&
9860             rack->rc_always_pace) {
9861                 /* Check if we are dragging bottom */
9862                 rack_check_bottom_drag(tp, rack, so, acked);
9863         }
9864         if (tp->snd_una == tp->snd_max) {
9865                 /* Nothing left outstanding */
9866                 tp->t_flags &= ~TF_PREVVALID;
9867                 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
9868                 rack->r_ctl.retran_during_recovery = 0;
9869                 rack->r_ctl.dsack_byte_cnt = 0;
9870                 if (rack->r_ctl.rc_went_idle_time == 0)
9871                         rack->r_ctl.rc_went_idle_time = 1;
9872                 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
9873                 if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
9874                         tp->t_acktime = 0;
9875                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
9876                 /* Set need output so persist might get set */
9877                 rack->r_wanted_output = 1;
9878                 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
9879                 if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
9880                     (sbavail(&so->so_snd) == 0) &&
9881                     (tp->t_flags2 & TF2_DROP_AF_DATA)) {
9882                         /*
9883                          * The socket was gone and the
9884                          * peer sent data (now or in the past), time to
9885                          * reset him.
9886                          */
9887                         *ret_val = 1;
9888                         /* tcp_close will kill the inp pre-log the Reset */
9889                         tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
9890                         tp = tcp_close(tp);
9891                         ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen);
9892                         return (1);
9893                 }
9894         }
9895         if (ofia)
9896                 *ofia = ourfinisacked;
9897         return (0);
9898 }
9899
9900 static void
9901 rack_collapsed_window(struct tcp_rack *rack)
9902 {
9903         /*
9904          * Now we must walk the
9905          * send map and divide the
9906          * ones left stranded. These
9907          * guys can't cause us to abort
9908          * the connection and are really
9909          * "unsent". However if a buggy
9910          * client actually did keep some
9911          * of the data i.e. collapsed the win
9912          * and refused to ack and then opened
9913          * the win and acked that data. We would
9914          * get into an ack war, the simplier
9915          * method then of just pretending we
9916          * did not send those segments something
9917          * won't work.
9918          */
9919         struct rack_sendmap *rsm, *nrsm, fe, *insret;
9920         tcp_seq max_seq;
9921
9922         max_seq = rack->rc_tp->snd_una + rack->rc_tp->snd_wnd;
9923         memset(&fe, 0, sizeof(fe));
9924         fe.r_start = max_seq;
9925         /* Find the first seq past or at maxseq */
9926         rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
9927         if (rsm == NULL) {
9928                 /* Nothing to do strange */
9929                 rack->rc_has_collapsed = 0;
9930                 return;
9931         }
9932         /*
9933          * Now do we need to split at
9934          * the collapse point?
9935          */
9936         if (SEQ_GT(max_seq, rsm->r_start)) {
9937                 nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
9938                 if (nrsm == NULL) {
9939                         /* We can't get a rsm, mark all? */
9940                         nrsm = rsm;
9941                         goto no_split;
9942                 }
9943                 /* Clone it */
9944                 rack_clone_rsm(rack, nrsm, rsm, max_seq);
9945                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
9946 #ifdef INVARIANTS
9947                 if (insret != NULL) {
9948                         panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
9949                               nrsm, insret, rack, rsm);
9950                 }
9951 #endif
9952                 rack_log_map_chg(rack->rc_tp, rack, NULL, rsm, nrsm, MAP_SPLIT, max_seq, __LINE__);
9953                 if (rsm->r_in_tmap) {
9954                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
9955                         nrsm->r_in_tmap = 1;
9956                 }
9957                 /*
9958                  * Set in the new RSM as the
9959                  * collapsed starting point
9960                  */
9961                 rsm = nrsm;
9962         }
9963 no_split:
9964         counter_u64_add(rack_collapsed_win, 1);
9965         RB_FOREACH_FROM(nrsm, rack_rb_tree_head, rsm) {
9966                 nrsm->r_flags |= RACK_RWND_COLLAPSED;
9967         }
9968         rack->rc_has_collapsed = 1;
9969 }
9970
9971 static void
9972 rack_un_collapse_window(struct tcp_rack *rack)
9973 {
9974         struct rack_sendmap *rsm;
9975
9976         RB_FOREACH_REVERSE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
9977                 if (rsm->r_flags & RACK_RWND_COLLAPSED)
9978                         rsm->r_flags &= ~RACK_RWND_COLLAPSED;
9979                 else
9980                         break;
9981         }
9982         rack->rc_has_collapsed = 0;
9983 }
9984
9985 static void
9986 rack_handle_delayed_ack(struct tcpcb *tp, struct tcp_rack *rack,
9987                         int32_t tlen, int32_t tfo_syn)
9988 {
9989         if (DELAY_ACK(tp, tlen) || tfo_syn) {
9990                 if (rack->rc_dack_mode &&
9991                     (tlen > 500) &&
9992                     (rack->rc_dack_toggle == 1)) {
9993                         goto no_delayed_ack;
9994                 }
9995                 rack_timer_cancel(tp, rack,
9996                                   rack->r_ctl.rc_rcvtime, __LINE__);
9997                 tp->t_flags |= TF_DELACK;
9998         } else {
9999 no_delayed_ack:
10000                 rack->r_wanted_output = 1;
10001                 tp->t_flags |= TF_ACKNOW;
10002                 if (rack->rc_dack_mode) {
10003                         if (tp->t_flags & TF_DELACK)
10004                                 rack->rc_dack_toggle = 1;
10005                         else
10006                                 rack->rc_dack_toggle = 0;
10007                 }
10008         }
10009 }
10010
10011 static void
10012 rack_validate_fo_sendwin_up(struct tcpcb *tp, struct tcp_rack *rack)
10013 {
10014         /*
10015          * If fast output is in progress, lets validate that
10016          * the new window did not shrink on us and make it
10017          * so fast output should end.
10018          */
10019         if (rack->r_fast_output) {
10020                 uint32_t out;
10021
10022                 /*
10023                  * Calculate what we will send if left as is
10024                  * and compare that to our send window.
10025                  */
10026                 out = ctf_outstanding(tp);
10027                 if ((out + rack->r_ctl.fsb.left_to_send) > tp->snd_wnd) {
10028                         /* ok we have an issue */
10029                         if (out >= tp->snd_wnd) {
10030                                 /* Turn off fast output the window is met or collapsed */
10031                                 rack->r_fast_output = 0;
10032                         } else {
10033                                 /* we have some room left */
10034                                 rack->r_ctl.fsb.left_to_send = tp->snd_wnd - out;
10035                                 if (rack->r_ctl.fsb.left_to_send < ctf_fixed_maxseg(tp)) {
10036                                         /* If not at least 1 full segment never mind */
10037                                         rack->r_fast_output = 0;
10038                                 }
10039                         }
10040                 }
10041         }
10042 }
10043
10044 /*
10045  * Return value of 1, the TCB is unlocked and most
10046  * likely gone, return value of 0, the TCP is still
10047  * locked.
10048  */
10049 static int
10050 rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
10051     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
10052     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
10053 {
10054         /*
10055          * Update window information. Don't look at window if no ACK: TAC's
10056          * send garbage on first SYN.
10057          */
10058         int32_t nsegs;
10059         int32_t tfo_syn;
10060         struct tcp_rack *rack;
10061
10062         rack = (struct tcp_rack *)tp->t_fb_ptr;
10063         INP_WLOCK_ASSERT(tp->t_inpcb);
10064         nsegs = max(1, m->m_pkthdr.lro_nsegs);
10065         if ((thflags & TH_ACK) &&
10066             (SEQ_LT(tp->snd_wl1, th->th_seq) ||
10067             (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
10068             (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
10069                 /* keep track of pure window updates */
10070                 if (tlen == 0 &&
10071                     tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
10072                         KMOD_TCPSTAT_INC(tcps_rcvwinupd);
10073                 tp->snd_wnd = tiwin;
10074                 rack_validate_fo_sendwin_up(tp, rack);
10075                 tp->snd_wl1 = th->th_seq;
10076                 tp->snd_wl2 = th->th_ack;
10077                 if (tp->snd_wnd > tp->max_sndwnd)
10078                         tp->max_sndwnd = tp->snd_wnd;
10079                 rack->r_wanted_output = 1;
10080         } else if (thflags & TH_ACK) {
10081                 if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
10082                         tp->snd_wnd = tiwin;
10083                         rack_validate_fo_sendwin_up(tp, rack);
10084                         tp->snd_wl1 = th->th_seq;
10085                         tp->snd_wl2 = th->th_ack;
10086                 }
10087         }
10088         if (tp->snd_wnd < ctf_outstanding(tp))
10089                 /* The peer collapsed the window */
10090                 rack_collapsed_window(rack);
10091         else if (rack->rc_has_collapsed)
10092                 rack_un_collapse_window(rack);
10093         /* Was persist timer active and now we have window space? */
10094         if ((rack->rc_in_persist != 0) &&
10095             (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
10096                                 rack->r_ctl.rc_pace_min_segs))) {
10097                 rack_exit_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10098                 tp->snd_nxt = tp->snd_max;
10099                 /* Make sure we output to start the timer */
10100                 rack->r_wanted_output = 1;
10101         }
10102         /* Do we enter persists? */
10103         if ((rack->rc_in_persist == 0) &&
10104             (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
10105             TCPS_HAVEESTABLISHED(tp->t_state) &&
10106             (tp->snd_max == tp->snd_una) &&
10107             sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
10108             (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
10109                 /*
10110                  * Here the rwnd is less than
10111                  * the pacing size, we are established,
10112                  * nothing is outstanding, and there is
10113                  * data to send. Enter persists.
10114                  */
10115                 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10116         }
10117         if (tp->t_flags2 & TF2_DROP_AF_DATA) {
10118                 m_freem(m);
10119                 return (0);
10120         }
10121         /*
10122          * don't process the URG bit, ignore them drag
10123          * along the up.
10124          */
10125         tp->rcv_up = tp->rcv_nxt;
10126         INP_WLOCK_ASSERT(tp->t_inpcb);
10127
10128         /*
10129          * Process the segment text, merging it into the TCP sequencing
10130          * queue, and arranging for acknowledgment of receipt if necessary.
10131          * This process logically involves adjusting tp->rcv_wnd as data is
10132          * presented to the user (this happens in tcp_usrreq.c, case
10133          * PRU_RCVD).  If a FIN has already been received on this connection
10134          * then we just ignore the text.
10135          */
10136         tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
10137                    IS_FASTOPEN(tp->t_flags));
10138         if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) &&
10139             TCPS_HAVERCVDFIN(tp->t_state) == 0) {
10140                 tcp_seq save_start = th->th_seq;
10141                 tcp_seq save_rnxt  = tp->rcv_nxt;
10142                 int     save_tlen  = tlen;
10143
10144                 m_adj(m, drop_hdrlen);  /* delayed header drop */
10145                 /*
10146                  * Insert segment which includes th into TCP reassembly
10147                  * queue with control block tp.  Set thflags to whether
10148                  * reassembly now includes a segment with FIN.  This handles
10149                  * the common case inline (segment is the next to be
10150                  * received on an established connection, and the queue is
10151                  * empty), avoiding linkage into and removal from the queue
10152                  * and repetition of various conversions. Set DELACK for
10153                  * segments received in order, but ack immediately when
10154                  * segments are out of order (so fast retransmit can work).
10155                  */
10156                 if (th->th_seq == tp->rcv_nxt &&
10157                     SEGQ_EMPTY(tp) &&
10158                     (TCPS_HAVEESTABLISHED(tp->t_state) ||
10159                     tfo_syn)) {
10160 #ifdef NETFLIX_SB_LIMITS
10161                         u_int mcnt, appended;
10162
10163                         if (so->so_rcv.sb_shlim) {
10164                                 mcnt = m_memcnt(m);
10165                                 appended = 0;
10166                                 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
10167                                     CFO_NOSLEEP, NULL) == false) {
10168                                         counter_u64_add(tcp_sb_shlim_fails, 1);
10169                                         m_freem(m);
10170                                         return (0);
10171                                 }
10172                         }
10173 #endif
10174                         rack_handle_delayed_ack(tp, rack, tlen, tfo_syn);
10175                         tp->rcv_nxt += tlen;
10176                         if (tlen &&
10177                             ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
10178                             (tp->t_fbyte_in == 0)) {
10179                                 tp->t_fbyte_in = ticks;
10180                                 if (tp->t_fbyte_in == 0)
10181                                         tp->t_fbyte_in = 1;
10182                                 if (tp->t_fbyte_out && tp->t_fbyte_in)
10183                                         tp->t_flags2 |= TF2_FBYTES_COMPLETE;
10184                         }
10185                         thflags = th->th_flags & TH_FIN;
10186                         KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
10187                         KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
10188                         SOCKBUF_LOCK(&so->so_rcv);
10189                         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10190                                 m_freem(m);
10191                         } else
10192 #ifdef NETFLIX_SB_LIMITS
10193                                 appended =
10194 #endif
10195                                         sbappendstream_locked(&so->so_rcv, m, 0);
10196
10197                         rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
10198                         SOCKBUF_UNLOCK(&so->so_rcv);
10199                         tp->t_flags |= TF_WAKESOR;
10200 #ifdef NETFLIX_SB_LIMITS
10201                         if (so->so_rcv.sb_shlim && appended != mcnt)
10202                                 counter_fo_release(so->so_rcv.sb_shlim,
10203                                     mcnt - appended);
10204 #endif
10205                 } else {
10206                         /*
10207                          * XXX: Due to the header drop above "th" is
10208                          * theoretically invalid by now.  Fortunately
10209                          * m_adj() doesn't actually frees any mbufs when
10210                          * trimming from the head.
10211                          */
10212                         tcp_seq temp = save_start;
10213
10214                         thflags = tcp_reass(tp, th, &temp, &tlen, m);
10215                         tp->t_flags |= TF_ACKNOW;
10216                 }
10217                 if ((tp->t_flags & TF_SACK_PERMIT) &&
10218                     (save_tlen > 0) &&
10219                     TCPS_HAVEESTABLISHED(tp->t_state)) {
10220                         if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) {
10221                                 /*
10222                                  * DSACK actually handled in the fastpath
10223                                  * above.
10224                                  */
10225                                 RACK_OPTS_INC(tcp_sack_path_1);
10226                                 tcp_update_sack_list(tp, save_start,
10227                                     save_start + save_tlen);
10228                         } else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) {
10229                                 if ((tp->rcv_numsacks >= 1) &&
10230                                     (tp->sackblks[0].end == save_start)) {
10231                                         /*
10232                                          * Partial overlap, recorded at todrop
10233                                          * above.
10234                                          */
10235                                         RACK_OPTS_INC(tcp_sack_path_2a);
10236                                         tcp_update_sack_list(tp,
10237                                             tp->sackblks[0].start,
10238                                             tp->sackblks[0].end);
10239                                 } else {
10240                                         RACK_OPTS_INC(tcp_sack_path_2b);
10241                                         tcp_update_dsack_list(tp, save_start,
10242                                             save_start + save_tlen);
10243                                 }
10244                         } else if (tlen >= save_tlen) {
10245                                 /* Update of sackblks. */
10246                                 RACK_OPTS_INC(tcp_sack_path_3);
10247                                 tcp_update_dsack_list(tp, save_start,
10248                                     save_start + save_tlen);
10249                         } else if (tlen > 0) {
10250                                 RACK_OPTS_INC(tcp_sack_path_4);
10251                                 tcp_update_dsack_list(tp, save_start,
10252                                     save_start + tlen);
10253                         }
10254                 }
10255         } else {
10256                 m_freem(m);
10257                 thflags &= ~TH_FIN;
10258         }
10259
10260         /*
10261          * If FIN is received ACK the FIN and let the user know that the
10262          * connection is closing.
10263          */
10264         if (thflags & TH_FIN) {
10265                 if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
10266                         socantrcvmore(so);
10267                         /* The socket upcall is handled by socantrcvmore. */
10268                         tp->t_flags &= ~TF_WAKESOR;
10269                         /*
10270                          * If connection is half-synchronized (ie NEEDSYN
10271                          * flag on) then delay ACK, so it may be piggybacked
10272                          * when SYN is sent. Otherwise, since we received a
10273                          * FIN then no more input can be expected, send ACK
10274                          * now.
10275                          */
10276                         if (tp->t_flags & TF_NEEDSYN) {
10277                                 rack_timer_cancel(tp, rack,
10278                                     rack->r_ctl.rc_rcvtime, __LINE__);
10279                                 tp->t_flags |= TF_DELACK;
10280                         } else {
10281                                 tp->t_flags |= TF_ACKNOW;
10282                         }
10283                         tp->rcv_nxt++;
10284                 }
10285                 switch (tp->t_state) {
10286                         /*
10287                          * In SYN_RECEIVED and ESTABLISHED STATES enter the
10288                          * CLOSE_WAIT state.
10289                          */
10290                 case TCPS_SYN_RECEIVED:
10291                         tp->t_starttime = ticks;
10292                         /* FALLTHROUGH */
10293                 case TCPS_ESTABLISHED:
10294                         rack_timer_cancel(tp, rack,
10295                             rack->r_ctl.rc_rcvtime, __LINE__);
10296                         tcp_state_change(tp, TCPS_CLOSE_WAIT);
10297                         break;
10298
10299                         /*
10300                          * If still in FIN_WAIT_1 STATE FIN has not been
10301                          * acked so enter the CLOSING state.
10302                          */
10303                 case TCPS_FIN_WAIT_1:
10304                         rack_timer_cancel(tp, rack,
10305                             rack->r_ctl.rc_rcvtime, __LINE__);
10306                         tcp_state_change(tp, TCPS_CLOSING);
10307                         break;
10308
10309                         /*
10310                          * In FIN_WAIT_2 state enter the TIME_WAIT state,
10311                          * starting the time-wait timer, turning off the
10312                          * other standard timers.
10313                          */
10314                 case TCPS_FIN_WAIT_2:
10315                         rack_timer_cancel(tp, rack,
10316                             rack->r_ctl.rc_rcvtime, __LINE__);
10317                         tcp_twstart(tp);
10318                         return (1);
10319                 }
10320         }
10321         /*
10322          * Return any desired output.
10323          */
10324         if ((tp->t_flags & TF_ACKNOW) ||
10325             (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) {
10326                 rack->r_wanted_output = 1;
10327         }
10328         INP_WLOCK_ASSERT(tp->t_inpcb);
10329         return (0);
10330 }
10331
10332 /*
10333  * Here nothing is really faster, its just that we
10334  * have broken out the fast-data path also just like
10335  * the fast-ack.
10336  */
10337 static int
10338 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
10339     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10340     uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos)
10341 {
10342         int32_t nsegs;
10343         int32_t newsize = 0;    /* automatic sockbuf scaling */
10344         struct tcp_rack *rack;
10345 #ifdef NETFLIX_SB_LIMITS
10346         u_int mcnt, appended;
10347 #endif
10348 #ifdef TCPDEBUG
10349         /*
10350          * The size of tcp_saveipgen must be the size of the max ip header,
10351          * now IPv6.
10352          */
10353         u_char tcp_saveipgen[IP6_HDR_LEN];
10354         struct tcphdr tcp_savetcp;
10355         short ostate = 0;
10356
10357 #endif
10358         /*
10359          * If last ACK falls within this segment's sequence numbers, record
10360          * the timestamp. NOTE that the test is modified according to the
10361          * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
10362          */
10363         if (__predict_false(th->th_seq != tp->rcv_nxt)) {
10364                 return (0);
10365         }
10366         if (__predict_false(tp->snd_nxt != tp->snd_max)) {
10367                 return (0);
10368         }
10369         if (tiwin && tiwin != tp->snd_wnd) {
10370                 return (0);
10371         }
10372         if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
10373                 return (0);
10374         }
10375         if (__predict_false((to->to_flags & TOF_TS) &&
10376             (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
10377                 return (0);
10378         }
10379         if (__predict_false((th->th_ack != tp->snd_una))) {
10380                 return (0);
10381         }
10382         if (__predict_false(tlen > sbspace(&so->so_rcv))) {
10383                 return (0);
10384         }
10385         if ((to->to_flags & TOF_TS) != 0 &&
10386             SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
10387                 tp->ts_recent_age = tcp_ts_getticks();
10388                 tp->ts_recent = to->to_tsval;
10389         }
10390         rack = (struct tcp_rack *)tp->t_fb_ptr;
10391         /*
10392          * This is a pure, in-sequence data packet with nothing on the
10393          * reassembly queue and we have enough buffer space to take it.
10394          */
10395         nsegs = max(1, m->m_pkthdr.lro_nsegs);
10396
10397 #ifdef NETFLIX_SB_LIMITS
10398         if (so->so_rcv.sb_shlim) {
10399                 mcnt = m_memcnt(m);
10400                 appended = 0;
10401                 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
10402                     CFO_NOSLEEP, NULL) == false) {
10403                         counter_u64_add(tcp_sb_shlim_fails, 1);
10404                         m_freem(m);
10405                         return (1);
10406                 }
10407         }
10408 #endif
10409         /* Clean receiver SACK report if present */
10410         if (tp->rcv_numsacks)
10411                 tcp_clean_sackreport(tp);
10412         KMOD_TCPSTAT_INC(tcps_preddat);
10413         tp->rcv_nxt += tlen;
10414         if (tlen &&
10415             ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
10416             (tp->t_fbyte_in == 0)) {
10417                 tp->t_fbyte_in = ticks;
10418                 if (tp->t_fbyte_in == 0)
10419                         tp->t_fbyte_in = 1;
10420                 if (tp->t_fbyte_out && tp->t_fbyte_in)
10421                         tp->t_flags2 |= TF2_FBYTES_COMPLETE;
10422         }
10423         /*
10424          * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
10425          */
10426         tp->snd_wl1 = th->th_seq;
10427         /*
10428          * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
10429          */
10430         tp->rcv_up = tp->rcv_nxt;
10431         KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
10432         KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
10433 #ifdef TCPDEBUG
10434         if (so->so_options & SO_DEBUG)
10435                 tcp_trace(TA_INPUT, ostate, tp,
10436                     (void *)tcp_saveipgen, &tcp_savetcp, 0);
10437 #endif
10438         newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
10439
10440         /* Add data to socket buffer. */
10441         SOCKBUF_LOCK(&so->so_rcv);
10442         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10443                 m_freem(m);
10444         } else {
10445                 /*
10446                  * Set new socket buffer size. Give up when limit is
10447                  * reached.
10448                  */
10449                 if (newsize)
10450                         if (!sbreserve_locked(&so->so_rcv,
10451                             newsize, so, NULL))
10452                                 so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
10453                 m_adj(m, drop_hdrlen);  /* delayed header drop */
10454 #ifdef NETFLIX_SB_LIMITS
10455                 appended =
10456 #endif
10457                         sbappendstream_locked(&so->so_rcv, m, 0);
10458                 ctf_calc_rwin(so, tp);
10459         }
10460         rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
10461         SOCKBUF_UNLOCK(&so->so_rcv);
10462         tp->t_flags |= TF_WAKESOR;
10463 #ifdef NETFLIX_SB_LIMITS
10464         if (so->so_rcv.sb_shlim && mcnt != appended)
10465                 counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended);
10466 #endif
10467         rack_handle_delayed_ack(tp, rack, tlen, 0);
10468         if (tp->snd_una == tp->snd_max)
10469                 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
10470         return (1);
10471 }
10472
10473 /*
10474  * This subfunction is used to try to highly optimize the
10475  * fast path. We again allow window updates that are
10476  * in sequence to remain in the fast-path. We also add
10477  * in the __predict's to attempt to help the compiler.
10478  * Note that if we return a 0, then we can *not* process
10479  * it and the caller should push the packet into the
10480  * slow-path.
10481  */
10482 static int
10483 rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
10484     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10485     uint32_t tiwin, int32_t nxt_pkt, uint32_t cts)
10486 {
10487         int32_t acked;
10488         int32_t nsegs;
10489 #ifdef TCPDEBUG
10490         /*
10491          * The size of tcp_saveipgen must be the size of the max ip header,
10492          * now IPv6.
10493          */
10494         u_char tcp_saveipgen[IP6_HDR_LEN];
10495         struct tcphdr tcp_savetcp;
10496         short ostate = 0;
10497 #endif
10498         int32_t under_pacing = 0;
10499         struct tcp_rack *rack;
10500
10501         if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
10502                 /* Old ack, behind (or duplicate to) the last one rcv'd */
10503                 return (0);
10504         }
10505         if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
10506                 /* Above what we have sent? */
10507                 return (0);
10508         }
10509         if (__predict_false(tp->snd_nxt != tp->snd_max)) {
10510                 /* We are retransmitting */
10511                 return (0);
10512         }
10513         if (__predict_false(tiwin == 0)) {
10514                 /* zero window */
10515                 return (0);
10516         }
10517         if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
10518                 /* We need a SYN or a FIN, unlikely.. */
10519                 return (0);
10520         }
10521         if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
10522                 /* Timestamp is behind .. old ack with seq wrap? */
10523                 return (0);
10524         }
10525         if (__predict_false(IN_RECOVERY(tp->t_flags))) {
10526                 /* Still recovering */
10527                 return (0);
10528         }
10529         rack = (struct tcp_rack *)tp->t_fb_ptr;
10530         if (rack->r_ctl.rc_sacked) {
10531                 /* We have sack holes on our scoreboard */
10532                 return (0);
10533         }
10534         /* Ok if we reach here, we can process a fast-ack */
10535         if (rack->gp_ready &&
10536             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
10537                 under_pacing = 1;
10538         }
10539         nsegs = max(1, m->m_pkthdr.lro_nsegs);
10540         rack_log_ack(tp, to, th, 0, 0);
10541         /* Did the window get updated? */
10542         if (tiwin != tp->snd_wnd) {
10543                 tp->snd_wnd = tiwin;
10544                 rack_validate_fo_sendwin_up(tp, rack);
10545                 tp->snd_wl1 = th->th_seq;
10546                 if (tp->snd_wnd > tp->max_sndwnd)
10547                         tp->max_sndwnd = tp->snd_wnd;
10548         }
10549         /* Do we exit persists? */
10550         if ((rack->rc_in_persist != 0) &&
10551             (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
10552                                rack->r_ctl.rc_pace_min_segs))) {
10553                 rack_exit_persist(tp, rack, cts);
10554         }
10555         /* Do we enter persists? */
10556         if ((rack->rc_in_persist == 0) &&
10557             (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
10558             TCPS_HAVEESTABLISHED(tp->t_state) &&
10559             (tp->snd_max == tp->snd_una) &&
10560             sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
10561             (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
10562                 /*
10563                  * Here the rwnd is less than
10564                  * the pacing size, we are established,
10565                  * nothing is outstanding, and there is
10566                  * data to send. Enter persists.
10567                  */
10568                 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10569         }
10570         /*
10571          * If last ACK falls within this segment's sequence numbers, record
10572          * the timestamp. NOTE that the test is modified according to the
10573          * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
10574          */
10575         if ((to->to_flags & TOF_TS) != 0 &&
10576             SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
10577                 tp->ts_recent_age = tcp_ts_getticks();
10578                 tp->ts_recent = to->to_tsval;
10579         }
10580         /*
10581          * This is a pure ack for outstanding data.
10582          */
10583         KMOD_TCPSTAT_INC(tcps_predack);
10584
10585         /*
10586          * "bad retransmit" recovery.
10587          */
10588         if ((tp->t_flags & TF_PREVVALID) &&
10589             ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
10590                 tp->t_flags &= ~TF_PREVVALID;
10591                 if (tp->t_rxtshift == 1 &&
10592                     (int)(ticks - tp->t_badrxtwin) < 0)
10593                         rack_cong_signal(tp, CC_RTO_ERR, th->th_ack);
10594         }
10595         /*
10596          * Recalculate the transmit timer / rtt.
10597          *
10598          * Some boxes send broken timestamp replies during the SYN+ACK
10599          * phase, ignore timestamps of 0 or we could calculate a huge RTT
10600          * and blow up the retransmit timer.
10601          */
10602         acked = BYTES_THIS_ACK(tp, th);
10603
10604 #ifdef TCP_HHOOK
10605         /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
10606         hhook_run_tcp_est_in(tp, th, to);
10607 #endif
10608         KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
10609         KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
10610         if (acked) {
10611                 struct mbuf *mfree;
10612
10613                 rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, 0);
10614                 SOCKBUF_LOCK(&so->so_snd);
10615                 mfree = sbcut_locked(&so->so_snd, acked);
10616                 tp->snd_una = th->th_ack;
10617                 /* Note we want to hold the sb lock through the sendmap adjust */
10618                 rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
10619                 /* Wake up the socket if we have room to write more */
10620                 rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
10621                 SOCKBUF_UNLOCK(&so->so_snd);
10622                 tp->t_flags |= TF_WAKESOW;
10623                 m_freem(mfree);
10624                 tp->t_rxtshift = 0;
10625                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
10626                               rack_rto_min, rack_rto_max);
10627                 rack->rc_tlp_in_progress = 0;
10628                 rack->r_ctl.rc_tlp_cnt_out = 0;
10629                 /*
10630                  * If it is the RXT timer we want to
10631                  * stop it, so we can restart a TLP.
10632                  */
10633                 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
10634                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
10635 #ifdef NETFLIX_HTTP_LOGGING
10636                 tcp_http_check_for_comp(rack->rc_tp, th->th_ack);
10637 #endif
10638         }
10639         /*
10640          * Let the congestion control algorithm update congestion control
10641          * related information. This typically means increasing the
10642          * congestion window.
10643          */
10644         if (tp->snd_wnd < ctf_outstanding(tp)) {
10645                 /* The peer collapsed the window */
10646                 rack_collapsed_window(rack);
10647         } else if (rack->rc_has_collapsed)
10648                 rack_un_collapse_window(rack);
10649
10650         /*
10651          * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
10652          */
10653         tp->snd_wl2 = th->th_ack;
10654         tp->t_dupacks = 0;
10655         m_freem(m);
10656         /* ND6_HINT(tp);         *//* Some progress has been made. */
10657
10658         /*
10659          * If all outstanding data are acked, stop retransmit timer,
10660          * otherwise restart timer using current (possibly backed-off)
10661          * value. If process is waiting for space, wakeup/selwakeup/signal.
10662          * If data are ready to send, let tcp_output decide between more
10663          * output or persist.
10664          */
10665 #ifdef TCPDEBUG
10666         if (so->so_options & SO_DEBUG)
10667                 tcp_trace(TA_INPUT, ostate, tp,
10668                     (void *)tcp_saveipgen,
10669                     &tcp_savetcp, 0);
10670 #endif
10671         if (under_pacing &&
10672             (rack->use_fixed_rate == 0) &&
10673             (rack->in_probe_rtt == 0) &&
10674             rack->rc_gp_dyn_mul &&
10675             rack->rc_always_pace) {
10676                 /* Check if we are dragging bottom */
10677                 rack_check_bottom_drag(tp, rack, so, acked);
10678         }
10679         if (tp->snd_una == tp->snd_max) {
10680                 tp->t_flags &= ~TF_PREVVALID;
10681                 rack->r_ctl.retran_during_recovery = 0;
10682                 rack->r_ctl.dsack_byte_cnt = 0;
10683                 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
10684                 if (rack->r_ctl.rc_went_idle_time == 0)
10685                         rack->r_ctl.rc_went_idle_time = 1;
10686                 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
10687                 if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
10688                         tp->t_acktime = 0;
10689                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
10690         }
10691         if (acked && rack->r_fast_output)
10692                 rack_gain_for_fastoutput(rack, tp, so, (uint32_t)acked);
10693         if (sbavail(&so->so_snd)) {
10694                 rack->r_wanted_output = 1;
10695         }
10696         return (1);
10697 }
10698
10699 /*
10700  * Return value of 1, the TCB is unlocked and most
10701  * likely gone, return value of 0, the TCP is still
10702  * locked.
10703  */
10704 static int
10705 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
10706     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10707     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
10708 {
10709         int32_t ret_val = 0;
10710         int32_t todrop;
10711         int32_t ourfinisacked = 0;
10712         struct tcp_rack *rack;
10713
10714         ctf_calc_rwin(so, tp);
10715         /*
10716          * If the state is SYN_SENT: if seg contains an ACK, but not for our
10717          * SYN, drop the input. if seg contains a RST, then drop the
10718          * connection. if seg does not contain SYN, then drop it. Otherwise
10719          * this is an acceptable SYN segment initialize tp->rcv_nxt and
10720          * tp->irs if seg contains ack then advance tp->snd_una if seg
10721          * contains an ECE and ECN support is enabled, the stream is ECN
10722          * capable. if SYN has been acked change to ESTABLISHED else
10723          * SYN_RCVD state arrange for segment to be acked (eventually)
10724          * continue processing rest of data/controls.
10725          */
10726         if ((thflags & TH_ACK) &&
10727             (SEQ_LEQ(th->th_ack, tp->iss) ||
10728             SEQ_GT(th->th_ack, tp->snd_max))) {
10729                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
10730                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
10731                 return (1);
10732         }
10733         if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
10734                 TCP_PROBE5(connect__refused, NULL, tp,
10735                     mtod(m, const char *), tp, th);
10736                 tp = tcp_drop(tp, ECONNREFUSED);
10737                 ctf_do_drop(m, tp);
10738                 return (1);
10739         }
10740         if (thflags & TH_RST) {
10741                 ctf_do_drop(m, tp);
10742                 return (1);
10743         }
10744         if (!(thflags & TH_SYN)) {
10745                 ctf_do_drop(m, tp);
10746                 return (1);
10747         }
10748         tp->irs = th->th_seq;
10749         tcp_rcvseqinit(tp);
10750         rack = (struct tcp_rack *)tp->t_fb_ptr;
10751         if (thflags & TH_ACK) {
10752                 int tfo_partial = 0;
10753
10754                 KMOD_TCPSTAT_INC(tcps_connects);
10755                 soisconnected(so);
10756 #ifdef MAC
10757                 mac_socketpeer_set_from_mbuf(m, so);
10758 #endif
10759                 /* Do window scaling on this connection? */
10760                 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
10761                     (TF_RCVD_SCALE | TF_REQ_SCALE)) {
10762                         tp->rcv_scale = tp->request_r_scale;
10763                 }
10764                 tp->rcv_adv += min(tp->rcv_wnd,
10765                     TCP_MAXWIN << tp->rcv_scale);
10766                 /*
10767                  * If not all the data that was sent in the TFO SYN
10768                  * has been acked, resend the remainder right away.
10769                  */
10770                 if (IS_FASTOPEN(tp->t_flags) &&
10771                     (tp->snd_una != tp->snd_max)) {
10772                         tp->snd_nxt = th->th_ack;
10773                         tfo_partial = 1;
10774                 }
10775                 /*
10776                  * If there's data, delay ACK; if there's also a FIN ACKNOW
10777                  * will be turned on later.
10778                  */
10779                 if (DELAY_ACK(tp, tlen) && tlen != 0 && !tfo_partial) {
10780                         rack_timer_cancel(tp, rack,
10781                                           rack->r_ctl.rc_rcvtime, __LINE__);
10782                         tp->t_flags |= TF_DELACK;
10783                 } else {
10784                         rack->r_wanted_output = 1;
10785                         tp->t_flags |= TF_ACKNOW;
10786                         rack->rc_dack_toggle = 0;
10787                 }
10788                 if (((thflags & (TH_CWR | TH_ECE)) == TH_ECE) &&
10789                     (V_tcp_do_ecn == 1)) {
10790                         tp->t_flags2 |= TF2_ECN_PERMIT;
10791                         KMOD_TCPSTAT_INC(tcps_ecn_shs);
10792                 }
10793                 if (SEQ_GT(th->th_ack, tp->snd_una)) {
10794                         /*
10795                          * We advance snd_una for the
10796                          * fast open case. If th_ack is
10797                          * acknowledging data beyond
10798                          * snd_una we can't just call
10799                          * ack-processing since the
10800                          * data stream in our send-map
10801                          * will start at snd_una + 1 (one
10802                          * beyond the SYN). If its just
10803                          * equal we don't need to do that
10804                          * and there is no send_map.
10805                          */
10806                         tp->snd_una++;
10807                 }
10808                 /*
10809                  * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
10810                  * SYN_SENT  --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
10811                  */
10812                 tp->t_starttime = ticks;
10813                 if (tp->t_flags & TF_NEEDFIN) {
10814                         tcp_state_change(tp, TCPS_FIN_WAIT_1);
10815                         tp->t_flags &= ~TF_NEEDFIN;
10816                         thflags &= ~TH_SYN;
10817                 } else {
10818                         tcp_state_change(tp, TCPS_ESTABLISHED);
10819                         TCP_PROBE5(connect__established, NULL, tp,
10820                             mtod(m, const char *), tp, th);
10821                         rack_cc_conn_init(tp);
10822                 }
10823         } else {
10824                 /*
10825                  * Received initial SYN in SYN-SENT[*] state => simultaneous
10826                  * open.  If segment contains CC option and there is a
10827                  * cached CC, apply TAO test. If it succeeds, connection is *
10828                  * half-synchronized. Otherwise, do 3-way handshake:
10829                  * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
10830                  * there was no CC option, clear cached CC value.
10831                  */
10832                 tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN);
10833                 tcp_state_change(tp, TCPS_SYN_RECEIVED);
10834         }
10835         INP_WLOCK_ASSERT(tp->t_inpcb);
10836         /*
10837          * Advance th->th_seq to correspond to first data byte. If data,
10838          * trim to stay within window, dropping FIN if necessary.
10839          */
10840         th->th_seq++;
10841         if (tlen > tp->rcv_wnd) {
10842                 todrop = tlen - tp->rcv_wnd;
10843                 m_adj(m, -todrop);
10844                 tlen = tp->rcv_wnd;
10845                 thflags &= ~TH_FIN;
10846                 KMOD_TCPSTAT_INC(tcps_rcvpackafterwin);
10847                 KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
10848         }
10849         tp->snd_wl1 = th->th_seq - 1;
10850         tp->rcv_up = th->th_seq;
10851         /*
10852          * Client side of transaction: already sent SYN and data. If the
10853          * remote host used T/TCP to validate the SYN, our data will be
10854          * ACK'd; if so, enter normal data segment processing in the middle
10855          * of step 5, ack processing. Otherwise, goto step 6.
10856          */
10857         if (thflags & TH_ACK) {
10858                 /* For syn-sent we need to possibly update the rtt */
10859                 if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
10860                         uint32_t t, mcts;
10861
10862                         mcts = tcp_ts_getticks();
10863                         t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
10864                         if (!tp->t_rttlow || tp->t_rttlow > t)
10865                                 tp->t_rttlow = t;
10866                         rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 4);
10867                         tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
10868                         tcp_rack_xmit_timer_commit(rack, tp);
10869                 }
10870                 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val))
10871                         return (ret_val);
10872                 /* We may have changed to FIN_WAIT_1 above */
10873                 if (tp->t_state == TCPS_FIN_WAIT_1) {
10874                         /*
10875                          * In FIN_WAIT_1 STATE in addition to the processing
10876                          * for the ESTABLISHED state if our FIN is now
10877                          * acknowledged then enter FIN_WAIT_2.
10878                          */
10879                         if (ourfinisacked) {
10880                                 /*
10881                                  * If we can't receive any more data, then
10882                                  * closing user can proceed. Starting the
10883                                  * timer is contrary to the specification,
10884                                  * but if we don't get a FIN we'll hang
10885                                  * forever.
10886                                  *
10887                                  * XXXjl: we should release the tp also, and
10888                                  * use a compressed state.
10889                                  */
10890                                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10891                                         soisdisconnected(so);
10892                                         tcp_timer_activate(tp, TT_2MSL,
10893                                             (tcp_fast_finwait2_recycle ?
10894                                             tcp_finwait2_timeout :
10895                                             TP_MAXIDLE(tp)));
10896                                 }
10897                                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
10898                         }
10899                 }
10900         }
10901         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
10902            tiwin, thflags, nxt_pkt));
10903 }
10904
10905 /*
10906  * Return value of 1, the TCB is unlocked and most
10907  * likely gone, return value of 0, the TCP is still
10908  * locked.
10909  */
10910 static int
10911 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
10912     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10913     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
10914 {
10915         struct tcp_rack *rack;
10916         int32_t ret_val = 0;
10917         int32_t ourfinisacked = 0;
10918
10919         ctf_calc_rwin(so, tp);
10920         if ((thflags & TH_ACK) &&
10921             (SEQ_LEQ(th->th_ack, tp->snd_una) ||
10922             SEQ_GT(th->th_ack, tp->snd_max))) {
10923                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
10924                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
10925                 return (1);
10926         }
10927         rack = (struct tcp_rack *)tp->t_fb_ptr;
10928         if (IS_FASTOPEN(tp->t_flags)) {
10929                 /*
10930                  * When a TFO connection is in SYN_RECEIVED, the
10931                  * only valid packets are the initial SYN, a
10932                  * retransmit/copy of the initial SYN (possibly with
10933                  * a subset of the original data), a valid ACK, a
10934                  * FIN, or a RST.
10935                  */
10936                 if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
10937                         tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
10938                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
10939                         return (1);
10940                 } else if (thflags & TH_SYN) {
10941                         /* non-initial SYN is ignored */
10942                         if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
10943                             (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
10944                             (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
10945                                 ctf_do_drop(m, NULL);
10946                                 return (0);
10947                         }
10948                 } else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
10949                         ctf_do_drop(m, NULL);
10950                         return (0);
10951                 }
10952         }
10953         if ((thflags & TH_RST) ||
10954             (tp->t_fin_is_rst && (thflags & TH_FIN)))
10955                 return (ctf_process_rst(m, th, so, tp));
10956         /*
10957          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
10958          * it's less than ts_recent, drop it.
10959          */
10960         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
10961             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
10962                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
10963                         return (ret_val);
10964         }
10965         /*
10966          * In the SYN-RECEIVED state, validate that the packet belongs to
10967          * this connection before trimming the data to fit the receive
10968          * window.  Check the sequence number versus IRS since we know the
10969          * sequence numbers haven't wrapped.  This is a partial fix for the
10970          * "LAND" DoS attack.
10971          */
10972         if (SEQ_LT(th->th_seq, tp->irs)) {
10973                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
10974                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
10975                 return (1);
10976         }
10977         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
10978                               &rack->r_ctl.challenge_ack_ts,
10979                               &rack->r_ctl.challenge_ack_cnt)) {
10980                 return (ret_val);
10981         }
10982         /*
10983          * If last ACK falls within this segment's sequence numbers, record
10984          * its timestamp. NOTE: 1) That the test incorporates suggestions
10985          * from the latest proposal of the tcplw@cray.com list (Braden
10986          * 1993/04/26). 2) That updating only on newer timestamps interferes
10987          * with our earlier PAWS tests, so this check should be solely
10988          * predicated on the sequence space of this segment. 3) That we
10989          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
10990          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
10991          * SEG.Len, This modified check allows us to overcome RFC1323's
10992          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
10993          * p.869. In such cases, we can still calculate the RTT correctly
10994          * when RCV.NXT == Last.ACK.Sent.
10995          */
10996         if ((to->to_flags & TOF_TS) != 0 &&
10997             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
10998             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
10999             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11000                 tp->ts_recent_age = tcp_ts_getticks();
11001                 tp->ts_recent = to->to_tsval;
11002         }
11003         tp->snd_wnd = tiwin;
11004         rack_validate_fo_sendwin_up(tp, rack);
11005         /*
11006          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11007          * is on (half-synchronized state), then queue data for later
11008          * processing; else drop segment and return.
11009          */
11010         if ((thflags & TH_ACK) == 0) {
11011                 if (IS_FASTOPEN(tp->t_flags)) {
11012                         rack_cc_conn_init(tp);
11013                 }
11014                 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11015                     tiwin, thflags, nxt_pkt));
11016         }
11017         KMOD_TCPSTAT_INC(tcps_connects);
11018         soisconnected(so);
11019         /* Do window scaling? */
11020         if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
11021             (TF_RCVD_SCALE | TF_REQ_SCALE)) {
11022                 tp->rcv_scale = tp->request_r_scale;
11023         }
11024         /*
11025          * Make transitions: SYN-RECEIVED  -> ESTABLISHED SYN-RECEIVED* ->
11026          * FIN-WAIT-1
11027          */
11028         tp->t_starttime = ticks;
11029         if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) {
11030                 tcp_fastopen_decrement_counter(tp->t_tfo_pending);
11031                 tp->t_tfo_pending = NULL;
11032         }
11033         if (tp->t_flags & TF_NEEDFIN) {
11034                 tcp_state_change(tp, TCPS_FIN_WAIT_1);
11035                 tp->t_flags &= ~TF_NEEDFIN;
11036         } else {
11037                 tcp_state_change(tp, TCPS_ESTABLISHED);
11038                 TCP_PROBE5(accept__established, NULL, tp,
11039                     mtod(m, const char *), tp, th);
11040                 /*
11041                  * TFO connections call cc_conn_init() during SYN
11042                  * processing.  Calling it again here for such connections
11043                  * is not harmless as it would undo the snd_cwnd reduction
11044                  * that occurs when a TFO SYN|ACK is retransmitted.
11045                  */
11046                 if (!IS_FASTOPEN(tp->t_flags))
11047                         rack_cc_conn_init(tp);
11048         }
11049         /*
11050          * Account for the ACK of our SYN prior to
11051          * regular ACK processing below, except for
11052          * simultaneous SYN, which is handled later.
11053          */
11054         if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN))
11055                 tp->snd_una++;
11056         /*
11057          * If segment contains data or ACK, will call tcp_reass() later; if
11058          * not, do so now to pass queued data to user.
11059          */
11060         if (tlen == 0 && (thflags & TH_FIN) == 0)
11061                 (void) tcp_reass(tp, (struct tcphdr *)0, NULL, 0,
11062                     (struct mbuf *)0);
11063         tp->snd_wl1 = th->th_seq - 1;
11064         /* For syn-recv we need to possibly update the rtt */
11065         if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
11066                 uint32_t t, mcts;
11067
11068                 mcts = tcp_ts_getticks();
11069                 t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
11070                 if (!tp->t_rttlow || tp->t_rttlow > t)
11071                         tp->t_rttlow = t;
11072                 rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 5);
11073                 tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
11074                 tcp_rack_xmit_timer_commit(rack, tp);
11075         }
11076         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11077                 return (ret_val);
11078         }
11079         if (tp->t_state == TCPS_FIN_WAIT_1) {
11080                 /* We could have went to FIN_WAIT_1 (or EST) above */
11081                 /*
11082                  * In FIN_WAIT_1 STATE in addition to the processing for the
11083                  * ESTABLISHED state if our FIN is now acknowledged then
11084                  * enter FIN_WAIT_2.
11085                  */
11086                 if (ourfinisacked) {
11087                         /*
11088                          * If we can't receive any more data, then closing
11089                          * user can proceed. Starting the timer is contrary
11090                          * to the specification, but if we don't get a FIN
11091                          * we'll hang forever.
11092                          *
11093                          * XXXjl: we should release the tp also, and use a
11094                          * compressed state.
11095                          */
11096                         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11097                                 soisdisconnected(so);
11098                                 tcp_timer_activate(tp, TT_2MSL,
11099                                     (tcp_fast_finwait2_recycle ?
11100                                     tcp_finwait2_timeout :
11101                                     TP_MAXIDLE(tp)));
11102                         }
11103                         tcp_state_change(tp, TCPS_FIN_WAIT_2);
11104                 }
11105         }
11106         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11107             tiwin, thflags, nxt_pkt));
11108 }
11109
11110 /*
11111  * Return value of 1, the TCB is unlocked and most
11112  * likely gone, return value of 0, the TCP is still
11113  * locked.
11114  */
11115 static int
11116 rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
11117     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11118     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11119 {
11120         int32_t ret_val = 0;
11121         struct tcp_rack *rack;
11122
11123         /*
11124          * Header prediction: check for the two common cases of a
11125          * uni-directional data xfer.  If the packet has no control flags,
11126          * is in-sequence, the window didn't change and we're not
11127          * retransmitting, it's a candidate.  If the length is zero and the
11128          * ack moved forward, we're the sender side of the xfer.  Just free
11129          * the data acked & wake any higher level process that was blocked
11130          * waiting for space.  If the length is non-zero and the ack didn't
11131          * move, we're the receiver side.  If we're getting packets in-order
11132          * (the reassembly queue is empty), add the data toc The socket
11133          * buffer and note that we need a delayed ack. Make sure that the
11134          * hidden state-flags are also off. Since we check for
11135          * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
11136          */
11137         rack = (struct tcp_rack *)tp->t_fb_ptr;
11138         if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
11139             __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_ACK)) == TH_ACK) &&
11140             __predict_true(SEGQ_EMPTY(tp)) &&
11141             __predict_true(th->th_seq == tp->rcv_nxt)) {
11142                 if (tlen == 0) {
11143                         if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
11144                             tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) {
11145                                 return (0);
11146                         }
11147                 } else {
11148                         if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
11149                             tiwin, nxt_pkt, iptos)) {
11150                                 return (0);
11151                         }
11152                 }
11153         }
11154         ctf_calc_rwin(so, tp);
11155
11156         if ((thflags & TH_RST) ||
11157             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11158                 return (ctf_process_rst(m, th, so, tp));
11159
11160         /*
11161          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11162          * synchronized state.
11163          */
11164         if (thflags & TH_SYN) {
11165                 ctf_challenge_ack(m, th, tp, &ret_val);
11166                 return (ret_val);
11167         }
11168         /*
11169          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11170          * it's less than ts_recent, drop it.
11171          */
11172         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11173             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11174                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11175                         return (ret_val);
11176         }
11177         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11178                               &rack->r_ctl.challenge_ack_ts,
11179                               &rack->r_ctl.challenge_ack_cnt)) {
11180                 return (ret_val);
11181         }
11182         /*
11183          * If last ACK falls within this segment's sequence numbers, record
11184          * its timestamp. NOTE: 1) That the test incorporates suggestions
11185          * from the latest proposal of the tcplw@cray.com list (Braden
11186          * 1993/04/26). 2) That updating only on newer timestamps interferes
11187          * with our earlier PAWS tests, so this check should be solely
11188          * predicated on the sequence space of this segment. 3) That we
11189          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11190          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11191          * SEG.Len, This modified check allows us to overcome RFC1323's
11192          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11193          * p.869. In such cases, we can still calculate the RTT correctly
11194          * when RCV.NXT == Last.ACK.Sent.
11195          */
11196         if ((to->to_flags & TOF_TS) != 0 &&
11197             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11198             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11199             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11200                 tp->ts_recent_age = tcp_ts_getticks();
11201                 tp->ts_recent = to->to_tsval;
11202         }
11203         /*
11204          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11205          * is on (half-synchronized state), then queue data for later
11206          * processing; else drop segment and return.
11207          */
11208         if ((thflags & TH_ACK) == 0) {
11209                 if (tp->t_flags & TF_NEEDSYN) {
11210                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11211                             tiwin, thflags, nxt_pkt));
11212
11213                 } else if (tp->t_flags & TF_ACKNOW) {
11214                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11215                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11216                         return (ret_val);
11217                 } else {
11218                         ctf_do_drop(m, NULL);
11219                         return (0);
11220                 }
11221         }
11222         /*
11223          * Ack processing.
11224          */
11225         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
11226                 return (ret_val);
11227         }
11228         if (sbavail(&so->so_snd)) {
11229                 if (ctf_progress_timeout_check(tp, true)) {
11230                         rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
11231                         tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11232                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11233                         return (1);
11234                 }
11235         }
11236         /* State changes only happen in rack_process_data() */
11237         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11238             tiwin, thflags, nxt_pkt));
11239 }
11240
11241 /*
11242  * Return value of 1, the TCB is unlocked and most
11243  * likely gone, return value of 0, the TCP is still
11244  * locked.
11245  */
11246 static int
11247 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
11248     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11249     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11250 {
11251         int32_t ret_val = 0;
11252         struct tcp_rack *rack;
11253
11254         rack = (struct tcp_rack *)tp->t_fb_ptr;
11255         ctf_calc_rwin(so, tp);
11256         if ((thflags & TH_RST) ||
11257             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11258                 return (ctf_process_rst(m, th, so, tp));
11259         /*
11260          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11261          * synchronized state.
11262          */
11263         if (thflags & TH_SYN) {
11264                 ctf_challenge_ack(m, th, tp, &ret_val);
11265                 return (ret_val);
11266         }
11267         /*
11268          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11269          * it's less than ts_recent, drop it.
11270          */
11271         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11272             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11273                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11274                         return (ret_val);
11275         }
11276         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11277                               &rack->r_ctl.challenge_ack_ts,
11278                               &rack->r_ctl.challenge_ack_cnt)) {
11279                 return (ret_val);
11280         }
11281         /*
11282          * If last ACK falls within this segment's sequence numbers, record
11283          * its timestamp. NOTE: 1) That the test incorporates suggestions
11284          * from the latest proposal of the tcplw@cray.com list (Braden
11285          * 1993/04/26). 2) That updating only on newer timestamps interferes
11286          * with our earlier PAWS tests, so this check should be solely
11287          * predicated on the sequence space of this segment. 3) That we
11288          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11289          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11290          * SEG.Len, This modified check allows us to overcome RFC1323's
11291          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11292          * p.869. In such cases, we can still calculate the RTT correctly
11293          * when RCV.NXT == Last.ACK.Sent.
11294          */
11295         if ((to->to_flags & TOF_TS) != 0 &&
11296             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11297             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11298             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11299                 tp->ts_recent_age = tcp_ts_getticks();
11300                 tp->ts_recent = to->to_tsval;
11301         }
11302         /*
11303          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11304          * is on (half-synchronized state), then queue data for later
11305          * processing; else drop segment and return.
11306          */
11307         if ((thflags & TH_ACK) == 0) {
11308                 if (tp->t_flags & TF_NEEDSYN) {
11309                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11310                             tiwin, thflags, nxt_pkt));
11311
11312                 } else if (tp->t_flags & TF_ACKNOW) {
11313                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11314                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11315                         return (ret_val);
11316                 } else {
11317                         ctf_do_drop(m, NULL);
11318                         return (0);
11319                 }
11320         }
11321         /*
11322          * Ack processing.
11323          */
11324         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
11325                 return (ret_val);
11326         }
11327         if (sbavail(&so->so_snd)) {
11328                 if (ctf_progress_timeout_check(tp, true)) {
11329                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11330                                                 tp, tick, PROGRESS_DROP, __LINE__);
11331                         tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11332                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11333                         return (1);
11334                 }
11335         }
11336         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11337             tiwin, thflags, nxt_pkt));
11338 }
11339
11340 static int
11341 rack_check_data_after_close(struct mbuf *m,
11342     struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so)
11343 {
11344         struct tcp_rack *rack;
11345
11346         rack = (struct tcp_rack *)tp->t_fb_ptr;
11347         if (rack->rc_allow_data_af_clo == 0) {
11348         close_now:
11349                 tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
11350                 /* tcp_close will kill the inp pre-log the Reset */
11351                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
11352                 tp = tcp_close(tp);
11353                 KMOD_TCPSTAT_INC(tcps_rcvafterclose);
11354                 ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen));
11355                 return (1);
11356         }
11357         if (sbavail(&so->so_snd) == 0)
11358                 goto close_now;
11359         /* Ok we allow data that is ignored and a followup reset */
11360         tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
11361         tp->rcv_nxt = th->th_seq + *tlen;
11362         tp->t_flags2 |= TF2_DROP_AF_DATA;
11363         rack->r_wanted_output = 1;
11364         *tlen = 0;
11365         return (0);
11366 }
11367
11368 /*
11369  * Return value of 1, the TCB is unlocked and most
11370  * likely gone, return value of 0, the TCP is still
11371  * locked.
11372  */
11373 static int
11374 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
11375     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11376     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11377 {
11378         int32_t ret_val = 0;
11379         int32_t ourfinisacked = 0;
11380         struct tcp_rack *rack;
11381
11382         rack = (struct tcp_rack *)tp->t_fb_ptr;
11383         ctf_calc_rwin(so, tp);
11384
11385         if ((thflags & TH_RST) ||
11386             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11387                 return (ctf_process_rst(m, th, so, tp));
11388         /*
11389          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11390          * synchronized state.
11391          */
11392         if (thflags & TH_SYN) {
11393                 ctf_challenge_ack(m, th, tp, &ret_val);
11394                 return (ret_val);
11395         }
11396         /*
11397          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11398          * it's less than ts_recent, drop it.
11399          */
11400         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11401             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11402                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11403                         return (ret_val);
11404         }
11405         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11406                               &rack->r_ctl.challenge_ack_ts,
11407                               &rack->r_ctl.challenge_ack_cnt)) {
11408                 return (ret_val);
11409         }
11410         /*
11411          * If new data are received on a connection after the user processes
11412          * are gone, then RST the other end.
11413          */
11414         if ((so->so_state & SS_NOFDREF) && tlen) {
11415                 if (rack_check_data_after_close(m, tp, &tlen, th, so))
11416                         return (1);
11417         }
11418         /*
11419          * If last ACK falls within this segment's sequence numbers, record
11420          * its timestamp. NOTE: 1) That the test incorporates suggestions
11421          * from the latest proposal of the tcplw@cray.com list (Braden
11422          * 1993/04/26). 2) That updating only on newer timestamps interferes
11423          * with our earlier PAWS tests, so this check should be solely
11424          * predicated on the sequence space of this segment. 3) That we
11425          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11426          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11427          * SEG.Len, This modified check allows us to overcome RFC1323's
11428          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11429          * p.869. In such cases, we can still calculate the RTT correctly
11430          * when RCV.NXT == Last.ACK.Sent.
11431          */
11432         if ((to->to_flags & TOF_TS) != 0 &&
11433             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11434             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11435             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11436                 tp->ts_recent_age = tcp_ts_getticks();
11437                 tp->ts_recent = to->to_tsval;
11438         }
11439         /*
11440          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11441          * is on (half-synchronized state), then queue data for later
11442          * processing; else drop segment and return.
11443          */
11444         if ((thflags & TH_ACK) == 0) {
11445                 if (tp->t_flags & TF_NEEDSYN) {
11446                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11447                             tiwin, thflags, nxt_pkt));
11448                 } else if (tp->t_flags & TF_ACKNOW) {
11449                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11450                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11451                         return (ret_val);
11452                 } else {
11453                         ctf_do_drop(m, NULL);
11454                         return (0);
11455                 }
11456         }
11457         /*
11458          * Ack processing.
11459          */
11460         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11461                 return (ret_val);
11462         }
11463         if (ourfinisacked) {
11464                 /*
11465                  * If we can't receive any more data, then closing user can
11466                  * proceed. Starting the timer is contrary to the
11467                  * specification, but if we don't get a FIN we'll hang
11468                  * forever.
11469                  *
11470                  * XXXjl: we should release the tp also, and use a
11471                  * compressed state.
11472                  */
11473                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11474                         soisdisconnected(so);
11475                         tcp_timer_activate(tp, TT_2MSL,
11476                             (tcp_fast_finwait2_recycle ?
11477                             tcp_finwait2_timeout :
11478                             TP_MAXIDLE(tp)));
11479                 }
11480                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
11481         }
11482         if (sbavail(&so->so_snd)) {
11483                 if (ctf_progress_timeout_check(tp, true)) {
11484                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11485                                                 tp, tick, PROGRESS_DROP, __LINE__);
11486                         tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11487                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11488                         return (1);
11489                 }
11490         }
11491         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11492             tiwin, thflags, nxt_pkt));
11493 }
11494
11495 /*
11496  * Return value of 1, the TCB is unlocked and most
11497  * likely gone, return value of 0, the TCP is still
11498  * locked.
11499  */
11500 static int
11501 rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so,
11502     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11503     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11504 {
11505         int32_t ret_val = 0;
11506         int32_t ourfinisacked = 0;
11507         struct tcp_rack *rack;
11508
11509         rack = (struct tcp_rack *)tp->t_fb_ptr;
11510         ctf_calc_rwin(so, tp);
11511
11512         if ((thflags & TH_RST) ||
11513             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11514                 return (ctf_process_rst(m, th, so, tp));
11515         /*
11516          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11517          * synchronized state.
11518          */
11519         if (thflags & TH_SYN) {
11520                 ctf_challenge_ack(m, th, tp, &ret_val);
11521                 return (ret_val);
11522         }
11523         /*
11524          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11525          * it's less than ts_recent, drop it.
11526          */
11527         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11528             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11529                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11530                         return (ret_val);
11531         }
11532         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11533                               &rack->r_ctl.challenge_ack_ts,
11534                               &rack->r_ctl.challenge_ack_cnt)) {
11535                 return (ret_val);
11536         }
11537         /*
11538          * If new data are received on a connection after the user processes
11539          * are gone, then RST the other end.
11540          */
11541         if ((so->so_state & SS_NOFDREF) && tlen) {
11542                 if (rack_check_data_after_close(m, tp, &tlen, th, so))
11543                         return (1);
11544         }
11545         /*
11546          * If last ACK falls within this segment's sequence numbers, record
11547          * its timestamp. NOTE: 1) That the test incorporates suggestions
11548          * from the latest proposal of the tcplw@cray.com list (Braden
11549          * 1993/04/26). 2) That updating only on newer timestamps interferes
11550          * with our earlier PAWS tests, so this check should be solely
11551          * predicated on the sequence space of this segment. 3) That we
11552          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11553          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11554          * SEG.Len, This modified check allows us to overcome RFC1323's
11555          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11556          * p.869. In such cases, we can still calculate the RTT correctly
11557          * when RCV.NXT == Last.ACK.Sent.
11558          */
11559         if ((to->to_flags & TOF_TS) != 0 &&
11560             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11561             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11562             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11563                 tp->ts_recent_age = tcp_ts_getticks();
11564                 tp->ts_recent = to->to_tsval;
11565         }
11566         /*
11567          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11568          * is on (half-synchronized state), then queue data for later
11569          * processing; else drop segment and return.
11570          */
11571         if ((thflags & TH_ACK) == 0) {
11572                 if (tp->t_flags & TF_NEEDSYN) {
11573                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11574                             tiwin, thflags, nxt_pkt));
11575                 } else if (tp->t_flags & TF_ACKNOW) {
11576                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11577                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11578                         return (ret_val);
11579                 } else {
11580                         ctf_do_drop(m, NULL);
11581                         return (0);
11582                 }
11583         }
11584         /*
11585          * Ack processing.
11586          */
11587         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11588                 return (ret_val);
11589         }
11590         if (ourfinisacked) {
11591                 tcp_twstart(tp);
11592                 m_freem(m);
11593                 return (1);
11594         }
11595         if (sbavail(&so->so_snd)) {
11596                 if (ctf_progress_timeout_check(tp, true)) {
11597                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11598                                                 tp, tick, PROGRESS_DROP, __LINE__);
11599                         tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11600                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11601                         return (1);
11602                 }
11603         }
11604         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11605             tiwin, thflags, nxt_pkt));
11606 }
11607
11608 /*
11609  * Return value of 1, the TCB is unlocked and most
11610  * likely gone, return value of 0, the TCP is still
11611  * locked.
11612  */
11613 static int
11614 rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
11615     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11616     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11617 {
11618         int32_t ret_val = 0;
11619         int32_t ourfinisacked = 0;
11620         struct tcp_rack *rack;
11621
11622         rack = (struct tcp_rack *)tp->t_fb_ptr;
11623         ctf_calc_rwin(so, tp);
11624
11625         if ((thflags & TH_RST) ||
11626             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11627                 return (ctf_process_rst(m, th, so, tp));
11628         /*
11629          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11630          * synchronized state.
11631          */
11632         if (thflags & TH_SYN) {
11633                 ctf_challenge_ack(m, th, tp, &ret_val);
11634                 return (ret_val);
11635         }
11636         /*
11637          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11638          * it's less than ts_recent, drop it.
11639          */
11640         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11641             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11642                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11643                         return (ret_val);
11644         }
11645         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11646                               &rack->r_ctl.challenge_ack_ts,
11647                               &rack->r_ctl.challenge_ack_cnt)) {
11648                 return (ret_val);
11649         }
11650         /*
11651          * If new data are received on a connection after the user processes
11652          * are gone, then RST the other end.
11653          */
11654         if ((so->so_state & SS_NOFDREF) && tlen) {
11655                 if (rack_check_data_after_close(m, tp, &tlen, th, so))
11656                         return (1);
11657         }
11658         /*
11659          * If last ACK falls within this segment's sequence numbers, record
11660          * its timestamp. NOTE: 1) That the test incorporates suggestions
11661          * from the latest proposal of the tcplw@cray.com list (Braden
11662          * 1993/04/26). 2) That updating only on newer timestamps interferes
11663          * with our earlier PAWS tests, so this check should be solely
11664          * predicated on the sequence space of this segment. 3) That we
11665          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11666          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11667          * SEG.Len, This modified check allows us to overcome RFC1323's
11668          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11669          * p.869. In such cases, we can still calculate the RTT correctly
11670          * when RCV.NXT == Last.ACK.Sent.
11671          */
11672         if ((to->to_flags & TOF_TS) != 0 &&
11673             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11674             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11675             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11676                 tp->ts_recent_age = tcp_ts_getticks();
11677                 tp->ts_recent = to->to_tsval;
11678         }
11679         /*
11680          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11681          * is on (half-synchronized state), then queue data for later
11682          * processing; else drop segment and return.
11683          */
11684         if ((thflags & TH_ACK) == 0) {
11685                 if (tp->t_flags & TF_NEEDSYN) {
11686                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11687                             tiwin, thflags, nxt_pkt));
11688                 } else if (tp->t_flags & TF_ACKNOW) {
11689                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11690                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11691                         return (ret_val);
11692                 } else {
11693                         ctf_do_drop(m, NULL);
11694                         return (0);
11695                 }
11696         }
11697         /*
11698          * case TCPS_LAST_ACK: Ack processing.
11699          */
11700         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11701                 return (ret_val);
11702         }
11703         if (ourfinisacked) {
11704                 tp = tcp_close(tp);
11705                 ctf_do_drop(m, tp);
11706                 return (1);
11707         }
11708         if (sbavail(&so->so_snd)) {
11709                 if (ctf_progress_timeout_check(tp, true)) {
11710                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11711                                                 tp, tick, PROGRESS_DROP, __LINE__);
11712                         tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11713                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11714                         return (1);
11715                 }
11716         }
11717         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11718             tiwin, thflags, nxt_pkt));
11719 }
11720
11721 /*
11722  * Return value of 1, the TCB is unlocked and most
11723  * likely gone, return value of 0, the TCP is still
11724  * locked.
11725  */
11726 static int
11727 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
11728     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11729     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11730 {
11731         int32_t ret_val = 0;
11732         int32_t ourfinisacked = 0;
11733         struct tcp_rack *rack;
11734
11735         rack = (struct tcp_rack *)tp->t_fb_ptr;
11736         ctf_calc_rwin(so, tp);
11737
11738         /* Reset receive buffer auto scaling when not in bulk receive mode. */
11739         if ((thflags & TH_RST) ||
11740             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11741                 return (ctf_process_rst(m, th, so, tp));
11742         /*
11743          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11744          * synchronized state.
11745          */
11746         if (thflags & TH_SYN) {
11747                 ctf_challenge_ack(m, th, tp, &ret_val);
11748                 return (ret_val);
11749         }
11750         /*
11751          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11752          * it's less than ts_recent, drop it.
11753          */
11754         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11755             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11756                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11757                         return (ret_val);
11758         }
11759         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11760                               &rack->r_ctl.challenge_ack_ts,
11761                               &rack->r_ctl.challenge_ack_cnt)) {
11762                 return (ret_val);
11763         }
11764         /*
11765          * If new data are received on a connection after the user processes
11766          * are gone, then RST the other end.
11767          */
11768         if ((so->so_state & SS_NOFDREF) &&
11769             tlen) {
11770                 if (rack_check_data_after_close(m, tp, &tlen, th, so))
11771                         return (1);
11772         }
11773         /*
11774          * If last ACK falls within this segment's sequence numbers, record
11775          * its timestamp. NOTE: 1) That the test incorporates suggestions
11776          * from the latest proposal of the tcplw@cray.com list (Braden
11777          * 1993/04/26). 2) That updating only on newer timestamps interferes
11778          * with our earlier PAWS tests, so this check should be solely
11779          * predicated on the sequence space of this segment. 3) That we
11780          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11781          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11782          * SEG.Len, This modified check allows us to overcome RFC1323's
11783          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11784          * p.869. In such cases, we can still calculate the RTT correctly
11785          * when RCV.NXT == Last.ACK.Sent.
11786          */
11787         if ((to->to_flags & TOF_TS) != 0 &&
11788             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11789             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11790             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11791                 tp->ts_recent_age = tcp_ts_getticks();
11792                 tp->ts_recent = to->to_tsval;
11793         }
11794         /*
11795          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11796          * is on (half-synchronized state), then queue data for later
11797          * processing; else drop segment and return.
11798          */
11799         if ((thflags & TH_ACK) == 0) {
11800                 if (tp->t_flags & TF_NEEDSYN) {
11801                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11802                             tiwin, thflags, nxt_pkt));
11803                 } else if (tp->t_flags & TF_ACKNOW) {
11804                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11805                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11806                         return (ret_val);
11807                 } else {
11808                         ctf_do_drop(m, NULL);
11809                         return (0);
11810                 }
11811         }
11812         /*
11813          * Ack processing.
11814          */
11815         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11816                 return (ret_val);
11817         }
11818         if (sbavail(&so->so_snd)) {
11819                 if (ctf_progress_timeout_check(tp, true)) {
11820                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11821                                                 tp, tick, PROGRESS_DROP, __LINE__);
11822                         tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11823                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11824                         return (1);
11825                 }
11826         }
11827         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11828             tiwin, thflags, nxt_pkt));
11829 }
11830
11831 static void inline
11832 rack_clear_rate_sample(struct tcp_rack *rack)
11833 {
11834         rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY;
11835         rack->r_ctl.rack_rs.rs_rtt_cnt = 0;
11836         rack->r_ctl.rack_rs.rs_rtt_tot = 0;
11837 }
11838
11839 static void
11840 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override)
11841 {
11842         uint64_t bw_est, rate_wanted;
11843         int chged = 0;
11844         uint32_t user_max, orig_min, orig_max;
11845
11846         orig_min = rack->r_ctl.rc_pace_min_segs;
11847         orig_max = rack->r_ctl.rc_pace_max_segs;
11848         user_max = ctf_fixed_maxseg(tp) * rack->rc_user_set_max_segs;
11849         if (ctf_fixed_maxseg(tp) != rack->r_ctl.rc_pace_min_segs)
11850                 chged = 1;
11851         rack->r_ctl.rc_pace_min_segs = ctf_fixed_maxseg(tp);
11852         if (rack->use_fixed_rate || rack->rc_force_max_seg) {
11853                 if (user_max != rack->r_ctl.rc_pace_max_segs)
11854                         chged = 1;
11855         }
11856         if (rack->rc_force_max_seg) {
11857                 rack->r_ctl.rc_pace_max_segs = user_max;
11858         } else if (rack->use_fixed_rate) {
11859                 bw_est = rack_get_bw(rack);
11860                 if ((rack->r_ctl.crte == NULL) ||
11861                     (bw_est != rack->r_ctl.crte->rate)) {
11862                         rack->r_ctl.rc_pace_max_segs = user_max;
11863                 } else {
11864                         /* We are pacing right at the hardware rate */
11865                         uint32_t segsiz;
11866
11867                         segsiz = min(ctf_fixed_maxseg(tp),
11868                                      rack->r_ctl.rc_pace_min_segs);
11869                         rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(
11870                                                            tp, bw_est, segsiz, 0,
11871                                                            rack->r_ctl.crte, NULL);
11872                 }
11873         } else if (rack->rc_always_pace) {
11874                 if (rack->r_ctl.gp_bw ||
11875 #ifdef NETFLIX_PEAKRATE
11876                     rack->rc_tp->t_maxpeakrate ||
11877 #endif
11878                     rack->r_ctl.init_rate) {
11879                         /* We have a rate of some sort set */
11880                         uint32_t  orig;
11881
11882                         bw_est = rack_get_bw(rack);
11883                         orig = rack->r_ctl.rc_pace_max_segs;
11884                         if (fill_override)
11885                                 rate_wanted = *fill_override;
11886                         else
11887                                 rate_wanted = rack_get_output_bw(rack, bw_est, NULL, NULL);
11888                         if (rate_wanted) {
11889                                 /* We have something */
11890                                 rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack,
11891                                                                                    rate_wanted,
11892                                                                                    ctf_fixed_maxseg(rack->rc_tp));
11893                         } else
11894                                 rack->r_ctl.rc_pace_max_segs = rack->r_ctl.rc_pace_min_segs;
11895                         if (orig != rack->r_ctl.rc_pace_max_segs)
11896                                 chged = 1;
11897                 } else if ((rack->r_ctl.gp_bw == 0) &&
11898                            (rack->r_ctl.rc_pace_max_segs == 0)) {
11899                         /*
11900                          * If we have nothing limit us to bursting
11901                          * out IW sized pieces.
11902                          */
11903                         chged = 1;
11904                         rack->r_ctl.rc_pace_max_segs = rc_init_window(rack);
11905                 }
11906         }
11907         if (rack->r_ctl.rc_pace_max_segs > PACE_MAX_IP_BYTES) {
11908                 chged = 1;
11909                 rack->r_ctl.rc_pace_max_segs = PACE_MAX_IP_BYTES;
11910         }
11911         if (chged)
11912                 rack_log_type_pacing_sizes(tp, rack, orig_min, orig_max, line, 2);
11913 }
11914
11915
11916 static void
11917 rack_init_fsb_block(struct tcpcb *tp, struct tcp_rack *rack)
11918 {
11919 #ifdef INET6
11920         struct ip6_hdr *ip6 = NULL;
11921 #endif
11922 #ifdef INET
11923         struct ip *ip = NULL;
11924 #endif
11925         struct udphdr *udp = NULL;
11926
11927         /* Ok lets fill in the fast block, it can only be used with no IP options! */
11928 #ifdef INET6
11929         if (rack->r_is_v6) {
11930                 rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
11931                 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
11932                 if (tp->t_port) {
11933                         rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
11934                         udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
11935                         udp->uh_sport = htons(V_tcp_udp_tunneling_port);
11936                         udp->uh_dport = tp->t_port;
11937                         rack->r_ctl.fsb.udp = udp;
11938                         rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
11939                 } else
11940                 {
11941                         rack->r_ctl.fsb.th = (struct tcphdr *)(ip6 + 1);
11942                         rack->r_ctl.fsb.udp = NULL;
11943                 }
11944                 tcpip_fillheaders(rack->rc_inp,
11945                                   tp->t_port,
11946                                   ip6, rack->r_ctl.fsb.th);
11947         } else
11948 #endif                          /* INET6 */
11949         {
11950                 rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr);
11951                 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
11952                 if (tp->t_port) {
11953                         rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
11954                         udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
11955                         udp->uh_sport = htons(V_tcp_udp_tunneling_port);
11956                         udp->uh_dport = tp->t_port;
11957                         rack->r_ctl.fsb.udp = udp;
11958                         rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
11959                 } else
11960                 {
11961                         rack->r_ctl.fsb.udp = NULL;
11962                         rack->r_ctl.fsb.th = (struct tcphdr *)(ip + 1);
11963                 }
11964                 tcpip_fillheaders(rack->rc_inp,
11965                                   tp->t_port,
11966                                   ip, rack->r_ctl.fsb.th);
11967         }
11968         rack->r_fsb_inited = 1;
11969 }
11970
11971 static int
11972 rack_init_fsb(struct tcpcb *tp, struct tcp_rack *rack)
11973 {
11974         /*
11975          * Allocate the larger of spaces V6 if available else just
11976          * V4 and include udphdr (overbook)
11977          */
11978 #ifdef INET6
11979         rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + sizeof(struct udphdr);
11980 #else
11981         rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr) + sizeof(struct udphdr);
11982 #endif
11983         rack->r_ctl.fsb.tcp_ip_hdr = malloc(rack->r_ctl.fsb.tcp_ip_hdr_len,
11984                                             M_TCPFSB, M_NOWAIT|M_ZERO);
11985         if (rack->r_ctl.fsb.tcp_ip_hdr == NULL) {
11986                 return (ENOMEM);
11987         }
11988         rack->r_fsb_inited = 0;
11989         return (0);
11990 }
11991
11992 static int
11993 rack_init(struct tcpcb *tp)
11994 {
11995         struct tcp_rack *rack = NULL;
11996         struct rack_sendmap *insret;
11997         uint32_t iwin, snt, us_cts;
11998         int err;
11999
12000         tp->t_fb_ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT);
12001         if (tp->t_fb_ptr == NULL) {
12002                 /*
12003                  * We need to allocate memory but cant. The INP and INP_INFO
12004                  * locks and they are recusive (happens during setup. So a
12005                  * scheme to drop the locks fails :(
12006                  *
12007                  */
12008                 return (ENOMEM);
12009         }
12010         memset(tp->t_fb_ptr, 0, sizeof(struct tcp_rack));
12011
12012         rack = (struct tcp_rack *)tp->t_fb_ptr;
12013         RB_INIT(&rack->r_ctl.rc_mtree);
12014         TAILQ_INIT(&rack->r_ctl.rc_free);
12015         TAILQ_INIT(&rack->r_ctl.rc_tmap);
12016         rack->rc_tp = tp;
12017         rack->rc_inp = tp->t_inpcb;
12018         /* Set the flag */
12019         rack->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
12020         /* Probably not needed but lets be sure */
12021         rack_clear_rate_sample(rack);
12022         /*
12023          * Save off the default values, socket options will poke
12024          * at these if pacing is not on or we have not yet
12025          * reached where pacing is on (gp_ready/fixed enabled).
12026          * When they get set into the CC module (when gp_ready
12027          * is enabled or we enable fixed) then we will set these
12028          * values into the CC and place in here the old values
12029          * so we have a restoral. Then we will set the flag
12030          * rc_pacing_cc_set. That way whenever we turn off pacing
12031          * or switch off this stack, we will know to go restore
12032          * the saved values.
12033          */
12034         rack->r_ctl.rc_saved_beta.beta = V_newreno_beta_ecn;
12035         rack->r_ctl.rc_saved_beta.beta_ecn = V_newreno_beta_ecn;
12036         /* We want abe like behavior as well */
12037         rack->r_ctl.rc_saved_beta.newreno_flags = CC_NEWRENO_BETA_ECN;
12038         rack->r_ctl.rc_reorder_fade = rack_reorder_fade;
12039         rack->rc_allow_data_af_clo = rack_ignore_data_after_close;
12040         rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh;
12041         if (use_rack_rr)
12042                 rack->use_rack_rr = 1;
12043         if (V_tcp_delack_enabled)
12044                 tp->t_delayed_ack = 1;
12045         else
12046                 tp->t_delayed_ack = 0;
12047 #ifdef TCP_ACCOUNTING
12048         if (rack_tcp_accounting) {
12049                 tp->t_flags2 |= TF2_TCP_ACCOUNTING;
12050         }
12051 #endif
12052         if (rack_enable_shared_cwnd)
12053                 rack->rack_enable_scwnd = 1;
12054         rack->rc_user_set_max_segs = rack_hptsi_segments;
12055         rack->rc_force_max_seg = 0;
12056         if (rack_use_imac_dack)
12057                 rack->rc_dack_mode = 1;
12058         TAILQ_INIT(&rack->r_ctl.opt_list);
12059         rack->r_ctl.rc_reorder_shift = rack_reorder_thresh;
12060         rack->r_ctl.rc_pkt_delay = rack_pkt_delay;
12061         rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp;
12062         rack->r_ctl.rc_lowest_us_rtt = 0xffffffff;
12063         rack->r_ctl.rc_highest_us_rtt = 0;
12064         rack->r_ctl.bw_rate_cap = rack_bw_rate_cap;
12065         if (rack_use_cmp_acks)
12066                 rack->r_use_cmp_ack = 1;
12067         if (rack_disable_prr)
12068                 rack->rack_no_prr = 1;
12069         if (rack_gp_no_rec_chg)
12070                 rack->rc_gp_no_rec_chg = 1;
12071         if (rack_pace_every_seg && tcp_can_enable_pacing()) {
12072                 rack->rc_always_pace = 1;
12073                 if (rack->use_fixed_rate || rack->gp_ready)
12074                         rack_set_cc_pacing(rack);
12075         } else
12076                 rack->rc_always_pace = 0;
12077         if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack)
12078                 rack->r_mbuf_queue = 1;
12079         else
12080                 rack->r_mbuf_queue = 0;
12081         if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
12082                 tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
12083         else
12084                 tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
12085         rack_set_pace_segments(tp, rack, __LINE__, NULL);
12086         if (rack_limits_scwnd)
12087                 rack->r_limit_scw = 1;
12088         else
12089                 rack->r_limit_scw = 0;
12090         rack->rc_labc = V_tcp_abc_l_var;
12091         rack->r_ctl.rc_high_rwnd = tp->snd_wnd;
12092         rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
12093         rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method;
12094         rack->rack_tlp_threshold_use = rack_tlp_threshold_use;
12095         rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr;
12096         rack->r_ctl.rc_min_to = rack_min_to;
12097         microuptime(&rack->r_ctl.act_rcv_time);
12098         rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
12099         rack->r_running_late = 0;
12100         rack->r_running_early = 0;
12101         rack->rc_init_win = rack_default_init_window;
12102         rack->r_ctl.rack_per_of_gp_ss = rack_per_of_gp_ss;
12103         if (rack_hw_up_only)
12104                 rack->r_up_only = 1;
12105         if (rack_do_dyn_mul) {
12106                 /* When dynamic adjustment is on CA needs to start at 100% */
12107                 rack->rc_gp_dyn_mul = 1;
12108                 if (rack_do_dyn_mul >= 100)
12109                         rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
12110         } else
12111                 rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
12112         rack->r_ctl.rack_per_of_gp_rec = rack_per_of_gp_rec;
12113         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
12114         rack->r_ctl.rc_tlp_rxt_last_time = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time);
12115         setup_time_filter_small(&rack->r_ctl.rc_gp_min_rtt, FILTER_TYPE_MIN,
12116                                 rack_probertt_filter_life);
12117         us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
12118         rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
12119         rack->r_ctl.rc_time_of_last_probertt = us_cts;
12120         rack->r_ctl.challenge_ack_ts = tcp_ts_getticks();
12121         rack->r_ctl.rc_time_probertt_starts = 0;
12122         /* We require at least one measurement, even if the sysctl is 0 */
12123         if (rack_req_measurements)
12124                 rack->r_ctl.req_measurements = rack_req_measurements;
12125         else
12126                 rack->r_ctl.req_measurements = 1;
12127         if (rack_enable_hw_pacing)
12128                 rack->rack_hdw_pace_ena = 1;
12129         if (rack_hw_rate_caps)
12130                 rack->r_rack_hw_rate_caps = 1;
12131         /* Do we force on detection? */
12132 #ifdef NETFLIX_EXP_DETECTION
12133         if (tcp_force_detection)
12134                 rack->do_detection = 1;
12135         else
12136 #endif
12137                 rack->do_detection = 0;
12138         if (rack_non_rxt_use_cr)
12139                 rack->rack_rec_nonrxt_use_cr = 1;
12140         err = rack_init_fsb(tp, rack);
12141         if (err) {
12142                 uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12143                 tp->t_fb_ptr = NULL;
12144                 return (err);
12145         }
12146         if (tp->snd_una != tp->snd_max) {
12147                 /* Create a send map for the current outstanding data */
12148                 struct rack_sendmap *rsm;
12149
12150                 rsm = rack_alloc(rack);
12151                 if (rsm == NULL) {
12152                         uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12153                         tp->t_fb_ptr = NULL;
12154                         return (ENOMEM);
12155                 }
12156                 rsm->r_no_rtt_allowed = 1;
12157                 rsm->r_tim_lastsent[0] = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
12158                 rsm->r_rtr_cnt = 1;
12159                 rsm->r_rtr_bytes = 0;
12160                 if (tp->t_flags & TF_SENTFIN) {
12161                         rsm->r_end = tp->snd_max - 1;
12162                         rsm->r_flags |= RACK_HAS_FIN;
12163                 } else {
12164                         rsm->r_end = tp->snd_max;
12165                 }
12166                 if (tp->snd_una == tp->iss) {
12167                         /* The data space is one beyond snd_una */
12168                         rsm->r_flags |= RACK_HAS_SYN;
12169                         rsm->r_start = tp->iss;
12170                         rsm->r_end = rsm->r_start + (tp->snd_max - tp->snd_una);
12171                 } else
12172                         rsm->r_start = tp->snd_una;
12173                 rsm->r_dupack = 0;
12174                 if (rack->rc_inp->inp_socket->so_snd.sb_mb != NULL) {
12175                         rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd, 0, &rsm->soff);
12176                         rsm->orig_m_len = rsm->m->m_len;
12177                 } else {
12178                         /*
12179                          * This can happen if we have a stand-alone FIN or
12180                          *  SYN.
12181                          */
12182                         rsm->m = NULL;
12183                         rsm->orig_m_len = 0;
12184                         rsm->soff = 0;
12185                 }
12186                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12187 #ifdef INVARIANTS
12188                 if (insret != NULL) {
12189                         panic("Insert in rb tree fails ret:%p rack:%p rsm:%p",
12190                               insret, rack, rsm);
12191                 }
12192 #endif
12193                 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
12194                 rsm->r_in_tmap = 1;
12195         }
12196         /*
12197          * Timers in Rack are kept in microseconds so lets
12198          * convert any initial incoming variables
12199          * from ticks into usecs. Note that we
12200          * also change the values of t_srtt and t_rttvar, if
12201          * they are non-zero. They are kept with a 5
12202          * bit decimal so we have to carefully convert
12203          * these to get the full precision.
12204          */
12205         rack_convert_rtts(tp);
12206         tp->t_rttlow = TICKS_2_USEC(tp->t_rttlow);
12207         if (rack_def_profile)
12208                 rack_set_profile(rack, rack_def_profile);
12209         /* Cancel the GP measurement in progress */
12210         tp->t_flags &= ~TF_GPUTINPROG;
12211         if (SEQ_GT(tp->snd_max, tp->iss))
12212                 snt = tp->snd_max - tp->iss;
12213         else
12214                 snt = 0;
12215         iwin = rc_init_window(rack);
12216         if (snt < iwin) {
12217                 /* We are not past the initial window
12218                  * so we need to make sure cwnd is
12219                  * correct.
12220                  */
12221                 if (tp->snd_cwnd < iwin)
12222                         tp->snd_cwnd = iwin;
12223                 /*
12224                  * If we are within the initial window
12225                  * we want ssthresh to be unlimited. Setting
12226                  * it to the rwnd (which the default stack does
12227                  * and older racks) is not really a good idea
12228                  * since we want to be in SS and grow both the
12229                  * cwnd and the rwnd (via dynamic rwnd growth). If
12230                  * we set it to the rwnd then as the peer grows its
12231                  * rwnd we will be stuck in CA and never hit SS.
12232                  *
12233                  * Its far better to raise it up high (this takes the
12234                  * risk that there as been a loss already, probably
12235                  * we should have an indicator in all stacks of loss
12236                  * but we don't), but considering the normal use this
12237                  * is a risk worth taking. The consequences of not
12238                  * hitting SS are far worse than going one more time
12239                  * into it early on (before we have sent even a IW).
12240                  * It is highly unlikely that we will have had a loss
12241                  * before getting the IW out.
12242                  */
12243                 tp->snd_ssthresh = 0xffffffff;
12244         }
12245         rack_stop_all_timers(tp);
12246         /* Lets setup the fsb block */
12247         rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
12248         rack_log_rtt_shrinks(rack,  us_cts,  tp->t_rxtcur,
12249                              __LINE__, RACK_RTTS_INIT);
12250         return (0);
12251 }
12252
12253 static int
12254 rack_handoff_ok(struct tcpcb *tp)
12255 {
12256         if ((tp->t_state == TCPS_CLOSED) ||
12257             (tp->t_state == TCPS_LISTEN)) {
12258                 /* Sure no problem though it may not stick */
12259                 return (0);
12260         }
12261         if ((tp->t_state == TCPS_SYN_SENT) ||
12262             (tp->t_state == TCPS_SYN_RECEIVED)) {
12263                 /*
12264                  * We really don't know if you support sack,
12265                  * you have to get to ESTAB or beyond to tell.
12266                  */
12267                 return (EAGAIN);
12268         }
12269         if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) > 1)) {
12270                 /*
12271                  * Rack will only send a FIN after all data is acknowledged.
12272                  * So in this case we have more data outstanding. We can't
12273                  * switch stacks until either all data and only the FIN
12274                  * is left (in which case rack_init() now knows how
12275                  * to deal with that) <or> all is acknowledged and we
12276                  * are only left with incoming data, though why you
12277                  * would want to switch to rack after all data is acknowledged
12278                  * I have no idea (rrs)!
12279                  */
12280                 return (EAGAIN);
12281         }
12282         if ((tp->t_flags & TF_SACK_PERMIT) || rack_sack_not_required){
12283                 return (0);
12284         }
12285         /*
12286          * If we reach here we don't do SACK on this connection so we can
12287          * never do rack.
12288          */
12289         return (EINVAL);
12290 }
12291
12292
12293 static void
12294 rack_fini(struct tcpcb *tp, int32_t tcb_is_purged)
12295 {
12296         int ack_cmp = 0;
12297
12298         if (tp->t_fb_ptr) {
12299                 struct tcp_rack *rack;
12300                 struct rack_sendmap *rsm, *nrsm, *rm;
12301
12302                 rack = (struct tcp_rack *)tp->t_fb_ptr;
12303                 if (tp->t_in_pkt) {
12304                         /*
12305                          * Since we are switching we need to process any
12306                          * inbound packets in case a compressed ack is
12307                          * in queue or the new stack does not support
12308                          * mbuf queuing. These packets in theory should
12309                          * have been handled by the old stack anyway.
12310                          */
12311                         if ((rack->rc_inp->inp_flags & (INP_DROPPED|INP_TIMEWAIT)) ||
12312                             (rack->rc_inp->inp_flags2 & INP_FREED)) {
12313                                 /* Kill all the packets */
12314                                 struct mbuf *save, *m;
12315
12316                                 m = tp->t_in_pkt;
12317                                 tp->t_in_pkt = NULL;
12318                                 tp->t_tail_pkt = NULL;
12319                                 while (m) {
12320                                         save = m->m_nextpkt;
12321                                         m->m_nextpkt = NULL;
12322                                         m_freem(m);
12323                                         m = save;
12324                                 }
12325                         } else {
12326                                 /* Process all the packets */
12327                                 ctf_do_queued_segments(rack->rc_inp->inp_socket, rack->rc_tp, 0);
12328                         }
12329                         if ((tp->t_inpcb) &&
12330                             (tp->t_inpcb->inp_flags2 & INP_MBUF_ACKCMP))
12331                                 ack_cmp = 1;
12332                         if (ack_cmp) {
12333                                 /* Total if we used large or small (if ack-cmp was used). */
12334                                 if (rack->rc_inp->inp_flags2 & INP_MBUF_L_ACKS)
12335                                         counter_u64_add(rack_large_ackcmp, 1);
12336                                 else
12337                                         counter_u64_add(rack_small_ackcmp, 1);
12338                         }
12339                 }
12340                 tp->t_flags &= ~TF_FORCEDATA;
12341 #ifdef NETFLIX_SHARED_CWND
12342                 if (rack->r_ctl.rc_scw) {
12343                         uint32_t limit;
12344
12345                         if (rack->r_limit_scw)
12346                                 limit = max(1, rack->r_ctl.rc_lowest_us_rtt);
12347                         else
12348                                 limit = 0;
12349                         tcp_shared_cwnd_free_full(tp, rack->r_ctl.rc_scw,
12350                                                   rack->r_ctl.rc_scw_index,
12351                                                   limit);
12352                         rack->r_ctl.rc_scw = NULL;
12353                 }
12354 #endif
12355                 if (rack->r_ctl.fsb.tcp_ip_hdr) {
12356                         free(rack->r_ctl.fsb.tcp_ip_hdr, M_TCPFSB);
12357                         rack->r_ctl.fsb.tcp_ip_hdr = NULL;
12358                         rack->r_ctl.fsb.th = NULL;
12359                 }
12360                 /* Convert back to ticks, with  */
12361                 if (tp->t_srtt > 1) {
12362                         uint32_t val, frac;
12363
12364                         val = USEC_2_TICKS(tp->t_srtt);
12365                         frac = tp->t_srtt % (HPTS_USEC_IN_SEC / hz);
12366                         tp->t_srtt = val << TCP_RTT_SHIFT;
12367                         /*
12368                          * frac is the fractional part here is left
12369                          * over from converting to hz and shifting.
12370                          * We need to convert this to the 5 bit
12371                          * remainder.
12372                          */
12373                         if (frac) {
12374                                 if (hz == 1000) {
12375                                         frac = (((uint64_t)frac *  (uint64_t)TCP_RTT_SCALE) / (uint64_t)HPTS_USEC_IN_MSEC);
12376                                 } else {
12377                                         frac = (((uint64_t)frac * (uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE) /(uint64_t)HPTS_USEC_IN_SEC);
12378                                 }
12379                                 tp->t_srtt += frac;
12380                         }
12381                 }
12382                 if (tp->t_rttvar) {
12383                         uint32_t val, frac;
12384
12385                         val = USEC_2_TICKS(tp->t_rttvar);
12386                         frac = tp->t_srtt % (HPTS_USEC_IN_SEC / hz);
12387                         tp->t_rttvar = val <<  TCP_RTTVAR_SHIFT;
12388                         /*
12389                          * frac is the fractional part here is left
12390                          * over from converting to hz and shifting.
12391                          * We need to convert this to the 5 bit
12392                          * remainder.
12393                          */
12394                         if (frac) {
12395                                 if (hz == 1000) {
12396                                         frac = (((uint64_t)frac *  (uint64_t)TCP_RTT_SCALE) / (uint64_t)HPTS_USEC_IN_MSEC);
12397                                 } else {
12398                                         frac = (((uint64_t)frac * (uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE) /(uint64_t)HPTS_USEC_IN_SEC);
12399                                 }
12400                                 tp->t_rttvar += frac;
12401                         }
12402                 }
12403                 tp->t_rxtcur = USEC_2_TICKS(tp->t_rxtcur);
12404                 tp->t_rttlow = USEC_2_TICKS(tp->t_rttlow);
12405                 if (rack->rc_always_pace) {
12406                         tcp_decrement_paced_conn();
12407                         rack_undo_cc_pacing(rack);
12408                         rack->rc_always_pace = 0;
12409                 }
12410                 /* Clean up any options if they were not applied */
12411                 while (!TAILQ_EMPTY(&rack->r_ctl.opt_list)) {
12412                         struct deferred_opt_list *dol;
12413
12414                         dol = TAILQ_FIRST(&rack->r_ctl.opt_list);
12415                         TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
12416                         free(dol, M_TCPDO);
12417                 }
12418                 /* rack does not use force data but other stacks may clear it */
12419                 if (rack->r_ctl.crte != NULL) {
12420                         tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
12421                         rack->rack_hdrw_pacing = 0;
12422                         rack->r_ctl.crte = NULL;
12423                 }
12424 #ifdef TCP_BLACKBOX
12425                 tcp_log_flowend(tp);
12426 #endif
12427                 RB_FOREACH_SAFE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm) {
12428                         rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12429 #ifdef INVARIANTS
12430                         if (rm != rsm) {
12431                                 panic("At fini, rack:%p rsm:%p rm:%p",
12432                                       rack, rsm, rm);
12433                         }
12434 #endif
12435                         uma_zfree(rack_zone, rsm);
12436                 }
12437                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
12438                 while (rsm) {
12439                         TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
12440                         uma_zfree(rack_zone, rsm);
12441                         rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
12442                 }
12443                 rack->rc_free_cnt = 0;
12444                 uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12445                 tp->t_fb_ptr = NULL;
12446         }
12447         if (tp->t_inpcb) {
12448                 tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
12449                 tp->t_inpcb->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
12450                 tp->t_inpcb->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
12451                 tp->t_inpcb->inp_flags2 &= ~INP_MBUF_ACKCMP;
12452                 /* Cancel the GP measurement in progress */
12453                 tp->t_flags &= ~TF_GPUTINPROG;
12454                 tp->t_inpcb->inp_flags2 &= ~INP_MBUF_L_ACKS;
12455         }
12456         /* Make sure snd_nxt is correctly set */
12457         tp->snd_nxt = tp->snd_max;
12458 }
12459
12460 static void
12461 rack_set_state(struct tcpcb *tp, struct tcp_rack *rack)
12462 {
12463         switch (tp->t_state) {
12464         case TCPS_SYN_SENT:
12465                 rack->r_state = TCPS_SYN_SENT;
12466                 rack->r_substate = rack_do_syn_sent;
12467                 break;
12468         case TCPS_SYN_RECEIVED:
12469                 rack->r_state = TCPS_SYN_RECEIVED;
12470                 rack->r_substate = rack_do_syn_recv;
12471                 break;
12472         case TCPS_ESTABLISHED:
12473                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12474                 rack->r_state = TCPS_ESTABLISHED;
12475                 rack->r_substate = rack_do_established;
12476                 break;
12477         case TCPS_CLOSE_WAIT:
12478                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12479                 rack->r_state = TCPS_CLOSE_WAIT;
12480                 rack->r_substate = rack_do_close_wait;
12481                 break;
12482         case TCPS_FIN_WAIT_1:
12483                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12484                 rack->r_state = TCPS_FIN_WAIT_1;
12485                 rack->r_substate = rack_do_fin_wait_1;
12486                 break;
12487         case TCPS_CLOSING:
12488                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12489                 rack->r_state = TCPS_CLOSING;
12490                 rack->r_substate = rack_do_closing;
12491                 break;
12492         case TCPS_LAST_ACK:
12493                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12494                 rack->r_state = TCPS_LAST_ACK;
12495                 rack->r_substate = rack_do_lastack;
12496                 break;
12497         case TCPS_FIN_WAIT_2:
12498                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12499                 rack->r_state = TCPS_FIN_WAIT_2;
12500                 rack->r_substate = rack_do_fin_wait_2;
12501                 break;
12502         case TCPS_LISTEN:
12503         case TCPS_CLOSED:
12504         case TCPS_TIME_WAIT:
12505         default:
12506                 break;
12507         };
12508         if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
12509                 rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
12510
12511 }
12512
12513 static void
12514 rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb)
12515 {
12516         /*
12517          * We received an ack, and then did not
12518          * call send or were bounced out due to the
12519          * hpts was running. Now a timer is up as well, is
12520          * it the right timer?
12521          */
12522         struct rack_sendmap *rsm;
12523         int tmr_up;
12524
12525         tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
12526         if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT))
12527                 return;
12528         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
12529         if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) &&
12530             (tmr_up == PACE_TMR_RXT)) {
12531                 /* Should be an RXT */
12532                 return;
12533         }
12534         if (rsm == NULL) {
12535                 /* Nothing outstanding? */
12536                 if (tp->t_flags & TF_DELACK) {
12537                         if (tmr_up == PACE_TMR_DELACK)
12538                                 /* We are supposed to have delayed ack up and we do */
12539                                 return;
12540                 } else if (sbavail(&tp->t_inpcb->inp_socket->so_snd) && (tmr_up == PACE_TMR_RXT)) {
12541                         /*
12542                          * if we hit enobufs then we would expect the possiblity
12543                          * of nothing outstanding and the RXT up (and the hptsi timer).
12544                          */
12545                         return;
12546                 } else if (((V_tcp_always_keepalive ||
12547                              rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
12548                             (tp->t_state <= TCPS_CLOSING)) &&
12549                            (tmr_up == PACE_TMR_KEEP) &&
12550                            (tp->snd_max == tp->snd_una)) {
12551                         /* We should have keep alive up and we do */
12552                         return;
12553                 }
12554         }
12555         if (SEQ_GT(tp->snd_max, tp->snd_una) &&
12556                    ((tmr_up == PACE_TMR_TLP) ||
12557                     (tmr_up == PACE_TMR_RACK) ||
12558                     (tmr_up == PACE_TMR_RXT))) {
12559                 /*
12560                  * Either a Rack, TLP or RXT is fine if  we
12561                  * have outstanding data.
12562                  */
12563                 return;
12564         } else if (tmr_up == PACE_TMR_DELACK) {
12565                 /*
12566                  * If the delayed ack was going to go off
12567                  * before the rtx/tlp/rack timer were going to
12568                  * expire, then that would be the timer in control.
12569                  * Note we don't check the time here trusting the
12570                  * code is correct.
12571                  */
12572                 return;
12573         }
12574         /*
12575          * Ok the timer originally started is not what we want now.
12576          * We will force the hpts to be stopped if any, and restart
12577          * with the slot set to what was in the saved slot.
12578          */
12579         if (rack->rc_inp->inp_in_hpts) {
12580                 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
12581                         uint32_t us_cts;
12582
12583                         us_cts = tcp_get_usecs(NULL);
12584                         if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
12585                                 rack->r_early = 1;
12586                                 rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
12587                         }
12588                         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
12589                 }
12590                 tcp_hpts_remove(tp->t_inpcb, HPTS_REMOVE_OUTPUT);
12591         }
12592         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
12593         rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
12594 }
12595
12596
12597 static void
12598 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)
12599 {
12600         tp->snd_wnd = tiwin;
12601         rack_validate_fo_sendwin_up(tp, rack);
12602         tp->snd_wl1 = seq;
12603         tp->snd_wl2 = ack;
12604         if (tp->snd_wnd > tp->max_sndwnd)
12605                 tp->max_sndwnd = tp->snd_wnd;
12606         if (tp->snd_wnd < (tp->snd_max - high_seq)) {
12607                 /* The peer collapsed the window */
12608                 rack_collapsed_window(rack);
12609         } else if (rack->rc_has_collapsed)
12610                 rack_un_collapse_window(rack);
12611         /* Do we exit persists? */
12612         if ((rack->rc_in_persist != 0) &&
12613             (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
12614                                 rack->r_ctl.rc_pace_min_segs))) {
12615                 rack_exit_persist(tp, rack, cts);
12616         }
12617         /* Do we enter persists? */
12618         if ((rack->rc_in_persist == 0) &&
12619             (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
12620             TCPS_HAVEESTABLISHED(tp->t_state) &&
12621             (tp->snd_max == tp->snd_una) &&
12622             sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
12623             (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
12624                 /*
12625                  * Here the rwnd is less than
12626                  * the pacing size, we are established,
12627                  * nothing is outstanding, and there is
12628                  * data to send. Enter persists.
12629                  */
12630                 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
12631         }
12632 }
12633
12634 static void
12635 rack_log_input_packet(struct tcpcb *tp, struct tcp_rack *rack, struct tcp_ackent *ae, int ackval, uint32_t high_seq)
12636 {
12637
12638         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
12639                 union tcp_log_stackspecific log;
12640                 struct timeval ltv;
12641                 char tcp_hdr_buf[60];
12642                 struct tcphdr *th;
12643                 struct timespec ts;
12644                 uint32_t orig_snd_una;
12645                 uint8_t xx = 0;
12646
12647 #ifdef NETFLIX_HTTP_LOGGING
12648                 struct http_sendfile_track *http_req;
12649
12650                 if (SEQ_GT(ae->ack, tp->snd_una)) {
12651                         http_req = tcp_http_find_req_for_seq(tp, (ae->ack-1));
12652                 } else {
12653                         http_req = tcp_http_find_req_for_seq(tp, ae->ack);
12654                 }
12655 #endif
12656                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
12657                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
12658                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
12659                 if (rack->rack_no_prr == 0)
12660                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
12661                 else
12662                         log.u_bbr.flex1 = 0;
12663                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
12664                 log.u_bbr.use_lt_bw <<= 1;
12665                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
12666                 log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
12667                 log.u_bbr.inflight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
12668                 log.u_bbr.pkts_out = tp->t_maxseg;
12669                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
12670                 log.u_bbr.flex7 = 1;
12671                 log.u_bbr.lost = ae->flags;
12672                 log.u_bbr.cwnd_gain = ackval;
12673                 log.u_bbr.pacing_gain = 0x2;
12674                 if (ae->flags & TSTMP_HDWR) {
12675                         /* Record the hardware timestamp if present */
12676                         log.u_bbr.flex3 = M_TSTMP;
12677                         ts.tv_sec = ae->timestamp / 1000000000;
12678                         ts.tv_nsec = ae->timestamp % 1000000000;
12679                         ltv.tv_sec = ts.tv_sec;
12680                         ltv.tv_usec = ts.tv_nsec / 1000;
12681                         log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
12682                 } else if (ae->flags & TSTMP_LRO) {
12683                         /* Record the LRO the arrival timestamp */
12684                         log.u_bbr.flex3 = M_TSTMP_LRO;
12685                         ts.tv_sec = ae->timestamp / 1000000000;
12686                         ts.tv_nsec = ae->timestamp % 1000000000;
12687                         ltv.tv_sec = ts.tv_sec;
12688                         ltv.tv_usec = ts.tv_nsec / 1000;
12689                         log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
12690                 }
12691                 log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
12692                 /* Log the rcv time */
12693                 log.u_bbr.delRate = ae->timestamp;
12694 #ifdef NETFLIX_HTTP_LOGGING
12695                 log.u_bbr.applimited = tp->t_http_closed;
12696                 log.u_bbr.applimited <<= 8;
12697                 log.u_bbr.applimited |= tp->t_http_open;
12698                 log.u_bbr.applimited <<= 8;
12699                 log.u_bbr.applimited |= tp->t_http_req;
12700                 if (http_req) {
12701                         /* Copy out any client req info */
12702                         /* seconds */
12703                         log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC);
12704                         /* useconds */
12705                         log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC);
12706                         log.u_bbr.rttProp = http_req->timestamp;
12707                         log.u_bbr.cur_del_rate = http_req->start;
12708                         if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) {
12709                                 log.u_bbr.flex8 |= 1;
12710                         } else {
12711                                 log.u_bbr.flex8 |= 2;
12712                                 log.u_bbr.bw_inuse = http_req->end;
12713                         }
12714                         log.u_bbr.flex6 = http_req->start_seq;
12715                         if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) {
12716                                 log.u_bbr.flex8 |= 4;
12717                                 log.u_bbr.epoch = http_req->end_seq;
12718                         }
12719                 }
12720 #endif
12721                 memset(tcp_hdr_buf, 0, sizeof(tcp_hdr_buf));
12722                 th = (struct tcphdr *)tcp_hdr_buf;
12723                 th->th_seq = ae->seq;
12724                 th->th_ack = ae->ack;
12725                 th->th_win = ae->win;
12726                 /* Now fill in the ports */
12727                 th->th_sport = tp->t_inpcb->inp_fport;
12728                 th->th_dport = tp->t_inpcb->inp_lport;
12729                 th->th_flags = ae->flags & 0xff;
12730                 /* Now do we have a timestamp option? */
12731                 if (ae->flags & HAS_TSTMP) {
12732                         u_char *cp;
12733                         uint32_t val;
12734
12735                         th->th_off = ((sizeof(struct tcphdr) + TCPOLEN_TSTAMP_APPA) >> 2);
12736                         cp = (u_char *)(th + 1);
12737                         *cp = TCPOPT_NOP;
12738                         cp++;
12739                         *cp = TCPOPT_NOP;
12740                         cp++;
12741                         *cp = TCPOPT_TIMESTAMP;
12742                         cp++;
12743                         *cp = TCPOLEN_TIMESTAMP;
12744                         cp++;
12745                         val = htonl(ae->ts_value);
12746                         bcopy((char *)&val,
12747                               (char *)cp, sizeof(uint32_t));
12748                         val = htonl(ae->ts_echo);
12749                         bcopy((char *)&val,
12750                               (char *)(cp + 4), sizeof(uint32_t));
12751                 } else
12752                         th->th_off = (sizeof(struct tcphdr) >> 2);
12753
12754                 /*
12755                  * For sane logging we need to play a little trick.
12756                  * If the ack were fully processed we would have moved
12757                  * snd_una to high_seq, but since compressed acks are
12758                  * processed in two phases, at this point (logging) snd_una
12759                  * won't be advanced. So we would see multiple acks showing
12760                  * the advancement. We can prevent that by "pretending" that
12761                  * snd_una was advanced and then un-advancing it so that the
12762                  * logging code has the right value for tlb_snd_una.
12763                  */
12764                 if (tp->snd_una != high_seq) {
12765                         orig_snd_una = tp->snd_una;
12766                         tp->snd_una = high_seq;
12767                         xx = 1;
12768                 } else
12769                         xx = 0;
12770                 TCP_LOG_EVENTP(tp, th,
12771                                &tp->t_inpcb->inp_socket->so_rcv,
12772                                &tp->t_inpcb->inp_socket->so_snd, TCP_LOG_IN, 0,
12773                                0, &log, true, &ltv);
12774                 if (xx) {
12775                         tp->snd_una = orig_snd_una;
12776                 }
12777         }
12778
12779 }
12780
12781 static int
12782 rack_do_compressed_ack_processing(struct tcpcb *tp, struct socket *so, struct mbuf *m, int nxt_pkt, struct timeval *tv)
12783 {
12784         /*
12785          * Handle a "special" compressed ack mbuf. Each incoming
12786          * ack has only four possible dispositions:
12787          *
12788          * A) It moves the cum-ack forward
12789          * B) It is behind the cum-ack.
12790          * C) It is a window-update ack.
12791          * D) It is a dup-ack.
12792          *
12793          * Note that we can have between 1 -> TCP_COMP_ACK_ENTRIES
12794          * in the incoming mbuf. We also need to still pay attention
12795          * to nxt_pkt since there may be another packet after this
12796          * one.
12797          */
12798 #ifdef TCP_ACCOUNTING
12799         uint64_t ts_val;
12800         uint64_t rdstc;
12801 #endif
12802         int segsiz;
12803         struct timespec ts;
12804         struct tcp_rack *rack;
12805         struct tcp_ackent *ae;
12806         uint32_t tiwin, us_cts, cts, acked, acked_amount, high_seq, win_seq, the_win, win_upd_ack;
12807         int cnt, i, did_out, ourfinisacked = 0;
12808         int win_up_req = 0;
12809         struct tcpopt to_holder, *to = NULL;
12810         int nsegs = 0;
12811         int under_pacing = 1;
12812         int recovery = 0;
12813         int idx;
12814 #ifdef TCP_ACCOUNTING
12815         sched_pin();
12816 #endif
12817         rack = (struct tcp_rack *)tp->t_fb_ptr;
12818         if (rack->gp_ready &&
12819             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT))
12820                 under_pacing = 0;
12821         else
12822                 under_pacing = 1;
12823
12824         if (rack->r_state != tp->t_state)
12825                 rack_set_state(tp, rack);
12826         to = &to_holder;
12827         to->to_flags = 0;
12828         KASSERT((m->m_len >= sizeof(struct tcp_ackent)),
12829                 ("tp:%p m_cmpack:%p with invalid len:%u", tp, m, m->m_len));
12830         cnt = m->m_len / sizeof(struct tcp_ackent);
12831         idx = cnt / 5;
12832         if (idx >= MAX_NUM_OF_CNTS)
12833                 idx = MAX_NUM_OF_CNTS - 1;
12834         counter_u64_add(rack_proc_comp_ack[idx], 1);
12835         counter_u64_add(rack_multi_single_eq, cnt);
12836         high_seq = tp->snd_una;
12837         the_win = tp->snd_wnd;
12838         win_seq = tp->snd_wl1;
12839         win_upd_ack = tp->snd_wl2;
12840         cts = us_cts = tcp_tv_to_usectick(tv);
12841         segsiz = ctf_fixed_maxseg(tp);
12842         if ((rack->rc_gp_dyn_mul) &&
12843             (rack->use_fixed_rate == 0) &&
12844             (rack->rc_always_pace)) {
12845                 /* Check in on probertt */
12846                 rack_check_probe_rtt(rack, us_cts);
12847         }
12848         for (i = 0; i < cnt; i++) {
12849 #ifdef TCP_ACCOUNTING
12850                 ts_val = get_cyclecount();
12851 #endif
12852                 rack_clear_rate_sample(rack);
12853                 ae = ((mtod(m, struct tcp_ackent *)) + i);
12854                 /* Setup the window */
12855                 tiwin = ae->win << tp->snd_scale;
12856                 /* figure out the type of ack */
12857                 if (SEQ_LT(ae->ack, high_seq)) {
12858                         /* Case B*/
12859                         ae->ack_val_set = ACK_BEHIND;
12860                 } else if (SEQ_GT(ae->ack, high_seq)) {
12861                         /* Case A */
12862                         ae->ack_val_set = ACK_CUMACK;
12863                 } else if (tiwin == the_win) {
12864                         /* Case D */
12865                         ae->ack_val_set = ACK_DUPACK;
12866                 } else {
12867                         /* Case C */
12868                         ae->ack_val_set = ACK_RWND;
12869                 }
12870                 rack_log_input_packet(tp, rack, ae, ae->ack_val_set, high_seq);
12871                 /* Validate timestamp */
12872                 if (ae->flags & HAS_TSTMP) {
12873                         /* Setup for a timestamp */
12874                         to->to_flags = TOF_TS;
12875                         ae->ts_echo -= tp->ts_offset;
12876                         to->to_tsecr = ae->ts_echo;
12877                         to->to_tsval = ae->ts_value;
12878                         /*
12879                          * If echoed timestamp is later than the current time, fall back to
12880                          * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
12881                          * were used when this connection was established.
12882                          */
12883                         if (TSTMP_GT(ae->ts_echo, cts))
12884                                 ae->ts_echo = 0;
12885                         if (tp->ts_recent &&
12886                             TSTMP_LT(ae->ts_value, tp->ts_recent)) {
12887                                 if (ctf_ts_check_ac(tp, (ae->flags & 0xff))) {
12888 #ifdef TCP_ACCOUNTING
12889                                         rdstc = get_cyclecount();
12890                                         if (rdstc > ts_val) {
12891                                                 counter_u64_add(tcp_proc_time[ae->ack_val_set] ,
12892                                                                 (rdstc - ts_val));
12893                                                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
12894                                                         tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
12895                                                 }
12896                                         }
12897 #endif
12898                                         continue;
12899                                 }
12900                         }
12901                         if (SEQ_LEQ(ae->seq, tp->last_ack_sent) &&
12902                             SEQ_LEQ(tp->last_ack_sent, ae->seq)) {
12903                                 tp->ts_recent_age = tcp_ts_getticks();
12904                                 tp->ts_recent = ae->ts_value;
12905                         }
12906                 } else {
12907                         /* Setup for a no options */
12908                         to->to_flags = 0;
12909                 }
12910                 /* Update the rcv time and perform idle reduction possibly */
12911                 if  (tp->t_idle_reduce &&
12912                      (tp->snd_max == tp->snd_una) &&
12913                      ((ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
12914                         counter_u64_add(rack_input_idle_reduces, 1);
12915                         rack_cc_after_idle(rack, tp);
12916                 }
12917                 tp->t_rcvtime = ticks;
12918                 /* Now what about ECN? */
12919                 if (tp->t_flags2 & TF2_ECN_PERMIT) {
12920                         if (ae->flags & TH_CWR) {
12921                                 tp->t_flags2 &= ~TF2_ECN_SND_ECE;
12922                                 tp->t_flags |= TF_ACKNOW;
12923                         }
12924                         switch (ae->codepoint & IPTOS_ECN_MASK) {
12925                         case IPTOS_ECN_CE:
12926                                 tp->t_flags2 |= TF2_ECN_SND_ECE;
12927                                 KMOD_TCPSTAT_INC(tcps_ecn_ce);
12928                                 break;
12929                         case IPTOS_ECN_ECT0:
12930                                 KMOD_TCPSTAT_INC(tcps_ecn_ect0);
12931                                 break;
12932                         case IPTOS_ECN_ECT1:
12933                                 KMOD_TCPSTAT_INC(tcps_ecn_ect1);
12934                                 break;
12935                         }
12936
12937                         /* Process a packet differently from RFC3168. */
12938                         cc_ecnpkt_handler_flags(tp, ae->flags, ae->codepoint);
12939                         /* Congestion experienced. */
12940                         if (ae->flags & TH_ECE) {
12941                                 rack_cong_signal(tp,  CC_ECN, ae->ack);
12942                         }
12943                 }
12944 #ifdef TCP_ACCOUNTING
12945                 /* Count for the specific type of ack in */
12946                 counter_u64_add(tcp_cnt_counters[ae->ack_val_set], 1);
12947                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
12948                         tp->tcp_cnt_counters[ae->ack_val_set]++;
12949                 }
12950 #endif
12951                 /*
12952                  * Note how we could move up these in the determination
12953                  * above, but we don't so that way the timestamp checks (and ECN)
12954                  * is done first before we do any processing on the ACK.
12955                  * The non-compressed path through the code has this
12956                  * weakness (noted by @jtl) that it actually does some
12957                  * processing before verifying the timestamp information.
12958                  * We don't take that path here which is why we set
12959                  * the ack_val_set first, do the timestamp and ecn
12960                  * processing, and then look at what we have setup.
12961                  */
12962                 if (ae->ack_val_set == ACK_BEHIND) {
12963                         /*
12964                          * Case B flag reordering, if window is not closed
12965                          * or it could be a keep-alive or persists
12966                          */
12967                         if (SEQ_LT(ae->ack, tp->snd_una) && (sbspace(&so->so_rcv) > segsiz)) {
12968                                 counter_u64_add(rack_reorder_seen, 1);
12969                                 rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
12970                         }
12971                 } else if (ae->ack_val_set == ACK_DUPACK) {
12972                         /* Case D */
12973
12974                         rack_strike_dupack(rack);
12975                 } else if (ae->ack_val_set == ACK_RWND) {
12976                         /* Case C */
12977
12978                         win_up_req = 1;
12979                         win_upd_ack = ae->ack;
12980                         win_seq = ae->seq;
12981                         the_win = tiwin;
12982                 } else {
12983                         /* Case A */
12984
12985                         if (SEQ_GT(ae->ack, tp->snd_max)) {
12986                                 /*
12987                                  * We just send an ack since the incoming
12988                                  * ack is beyond the largest seq we sent.
12989                                  */
12990                                 if ((tp->t_flags & TF_ACKNOW) == 0) {
12991                                         ctf_ack_war_checks(tp, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt);
12992                                         if (tp->t_flags && TF_ACKNOW)
12993                                                 rack->r_wanted_output = 1;
12994                                 }
12995                         } else {
12996                                 nsegs++;
12997                                 /* If the window changed setup to update */
12998                                 if (tiwin != tp->snd_wnd) {
12999                                         win_up_req = 1;
13000                                         win_upd_ack = ae->ack;
13001                                         win_seq = ae->seq;
13002                                         the_win = tiwin;
13003                                 }
13004 #ifdef TCP_ACCOUNTING
13005                                 /* Account for the acks */
13006                                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13007                                         tp->tcp_cnt_counters[CNT_OF_ACKS_IN] += (((ae->ack - high_seq) + segsiz - 1) / segsiz);
13008                                 }
13009                                 counter_u64_add(tcp_cnt_counters[CNT_OF_ACKS_IN],
13010                                                 (((ae->ack - high_seq) + segsiz - 1) / segsiz));
13011 #endif
13012                                 high_seq = ae->ack;
13013                                 /* Setup our act_rcv_time */
13014                                 if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
13015                                         ts.tv_sec = ae->timestamp / 1000000000;
13016                                         ts.tv_nsec = ae->timestamp % 1000000000;
13017                                         rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
13018                                         rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
13019                                 } else {
13020                                         rack->r_ctl.act_rcv_time = *tv;
13021                                 }
13022                                 rack_process_to_cumack(tp, rack, ae->ack, cts, to);
13023                         }
13024                 }
13025                 /* And lets be sure to commit the rtt measurements for this ack */
13026                 tcp_rack_xmit_timer_commit(rack, tp);
13027 #ifdef TCP_ACCOUNTING
13028                 rdstc = get_cyclecount();
13029                 if (rdstc > ts_val) {
13030                         counter_u64_add(tcp_proc_time[ae->ack_val_set] , (rdstc - ts_val));
13031                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13032                                 tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
13033                                 if (ae->ack_val_set == ACK_CUMACK)
13034                                         tp->tcp_proc_time[CYC_HANDLE_MAP] += (rdstc - ts_val);
13035                         }
13036                 }
13037 #endif
13038         }
13039 #ifdef TCP_ACCOUNTING
13040         ts_val = get_cyclecount();
13041 #endif
13042         acked_amount = acked = (high_seq - tp->snd_una);
13043         if (win_up_req) {
13044                 rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts, high_seq);
13045         }
13046         if (acked) {
13047                 if (rack->sack_attack_disable == 0)
13048                         rack_do_decay(rack);
13049                 if (acked >= segsiz) {
13050                         /*
13051                          * You only get credit for
13052                          * MSS and greater (and you get extra
13053                          * credit for larger cum-ack moves).
13054                          */
13055                         int ac;
13056
13057                         ac = acked / segsiz;
13058                         rack->r_ctl.ack_count += ac;
13059                         counter_u64_add(rack_ack_total, ac);
13060                 }
13061                 if (rack->r_ctl.ack_count > 0xfff00000) {
13062                         /*
13063                          * reduce the number to keep us under
13064                          * a uint32_t.
13065                          */
13066                         rack->r_ctl.ack_count /= 2;
13067                         rack->r_ctl.sack_count /= 2;
13068                 }
13069                 if (tp->t_flags & TF_NEEDSYN) {
13070                         /*
13071                          * T/TCP: Connection was half-synchronized, and our SYN has
13072                          * been ACK'd (so connection is now fully synchronized).  Go
13073                          * to non-starred state, increment snd_una for ACK of SYN,
13074                          * and check if we can do window scaling.
13075                          */
13076                         tp->t_flags &= ~TF_NEEDSYN;
13077                         tp->snd_una++;
13078                         acked_amount = acked = (high_seq - tp->snd_una);
13079                 }
13080                 if (acked > sbavail(&so->so_snd))
13081                         acked_amount = sbavail(&so->so_snd);
13082 #ifdef NETFLIX_EXP_DETECTION
13083                 /*
13084                  * We only care on a cum-ack move if we are in a sack-disabled
13085                  * state. We have already added in to the ack_count, and we never
13086                  * would disable on a cum-ack move, so we only care to do the
13087                  * detection if it may "undo" it, i.e. we were in disabled already.
13088                  */
13089                 if (rack->sack_attack_disable)
13090                         rack_do_detection(tp, rack, acked_amount, segsiz);
13091 #endif
13092                 if (IN_FASTRECOVERY(tp->t_flags) &&
13093                     (rack->rack_no_prr == 0))
13094                         rack_update_prr(tp, rack, acked_amount, high_seq);
13095                 if (IN_RECOVERY(tp->t_flags)) {
13096                         if (SEQ_LT(high_seq, tp->snd_recover) &&
13097                             (SEQ_LT(high_seq, tp->snd_max))) {
13098                                 tcp_rack_partialack(tp);
13099                         } else {
13100                                 rack_post_recovery(tp, high_seq);
13101                                 recovery = 1;
13102                         }
13103                 }
13104                 /* Handle the rack-log-ack part (sendmap) */
13105                 if ((sbused(&so->so_snd) == 0) &&
13106                     (acked > acked_amount) &&
13107                     (tp->t_state >= TCPS_FIN_WAIT_1) &&
13108                     (tp->t_flags & TF_SENTFIN)) {
13109                         /*
13110                          * We must be sure our fin
13111                          * was sent and acked (we can be
13112                          * in FIN_WAIT_1 without having
13113                          * sent the fin).
13114                          */
13115                         ourfinisacked = 1;
13116                         /*
13117                          * Lets make sure snd_una is updated
13118                          * since most likely acked_amount = 0 (it
13119                          * should be).
13120                          */
13121                         tp->snd_una = high_seq;
13122                 }
13123                 /* Did we make a RTO error? */
13124                 if ((tp->t_flags & TF_PREVVALID) &&
13125                     ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
13126                         tp->t_flags &= ~TF_PREVVALID;
13127                         if (tp->t_rxtshift == 1 &&
13128                             (int)(ticks - tp->t_badrxtwin) < 0)
13129                                 rack_cong_signal(tp, CC_RTO_ERR, high_seq);
13130                 }
13131                 /* Handle the data in the socket buffer */
13132                 KMOD_TCPSTAT_ADD(tcps_rcvackpack, 1);
13133                 KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
13134                 if (acked_amount > 0) {
13135                         struct mbuf *mfree;
13136
13137                         rack_ack_received(tp, rack, high_seq, nsegs, CC_ACK, recovery);
13138                         SOCKBUF_LOCK(&so->so_snd);
13139                         mfree = sbcut_locked(&so->so_snd, acked);
13140                         tp->snd_una = high_seq;
13141                         /* Note we want to hold the sb lock through the sendmap adjust */
13142                         rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
13143                         /* Wake up the socket if we have room to write more */
13144                         rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
13145                         SOCKBUF_UNLOCK(&so->so_snd);
13146                         tp->t_flags |= TF_WAKESOW;
13147                         m_freem(mfree);
13148                 }
13149                 /* update progress */
13150                 tp->t_acktime = ticks;
13151                 rack_log_progress_event(rack, tp, tp->t_acktime,
13152                                         PROGRESS_UPDATE, __LINE__);
13153                 /* Clear out shifts and such */
13154                 tp->t_rxtshift = 0;
13155                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
13156                                    rack_rto_min, rack_rto_max);
13157                 rack->rc_tlp_in_progress = 0;
13158                 rack->r_ctl.rc_tlp_cnt_out = 0;
13159                 /* Send recover and snd_nxt must be dragged along */
13160                 if (SEQ_GT(tp->snd_una, tp->snd_recover))
13161                         tp->snd_recover = tp->snd_una;
13162                 if (SEQ_LT(tp->snd_nxt, tp->snd_una))
13163                         tp->snd_nxt = tp->snd_una;
13164                 /*
13165                  * If the RXT timer is running we want to
13166                  * stop it, so we can restart a TLP (or new RXT).
13167                  */
13168                 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
13169                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13170 #ifdef NETFLIX_HTTP_LOGGING
13171                 tcp_http_check_for_comp(rack->rc_tp, high_seq);
13172 #endif
13173                 tp->snd_wl2 = high_seq;
13174                 tp->t_dupacks = 0;
13175                 if (under_pacing &&
13176                     (rack->use_fixed_rate == 0) &&
13177                     (rack->in_probe_rtt == 0) &&
13178                     rack->rc_gp_dyn_mul &&
13179                     rack->rc_always_pace) {
13180                         /* Check if we are dragging bottom */
13181                         rack_check_bottom_drag(tp, rack, so, acked);
13182                 }
13183                 if (tp->snd_una == tp->snd_max) {
13184                         tp->t_flags &= ~TF_PREVVALID;
13185                         rack->r_ctl.retran_during_recovery = 0;
13186                         rack->r_ctl.dsack_byte_cnt = 0;
13187                         rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
13188                         if (rack->r_ctl.rc_went_idle_time == 0)
13189                                 rack->r_ctl.rc_went_idle_time = 1;
13190                         rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
13191                         if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
13192                                 tp->t_acktime = 0;
13193                         /* Set so we might enter persists... */
13194                         rack->r_wanted_output = 1;
13195                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13196                         sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
13197                         if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
13198                             (sbavail(&so->so_snd) == 0) &&
13199                             (tp->t_flags2 & TF2_DROP_AF_DATA)) {
13200                                 /*
13201                                  * The socket was gone and the
13202                                  * peer sent data (not now in the past), time to
13203                                  * reset him.
13204                                  */
13205                                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13206                                 /* tcp_close will kill the inp pre-log the Reset */
13207                                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
13208 #ifdef TCP_ACCOUNTING
13209                                 rdstc = get_cyclecount();
13210                                 if (rdstc > ts_val) {
13211                                         counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13212                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13213                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13214                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13215                                         }
13216                                 }
13217 #endif
13218                                 m_freem(m);
13219                                 tp = tcp_close(tp);
13220                                 if (tp == NULL) {
13221 #ifdef TCP_ACCOUNTING
13222                                         sched_unpin();
13223 #endif
13224                                         return (1);
13225                                 }
13226                                 /*
13227                                  * We would normally do drop-with-reset which would
13228                                  * send back a reset. We can't since we don't have
13229                                  * all the needed bits. Instead lets arrange for
13230                                  * a call to tcp_output(). That way since we
13231                                  * are in the closed state we will generate a reset.
13232                                  *
13233                                  * Note if tcp_accounting is on we don't unpin since
13234                                  * we do that after the goto label.
13235                                  */
13236                                 goto send_out_a_rst;
13237                         }
13238                         if ((sbused(&so->so_snd) == 0) &&
13239                             (tp->t_state >= TCPS_FIN_WAIT_1) &&
13240                             (tp->t_flags & TF_SENTFIN)) {
13241                                 /*
13242                                  * If we can't receive any more data, then closing user can
13243                                  * proceed. Starting the timer is contrary to the
13244                                  * specification, but if we don't get a FIN we'll hang
13245                                  * forever.
13246                                  *
13247                                  */
13248                                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13249                                         soisdisconnected(so);
13250                                         tcp_timer_activate(tp, TT_2MSL,
13251                                                            (tcp_fast_finwait2_recycle ?
13252                                                             tcp_finwait2_timeout :
13253                                                             TP_MAXIDLE(tp)));
13254                                 }
13255                                 if (ourfinisacked == 0) {
13256                                         /*
13257                                          * We don't change to fin-wait-2 if we have our fin acked
13258                                          * which means we are probably in TCPS_CLOSING.
13259                                          */
13260                                         tcp_state_change(tp, TCPS_FIN_WAIT_2);
13261                                 }
13262                         }
13263                 }
13264                 /* Wake up the socket if we have room to write more */
13265                 if (sbavail(&so->so_snd)) {
13266                         rack->r_wanted_output = 1;
13267                         if (ctf_progress_timeout_check(tp, true)) {
13268                                 rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13269                                                         tp, tick, PROGRESS_DROP, __LINE__);
13270                                 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
13271                                 /*
13272                                  * We cheat here and don't send a RST, we should send one
13273                                  * when the pacer drops the connection.
13274                                  */
13275 #ifdef TCP_ACCOUNTING
13276                                 rdstc = get_cyclecount();
13277                                 if (rdstc > ts_val) {
13278                                         counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13279                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13280                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13281                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13282                                         }
13283                                 }
13284                                 sched_unpin();
13285 #endif
13286                                 INP_WUNLOCK(rack->rc_inp);
13287                                 m_freem(m);
13288                                 return (1);
13289                         }
13290                 }
13291                 if (ourfinisacked) {
13292                         switch(tp->t_state) {
13293                         case TCPS_CLOSING:
13294 #ifdef TCP_ACCOUNTING
13295                                 rdstc = get_cyclecount();
13296                                 if (rdstc > ts_val) {
13297                                         counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13298                                                         (rdstc - ts_val));
13299                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13300                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13301                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13302                                         }
13303                                 }
13304                                 sched_unpin();
13305 #endif
13306                                 tcp_twstart(tp);
13307                                 m_freem(m);
13308                                 return (1);
13309                                 break;
13310                         case TCPS_LAST_ACK:
13311 #ifdef TCP_ACCOUNTING
13312                                 rdstc = get_cyclecount();
13313                                 if (rdstc > ts_val) {
13314                                         counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13315                                                         (rdstc - ts_val));
13316                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13317                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13318                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13319                                         }
13320                                 }
13321                                 sched_unpin();
13322 #endif
13323                                 tp = tcp_close(tp);
13324                                 ctf_do_drop(m, tp);
13325                                 return (1);
13326                                 break;
13327                         case TCPS_FIN_WAIT_1:
13328 #ifdef TCP_ACCOUNTING
13329                                 rdstc = get_cyclecount();
13330                                 if (rdstc > ts_val) {
13331                                         counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13332                                                         (rdstc - ts_val));
13333                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13334                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13335                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13336                                         }
13337                                 }
13338 #endif
13339                                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13340                                         soisdisconnected(so);
13341                                         tcp_timer_activate(tp, TT_2MSL,
13342                                                            (tcp_fast_finwait2_recycle ?
13343                                                             tcp_finwait2_timeout :
13344                                                             TP_MAXIDLE(tp)));
13345                                 }
13346                                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
13347                                 break;
13348                         default:
13349                                 break;
13350                         }
13351                 }
13352                 if (rack->r_fast_output) {
13353                         /*
13354                          * We re doing fast output.. can we expand that?
13355                          */
13356                         rack_gain_for_fastoutput(rack, tp, so, acked_amount);
13357                 }
13358 #ifdef TCP_ACCOUNTING
13359                 rdstc = get_cyclecount();
13360                 if (rdstc > ts_val) {
13361                         counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13362                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13363                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13364                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13365                         }
13366                 }
13367
13368         } else if (win_up_req) {
13369                 rdstc = get_cyclecount();
13370                 if (rdstc > ts_val) {
13371                         counter_u64_add(tcp_proc_time[ACK_RWND] , (rdstc - ts_val));
13372                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13373                                 tp->tcp_proc_time[ACK_RWND] += (rdstc - ts_val);
13374                         }
13375                 }
13376 #endif
13377         }
13378         /* Now is there a next packet, if so we are done */
13379         m_freem(m);
13380         did_out = 0;
13381         if (nxt_pkt) {
13382 #ifdef TCP_ACCOUNTING
13383                 sched_unpin();
13384 #endif
13385                 rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 5, nsegs);
13386                 return (0);
13387         }
13388         rack_handle_might_revert(tp, rack);
13389         ctf_calc_rwin(so, tp);
13390         if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
13391         send_out_a_rst:
13392                 (void)tp->t_fb->tfb_tcp_output(tp);
13393                 did_out = 1;
13394         }
13395         rack_free_trim(rack);
13396 #ifdef TCP_ACCOUNTING
13397         sched_unpin();
13398 #endif
13399         rack_timer_audit(tp, rack, &so->so_snd);
13400         rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 6, nsegs);
13401         return (0);
13402 }
13403
13404
13405 static int
13406 rack_do_segment_nounlock(struct mbuf *m, struct tcphdr *th, struct socket *so,
13407     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos,
13408     int32_t nxt_pkt, struct timeval *tv)
13409 {
13410 #ifdef TCP_ACCOUNTING
13411         uint64_t ts_val;
13412 #endif
13413         int32_t thflags, retval, did_out = 0;
13414         int32_t way_out = 0;
13415         uint32_t cts;
13416         uint32_t tiwin;
13417         struct timespec ts;
13418         struct tcpopt to;
13419         struct tcp_rack *rack;
13420         struct rack_sendmap *rsm;
13421         int32_t prev_state = 0;
13422 #ifdef TCP_ACCOUNTING
13423         int ack_val_set = 0xf;
13424 #endif
13425         uint32_t us_cts;
13426         /*
13427          * tv passed from common code is from either M_TSTMP_LRO or
13428          * tcp_get_usecs() if no LRO m_pkthdr timestamp is present.
13429          */
13430         if (m->m_flags & M_ACKCMP) {
13431                 return (rack_do_compressed_ack_processing(tp, so, m, nxt_pkt, tv));
13432         }
13433         if (m->m_flags & M_ACKCMP) {
13434                 panic("Impossible reach m has ackcmp? m:%p tp:%p", m, tp);
13435         }
13436         counter_u64_add(rack_proc_non_comp_ack, 1);
13437         thflags = th->th_flags;
13438 #ifdef TCP_ACCOUNTING
13439         sched_pin();
13440         if (thflags & TH_ACK)
13441                 ts_val = get_cyclecount();
13442 #endif
13443         cts = tcp_tv_to_usectick(tv);
13444         rack = (struct tcp_rack *)tp->t_fb_ptr;
13445
13446         if ((m->m_flags & M_TSTMP) ||
13447             (m->m_flags & M_TSTMP_LRO)) {
13448                 mbuf_tstmp2timespec(m, &ts);
13449                 rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
13450                 rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
13451         } else
13452                 rack->r_ctl.act_rcv_time = *tv;
13453         kern_prefetch(rack, &prev_state);
13454         prev_state = 0;
13455         /*
13456          * Unscale the window into a 32-bit value. For the SYN_SENT state
13457          * the scale is zero.
13458          */
13459         tiwin = th->th_win << tp->snd_scale;
13460         /*
13461          * Parse options on any incoming segment.
13462          */
13463         memset(&to, 0, sizeof(to));
13464         tcp_dooptions(&to, (u_char *)(th + 1),
13465             (th->th_off << 2) - sizeof(struct tcphdr),
13466             (thflags & TH_SYN) ? TO_SYN : 0);
13467 #ifdef TCP_ACCOUNTING
13468         if (thflags & TH_ACK) {
13469                 /*
13470                  * We have a tradeoff here. We can either do what we are
13471                  * doing i.e. pinning to this CPU and then doing the accounting
13472                  * <or> we could do a critical enter, setup the rdtsc and cpu
13473                  * as in below, and then validate we are on the same CPU on
13474                  * exit. I have choosen to not do the critical enter since
13475                  * that often will gain you a context switch, and instead lock
13476                  * us (line above this if) to the same CPU with sched_pin(). This
13477                  * means we may be context switched out for a higher priority
13478                  * interupt but we won't be moved to another CPU.
13479                  *
13480                  * If this occurs (which it won't very often since we most likely
13481                  * are running this code in interupt context and only a higher
13482                  * priority will bump us ... clock?) we will falsely add in
13483                  * to the time the interupt processing time plus the ack processing
13484                  * time. This is ok since its a rare event.
13485                  */
13486                 ack_val_set = tcp_do_ack_accounting(tp, th, &to, tiwin,
13487                                                     ctf_fixed_maxseg(tp));
13488         }
13489 #endif
13490         NET_EPOCH_ASSERT();
13491         INP_WLOCK_ASSERT(tp->t_inpcb);
13492         KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
13493             __func__));
13494         KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
13495             __func__));
13496         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
13497                 union tcp_log_stackspecific log;
13498                 struct timeval ltv;
13499 #ifdef NETFLIX_HTTP_LOGGING
13500                 struct http_sendfile_track *http_req;
13501
13502                 if (SEQ_GT(th->th_ack, tp->snd_una)) {
13503                         http_req = tcp_http_find_req_for_seq(tp, (th->th_ack-1));
13504                 } else {
13505                         http_req = tcp_http_find_req_for_seq(tp, th->th_ack);
13506                 }
13507 #endif
13508                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
13509                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
13510                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
13511                 if (rack->rack_no_prr == 0)
13512                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
13513                 else
13514                         log.u_bbr.flex1 = 0;
13515                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
13516                 log.u_bbr.use_lt_bw <<= 1;
13517                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
13518                 log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
13519                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
13520                 log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
13521                 log.u_bbr.flex3 = m->m_flags;
13522                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
13523                 log.u_bbr.lost = thflags;
13524                 log.u_bbr.pacing_gain = 0x1;
13525 #ifdef TCP_ACCOUNTING
13526                 log.u_bbr.cwnd_gain = ack_val_set;
13527 #endif
13528                 log.u_bbr.flex7 = 2;
13529                 if (m->m_flags & M_TSTMP) {
13530                         /* Record the hardware timestamp if present */
13531                         mbuf_tstmp2timespec(m, &ts);
13532                         ltv.tv_sec = ts.tv_sec;
13533                         ltv.tv_usec = ts.tv_nsec / 1000;
13534                         log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
13535                 } else if (m->m_flags & M_TSTMP_LRO) {
13536                         /* Record the LRO the arrival timestamp */
13537                         mbuf_tstmp2timespec(m, &ts);
13538                         ltv.tv_sec = ts.tv_sec;
13539                         ltv.tv_usec = ts.tv_nsec / 1000;
13540                         log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
13541                 }
13542                 log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
13543                 /* Log the rcv time */
13544                 log.u_bbr.delRate = m->m_pkthdr.rcv_tstmp;
13545 #ifdef NETFLIX_HTTP_LOGGING
13546                 log.u_bbr.applimited = tp->t_http_closed;
13547                 log.u_bbr.applimited <<= 8;
13548                 log.u_bbr.applimited |= tp->t_http_open;
13549                 log.u_bbr.applimited <<= 8;
13550                 log.u_bbr.applimited |= tp->t_http_req;
13551                 if (http_req) {
13552                         /* Copy out any client req info */
13553                         /* seconds */
13554                         log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC);
13555                         /* useconds */
13556                         log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC);
13557                         log.u_bbr.rttProp = http_req->timestamp;
13558                         log.u_bbr.cur_del_rate = http_req->start;
13559                         if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) {
13560                                 log.u_bbr.flex8 |= 1;
13561                         } else {
13562                                 log.u_bbr.flex8 |= 2;
13563                                 log.u_bbr.bw_inuse = http_req->end;
13564                         }
13565                         log.u_bbr.flex6 = http_req->start_seq;
13566                         if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) {
13567                                 log.u_bbr.flex8 |= 4;
13568                                 log.u_bbr.epoch = http_req->end_seq;
13569                         }
13570                 }
13571 #endif
13572                 TCP_LOG_EVENTP(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0,
13573                     tlen, &log, true, &ltv);
13574         }
13575         if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) {
13576                 way_out = 4;
13577                 retval = 0;
13578                 goto done_with_input;
13579         }
13580         /*
13581          * If a segment with the ACK-bit set arrives in the SYN-SENT state
13582          * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9.
13583          */
13584         if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) &&
13585             (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) {
13586                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13587                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13588 #ifdef TCP_ACCOUNTING
13589                 sched_unpin();
13590 #endif
13591                 return (1);
13592         }
13593
13594         /*
13595          * Parse options on any incoming segment.
13596          */
13597         tcp_dooptions(&to, (u_char *)(th + 1),
13598             (th->th_off << 2) - sizeof(struct tcphdr),
13599             (thflags & TH_SYN) ? TO_SYN : 0);
13600
13601         /*
13602          * If timestamps were negotiated during SYN/ACK and a
13603          * segment without a timestamp is received, silently drop
13604          * the segment, unless it is a RST segment or missing timestamps are
13605          * tolerated.
13606          * See section 3.2 of RFC 7323.
13607          */
13608         if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) &&
13609             ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) {
13610                 way_out = 5;
13611                 retval = 0;
13612                 goto done_with_input;
13613         }
13614
13615         /*
13616          * Segment received on connection. Reset idle time and keep-alive
13617          * timer. XXX: This should be done after segment validation to
13618          * ignore broken/spoofed segs.
13619          */
13620         if  (tp->t_idle_reduce &&
13621              (tp->snd_max == tp->snd_una) &&
13622              ((ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
13623                 counter_u64_add(rack_input_idle_reduces, 1);
13624                 rack_cc_after_idle(rack, tp);
13625         }
13626         tp->t_rcvtime = ticks;
13627 #ifdef STATS
13628         stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
13629 #endif
13630         if (tiwin > rack->r_ctl.rc_high_rwnd)
13631                 rack->r_ctl.rc_high_rwnd = tiwin;
13632         /*
13633          * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move
13634          * this to occur after we've validated the segment.
13635          */
13636         if (tp->t_flags2 & TF2_ECN_PERMIT) {
13637                 if (thflags & TH_CWR) {
13638                         tp->t_flags2 &= ~TF2_ECN_SND_ECE;
13639                         tp->t_flags |= TF_ACKNOW;
13640                 }
13641                 switch (iptos & IPTOS_ECN_MASK) {
13642                 case IPTOS_ECN_CE:
13643                         tp->t_flags2 |= TF2_ECN_SND_ECE;
13644                         KMOD_TCPSTAT_INC(tcps_ecn_ce);
13645                         break;
13646                 case IPTOS_ECN_ECT0:
13647                         KMOD_TCPSTAT_INC(tcps_ecn_ect0);
13648                         break;
13649                 case IPTOS_ECN_ECT1:
13650                         KMOD_TCPSTAT_INC(tcps_ecn_ect1);
13651                         break;
13652                 }
13653
13654                 /* Process a packet differently from RFC3168. */
13655                 cc_ecnpkt_handler(tp, th, iptos);
13656
13657                 /* Congestion experienced. */
13658                 if (thflags & TH_ECE) {
13659                         rack_cong_signal(tp, CC_ECN, th->th_ack);
13660                 }
13661         }
13662
13663         /*
13664          * If echoed timestamp is later than the current time, fall back to
13665          * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
13666          * were used when this connection was established.
13667          */
13668         if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
13669                 to.to_tsecr -= tp->ts_offset;
13670                 if (TSTMP_GT(to.to_tsecr, cts))
13671                         to.to_tsecr = 0;
13672         }
13673
13674         /*
13675          * If its the first time in we need to take care of options and
13676          * verify we can do SACK for rack!
13677          */
13678         if (rack->r_state == 0) {
13679                 /* Should be init'd by rack_init() */
13680                 KASSERT(rack->rc_inp != NULL,
13681                     ("%s: rack->rc_inp unexpectedly NULL", __func__));
13682                 if (rack->rc_inp == NULL) {
13683                         rack->rc_inp = tp->t_inpcb;
13684                 }
13685
13686                 /*
13687                  * Process options only when we get SYN/ACK back. The SYN
13688                  * case for incoming connections is handled in tcp_syncache.
13689                  * According to RFC1323 the window field in a SYN (i.e., a
13690                  * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
13691                  * this is traditional behavior, may need to be cleaned up.
13692                  */
13693                 if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
13694                         /* Handle parallel SYN for ECN */
13695                         if (!(thflags & TH_ACK) &&
13696                             ((thflags & (TH_CWR | TH_ECE)) == (TH_CWR | TH_ECE)) &&
13697                             ((V_tcp_do_ecn == 1) || (V_tcp_do_ecn == 2))) {
13698                                 tp->t_flags2 |= TF2_ECN_PERMIT;
13699                                 tp->t_flags2 |= TF2_ECN_SND_ECE;
13700                                 TCPSTAT_INC(tcps_ecn_shs);
13701                         }
13702                         if ((to.to_flags & TOF_SCALE) &&
13703                             (tp->t_flags & TF_REQ_SCALE)) {
13704                                 tp->t_flags |= TF_RCVD_SCALE;
13705                                 tp->snd_scale = to.to_wscale;
13706                         } else
13707                                 tp->t_flags &= ~TF_REQ_SCALE;
13708                         /*
13709                          * Initial send window.  It will be updated with the
13710                          * next incoming segment to the scaled value.
13711                          */
13712                         tp->snd_wnd = th->th_win;
13713                         rack_validate_fo_sendwin_up(tp, rack);
13714                         if ((to.to_flags & TOF_TS) &&
13715                             (tp->t_flags & TF_REQ_TSTMP)) {
13716                                 tp->t_flags |= TF_RCVD_TSTMP;
13717                                 tp->ts_recent = to.to_tsval;
13718                                 tp->ts_recent_age = cts;
13719                         } else
13720                                 tp->t_flags &= ~TF_REQ_TSTMP;
13721                         if (to.to_flags & TOF_MSS) {
13722                                 tcp_mss(tp, to.to_mss);
13723                         }
13724                         if ((tp->t_flags & TF_SACK_PERMIT) &&
13725                             (to.to_flags & TOF_SACKPERM) == 0)
13726                                 tp->t_flags &= ~TF_SACK_PERMIT;
13727                         if (IS_FASTOPEN(tp->t_flags)) {
13728                                 if (to.to_flags & TOF_FASTOPEN) {
13729                                         uint16_t mss;
13730
13731                                         if (to.to_flags & TOF_MSS)
13732                                                 mss = to.to_mss;
13733                                         else
13734                                                 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0)
13735                                                         mss = TCP6_MSS;
13736                                                 else
13737                                                         mss = TCP_MSS;
13738                                         tcp_fastopen_update_cache(tp, mss,
13739                                             to.to_tfo_len, to.to_tfo_cookie);
13740                                 } else
13741                                         tcp_fastopen_disable_path(tp);
13742                         }
13743                 }
13744                 /*
13745                  * At this point we are at the initial call. Here we decide
13746                  * if we are doing RACK or not. We do this by seeing if
13747                  * TF_SACK_PERMIT is set and the sack-not-required is clear.
13748                  * The code now does do dup-ack counting so if you don't
13749                  * switch back you won't get rack & TLP, but you will still
13750                  * get this stack.
13751                  */
13752
13753                 if ((rack_sack_not_required == 0) &&
13754                     ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
13755                         tcp_switch_back_to_default(tp);
13756                         (*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen,
13757                             tlen, iptos);
13758 #ifdef TCP_ACCOUNTING
13759                         sched_unpin();
13760 #endif
13761                         return (1);
13762                 }
13763                 tcp_set_hpts(tp->t_inpcb);
13764                 sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack);
13765         }
13766         if (thflags & TH_FIN)
13767                 tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN);
13768         us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
13769         if ((rack->rc_gp_dyn_mul) &&
13770             (rack->use_fixed_rate == 0) &&
13771             (rack->rc_always_pace)) {
13772                 /* Check in on probertt */
13773                 rack_check_probe_rtt(rack, us_cts);
13774         }
13775         if (rack->forced_ack) {
13776                 uint32_t us_rtt;
13777
13778                 /*
13779                  * A persist or keep-alive was forced out, update our
13780                  * min rtt time. Note we do not worry about lost
13781                  * retransmissions since KEEP-ALIVES and persists
13782                  * are usually way long on times of sending (though
13783                  * if we were really paranoid or worried we could
13784                  * at least use timestamps if available to validate).
13785                  */
13786                 rack->forced_ack = 0;
13787                 us_rtt = us_cts - rack->r_ctl.forced_ack_ts;
13788                 if (us_rtt == 0)
13789                         us_rtt = 1;
13790                 rack_log_rtt_upd(tp, rack, us_rtt, 0, NULL, 3);
13791                 rack_apply_updated_usrtt(rack, us_rtt, us_cts);
13792         }
13793         /*
13794          * This is the one exception case where we set the rack state
13795          * always. All other times (timers etc) we must have a rack-state
13796          * set (so we assure we have done the checks above for SACK).
13797          */
13798         rack->r_ctl.rc_rcvtime = cts;
13799         if (rack->r_state != tp->t_state)
13800                 rack_set_state(tp, rack);
13801         if (SEQ_GT(th->th_ack, tp->snd_una) &&
13802             (rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree)) != NULL)
13803                 kern_prefetch(rsm, &prev_state);
13804         prev_state = rack->r_state;
13805         rack_clear_rate_sample(rack);
13806         retval = (*rack->r_substate) (m, th, so,
13807             tp, &to, drop_hdrlen,
13808             tlen, tiwin, thflags, nxt_pkt, iptos);
13809 #ifdef INVARIANTS
13810         if ((retval == 0) &&
13811             (tp->t_inpcb == NULL)) {
13812                 panic("retval:%d tp:%p t_inpcb:NULL state:%d",
13813                     retval, tp, prev_state);
13814         }
13815 #endif
13816         if (retval == 0) {
13817                 /*
13818                  * If retval is 1 the tcb is unlocked and most likely the tp
13819                  * is gone.
13820                  */
13821                 INP_WLOCK_ASSERT(tp->t_inpcb);
13822                 if ((rack->rc_gp_dyn_mul) &&
13823                     (rack->rc_always_pace) &&
13824                     (rack->use_fixed_rate == 0) &&
13825                     rack->in_probe_rtt &&
13826                     (rack->r_ctl.rc_time_probertt_starts == 0)) {
13827                         /*
13828                          * If we are going for target, lets recheck before
13829                          * we output.
13830                          */
13831                         rack_check_probe_rtt(rack, us_cts);
13832                 }
13833                 if (rack->set_pacing_done_a_iw == 0) {
13834                         /* How much has been acked? */
13835                         if ((tp->snd_una - tp->iss) > (ctf_fixed_maxseg(tp) * 10)) {
13836                                 /* We have enough to set in the pacing segment size */
13837                                 rack->set_pacing_done_a_iw = 1;
13838                                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
13839                         }
13840                 }
13841                 tcp_rack_xmit_timer_commit(rack, tp);
13842 #ifdef TCP_ACCOUNTING
13843                 /*
13844                  * If we set the ack_val_se to what ack processing we are doing
13845                  * we also want to track how many cycles we burned. Note
13846                  * the bits after tcp_output we let be "free". This is because
13847                  * we are also tracking the tcp_output times as well. Note the
13848                  * use of 0xf here since we only have 11 counter (0 - 0xa) and
13849                  * 0xf cannot be returned and is what we initialize it too to
13850                  * indicate we are not doing the tabulations.
13851                  */
13852                 if (ack_val_set != 0xf) {
13853                         uint64_t crtsc;
13854
13855                         crtsc = get_cyclecount();
13856                         counter_u64_add(tcp_proc_time[ack_val_set] , (crtsc - ts_val));
13857                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13858                                 tp->tcp_proc_time[ack_val_set] += (crtsc - ts_val);
13859                         }
13860                 }
13861 #endif
13862                 if (nxt_pkt == 0) {
13863                         if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
13864 do_output_now:
13865                                 did_out = 1;
13866                                 (void)tp->t_fb->tfb_tcp_output(tp);
13867                         }
13868                         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
13869                         rack_free_trim(rack);
13870                 }
13871                 if ((nxt_pkt == 0) &&
13872                     ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
13873                     (SEQ_GT(tp->snd_max, tp->snd_una) ||
13874                      (tp->t_flags & TF_DELACK) ||
13875                      ((V_tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
13876                       (tp->t_state <= TCPS_CLOSING)))) {
13877                         /* We could not send (probably in the hpts but stopped the timer earlier)? */
13878                         if ((tp->snd_max == tp->snd_una) &&
13879                             ((tp->t_flags & TF_DELACK) == 0) &&
13880                             (rack->rc_inp->inp_in_hpts) &&
13881                             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
13882                                 /* keep alive not needed if we are hptsi output yet */
13883                                 ;
13884                         } else {
13885                                 int late = 0;
13886                                 if (rack->rc_inp->inp_in_hpts) {
13887                                         if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
13888                                                 us_cts = tcp_get_usecs(NULL);
13889                                                 if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
13890                                                         rack->r_early = 1;
13891                                                         rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
13892                                                 } else
13893                                                         late = 1;
13894                                                 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
13895                                         }
13896                                         tcp_hpts_remove(tp->t_inpcb, HPTS_REMOVE_OUTPUT);
13897                                 }
13898                                 if (late && (did_out == 0)) {
13899                                         /*
13900                                          * We are late in the sending
13901                                          * and we did not call the output
13902                                          * (this probably should not happen).
13903                                          */
13904                                         goto do_output_now;
13905                                 }
13906                                 rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
13907                         }
13908                         way_out = 1;
13909                 } else if (nxt_pkt == 0) {
13910                         /* Do we have the correct timer running? */
13911                         rack_timer_audit(tp, rack, &so->so_snd);
13912                         way_out = 2;
13913                 }
13914         done_with_input:
13915                 rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out, max(1, m->m_pkthdr.lro_nsegs));
13916                 if (did_out)
13917                         rack->r_wanted_output = 0;
13918 #ifdef INVARIANTS
13919                 if (tp->t_inpcb == NULL) {
13920                         panic("OP:%d retval:%d tp:%p t_inpcb:NULL state:%d",
13921                               did_out,
13922                               retval, tp, prev_state);
13923                 }
13924 #endif
13925 #ifdef TCP_ACCOUNTING
13926         } else {
13927                 /*
13928                  * Track the time (see above).
13929                  */
13930                 if (ack_val_set != 0xf) {
13931                         uint64_t crtsc;
13932
13933                         crtsc = get_cyclecount();
13934                         counter_u64_add(tcp_proc_time[ack_val_set] , (crtsc - ts_val));
13935                         /*
13936                          * Note we *DO NOT* increment the per-tcb counters since
13937                          * in the else the TP may be gone!!
13938                          */
13939                 }
13940 #endif
13941         }
13942 #ifdef TCP_ACCOUNTING
13943         sched_unpin();
13944 #endif
13945         return (retval);
13946 }
13947
13948 void
13949 rack_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
13950     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
13951 {
13952         struct timeval tv;
13953
13954         /* First lets see if we have old packets */
13955         if (tp->t_in_pkt) {
13956                 if (ctf_do_queued_segments(so, tp, 1)) {
13957                         m_freem(m);
13958                         return;
13959                 }
13960         }
13961         if (m->m_flags & M_TSTMP_LRO) {
13962                 tv.tv_sec = m->m_pkthdr.rcv_tstmp /1000000000;
13963                 tv.tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000)/1000;
13964         } else {
13965                 /* Should not be should we kassert instead? */
13966                 tcp_get_usecs(&tv);
13967         }
13968         if (rack_do_segment_nounlock(m, th, so, tp,
13969                                      drop_hdrlen, tlen, iptos, 0, &tv) == 0) {
13970                 tcp_handle_wakeup(tp, so);
13971                 INP_WUNLOCK(tp->t_inpcb);
13972         }
13973 }
13974
13975 struct rack_sendmap *
13976 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused)
13977 {
13978         struct rack_sendmap *rsm = NULL;
13979         int32_t idx;
13980         uint32_t srtt = 0, thresh = 0, ts_low = 0;
13981
13982         /* Return the next guy to be re-transmitted */
13983         if (RB_EMPTY(&rack->r_ctl.rc_mtree)) {
13984                 return (NULL);
13985         }
13986         if (tp->t_flags & TF_SENTFIN) {
13987                 /* retran the end FIN? */
13988                 return (NULL);
13989         }
13990         /* ok lets look at this one */
13991         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
13992         if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) {
13993                 goto check_it;
13994         }
13995         rsm = rack_find_lowest_rsm(rack);
13996         if (rsm == NULL) {
13997                 return (NULL);
13998         }
13999 check_it:
14000         if (((rack->rc_tp->t_flags & TF_SACK_PERMIT) == 0) &&
14001             (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
14002                 /*
14003                  * No sack so we automatically do the 3 strikes and
14004                  * retransmit (no rack timer would be started).
14005                  */
14006
14007                 return (rsm);
14008         }
14009         if (rsm->r_flags & RACK_ACKED) {
14010                 return (NULL);
14011         }
14012         if (((rsm->r_flags & RACK_SACK_PASSED) == 0) &&
14013             (rsm->r_dupack < DUP_ACK_THRESHOLD)) {
14014                 /* Its not yet ready */
14015                 return (NULL);
14016         }
14017         srtt = rack_grab_rtt(tp, rack);
14018         idx = rsm->r_rtr_cnt - 1;
14019         ts_low = (uint32_t)rsm->r_tim_lastsent[idx];
14020         thresh = rack_calc_thresh_rack(rack, srtt, tsused);
14021         if ((tsused == ts_low) ||
14022             (TSTMP_LT(tsused, ts_low))) {
14023                 /* No time since sending */
14024                 return (NULL);
14025         }
14026         if ((tsused - ts_low) < thresh) {
14027                 /* It has not been long enough yet */
14028                 return (NULL);
14029         }
14030         if ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
14031             ((rsm->r_flags & RACK_SACK_PASSED) &&
14032              (rack->sack_attack_disable == 0))) {
14033                 /*
14034                  * We have passed the dup-ack threshold <or>
14035                  * a SACK has indicated this is missing.
14036                  * Note that if you are a declared attacker
14037                  * it is only the dup-ack threshold that
14038                  * will cause retransmits.
14039                  */
14040                 /* log retransmit reason */
14041                 rack_log_retran_reason(rack, rsm, (tsused - ts_low), thresh, 1);
14042                 rack->r_fast_output = 0;
14043                 return (rsm);
14044         }
14045         return (NULL);
14046 }
14047
14048 static void
14049 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
14050                            uint64_t bw_est, uint64_t bw, uint64_t len_time, int method,
14051                            int line, struct rack_sendmap *rsm)
14052 {
14053         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
14054                 union tcp_log_stackspecific log;
14055                 struct timeval tv;
14056
14057                 memset(&log, 0, sizeof(log));
14058                 log.u_bbr.flex1 = slot;
14059                 log.u_bbr.flex2 = len;
14060                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_min_segs;
14061                 log.u_bbr.flex4 = rack->r_ctl.rc_pace_max_segs;
14062                 log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ss;
14063                 log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_ca;
14064                 log.u_bbr.use_lt_bw = rack->rc_ack_can_sendout_data;
14065                 log.u_bbr.use_lt_bw <<= 1;
14066                 log.u_bbr.use_lt_bw |= rack->r_late;
14067                 log.u_bbr.use_lt_bw <<= 1;
14068                 log.u_bbr.use_lt_bw |= rack->r_early;
14069                 log.u_bbr.use_lt_bw <<= 1;
14070                 log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
14071                 log.u_bbr.use_lt_bw <<= 1;
14072                 log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
14073                 log.u_bbr.use_lt_bw <<= 1;
14074                 log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
14075                 log.u_bbr.use_lt_bw <<= 1;
14076                 log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
14077                 log.u_bbr.use_lt_bw <<= 1;
14078                 log.u_bbr.use_lt_bw |= rack->gp_ready;
14079                 log.u_bbr.pkt_epoch = line;
14080                 log.u_bbr.epoch = rack->r_ctl.rc_agg_delayed;
14081                 log.u_bbr.lt_epoch = rack->r_ctl.rc_agg_early;
14082                 log.u_bbr.applimited = rack->r_ctl.rack_per_of_gp_rec;
14083                 log.u_bbr.bw_inuse = bw_est;
14084                 log.u_bbr.delRate = bw;
14085                 if (rack->r_ctl.gp_bw == 0)
14086                         log.u_bbr.cur_del_rate = 0;
14087                 else
14088                         log.u_bbr.cur_del_rate = rack_get_bw(rack);
14089                 log.u_bbr.rttProp = len_time;
14090                 log.u_bbr.pkts_out = rack->r_ctl.rc_rack_min_rtt;
14091                 log.u_bbr.lost = rack->r_ctl.rc_probertt_sndmax_atexit;
14092                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
14093                 if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) {
14094                         /* We are in slow start */
14095                         log.u_bbr.flex7 = 1;
14096                 } else {
14097                         /* we are on congestion avoidance */
14098                         log.u_bbr.flex7 = 0;
14099                 }
14100                 log.u_bbr.flex8 = method;
14101                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14102                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14103                 log.u_bbr.cwnd_gain = rack->rc_gp_saw_rec;
14104                 log.u_bbr.cwnd_gain <<= 1;
14105                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
14106                 log.u_bbr.cwnd_gain <<= 1;
14107                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
14108                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
14109                     &rack->rc_inp->inp_socket->so_rcv,
14110                     &rack->rc_inp->inp_socket->so_snd,
14111                     BBR_LOG_HPTSI_CALC, 0,
14112                     0, &log, false, &tv);
14113         }
14114 }
14115
14116 static uint32_t
14117 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss)
14118 {
14119         uint32_t new_tso, user_max;
14120
14121         user_max = rack->rc_user_set_max_segs * mss;
14122         if (rack->rc_force_max_seg) {
14123                 return (user_max);
14124         }
14125         if (rack->use_fixed_rate &&
14126             ((rack->r_ctl.crte == NULL) ||
14127              (bw != rack->r_ctl.crte->rate))) {
14128                 /* Use the user mss since we are not exactly matched */
14129                 return (user_max);
14130         }
14131         new_tso = tcp_get_pacing_burst_size(rack->rc_tp, bw, mss, rack_pace_one_seg, rack->r_ctl.crte, NULL);
14132         if (new_tso > user_max)
14133                 new_tso = user_max;
14134         return (new_tso);
14135 }
14136
14137 static int32_t
14138 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)
14139 {
14140         uint64_t lentim, fill_bw;
14141
14142         /* Lets first see if we are full, if so continue with normal rate */
14143         rack->r_via_fill_cw = 0;
14144         if (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.cwnd_to_use)
14145                 return (slot);
14146         if ((ctf_outstanding(rack->rc_tp) + (segsiz-1)) > rack->rc_tp->snd_wnd)
14147                 return (slot);
14148         if (rack->r_ctl.rc_last_us_rtt == 0)
14149                 return (slot);
14150         if (rack->rc_pace_fill_if_rttin_range &&
14151             (rack->r_ctl.rc_last_us_rtt >=
14152              (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack->rtt_limit_mul))) {
14153                 /* The rtt is huge, N * smallest, lets not fill */
14154                 return (slot);
14155         }
14156         /*
14157          * first lets calculate the b/w based on the last us-rtt
14158          * and the sndwnd.
14159          */
14160         fill_bw = rack->r_ctl.cwnd_to_use;
14161         /* Take the rwnd if its smaller */
14162         if (fill_bw > rack->rc_tp->snd_wnd)
14163                 fill_bw = rack->rc_tp->snd_wnd;
14164         if (rack->r_fill_less_agg) {
14165                 /*
14166                  * Now take away the inflight (this will reduce our
14167                  * aggressiveness and yeah, if we get that much out in 1RTT
14168                  * we will have had acks come back and still be behind).
14169                  */
14170                 fill_bw -= ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14171         }
14172         /* Now lets make it into a b/w */
14173         fill_bw *= (uint64_t)HPTS_USEC_IN_SEC;
14174         fill_bw /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
14175         /* We are below the min b/w */
14176         if (non_paced)
14177                 *rate_wanted = fill_bw;
14178         if ((fill_bw < RACK_MIN_BW) || (fill_bw < *rate_wanted))
14179                 return (slot);
14180         if (rack->r_ctl.bw_rate_cap && (fill_bw > rack->r_ctl.bw_rate_cap))
14181                 fill_bw = rack->r_ctl.bw_rate_cap;
14182         rack->r_via_fill_cw = 1;
14183         if (rack->r_rack_hw_rate_caps &&
14184             (rack->r_ctl.crte != NULL)) {
14185                 uint64_t high_rate;
14186
14187                 high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
14188                 if (fill_bw > high_rate) {
14189                         /* We are capping bw at the highest rate table entry */
14190                         if (*rate_wanted > high_rate) {
14191                                 /* The original rate was also capped */
14192                                 rack->r_via_fill_cw = 0;
14193                         }
14194                         rack_log_hdwr_pacing(rack,
14195                                              fill_bw, high_rate, __LINE__,
14196                                              0, 3);
14197                         fill_bw = high_rate;
14198                         if (capped)
14199                                 *capped = 1;
14200                 }
14201         } else if ((rack->r_ctl.crte == NULL) &&
14202                    (rack->rack_hdrw_pacing == 0) &&
14203                    (rack->rack_hdw_pace_ena) &&
14204                    rack->r_rack_hw_rate_caps &&
14205                    (rack->rack_attempt_hdwr_pace == 0) &&
14206                    (rack->rc_inp->inp_route.ro_nh != NULL) &&
14207                    (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
14208                 /*
14209                  * Ok we may have a first attempt that is greater than our top rate
14210                  * lets check.
14211                  */
14212                 uint64_t high_rate;
14213
14214                 high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
14215                 if (high_rate) {
14216                         if (fill_bw > high_rate) {
14217                                 fill_bw = high_rate;
14218                                 if (capped)
14219                                         *capped = 1;
14220                         }
14221                 }
14222         }
14223         /*
14224          * Ok fill_bw holds our mythical b/w to fill the cwnd
14225          * in a rtt, what does that time wise equate too?
14226          */
14227         lentim = (uint64_t)(len) * (uint64_t)HPTS_USEC_IN_SEC;
14228         lentim /= fill_bw;
14229         *rate_wanted = fill_bw;
14230         if (non_paced || (lentim < slot)) {
14231                 rack_log_pacing_delay_calc(rack, len, slot, fill_bw,
14232                                            0, lentim, 12, __LINE__, NULL);
14233                 return ((int32_t)lentim);
14234         } else
14235                 return (slot);
14236 }
14237
14238 static int32_t
14239 rack_get_pacing_delay(struct tcp_rack *rack, struct tcpcb *tp, uint32_t len, struct rack_sendmap *rsm, uint32_t segsiz)
14240 {
14241         struct rack_sendmap *lrsm;
14242         int32_t slot = 0;
14243         int can_start_hw_pacing = 1;
14244         int err;
14245
14246         if (rack->rc_always_pace == 0) {
14247                 /*
14248                  * We use the most optimistic possible cwnd/srtt for
14249                  * sending calculations. This will make our
14250                  * calculation anticipate getting more through
14251                  * quicker then possible. But thats ok we don't want
14252                  * the peer to have a gap in data sending.
14253                  */
14254                 uint32_t srtt, cwnd, tr_perms = 0;
14255                 int32_t reduce = 0;
14256
14257         old_method:
14258                 /*
14259                  * We keep no precise pacing with the old method
14260                  * instead we use the pacer to mitigate bursts.
14261                  */
14262                 if (rack->r_ctl.rc_rack_min_rtt)
14263                         srtt = rack->r_ctl.rc_rack_min_rtt;
14264                 else
14265                         srtt = max(tp->t_srtt, 1);
14266                 if (rack->r_ctl.rc_rack_largest_cwnd)
14267                         cwnd = rack->r_ctl.rc_rack_largest_cwnd;
14268                 else
14269                         cwnd = rack->r_ctl.cwnd_to_use;
14270                 /* Inflate cwnd by 1000 so srtt of usecs is in ms */
14271                 tr_perms = (cwnd * 1000) / srtt;
14272                 if (tr_perms == 0) {
14273                         tr_perms = ctf_fixed_maxseg(tp);
14274                 }
14275                 /*
14276                  * Calculate how long this will take to drain, if
14277                  * the calculation comes out to zero, thats ok we
14278                  * will use send_a_lot to possibly spin around for
14279                  * more increasing tot_len_this_send to the point
14280                  * that its going to require a pace, or we hit the
14281                  * cwnd. Which in that case we are just waiting for
14282                  * a ACK.
14283                  */
14284                 slot = len / tr_perms;
14285                 /* Now do we reduce the time so we don't run dry? */
14286                 if (slot && rack_slot_reduction) {
14287                         reduce = (slot / rack_slot_reduction);
14288                         if (reduce < slot) {
14289                                 slot -= reduce;
14290                         } else
14291                                 slot = 0;
14292                 }
14293                 slot *= HPTS_USEC_IN_MSEC;
14294                 if (rsm == NULL) {
14295                         /*
14296                          * We always consider ourselves app limited with old style
14297                          * that are not retransmits. This could be the initial
14298                          * measurement, but thats ok its all setup and specially
14299                          * handled. If another send leaks out, then that too will
14300                          * be mark app-limited.
14301                          */
14302                         lrsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
14303                         if (lrsm && ((lrsm->r_flags & RACK_APP_LIMITED) == 0)) {
14304                                 rack->r_ctl.rc_first_appl = lrsm;
14305                                 lrsm->r_flags |= RACK_APP_LIMITED;
14306                                 rack->r_ctl.rc_app_limited_cnt++;
14307                         }
14308                 }
14309                 if (rack->rc_pace_to_cwnd) {
14310                         uint64_t rate_wanted = 0;
14311
14312                         slot = pace_to_fill_cwnd(rack, slot, len, segsiz, NULL, &rate_wanted, 1);
14313                         rack->rc_ack_can_sendout_data = 1;
14314                         rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, 0, 0, 14, __LINE__, NULL);
14315                 } else
14316                         rack_log_pacing_delay_calc(rack, len, slot, tr_perms, reduce, 0, 7, __LINE__, NULL);
14317         } else {
14318                 uint64_t bw_est, res, lentim, rate_wanted;
14319                 uint32_t orig_val, srtt, segs, oh;
14320                 int capped = 0;
14321                 int prev_fill;
14322
14323                 if ((rack->r_rr_config == 1) && rsm) {
14324                         return (rack->r_ctl.rc_min_to);
14325                 }
14326                 if (rack->use_fixed_rate) {
14327                         rate_wanted = bw_est = rack_get_fixed_pacing_bw(rack);
14328                 } else if ((rack->r_ctl.init_rate == 0) &&
14329 #ifdef NETFLIX_PEAKRATE
14330                            (rack->rc_tp->t_maxpeakrate == 0) &&
14331 #endif
14332                            (rack->r_ctl.gp_bw == 0)) {
14333                         /* no way to yet do an estimate */
14334                         bw_est = rate_wanted = 0;
14335                 } else {
14336                         bw_est = rack_get_bw(rack);
14337                         rate_wanted = rack_get_output_bw(rack, bw_est, rsm, &capped);
14338                 }
14339                 if ((bw_est == 0) || (rate_wanted == 0) ||
14340                     ((rack->gp_ready == 0) && (rack->use_fixed_rate == 0))) {
14341                         /*
14342                          * No way yet to make a b/w estimate or
14343                          * our raise is set incorrectly.
14344                          */
14345                         goto old_method;
14346                 }
14347                 /* We need to account for all the overheads */
14348                 segs = (len + segsiz - 1) / segsiz;
14349                 /*
14350                  * We need the diff between 1514 bytes (e-mtu with e-hdr)
14351                  * and how much data we put in each packet. Yes this
14352                  * means we may be off if we are larger than 1500 bytes
14353                  * or smaller. But this just makes us more conservative.
14354                  */
14355                 if (rack_hw_rate_min &&
14356                     (bw_est < rack_hw_rate_min))
14357                         can_start_hw_pacing = 0;
14358                 if (ETHERNET_SEGMENT_SIZE > segsiz)
14359                         oh = ETHERNET_SEGMENT_SIZE - segsiz;
14360                 else
14361                         oh = 0;
14362                 segs *= oh;
14363                 lentim = (uint64_t)(len + segs) * (uint64_t)HPTS_USEC_IN_SEC;
14364                 res = lentim / rate_wanted;
14365                 slot = (uint32_t)res;
14366                 orig_val = rack->r_ctl.rc_pace_max_segs;
14367                 if (rack->r_ctl.crte == NULL) {
14368                         /*
14369                          * Only do this if we are not hardware pacing
14370                          * since if we are doing hw-pacing below we will
14371                          * set make a call after setting up or changing
14372                          * the rate.
14373                          */
14374                         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
14375                 } else if (rack->rc_inp->inp_snd_tag == NULL) {
14376                         /*
14377                          * We lost our rate somehow, this can happen
14378                          * if the interface changed underneath us.
14379                          */
14380                         tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
14381                         rack->r_ctl.crte = NULL;
14382                         /* Lets re-allow attempting to setup pacing */
14383                         rack->rack_hdrw_pacing = 0;
14384                         rack->rack_attempt_hdwr_pace = 0;
14385                         rack_log_hdwr_pacing(rack,
14386                                              rate_wanted, bw_est, __LINE__,
14387                                              0, 6);
14388                 }
14389                 /* Did we change the TSO size, if so log it */
14390                 if (rack->r_ctl.rc_pace_max_segs != orig_val)
14391                         rack_log_pacing_delay_calc(rack, len, slot, orig_val, 0, 0, 15, __LINE__, NULL);
14392                 prev_fill = rack->r_via_fill_cw;
14393                 if ((rack->rc_pace_to_cwnd) &&
14394                     (capped == 0) &&
14395                     (rack->use_fixed_rate == 0) &&
14396                     (rack->in_probe_rtt == 0) &&
14397                     (IN_FASTRECOVERY(rack->rc_tp->t_flags) == 0)) {
14398                         /*
14399                          * We want to pace at our rate *or* faster to
14400                          * fill the cwnd to the max if its not full.
14401                          */
14402                         slot = pace_to_fill_cwnd(rack, slot, (len+segs), segsiz, &capped, &rate_wanted, 0);
14403                 }
14404                 if ((rack->rc_inp->inp_route.ro_nh != NULL) &&
14405                     (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
14406                         if ((rack->rack_hdw_pace_ena) &&
14407                             (can_start_hw_pacing > 0) &&
14408                             (rack->rack_hdrw_pacing == 0) &&
14409                             (rack->rack_attempt_hdwr_pace == 0)) {
14410                                 /*
14411                                  * Lets attempt to turn on hardware pacing
14412                                  * if we can.
14413                                  */
14414                                 rack->rack_attempt_hdwr_pace = 1;
14415                                 rack->r_ctl.crte = tcp_set_pacing_rate(rack->rc_tp,
14416                                                                        rack->rc_inp->inp_route.ro_nh->nh_ifp,
14417                                                                        rate_wanted,
14418                                                                        RS_PACING_GEQ,
14419                                                                        &err, &rack->r_ctl.crte_prev_rate);
14420                                 if (rack->r_ctl.crte) {
14421                                         rack->rack_hdrw_pacing = 1;
14422                                         rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted, segsiz,
14423                                                                                                  0, rack->r_ctl.crte,
14424                                                                                                  NULL);
14425                                         rack_log_hdwr_pacing(rack,
14426                                                              rate_wanted, rack->r_ctl.crte->rate, __LINE__,
14427                                                              err, 0);
14428                                         rack->r_ctl.last_hw_bw_req = rate_wanted;
14429                                 } else {
14430                                         counter_u64_add(rack_hw_pace_init_fail, 1);
14431                                 }
14432                         } else if (rack->rack_hdrw_pacing &&
14433                                    (rack->r_ctl.last_hw_bw_req != rate_wanted)) {
14434                                 /* Do we need to adjust our rate? */
14435                                 const struct tcp_hwrate_limit_table *nrte;
14436
14437                                 if (rack->r_up_only &&
14438                                     (rate_wanted < rack->r_ctl.crte->rate)) {
14439                                         /**
14440                                          * We have four possible states here
14441                                          * having to do with the previous time
14442                                          * and this time.
14443                                          *   previous  |  this-time
14444                                          * A)     0      |     0   -- fill_cw not in the picture
14445                                          * B)     1      |     0   -- we were doing a fill-cw but now are not
14446                                          * C)     1      |     1   -- all rates from fill_cw
14447                                          * D)     0      |     1   -- we were doing non-fill and now we are filling
14448                                          *
14449                                          * For case A, C and D we don't allow a drop. But for
14450                                          * case B where we now our on our steady rate we do
14451                                          * allow a drop.
14452                                          *
14453                                          */
14454                                         if (!((prev_fill == 1) && (rack->r_via_fill_cw == 0)))
14455                                                 goto done_w_hdwr;
14456                                 }
14457                                 if ((rate_wanted > rack->r_ctl.crte->rate) ||
14458                                     (rate_wanted <= rack->r_ctl.crte_prev_rate)) {
14459                                         if (rack_hw_rate_to_low &&
14460                                             (bw_est < rack_hw_rate_to_low)) {
14461                                                 /*
14462                                                  * The pacing rate is too low for hardware, but
14463                                                  * do allow hardware pacing to be restarted.
14464                                                  */
14465                                                 rack_log_hdwr_pacing(rack,
14466                                                              bw_est, rack->r_ctl.crte->rate, __LINE__,
14467                                                              0, 5);
14468                                                 tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
14469                                                 rack->r_ctl.crte = NULL;
14470                                                 rack->rack_attempt_hdwr_pace = 0;
14471                                                 rack->rack_hdrw_pacing = 0;
14472                                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
14473                                                 goto done_w_hdwr;
14474                                         }
14475                                         nrte = tcp_chg_pacing_rate(rack->r_ctl.crte,
14476                                                                    rack->rc_tp,
14477                                                                    rack->rc_inp->inp_route.ro_nh->nh_ifp,
14478                                                                    rate_wanted,
14479                                                                    RS_PACING_GEQ,
14480                                                                    &err, &rack->r_ctl.crte_prev_rate);
14481                                         if (nrte == NULL) {
14482                                                 /* Lost the rate */
14483                                                 rack->rack_hdrw_pacing = 0;
14484                                                 rack->r_ctl.crte = NULL;
14485                                                 rack_log_hdwr_pacing(rack,
14486                                                                      rate_wanted, 0, __LINE__,
14487                                                                      err, 1);
14488                                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
14489                                                 counter_u64_add(rack_hw_pace_lost, 1);
14490                                         } else if (nrte != rack->r_ctl.crte) {
14491                                                 rack->r_ctl.crte = nrte;
14492                                                 rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted,
14493                                                                                                          segsiz, 0,
14494                                                                                                          rack->r_ctl.crte,
14495                                                                                                          NULL);
14496                                                 rack_log_hdwr_pacing(rack,
14497                                                                      rate_wanted, rack->r_ctl.crte->rate, __LINE__,
14498                                                                      err, 2);
14499                                                 rack->r_ctl.last_hw_bw_req = rate_wanted;
14500                                         }
14501                                 } else {
14502                                         /* We just need to adjust the segment size */
14503                                         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
14504                                         rack_log_hdwr_pacing(rack,
14505                                                              rate_wanted, rack->r_ctl.crte->rate, __LINE__,
14506                                                              0, 4);
14507                                         rack->r_ctl.last_hw_bw_req = rate_wanted;
14508                                 }
14509                         }
14510                 }
14511                 if ((rack->r_ctl.crte != NULL) &&
14512                     (rack->r_ctl.crte->rate == rate_wanted)) {
14513                         /*
14514                          * We need to add a extra if the rates
14515                          * are exactly matched. The idea is
14516                          * we want the software to make sure the
14517                          * queue is empty before adding more, this
14518                          * gives us N MSS extra pace times where
14519                          * N is our sysctl
14520                          */
14521                         slot += (rack->r_ctl.crte->time_between * rack_hw_pace_extra_slots);
14522                 }
14523 done_w_hdwr:
14524                 if (rack_limit_time_with_srtt &&
14525                     (rack->use_fixed_rate == 0) &&
14526 #ifdef NETFLIX_PEAKRATE
14527                     (rack->rc_tp->t_maxpeakrate == 0) &&
14528 #endif
14529                     (rack->rack_hdrw_pacing == 0)) {
14530                         /*
14531                          * Sanity check, we do not allow the pacing delay
14532                          * to be longer than the SRTT of the path. If it is
14533                          * a slow path, then adding a packet should increase
14534                          * the RTT and compensate for this i.e. the srtt will
14535                          * be greater so the allowed pacing time will be greater.
14536                          *
14537                          * Note this restriction is not for where a peak rate
14538                          * is set, we are doing fixed pacing or hardware pacing.
14539                          */
14540                         if (rack->rc_tp->t_srtt)
14541                                 srtt = rack->rc_tp->t_srtt;
14542                         else
14543                                 srtt = RACK_INITIAL_RTO * HPTS_USEC_IN_MSEC;    /* its in ms convert */
14544                         if (srtt < slot) {
14545                                 rack_log_pacing_delay_calc(rack, srtt, slot, rate_wanted, bw_est, lentim, 99, __LINE__, NULL);
14546                                 slot = srtt;
14547                         }
14548                 }
14549                 rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, bw_est, lentim, 2, __LINE__, rsm);
14550         }
14551         if (rack->r_ctl.crte && (rack->r_ctl.crte->rs_num_enobufs > 0)) {
14552                 /*
14553                  * If this rate is seeing enobufs when it
14554                  * goes to send then either the nic is out
14555                  * of gas or we are mis-estimating the time
14556                  * somehow and not letting the queue empty
14557                  * completely. Lets add to the pacing time.
14558                  */
14559                 int hw_boost_delay;
14560
14561                 hw_boost_delay = rack->r_ctl.crte->time_between * rack_enobuf_hw_boost_mult;
14562                 if (hw_boost_delay > rack_enobuf_hw_max)
14563                         hw_boost_delay = rack_enobuf_hw_max;
14564                 else if (hw_boost_delay < rack_enobuf_hw_min)
14565                         hw_boost_delay = rack_enobuf_hw_min;
14566                 slot += hw_boost_delay;
14567         }
14568         if (slot)
14569                 counter_u64_add(rack_calc_nonzero, 1);
14570         else
14571                 counter_u64_add(rack_calc_zero, 1);
14572         return (slot);
14573 }
14574
14575 static void
14576 rack_start_gp_measurement(struct tcpcb *tp, struct tcp_rack *rack,
14577     tcp_seq startseq, uint32_t sb_offset)
14578 {
14579         struct rack_sendmap *my_rsm = NULL;
14580         struct rack_sendmap fe;
14581
14582         if (tp->t_state < TCPS_ESTABLISHED) {
14583                 /*
14584                  * We don't start any measurements if we are
14585                  * not at least established.
14586                  */
14587                 return;
14588         }
14589         tp->t_flags |= TF_GPUTINPROG;
14590         rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
14591         rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
14592         tp->gput_seq = startseq;
14593         rack->app_limited_needs_set = 0;
14594         if (rack->in_probe_rtt)
14595                 rack->measure_saw_probe_rtt = 1;
14596         else if ((rack->measure_saw_probe_rtt) &&
14597                  (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
14598                 rack->measure_saw_probe_rtt = 0;
14599         if (rack->rc_gp_filled)
14600                 tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
14601         else {
14602                 /* Special case initial measurement */
14603                 struct timeval tv;
14604
14605                 tp->gput_ts = tcp_get_usecs(&tv);
14606                 rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
14607         }
14608         /*
14609          * We take a guess out into the future,
14610          * if we have no measurement and no
14611          * initial rate, we measure the first
14612          * initial-windows worth of data to
14613          * speed up getting some GP measurement and
14614          * thus start pacing.
14615          */
14616         if ((rack->rc_gp_filled == 0) && (rack->r_ctl.init_rate == 0)) {
14617                 rack->app_limited_needs_set = 1;
14618                 tp->gput_ack = startseq + max(rc_init_window(rack),
14619                                               (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
14620                 rack_log_pacing_delay_calc(rack,
14621                                            tp->gput_seq,
14622                                            tp->gput_ack,
14623                                            0,
14624                                            tp->gput_ts,
14625                                            rack->r_ctl.rc_app_limited_cnt,
14626                                            9,
14627                                            __LINE__, NULL);
14628                 return;
14629         }
14630         if (sb_offset) {
14631                 /*
14632                  * We are out somewhere in the sb
14633                  * can we use the already outstanding data?
14634                  */
14635
14636                 if (rack->r_ctl.rc_app_limited_cnt == 0) {
14637                         /*
14638                          * Yes first one is good and in this case
14639                          * the tp->gput_ts is correctly set based on
14640                          * the last ack that arrived (no need to
14641                          * set things up when an ack comes in).
14642                          */
14643                         my_rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
14644                         if ((my_rsm == NULL) ||
14645                             (my_rsm->r_rtr_cnt != 1)) {
14646                                 /* retransmission? */
14647                                 goto use_latest;
14648                         }
14649                 } else {
14650                         if (rack->r_ctl.rc_first_appl == NULL) {
14651                                 /*
14652                                  * If rc_first_appl is NULL
14653                                  * then the cnt should be 0.
14654                                  * This is probably an error, maybe
14655                                  * a KASSERT would be approprate.
14656                                  */
14657                                 goto use_latest;
14658                         }
14659                         /*
14660                          * If we have a marker pointer to the last one that is
14661                          * app limited we can use that, but we need to set
14662                          * things up so that when it gets ack'ed we record
14663                          * the ack time (if its not already acked).
14664                          */
14665                         rack->app_limited_needs_set = 1;
14666                         /*
14667                          * We want to get to the rsm that is either
14668                          * next with space i.e. over 1 MSS or the one
14669                          * after that (after the app-limited).
14670                          */
14671                         my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
14672                                          rack->r_ctl.rc_first_appl);
14673                         if (my_rsm) {
14674                                 if ((my_rsm->r_end - my_rsm->r_start) <= ctf_fixed_maxseg(tp))
14675                                         /* Have to use the next one */
14676                                         my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
14677                                                          my_rsm);
14678                                 else {
14679                                         /* Use after the first MSS of it is acked */
14680                                         tp->gput_seq = my_rsm->r_start + ctf_fixed_maxseg(tp);
14681                                         goto start_set;
14682                                 }
14683                         }
14684                         if ((my_rsm == NULL) ||
14685                             (my_rsm->r_rtr_cnt != 1)) {
14686                                 /*
14687                                  * Either its a retransmit or
14688                                  * the last is the app-limited one.
14689                                  */
14690                                 goto use_latest;
14691                         }
14692                 }
14693                 tp->gput_seq = my_rsm->r_start;
14694 start_set:
14695                 if (my_rsm->r_flags & RACK_ACKED) {
14696                         /*
14697                          * This one has been acked use the arrival ack time
14698                          */
14699                         tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
14700                         rack->app_limited_needs_set = 0;
14701                 }
14702                 rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)];
14703                 tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
14704                 rack_log_pacing_delay_calc(rack,
14705                                            tp->gput_seq,
14706                                            tp->gput_ack,
14707                                            (uint64_t)my_rsm,
14708                                            tp->gput_ts,
14709                                            rack->r_ctl.rc_app_limited_cnt,
14710                                            9,
14711                                            __LINE__, NULL);
14712                 return;
14713         }
14714
14715 use_latest:
14716         /*
14717          * We don't know how long we may have been
14718          * idle or if this is the first-send. Lets
14719          * setup the flag so we will trim off
14720          * the first ack'd data so we get a true
14721          * measurement.
14722          */
14723         rack->app_limited_needs_set = 1;
14724         tp->gput_ack = startseq + rack_get_measure_window(tp, rack);
14725         /* Find this guy so we can pull the send time */
14726         fe.r_start = startseq;
14727         my_rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
14728         if (my_rsm) {
14729                 rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)];
14730                 if (my_rsm->r_flags & RACK_ACKED) {
14731                         /*
14732                          * Unlikely since its probably what was
14733                          * just transmitted (but I am paranoid).
14734                          */
14735                         tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
14736                         rack->app_limited_needs_set = 0;
14737                 }
14738                 if (SEQ_LT(my_rsm->r_start, tp->gput_seq)) {
14739                         /* This also is unlikely */
14740                         tp->gput_seq = my_rsm->r_start;
14741                 }
14742         } else {
14743                 /*
14744                  * TSNH unless we have some send-map limit,
14745                  * and even at that it should not be hitting
14746                  * that limit (we should have stopped sending).
14747                  */
14748                 struct timeval tv;
14749
14750                 microuptime(&tv);
14751                 rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
14752         }
14753         rack_log_pacing_delay_calc(rack,
14754                                    tp->gput_seq,
14755                                    tp->gput_ack,
14756                                    (uint64_t)my_rsm,
14757                                    tp->gput_ts,
14758                                    rack->r_ctl.rc_app_limited_cnt,
14759                                    9, __LINE__, NULL);
14760 }
14761
14762 static inline uint32_t
14763 rack_what_can_we_send(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t cwnd_to_use,
14764     uint32_t avail, int32_t sb_offset)
14765 {
14766         uint32_t len;
14767         uint32_t sendwin;
14768
14769         if (tp->snd_wnd > cwnd_to_use)
14770                 sendwin = cwnd_to_use;
14771         else
14772                 sendwin = tp->snd_wnd;
14773         if (ctf_outstanding(tp) >= tp->snd_wnd) {
14774                 /* We never want to go over our peers rcv-window */
14775                 len = 0;
14776         } else {
14777                 uint32_t flight;
14778
14779                 flight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
14780                 if (flight >= sendwin) {
14781                         /*
14782                          * We have in flight what we are allowed by cwnd (if
14783                          * it was rwnd blocking it would have hit above out
14784                          * >= tp->snd_wnd).
14785                          */
14786                         return (0);
14787                 }
14788                 len = sendwin - flight;
14789                 if ((len + ctf_outstanding(tp)) > tp->snd_wnd) {
14790                         /* We would send too much (beyond the rwnd) */
14791                         len = tp->snd_wnd - ctf_outstanding(tp);
14792                 }
14793                 if ((len + sb_offset) > avail) {
14794                         /*
14795                          * We don't have that much in the SB, how much is
14796                          * there?
14797                          */
14798                         len = avail - sb_offset;
14799                 }
14800         }
14801         return (len);
14802 }
14803
14804 static void
14805 rack_log_fsb(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t flags,
14806              unsigned ipoptlen, int32_t orig_len, int32_t len, int error,
14807              int rsm_is_null, int optlen, int line, uint16_t mode)
14808 {
14809         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
14810                 union tcp_log_stackspecific log;
14811                 struct timeval tv;
14812
14813                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
14814                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
14815                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
14816                 log.u_bbr.flex1 = error;
14817                 log.u_bbr.flex2 = flags;
14818                 log.u_bbr.flex3 = rsm_is_null;
14819                 log.u_bbr.flex4 = ipoptlen;
14820                 log.u_bbr.flex5 = tp->rcv_numsacks;
14821                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
14822                 log.u_bbr.flex7 = optlen;
14823                 log.u_bbr.flex8 = rack->r_fsb_inited;
14824                 log.u_bbr.applimited = rack->r_fast_output;
14825                 log.u_bbr.bw_inuse = rack_get_bw(rack);
14826                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
14827                 log.u_bbr.cwnd_gain = mode;
14828                 log.u_bbr.pkts_out = orig_len;
14829                 log.u_bbr.lt_epoch = len;
14830                 log.u_bbr.delivered = line;
14831                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14832                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14833                 tcp_log_event_(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_FSB, 0,
14834                                len, &log, false, NULL, NULL, 0, &tv);
14835         }
14836 }
14837
14838
14839 static struct mbuf *
14840 rack_fo_base_copym(struct mbuf *the_m, uint32_t the_off, int32_t *plen,
14841                    struct rack_fast_send_blk *fsb,
14842                    int32_t seglimit, int32_t segsize)
14843 {
14844 #ifdef KERN_TLS
14845         struct ktls_session *tls, *ntls;
14846         struct mbuf *start;
14847 #endif
14848         struct mbuf *m, *n, **np, *smb;
14849         struct mbuf *top;
14850         int32_t off, soff;
14851         int32_t len = *plen;
14852         int32_t fragsize;
14853         int32_t len_cp = 0;
14854         uint32_t mlen, frags;
14855
14856         soff = off = the_off;
14857         smb = m = the_m;
14858         np = &top;
14859         top = NULL;
14860 #ifdef KERN_TLS
14861         if (hw_tls && (m->m_flags & M_EXTPG))
14862                 tls = m->m_epg_tls;
14863         else
14864                 tls = NULL;
14865         start = m;
14866 #endif
14867         while (len > 0) {
14868                 if (m == NULL) {
14869                         *plen = len_cp;
14870                         break;
14871                 }
14872 #ifdef KERN_TLS
14873                 if (hw_tls) {
14874                         if (m->m_flags & M_EXTPG)
14875                                 ntls = m->m_epg_tls;
14876                         else
14877                                 ntls = NULL;
14878
14879                         /*
14880                          * Avoid mixing TLS records with handshake
14881                          * data or TLS records from different
14882                          * sessions.
14883                          */
14884                         if (tls != ntls) {
14885                                 MPASS(m != start);
14886                                 *plen = len_cp;
14887                                 break;
14888                         }
14889                 }
14890 #endif
14891                 mlen = min(len, m->m_len - off);
14892                 if (seglimit) {
14893                         /*
14894                          * For M_EXTPG mbufs, add 3 segments
14895                          * + 1 in case we are crossing page boundaries
14896                          * + 2 in case the TLS hdr/trailer are used
14897                          * It is cheaper to just add the segments
14898                          * than it is to take the cache miss to look
14899                          * at the mbuf ext_pgs state in detail.
14900                          */
14901                         if (m->m_flags & M_EXTPG) {
14902                                 fragsize = min(segsize, PAGE_SIZE);
14903                                 frags = 3;
14904                         } else {
14905                                 fragsize = segsize;
14906                                 frags = 0;
14907                         }
14908
14909                         /* Break if we really can't fit anymore. */
14910                         if ((frags + 1) >= seglimit) {
14911                                 *plen = len_cp;
14912                                 break;
14913                         }
14914
14915                         /*
14916                          * Reduce size if you can't copy the whole
14917                          * mbuf. If we can't copy the whole mbuf, also
14918                          * adjust len so the loop will end after this
14919                          * mbuf.
14920                          */
14921                         if ((frags + howmany(mlen, fragsize)) >= seglimit) {
14922                                 mlen = (seglimit - frags - 1) * fragsize;
14923                                 len = mlen;
14924                                 *plen = len_cp + len;
14925                         }
14926                         frags += howmany(mlen, fragsize);
14927                         if (frags == 0)
14928                                 frags++;
14929                         seglimit -= frags;
14930                         KASSERT(seglimit > 0,
14931                             ("%s: seglimit went too low", __func__));
14932                 }
14933                 n = m_get(M_NOWAIT, m->m_type);
14934                 *np = n;
14935                 if (n == NULL)
14936                         goto nospace;
14937                 n->m_len = mlen;
14938                 soff += mlen;
14939                 len_cp += n->m_len;
14940                 if (m->m_flags & (M_EXT|M_EXTPG)) {
14941                         n->m_data = m->m_data + off;
14942                         mb_dupcl(n, m);
14943                 } else {
14944                         bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
14945                             (u_int)n->m_len);
14946                 }
14947                 len -= n->m_len;
14948                 off = 0;
14949                 m = m->m_next;
14950                 np = &n->m_next;
14951                 if (len || (soff == smb->m_len)) {
14952                         /*
14953                          * We have more so we move forward  or
14954                          * we have consumed the entire mbuf and
14955                          * len has fell to 0.
14956                          */
14957                         soff = 0;
14958                         smb = m;
14959                 }
14960
14961         }
14962         if (fsb != NULL) {
14963                 fsb->m = smb;
14964                 fsb->off = soff;
14965                 if (smb) {
14966                         /*
14967                          * Save off the size of the mbuf. We do
14968                          * this so that we can recognize when it
14969                          * has been trimmed by sbcut() as acks
14970                          * come in.
14971                          */
14972                         fsb->o_m_len = smb->m_len;
14973                 } else {
14974                         /*
14975                          * This is the case where the next mbuf went to NULL. This
14976                          * means with this copy we have sent everything in the sb.
14977                          * In theory we could clear the fast_output flag, but lets
14978                          * not since its possible that we could get more added
14979                          * and acks that call the extend function which would let
14980                          * us send more.
14981                          */
14982                         fsb->o_m_len = 0;
14983                 }
14984         }
14985         return (top);
14986 nospace:
14987         if (top)
14988                 m_freem(top);
14989         return (NULL);
14990
14991 }
14992
14993 /*
14994  * This is a copy of m_copym(), taking the TSO segment size/limit
14995  * constraints into account, and advancing the sndptr as it goes.
14996  */
14997 static struct mbuf *
14998 rack_fo_m_copym(struct tcp_rack *rack, int32_t *plen,
14999                 int32_t seglimit, int32_t segsize, struct mbuf **s_mb, int *s_soff)
15000 {
15001         struct mbuf *m, *n;
15002         int32_t soff;
15003
15004         soff = rack->r_ctl.fsb.off;
15005         m = rack->r_ctl.fsb.m;
15006         if (rack->r_ctl.fsb.o_m_len != m->m_len) {
15007                 /*
15008                  * The mbuf had the front of it chopped off by an ack
15009                  * we need to adjust the soff/off by that difference.
15010                  */
15011                 uint32_t delta;
15012
15013                 delta = rack->r_ctl.fsb.o_m_len - m->m_len;
15014                 soff -= delta;
15015         }
15016         KASSERT(soff >= 0, ("%s, negative off %d", __FUNCTION__, soff));
15017         KASSERT(*plen >= 0, ("%s, negative len %d", __FUNCTION__, *plen));
15018         KASSERT(soff < m->m_len, ("%s rack:%p len:%u m:%p m->m_len:%u < off?",
15019                                  __FUNCTION__,
15020                                  rack, *plen, m, m->m_len));
15021         /* Save off the right location before we copy and advance */
15022         *s_soff = soff;
15023         *s_mb = rack->r_ctl.fsb.m;
15024         n = rack_fo_base_copym(m, soff, plen,
15025                                &rack->r_ctl.fsb,
15026                                seglimit, segsize);
15027         return (n);
15028 }
15029
15030 static int
15031 rack_fast_rsm_output(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm,
15032                      uint64_t ts_val, uint32_t cts, uint32_t ms_cts, struct timeval *tv, int len)
15033 {
15034         /*
15035          * Enter the fast retransmit path. We are given that a sched_pin is
15036          * in place (if accounting is compliled in) and the cycle count taken
15037          * at the entry is in the ts_val. The concept her is that the rsm
15038          * now holds the mbuf offsets and such so we can directly transmit
15039          * without a lot of overhead, the len field is already set for
15040          * us to prohibit us from sending too much (usually its 1MSS).
15041          */
15042         struct ip *ip = NULL;
15043         struct udphdr *udp = NULL;
15044         struct tcphdr *th = NULL;
15045         struct mbuf *m = NULL;
15046         struct inpcb *inp;
15047         uint8_t *cpto;
15048         struct tcp_log_buffer *lgb;
15049 #ifdef TCP_ACCOUNTING
15050         uint64_t crtsc;
15051         int cnt_thru = 1;
15052 #endif
15053         int doing_tlp = 0;
15054         struct tcpopt to;
15055         u_char opt[TCP_MAXOLEN];
15056         uint32_t hdrlen, optlen;
15057         int32_t slot, segsiz, max_val, tso = 0, error, flags, ulen = 0;
15058         uint32_t us_cts;
15059         uint32_t if_hw_tsomaxsegcount = 0, startseq;
15060         uint32_t if_hw_tsomaxsegsize;
15061 #ifdef INET6
15062         struct ip6_hdr *ip6 = NULL;
15063
15064         if (rack->r_is_v6) {
15065                 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
15066                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
15067         } else
15068 #endif                          /* INET6 */
15069         {
15070                 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
15071                 hdrlen = sizeof(struct tcpiphdr);
15072         }
15073         if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
15074                 goto failed;
15075         }
15076         if (rsm->r_flags & RACK_TLP)
15077                 doing_tlp = 1;
15078         startseq = rsm->r_start;
15079         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
15080         inp = rack->rc_inp;
15081         to.to_flags = 0;
15082         flags = tcp_outflags[tp->t_state];
15083         if (flags & (TH_SYN|TH_RST)) {
15084                 goto failed;
15085         }
15086         if (rsm->r_flags & RACK_HAS_FIN) {
15087                 /* We can't send a FIN here */
15088                 goto failed;
15089         }
15090         if (flags & TH_FIN) {
15091                 /* We never send a FIN */
15092                 flags &= ~TH_FIN;
15093         }
15094         if (tp->t_flags & TF_RCVD_TSTMP) {
15095                 to.to_tsval = ms_cts + tp->ts_offset;
15096                 to.to_tsecr = tp->ts_recent;
15097                 to.to_flags = TOF_TS;
15098         }
15099         optlen = tcp_addoptions(&to, opt);
15100         hdrlen += optlen;
15101         udp = rack->r_ctl.fsb.udp;
15102         if (udp)
15103                 hdrlen += sizeof(struct udphdr);
15104         if (rack->r_ctl.rc_pace_max_segs)
15105                 max_val = rack->r_ctl.rc_pace_max_segs;
15106         else if (rack->rc_user_set_max_segs)
15107                 max_val = rack->rc_user_set_max_segs * segsiz;
15108         else
15109                 max_val = len;
15110         if ((tp->t_flags & TF_TSO) &&
15111             V_tcp_do_tso &&
15112             (len > segsiz) &&
15113             (tp->t_port == 0))
15114                 tso = 1;
15115 #ifdef INET6
15116         if (MHLEN < hdrlen + max_linkhdr)
15117                 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
15118         else
15119 #endif
15120                 m = m_gethdr(M_NOWAIT, MT_DATA);
15121         if (m == NULL)
15122                 goto failed;
15123         m->m_data += max_linkhdr;
15124         m->m_len = hdrlen;
15125         th = rack->r_ctl.fsb.th;
15126         /* Establish the len to send */
15127         if (len > max_val)
15128                 len = max_val;
15129         if ((tso) && (len + optlen > tp->t_maxseg)) {
15130                 uint32_t if_hw_tsomax;
15131                 int32_t max_len;
15132
15133                 /* extract TSO information */
15134                 if_hw_tsomax = tp->t_tsomax;
15135                 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
15136                 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
15137                 /*
15138                  * Check if we should limit by maximum payload
15139                  * length:
15140                  */
15141                 if (if_hw_tsomax != 0) {
15142                         /* compute maximum TSO length */
15143                         max_len = (if_hw_tsomax - hdrlen -
15144                                    max_linkhdr);
15145                         if (max_len <= 0) {
15146                                 goto failed;
15147                         } else if (len > max_len) {
15148                                 len = max_len;
15149                         }
15150                 }
15151                 if (len <= segsiz) {
15152                         /*
15153                          * In case there are too many small fragments don't
15154                          * use TSO:
15155                          */
15156                         tso = 0;
15157                 }
15158         } else {
15159                 tso = 0;
15160         }
15161         if ((tso == 0) && (len > segsiz))
15162                 len = segsiz;
15163         us_cts = tcp_get_usecs(tv);
15164         if ((len == 0) ||
15165             (len <= MHLEN - hdrlen - max_linkhdr)) {
15166                 goto failed;
15167         }
15168         th->th_seq = htonl(rsm->r_start);
15169         th->th_ack = htonl(tp->rcv_nxt);
15170         if(rsm->r_flags & RACK_HAD_PUSH)
15171                 flags |= TH_PUSH;
15172         th->th_flags = flags;
15173         th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
15174         if (th->th_win == 0) {
15175                 tp->t_sndzerowin++;
15176                 tp->t_flags |= TF_RXWIN0SENT;
15177         } else
15178                 tp->t_flags &= ~TF_RXWIN0SENT;
15179         if (rsm->r_flags & RACK_TLP) {
15180                 /*
15181                  * TLP should not count in retran count, but
15182                  * in its own bin
15183                  */
15184                 counter_u64_add(rack_tlp_retran, 1);
15185                 counter_u64_add(rack_tlp_retran_bytes, len);
15186         } else {
15187                 tp->t_sndrexmitpack++;
15188                 KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
15189                 KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
15190         }
15191 #ifdef STATS
15192         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
15193                                  len);
15194 #endif
15195         if (rsm->m == NULL)
15196                 goto failed;
15197         if (rsm->orig_m_len != rsm->m->m_len) {
15198                 /* Fix up the orig_m_len and possibly the mbuf offset */
15199                 rack_adjust_orig_mlen(rsm);
15200         }
15201         m->m_next = rack_fo_base_copym(rsm->m, rsm->soff, &len, NULL, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize);
15202         if (len <= segsiz) {
15203                 /*
15204                  * Must have ran out of mbufs for the copy
15205                  * shorten it to no longer need tso. Lets
15206                  * not put on sendalot since we are low on
15207                  * mbufs.
15208                  */
15209                 tso = 0;
15210         }
15211         if ((m->m_next == NULL) || (len <= 0)){
15212                 goto failed;
15213         }
15214         if (udp) {
15215                 if (rack->r_is_v6)
15216                         ulen = hdrlen + len - sizeof(struct ip6_hdr);
15217                 else
15218                         ulen = hdrlen + len - sizeof(struct ip);
15219                 udp->uh_ulen = htons(ulen);
15220         }
15221         m->m_pkthdr.rcvif = (struct ifnet *)0;
15222         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
15223 #ifdef INET6
15224         if (rack->r_is_v6) {
15225                 if (tp->t_port) {
15226                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
15227                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15228                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
15229                         th->th_sum = htons(0);
15230                         UDPSTAT_INC(udps_opackets);
15231                 } else {
15232                         m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
15233                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15234                         th->th_sum = in6_cksum_pseudo(ip6,
15235                                                       sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
15236                                                       0);
15237                 }
15238         }
15239 #endif
15240 #if defined(INET6) && defined(INET)
15241         else
15242 #endif
15243 #ifdef INET
15244         {
15245                 if (tp->t_port) {
15246                         m->m_pkthdr.csum_flags = CSUM_UDP;
15247                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15248                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
15249                                                 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
15250                         th->th_sum = htons(0);
15251                         UDPSTAT_INC(udps_opackets);
15252                 } else {
15253                         m->m_pkthdr.csum_flags = CSUM_TCP;
15254                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15255                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
15256                                                ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
15257                                                                         IPPROTO_TCP + len + optlen));
15258                 }
15259                 /* IP version must be set here for ipv4/ipv6 checking later */
15260                 KASSERT(ip->ip_v == IPVERSION,
15261                         ("%s: IP version incorrect: %d", __func__, ip->ip_v));
15262         }
15263 #endif
15264         if (tso) {
15265                 KASSERT(len > tp->t_maxseg - optlen,
15266                         ("%s: len <= tso_segsz tp:%p", __func__, tp));
15267                 m->m_pkthdr.csum_flags |= CSUM_TSO;
15268                 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
15269         }
15270 #ifdef INET6
15271         if (rack->r_is_v6) {
15272                 ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
15273                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
15274                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
15275                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15276                 else
15277                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15278         }
15279 #endif
15280 #if defined(INET) && defined(INET6)
15281         else
15282 #endif
15283 #ifdef INET
15284         {
15285                 ip->ip_len = htons(m->m_pkthdr.len);
15286                 ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
15287                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
15288                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15289                         if (tp->t_port == 0 || len < V_tcp_minmss) {
15290                                 ip->ip_off |= htons(IP_DF);
15291                         }
15292                 } else {
15293                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15294                 }
15295         }
15296 #endif
15297         /* Time to copy in our header */
15298         cpto = mtod(m, uint8_t *);
15299         memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
15300         th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
15301         if (optlen) {
15302                 bcopy(opt, th + 1, optlen);
15303                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
15304         } else {
15305                 th->th_off = sizeof(struct tcphdr) >> 2;
15306         }
15307         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
15308                 union tcp_log_stackspecific log;
15309
15310                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15311                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
15312                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
15313                 if (rack->rack_no_prr)
15314                         log.u_bbr.flex1 = 0;
15315                 else
15316                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
15317                 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
15318                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
15319                 log.u_bbr.flex4 = max_val;
15320                 log.u_bbr.flex5 = 0;
15321                 /* Save off the early/late values */
15322                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
15323                 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
15324                 log.u_bbr.bw_inuse = rack_get_bw(rack);
15325                 log.u_bbr.flex8 = 1;
15326                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
15327                 log.u_bbr.flex7 = 55;
15328                 log.u_bbr.pkts_out = tp->t_maxseg;
15329                 log.u_bbr.timeStamp = cts;
15330                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15331                 log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
15332                 log.u_bbr.delivered = 0;
15333                 lgb = tcp_log_event_(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
15334                                      len, &log, false, NULL, NULL, 0, tv);
15335         } else
15336                 lgb = NULL;
15337 #ifdef INET6
15338         if (rack->r_is_v6) {
15339                 error = ip6_output(m, NULL,
15340                                    &inp->inp_route6,
15341                                    0, NULL, NULL, inp);
15342         }
15343 #endif
15344 #if defined(INET) && defined(INET6)
15345         else
15346 #endif
15347 #ifdef INET
15348         {
15349                 error = ip_output(m, NULL,
15350                                   &inp->inp_route,
15351                                   0, 0, inp);
15352         }
15353 #endif
15354         m = NULL;
15355         if (lgb) {
15356                 lgb->tlb_errno = error;
15357                 lgb = NULL;
15358         }
15359         if (error) {
15360                 goto failed;
15361         }
15362         rack_log_output(tp, &to, len, rsm->r_start, flags, error, rack_to_usec_ts(tv),
15363                         rsm, RACK_SENT_FP, rsm->m, rsm->soff);
15364         if (doing_tlp && (rack->fast_rsm_hack == 0)) {
15365                 rack->rc_tlp_in_progress = 1;
15366                 rack->r_ctl.rc_tlp_cnt_out++;
15367         }
15368         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
15369         rack->forced_ack = 0;   /* If we send something zap the FA flag */
15370         if (IN_FASTRECOVERY(tp->t_flags) && rsm)
15371                 rack->r_ctl.retran_during_recovery += len;
15372         {
15373                 int idx;
15374
15375                 idx = (len / segsiz) + 3;
15376                 if (idx >= TCP_MSS_ACCT_ATIMER)
15377                         counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
15378                 else
15379                         counter_u64_add(rack_out_size[idx], 1);
15380         }
15381         if (tp->t_rtttime == 0) {
15382                 tp->t_rtttime = ticks;
15383                 tp->t_rtseq = startseq;
15384                 KMOD_TCPSTAT_INC(tcps_segstimed);
15385         }
15386         counter_u64_add(rack_fto_rsm_send, 1);
15387         if (error && (error == ENOBUFS)) {
15388                 slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
15389                 if (rack->rc_enobuf < 0x7f)
15390                         rack->rc_enobuf++;
15391                 if (slot < (10 * HPTS_USEC_IN_MSEC))
15392                         slot = 10 * HPTS_USEC_IN_MSEC;
15393         } else
15394                 slot = rack_get_pacing_delay(rack, tp, len, NULL, segsiz);
15395         if ((slot == 0) ||
15396             (rack->rc_always_pace == 0) ||
15397             (rack->r_rr_config == 1)) {
15398                 /*
15399                  * We have no pacing set or we
15400                  * are using old-style rack or
15401                  * we are overriden to use the old 1ms pacing.
15402                  */
15403                 slot = rack->r_ctl.rc_min_to;
15404         }
15405         rack_start_hpts_timer(rack, tp, cts, slot, len, 0);
15406         if (rack->r_must_retran) {
15407                 rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
15408                 if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
15409                         /*
15410                          * We have retransmitted all we need.
15411                          */
15412                         rack->r_must_retran = 0;
15413                         rack->r_ctl.rc_out_at_rto = 0;
15414                 }
15415         }
15416 #ifdef TCP_ACCOUNTING
15417         crtsc = get_cyclecount();
15418         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15419                 tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
15420         }
15421         counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], cnt_thru);
15422         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15423                 tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
15424         }
15425         counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
15426         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15427                 tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((len + segsiz - 1) / segsiz);
15428         }
15429         counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((len + segsiz - 1) / segsiz));
15430         sched_unpin();
15431 #endif
15432         return (0);
15433 failed:
15434         if (m)
15435                 m_free(m);
15436         return (-1);
15437 }
15438
15439 static void
15440 rack_sndbuf_autoscale(struct tcp_rack *rack)
15441 {
15442         /*
15443          * Automatic sizing of send socket buffer.  Often the send buffer
15444          * size is not optimally adjusted to the actual network conditions
15445          * at hand (delay bandwidth product).  Setting the buffer size too
15446          * small limits throughput on links with high bandwidth and high
15447          * delay (eg. trans-continental/oceanic links).  Setting the
15448          * buffer size too big consumes too much real kernel memory,
15449          * especially with many connections on busy servers.
15450          *
15451          * The criteria to step up the send buffer one notch are:
15452          *  1. receive window of remote host is larger than send buffer
15453          *     (with a fudge factor of 5/4th);
15454          *  2. send buffer is filled to 7/8th with data (so we actually
15455          *     have data to make use of it);
15456          *  3. send buffer fill has not hit maximal automatic size;
15457          *  4. our send window (slow start and cogestion controlled) is
15458          *     larger than sent but unacknowledged data in send buffer.
15459          *
15460          * Note that the rack version moves things much faster since
15461          * we want to avoid hitting cache lines in the rack_fast_output()
15462          * path so this is called much less often and thus moves
15463          * the SB forward by a percentage.
15464          */
15465         struct socket *so;
15466         struct tcpcb *tp;
15467         uint32_t sendwin, scaleup;
15468
15469         tp = rack->rc_tp;
15470         so = rack->rc_inp->inp_socket;
15471         sendwin = min(rack->r_ctl.cwnd_to_use, tp->snd_wnd);
15472         if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
15473                 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
15474                     sbused(&so->so_snd) >=
15475                     (so->so_snd.sb_hiwat / 8 * 7) &&
15476                     sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
15477                     sendwin >= (sbused(&so->so_snd) -
15478                     (tp->snd_nxt - tp->snd_una))) {
15479                         if (rack_autosndbuf_inc)
15480                                 scaleup = (rack_autosndbuf_inc * so->so_snd.sb_hiwat) / 100;
15481                         else
15482                                 scaleup = V_tcp_autosndbuf_inc;
15483                         if (scaleup < V_tcp_autosndbuf_inc)
15484                                 scaleup = V_tcp_autosndbuf_inc;
15485                         scaleup += so->so_snd.sb_hiwat;
15486                         if (scaleup > V_tcp_autosndbuf_max)
15487                                 scaleup = V_tcp_autosndbuf_max;
15488                         if (!sbreserve_locked(&so->so_snd, scaleup, so, curthread))
15489                                 so->so_snd.sb_flags &= ~SB_AUTOSIZE;
15490                 }
15491         }
15492 }
15493
15494 static int
15495 rack_fast_output(struct tcpcb *tp, struct tcp_rack *rack, uint64_t ts_val,
15496                  uint32_t cts, uint32_t ms_cts, struct timeval *tv, long tot_len, int *send_err)
15497 {
15498         /*
15499          * Enter to do fast output. We are given that the sched_pin is
15500          * in place (if accounting is compiled in) and the cycle count taken
15501          * at entry is in place in ts_val. The idea here is that
15502          * we know how many more bytes needs to be sent (presumably either
15503          * during pacing or to fill the cwnd and that was greater than
15504          * the max-burst). We have how much to send and all the info we
15505          * need to just send.
15506          */
15507         struct ip *ip = NULL;
15508         struct udphdr *udp = NULL;
15509         struct tcphdr *th = NULL;
15510         struct mbuf *m, *s_mb;
15511         struct inpcb *inp;
15512         uint8_t *cpto;
15513         struct tcp_log_buffer *lgb;
15514 #ifdef TCP_ACCOUNTING
15515         uint64_t crtsc;
15516 #endif
15517         struct tcpopt to;
15518         u_char opt[TCP_MAXOLEN];
15519         uint32_t hdrlen, optlen;
15520         int cnt_thru = 1;
15521         int32_t slot, segsiz, len, max_val, tso = 0, sb_offset, error, flags, ulen = 0;
15522         uint32_t us_cts, s_soff;
15523         uint32_t if_hw_tsomaxsegcount = 0, startseq;
15524         uint32_t if_hw_tsomaxsegsize;
15525         uint16_t add_flag = RACK_SENT_FP;
15526 #ifdef INET6
15527         struct ip6_hdr *ip6 = NULL;
15528
15529         if (rack->r_is_v6) {
15530                 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
15531                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
15532         } else
15533 #endif                          /* INET6 */
15534         {
15535                 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
15536                 hdrlen = sizeof(struct tcpiphdr);
15537         }
15538         if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
15539                 m = NULL;
15540                 goto failed;
15541         }
15542         startseq = tp->snd_max;
15543         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
15544         inp = rack->rc_inp;
15545         len = rack->r_ctl.fsb.left_to_send;
15546         to.to_flags = 0;
15547         flags = rack->r_ctl.fsb.tcp_flags;
15548         if (tp->t_flags & TF_RCVD_TSTMP) {
15549                 to.to_tsval = ms_cts + tp->ts_offset;
15550                 to.to_tsecr = tp->ts_recent;
15551                 to.to_flags = TOF_TS;
15552         }
15553         optlen = tcp_addoptions(&to, opt);
15554         hdrlen += optlen;
15555         udp = rack->r_ctl.fsb.udp;
15556         if (udp)
15557                 hdrlen += sizeof(struct udphdr);
15558         if (rack->r_ctl.rc_pace_max_segs)
15559                 max_val = rack->r_ctl.rc_pace_max_segs;
15560         else if (rack->rc_user_set_max_segs)
15561                 max_val = rack->rc_user_set_max_segs * segsiz;
15562         else
15563                 max_val = len;
15564         if ((tp->t_flags & TF_TSO) &&
15565             V_tcp_do_tso &&
15566             (len > segsiz) &&
15567             (tp->t_port == 0))
15568                 tso = 1;
15569 again:
15570 #ifdef INET6
15571         if (MHLEN < hdrlen + max_linkhdr)
15572                 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
15573         else
15574 #endif
15575                 m = m_gethdr(M_NOWAIT, MT_DATA);
15576         if (m == NULL)
15577                 goto failed;
15578         m->m_data += max_linkhdr;
15579         m->m_len = hdrlen;
15580         th = rack->r_ctl.fsb.th;
15581         /* Establish the len to send */
15582         if (len > max_val)
15583                 len = max_val;
15584         if ((tso) && (len + optlen > tp->t_maxseg)) {
15585                 uint32_t if_hw_tsomax;
15586                 int32_t max_len;
15587
15588                 /* extract TSO information */
15589                 if_hw_tsomax = tp->t_tsomax;
15590                 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
15591                 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
15592                 /*
15593                  * Check if we should limit by maximum payload
15594                  * length:
15595                  */
15596                 if (if_hw_tsomax != 0) {
15597                         /* compute maximum TSO length */
15598                         max_len = (if_hw_tsomax - hdrlen -
15599                                    max_linkhdr);
15600                         if (max_len <= 0) {
15601                                 goto failed;
15602                         } else if (len > max_len) {
15603                                 len = max_len;
15604                         }
15605                 }
15606                 if (len <= segsiz) {
15607                         /*
15608                          * In case there are too many small fragments don't
15609                          * use TSO:
15610                          */
15611                         tso = 0;
15612                 }
15613         } else {
15614                 tso = 0;
15615         }
15616         if ((tso == 0) && (len > segsiz))
15617                 len = segsiz;
15618         us_cts = tcp_get_usecs(tv);
15619         if ((len == 0) ||
15620             (len <= MHLEN - hdrlen - max_linkhdr)) {
15621                 goto failed;
15622         }
15623         sb_offset = tp->snd_max - tp->snd_una;
15624         th->th_seq = htonl(tp->snd_max);
15625         th->th_ack = htonl(tp->rcv_nxt);
15626         th->th_flags = flags;
15627         th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
15628         if (th->th_win == 0) {
15629                 tp->t_sndzerowin++;
15630                 tp->t_flags |= TF_RXWIN0SENT;
15631         } else
15632                 tp->t_flags &= ~TF_RXWIN0SENT;
15633         tp->snd_up = tp->snd_una;       /* drag it along, its deprecated */
15634         KMOD_TCPSTAT_INC(tcps_sndpack);
15635         KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
15636 #ifdef STATS
15637         stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
15638                                  len);
15639 #endif
15640         if (rack->r_ctl.fsb.m == NULL)
15641                 goto failed;
15642
15643         /* s_mb and s_soff are saved for rack_log_output */
15644         m->m_next = rack_fo_m_copym(rack, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, &s_mb, &s_soff);
15645         if (len <= segsiz) {
15646                 /*
15647                  * Must have ran out of mbufs for the copy
15648                  * shorten it to no longer need tso. Lets
15649                  * not put on sendalot since we are low on
15650                  * mbufs.
15651                  */
15652                 tso = 0;
15653         }
15654         if (rack->r_ctl.fsb.rfo_apply_push &&
15655             (len == rack->r_ctl.fsb.left_to_send)) {
15656                 th->th_flags |= TH_PUSH;
15657                 add_flag |= RACK_HAD_PUSH;
15658         }
15659         if ((m->m_next == NULL) || (len <= 0)){
15660                 goto failed;
15661         }
15662         if (udp) {
15663                 if (rack->r_is_v6)
15664                         ulen = hdrlen + len - sizeof(struct ip6_hdr);
15665                 else
15666                         ulen = hdrlen + len - sizeof(struct ip);
15667                 udp->uh_ulen = htons(ulen);
15668         }
15669         m->m_pkthdr.rcvif = (struct ifnet *)0;
15670         if (tp->t_state == TCPS_ESTABLISHED &&
15671             (tp->t_flags2 & TF2_ECN_PERMIT)) {
15672                 /*
15673                  * If the peer has ECN, mark data packets with ECN capable
15674                  * transmission (ECT). Ignore pure ack packets,
15675                  * retransmissions.
15676                  */
15677                 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max)) {
15678 #ifdef INET6
15679                         if (rack->r_is_v6)
15680                                 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
15681                         else
15682 #endif
15683                                 ip->ip_tos |= IPTOS_ECN_ECT0;
15684                         KMOD_TCPSTAT_INC(tcps_ecn_ect0);
15685                         /*
15686                          * Reply with proper ECN notifications.
15687                          * Only set CWR on new data segments.
15688                          */
15689                         if (tp->t_flags2 & TF2_ECN_SND_CWR) {
15690                                 flags |= TH_CWR;
15691                                 tp->t_flags2 &= ~TF2_ECN_SND_CWR;
15692                         }
15693                 }
15694                 if (tp->t_flags2 & TF2_ECN_SND_ECE)
15695                         flags |= TH_ECE;
15696         }
15697         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
15698 #ifdef INET6
15699         if (rack->r_is_v6) {
15700                 if (tp->t_port) {
15701                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
15702                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15703                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
15704                         th->th_sum = htons(0);
15705                         UDPSTAT_INC(udps_opackets);
15706                 } else {
15707                         m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
15708                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15709                         th->th_sum = in6_cksum_pseudo(ip6,
15710                                                       sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
15711                                                       0);
15712                 }
15713         }
15714 #endif
15715 #if defined(INET6) && defined(INET)
15716         else
15717 #endif
15718 #ifdef INET
15719         {
15720                 if (tp->t_port) {
15721                         m->m_pkthdr.csum_flags = CSUM_UDP;
15722                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15723                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
15724                                                 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
15725                         th->th_sum = htons(0);
15726                         UDPSTAT_INC(udps_opackets);
15727                 } else {
15728                         m->m_pkthdr.csum_flags = CSUM_TCP;
15729                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15730                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
15731                                                ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
15732                                                                         IPPROTO_TCP + len + optlen));
15733                 }
15734                 /* IP version must be set here for ipv4/ipv6 checking later */
15735                 KASSERT(ip->ip_v == IPVERSION,
15736                         ("%s: IP version incorrect: %d", __func__, ip->ip_v));
15737         }
15738 #endif
15739         if (tso) {
15740                 KASSERT(len > tp->t_maxseg - optlen,
15741                         ("%s: len <= tso_segsz tp:%p", __func__, tp));
15742                 m->m_pkthdr.csum_flags |= CSUM_TSO;
15743                 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
15744         }
15745 #ifdef INET6
15746         if (rack->r_is_v6) {
15747                 ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
15748                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
15749                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
15750                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15751                 else
15752                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15753         }
15754 #endif
15755 #if defined(INET) && defined(INET6)
15756         else
15757 #endif
15758 #ifdef INET
15759         {
15760                 ip->ip_len = htons(m->m_pkthdr.len);
15761                 ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
15762                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
15763                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15764                         if (tp->t_port == 0 || len < V_tcp_minmss) {
15765                                 ip->ip_off |= htons(IP_DF);
15766                         }
15767                 } else {
15768                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15769                 }
15770         }
15771 #endif
15772         /* Time to copy in our header */
15773         cpto = mtod(m, uint8_t *);
15774         memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
15775         th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
15776         if (optlen) {
15777                 bcopy(opt, th + 1, optlen);
15778                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
15779         } else {
15780                 th->th_off = sizeof(struct tcphdr) >> 2;
15781         }
15782         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
15783                 union tcp_log_stackspecific log;
15784
15785                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15786                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
15787                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
15788                 if (rack->rack_no_prr)
15789                         log.u_bbr.flex1 = 0;
15790                 else
15791                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
15792                 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
15793                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
15794                 log.u_bbr.flex4 = max_val;
15795                 log.u_bbr.flex5 = 0;
15796                 /* Save off the early/late values */
15797                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
15798                 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
15799                 log.u_bbr.bw_inuse = rack_get_bw(rack);
15800                 log.u_bbr.flex8 = 0;
15801                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
15802                 log.u_bbr.flex7 = 44;
15803                 log.u_bbr.pkts_out = tp->t_maxseg;
15804                 log.u_bbr.timeStamp = cts;
15805                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15806                 log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
15807                 log.u_bbr.delivered = 0;
15808                 lgb = tcp_log_event_(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
15809                                      len, &log, false, NULL, NULL, 0, tv);
15810         } else
15811                 lgb = NULL;
15812 #ifdef INET6
15813         if (rack->r_is_v6) {
15814                 error = ip6_output(m, NULL,
15815                                    &inp->inp_route6,
15816                                    0, NULL, NULL, inp);
15817         }
15818 #endif
15819 #if defined(INET) && defined(INET6)
15820         else
15821 #endif
15822 #ifdef INET
15823         {
15824                 error = ip_output(m, NULL,
15825                                   &inp->inp_route,
15826                                   0, 0, inp);
15827         }
15828 #endif
15829         if (lgb) {
15830                 lgb->tlb_errno = error;
15831                 lgb = NULL;
15832         }
15833         if (error) {
15834                 *send_err = error;
15835                 m = NULL;
15836                 goto failed;
15837         }
15838         rack_log_output(tp, &to, len, tp->snd_max, flags, error, rack_to_usec_ts(tv),
15839                         NULL, add_flag, s_mb, s_soff);
15840         m = NULL;
15841         if (tp->snd_una == tp->snd_max) {
15842                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
15843                 rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
15844                 tp->t_acktime = ticks;
15845         }
15846         rack->forced_ack = 0;   /* If we send something zap the FA flag */
15847         tot_len += len;
15848         if ((tp->t_flags & TF_GPUTINPROG) == 0)
15849                 rack_start_gp_measurement(tp, rack, tp->snd_max, sb_offset);
15850         tp->snd_max += len;
15851         tp->snd_nxt = tp->snd_max;
15852         {
15853                 int idx;
15854
15855                 idx = (len / segsiz) + 3;
15856                 if (idx >= TCP_MSS_ACCT_ATIMER)
15857                         counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
15858                 else
15859                         counter_u64_add(rack_out_size[idx], 1);
15860         }
15861         if (len <= rack->r_ctl.fsb.left_to_send)
15862                 rack->r_ctl.fsb.left_to_send -= len;
15863         else
15864                 rack->r_ctl.fsb.left_to_send = 0;
15865         if (rack->r_ctl.fsb.left_to_send < segsiz) {
15866                 rack->r_fast_output = 0;
15867                 rack->r_ctl.fsb.left_to_send = 0;
15868                 /* At the end of fast_output scale up the sb */
15869                 SOCKBUF_LOCK(&rack->rc_inp->inp_socket->so_snd);
15870                 rack_sndbuf_autoscale(rack);
15871                 SOCKBUF_UNLOCK(&rack->rc_inp->inp_socket->so_snd);
15872         }
15873         if (tp->t_rtttime == 0) {
15874                 tp->t_rtttime = ticks;
15875                 tp->t_rtseq = startseq;
15876                 KMOD_TCPSTAT_INC(tcps_segstimed);
15877         }
15878         if ((rack->r_ctl.fsb.left_to_send >= segsiz) &&
15879             (max_val > len) &&
15880             (tso == 0)) {
15881                 max_val -= len;
15882                 len = segsiz;
15883                 th = rack->r_ctl.fsb.th;
15884                 cnt_thru++;
15885                 goto again;
15886         }
15887         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
15888         counter_u64_add(rack_fto_send, 1);
15889         slot = rack_get_pacing_delay(rack, tp, tot_len, NULL, segsiz);
15890         rack_start_hpts_timer(rack, tp, cts, slot, tot_len, 0);
15891 #ifdef TCP_ACCOUNTING
15892         crtsc = get_cyclecount();
15893         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15894                 tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
15895         }
15896         counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], cnt_thru);
15897         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15898                 tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
15899         }
15900         counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
15901         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15902                 tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len + segsiz - 1) / segsiz);
15903         }
15904         counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len + segsiz - 1) / segsiz));
15905         sched_unpin();
15906 #endif
15907         return (0);
15908 failed:
15909         if (m)
15910                 m_free(m);
15911         rack->r_fast_output = 0;
15912         return (-1);
15913 }
15914
15915 static int
15916 rack_output(struct tcpcb *tp)
15917 {
15918         struct socket *so;
15919         uint32_t recwin;
15920         uint32_t sb_offset, s_moff = 0;
15921         int32_t len, flags, error = 0;
15922         struct mbuf *m, *s_mb = NULL;
15923         struct mbuf *mb;
15924         uint32_t if_hw_tsomaxsegcount = 0;
15925         uint32_t if_hw_tsomaxsegsize;
15926         int32_t segsiz, minseg;
15927         long tot_len_this_send = 0;
15928 #ifdef INET
15929         struct ip *ip = NULL;
15930 #endif
15931 #ifdef TCPDEBUG
15932         struct ipovly *ipov = NULL;
15933 #endif
15934         struct udphdr *udp = NULL;
15935         struct tcp_rack *rack;
15936         struct tcphdr *th;
15937         uint8_t pass = 0;
15938         uint8_t mark = 0;
15939         uint8_t wanted_cookie = 0;
15940         u_char opt[TCP_MAXOLEN];
15941         unsigned ipoptlen, optlen, hdrlen, ulen=0;
15942         uint32_t rack_seq;
15943
15944 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
15945         unsigned ipsec_optlen = 0;
15946
15947 #endif
15948         int32_t idle, sendalot;
15949         int32_t sub_from_prr = 0;
15950         volatile int32_t sack_rxmit;
15951         struct rack_sendmap *rsm = NULL;
15952         int32_t tso, mtu;
15953         struct tcpopt to;
15954         int32_t slot = 0;
15955         int32_t sup_rack = 0;
15956         uint32_t cts, ms_cts, delayed, early;
15957         uint16_t add_flag = RACK_SENT_SP;
15958         uint8_t hpts_calling,  doing_tlp = 0;
15959         uint32_t cwnd_to_use, pace_max_seg;
15960         int32_t do_a_prefetch = 0;
15961         int32_t prefetch_rsm = 0;
15962         int32_t orig_len = 0;
15963         struct timeval tv;
15964         int32_t prefetch_so_done = 0;
15965         struct tcp_log_buffer *lgb;
15966         struct inpcb *inp;
15967         struct sockbuf *sb;
15968         uint64_t ts_val = 0;
15969 #ifdef TCP_ACCOUNTING
15970         uint64_t crtsc;
15971 #endif
15972 #ifdef INET6
15973         struct ip6_hdr *ip6 = NULL;
15974         int32_t isipv6;
15975 #endif
15976         uint8_t filled_all = 0;
15977         bool hw_tls = false;
15978
15979         /* setup and take the cache hits here */
15980         rack = (struct tcp_rack *)tp->t_fb_ptr;
15981 #ifdef TCP_ACCOUNTING
15982         sched_pin();
15983         ts_val = get_cyclecount();
15984 #endif
15985         hpts_calling = rack->rc_inp->inp_hpts_calls;
15986         NET_EPOCH_ASSERT();
15987         INP_WLOCK_ASSERT(rack->rc_inp);
15988 #ifdef TCP_OFFLOAD
15989         if (tp->t_flags & TF_TOE) {
15990 #ifdef TCP_ACCOUNTING
15991                 sched_unpin();
15992 #endif
15993                 return (tcp_offload_output(tp));
15994         }
15995 #endif
15996         /*
15997          * For TFO connections in SYN_RECEIVED, only allow the initial
15998          * SYN|ACK and those sent by the retransmit timer.
15999          */
16000         if (IS_FASTOPEN(tp->t_flags) &&
16001             (tp->t_state == TCPS_SYN_RECEIVED) &&
16002             SEQ_GT(tp->snd_max, tp->snd_una) &&    /* initial SYN|ACK sent */
16003             (rack->r_ctl.rc_resend == NULL)) {         /* not a retransmit */
16004 #ifdef TCP_ACCOUNTING
16005                 sched_unpin();
16006 #endif
16007                 return (0);
16008         }
16009 #ifdef INET6
16010         if (rack->r_state) {
16011                 /* Use the cache line loaded if possible */
16012                 isipv6 = rack->r_is_v6;
16013         } else {
16014                 isipv6 = (rack->rc_inp->inp_vflag & INP_IPV6) != 0;
16015         }
16016 #endif
16017         early = 0;
16018         cts = tcp_get_usecs(&tv);
16019         ms_cts = tcp_tv_to_mssectick(&tv);
16020         if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
16021             rack->rc_inp->inp_in_hpts) {
16022                 /*
16023                  * We are on the hpts for some timer but not hptsi output.
16024                  * Remove from the hpts unconditionally.
16025                  */
16026                 rack_timer_cancel(tp, rack, cts, __LINE__);
16027         }
16028         /* Are we pacing and late? */
16029         if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16030             TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to)) {
16031                 /* We are delayed */
16032                 delayed = cts - rack->r_ctl.rc_last_output_to;
16033         } else {
16034                 delayed = 0;
16035         }
16036         /* Do the timers, which may override the pacer */
16037         if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
16038                 if (rack_process_timers(tp, rack, cts, hpts_calling)) {
16039                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1);
16040 #ifdef TCP_ACCOUNTING
16041                         sched_unpin();
16042 #endif
16043                         return (0);
16044                 }
16045         }
16046         if (rack->rc_in_persist) {
16047                 if (rack->rc_inp->inp_in_hpts == 0) {
16048                         /* Timer is not running */
16049                         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
16050                 }
16051 #ifdef TCP_ACCOUNTING
16052                 sched_unpin();
16053 #endif
16054                 return (0);
16055         }
16056         if ((rack->r_timer_override) ||
16057             (rack->rc_ack_can_sendout_data) ||
16058             (delayed) ||
16059             (tp->t_state < TCPS_ESTABLISHED)) {
16060                 rack->rc_ack_can_sendout_data = 0;
16061                 if (rack->rc_inp->inp_in_hpts)
16062                         tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
16063         } else if (rack->rc_inp->inp_in_hpts) {
16064                 /*
16065                  * On the hpts you can't pass even if ACKNOW is on, we will
16066                  * when the hpts fires.
16067                  */
16068 #ifdef TCP_ACCOUNTING
16069                 crtsc = get_cyclecount();
16070                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16071                         tp->tcp_proc_time[SND_BLOCKED] += (crtsc - ts_val);
16072                 }
16073                 counter_u64_add(tcp_proc_time[SND_BLOCKED], (crtsc - ts_val));
16074                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16075                         tp->tcp_cnt_counters[SND_BLOCKED]++;
16076                 }
16077                 counter_u64_add(tcp_cnt_counters[SND_BLOCKED], 1);
16078                 sched_unpin();
16079 #endif
16080                 counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1);
16081                 return (0);
16082         }
16083         rack->rc_inp->inp_hpts_calls = 0;
16084         /* Finish out both pacing early and late accounting */
16085         if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16086             TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
16087                 early = rack->r_ctl.rc_last_output_to - cts;
16088         } else
16089                 early = 0;
16090         if (delayed) {
16091                 rack->r_ctl.rc_agg_delayed += delayed;
16092                 rack->r_late = 1;
16093         } else if (early) {
16094                 rack->r_ctl.rc_agg_early += early;
16095                 rack->r_early = 1;
16096         }
16097         /* Now that early/late accounting is done turn off the flag */
16098         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
16099         rack->r_wanted_output = 0;
16100         rack->r_timer_override = 0;
16101         if ((tp->t_state != rack->r_state) &&
16102             TCPS_HAVEESTABLISHED(tp->t_state)) {
16103                 rack_set_state(tp, rack);
16104         }
16105         if ((rack->r_fast_output) &&
16106             (tp->rcv_numsacks == 0)) {
16107                 int ret;
16108
16109                 error = 0;
16110                 ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
16111                 if (ret >= 0)
16112                         return(ret);
16113                 else if (error) {
16114                         inp = rack->rc_inp;
16115                         so = inp->inp_socket;
16116                         sb = &so->so_snd;
16117                         goto nomore;
16118                 }
16119         }
16120         inp = rack->rc_inp;
16121         /*
16122          * For TFO connections in SYN_SENT or SYN_RECEIVED,
16123          * only allow the initial SYN or SYN|ACK and those sent
16124          * by the retransmit timer.
16125          */
16126         if (IS_FASTOPEN(tp->t_flags) &&
16127             ((tp->t_state == TCPS_SYN_RECEIVED) ||
16128              (tp->t_state == TCPS_SYN_SENT)) &&
16129             SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
16130             (tp->t_rxtshift == 0)) {              /* not a retransmit */
16131                 cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
16132                 so = inp->inp_socket;
16133                 sb = &so->so_snd;
16134                 goto just_return_nolock;
16135         }
16136         /*
16137          * Determine length of data that should be transmitted, and flags
16138          * that will be used. If there is some data or critical controls
16139          * (SYN, RST) to send, then transmit; otherwise, investigate
16140          * further.
16141          */
16142         idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
16143         if (tp->t_idle_reduce) {
16144                 if (idle && ((ticks - tp->t_rcvtime) >= tp->t_rxtcur))
16145                         rack_cc_after_idle(rack, tp);
16146         }
16147         tp->t_flags &= ~TF_LASTIDLE;
16148         if (idle) {
16149                 if (tp->t_flags & TF_MORETOCOME) {
16150                         tp->t_flags |= TF_LASTIDLE;
16151                         idle = 0;
16152                 }
16153         }
16154         if ((tp->snd_una == tp->snd_max) &&
16155             rack->r_ctl.rc_went_idle_time &&
16156             TSTMP_GT(cts, rack->r_ctl.rc_went_idle_time)) {
16157                 idle = cts - rack->r_ctl.rc_went_idle_time;
16158                 if (idle > rack_min_probertt_hold) {
16159                         /* Count as a probe rtt */
16160                         if (rack->in_probe_rtt == 0) {
16161                                 rack->r_ctl.rc_lower_rtt_us_cts = cts;
16162                                 rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
16163                                 rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
16164                                 rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
16165                         } else {
16166                                 rack_exit_probertt(rack, cts);
16167                         }
16168                 }
16169                 idle = 0;
16170         }
16171         if (rack_use_fsb && (rack->r_fsb_inited == 0))
16172                 rack_init_fsb_block(tp, rack);
16173 again:
16174         /*
16175          * If we've recently taken a timeout, snd_max will be greater than
16176          * snd_nxt.  There may be SACK information that allows us to avoid
16177          * resending already delivered data.  Adjust snd_nxt accordingly.
16178          */
16179         sendalot = 0;
16180         cts = tcp_get_usecs(&tv);
16181         ms_cts = tcp_tv_to_mssectick(&tv);
16182         tso = 0;
16183         mtu = 0;
16184         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
16185         minseg = segsiz;
16186         if (rack->r_ctl.rc_pace_max_segs == 0)
16187                 pace_max_seg = rack->rc_user_set_max_segs * segsiz;
16188         else
16189                 pace_max_seg = rack->r_ctl.rc_pace_max_segs;
16190         sb_offset = tp->snd_max - tp->snd_una;
16191         cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
16192         flags = tcp_outflags[tp->t_state];
16193         while (rack->rc_free_cnt < rack_free_cache) {
16194                 rsm = rack_alloc(rack);
16195                 if (rsm == NULL) {
16196                         if (inp->inp_hpts_calls)
16197                                 /* Retry in a ms */
16198                                 slot = (1 * HPTS_USEC_IN_MSEC);
16199                         so = inp->inp_socket;
16200                         sb = &so->so_snd;
16201                         goto just_return_nolock;
16202                 }
16203                 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_tnext);
16204                 rack->rc_free_cnt++;
16205                 rsm = NULL;
16206         }
16207         if (inp->inp_hpts_calls)
16208                 inp->inp_hpts_calls = 0;
16209         sack_rxmit = 0;
16210         len = 0;
16211         rsm = NULL;
16212         if (flags & TH_RST) {
16213                 SOCKBUF_LOCK(&inp->inp_socket->so_snd);
16214                 so = inp->inp_socket;
16215                 sb = &so->so_snd;
16216                 goto send;
16217         }
16218         if (rack->r_ctl.rc_resend) {
16219                 /* Retransmit timer */
16220                 rsm = rack->r_ctl.rc_resend;
16221                 rack->r_ctl.rc_resend = NULL;
16222                 rsm->r_flags &= ~RACK_TLP;
16223                 len = rsm->r_end - rsm->r_start;
16224                 sack_rxmit = 1;
16225                 sendalot = 0;
16226                 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16227                         ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16228                          __func__, __LINE__,
16229                          rsm->r_start, tp->snd_una, tp, rack, rsm));
16230                 sb_offset = rsm->r_start - tp->snd_una;
16231                 if (len >= segsiz)
16232                         len = segsiz;
16233         } else if ((rsm = tcp_rack_output(tp, rack, cts)) != NULL) {
16234                 /* We have a retransmit that takes precedence */
16235                 rsm->r_flags &= ~RACK_TLP;
16236                 if ((!IN_FASTRECOVERY(tp->t_flags)) &&
16237                     ((tp->t_flags & TF_WASFRECOVERY) == 0)) {
16238                         /* Enter recovery if not induced by a time-out */
16239                         rack->r_ctl.rc_rsm_start = rsm->r_start;
16240                         rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
16241                         rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
16242                         rack_cong_signal(tp, CC_NDUPACK, tp->snd_una);
16243                 }
16244 #ifdef INVARIANTS
16245                 if (SEQ_LT(rsm->r_start, tp->snd_una)) {
16246                         panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n",
16247                               tp, rack, rsm, rsm->r_start, tp->snd_una);
16248                 }
16249 #endif
16250                 len = rsm->r_end - rsm->r_start;
16251                 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16252                         ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16253                          __func__, __LINE__,
16254                          rsm->r_start, tp->snd_una, tp, rack, rsm));
16255                 sb_offset = rsm->r_start - tp->snd_una;
16256                 sendalot = 0;
16257                 if (len >= segsiz)
16258                         len = segsiz;
16259                 if (len > 0) {
16260                         sack_rxmit = 1;
16261                         KMOD_TCPSTAT_INC(tcps_sack_rexmits);
16262                         KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes,
16263                             min(len, segsiz));
16264                         counter_u64_add(rack_rtm_prr_retran, 1);
16265                 }
16266         } else if (rack->r_ctl.rc_tlpsend) {
16267                 /* Tail loss probe */
16268                 long cwin;
16269                 long tlen;
16270
16271                 doing_tlp = 1;
16272                 /*
16273                  * Check if we can do a TLP with a RACK'd packet
16274                  * this can happen if we are not doing the rack
16275                  * cheat and we skipped to a TLP and it
16276                  * went off.
16277                  */
16278                 rsm = rack->r_ctl.rc_tlpsend;
16279                 rsm->r_flags |= RACK_TLP;
16280
16281                 rack->r_ctl.rc_tlpsend = NULL;
16282                 sack_rxmit = 1;
16283                 tlen = rsm->r_end - rsm->r_start;
16284                 if (tlen > segsiz)
16285                         tlen = segsiz;
16286                 tp->t_sndtlppack++;
16287                 tp->t_sndtlpbyte += tlen;
16288                 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16289                         ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16290                          __func__, __LINE__,
16291                          rsm->r_start, tp->snd_una, tp, rack, rsm));
16292                 sb_offset = rsm->r_start - tp->snd_una;
16293                 cwin = min(tp->snd_wnd, tlen);
16294                 len = cwin;
16295         }
16296         if (rack->r_must_retran &&
16297             (rsm == NULL)) {
16298                 /*
16299                  * Non-Sack and we had a RTO or MTU change, we
16300                  * need to retransmit until we reach
16301                  * the former snd_max (rack->r_ctl.rc_snd_max_at_rto).
16302                  */
16303                 if (SEQ_GT(tp->snd_max, tp->snd_una)) {
16304                         int sendwin, flight;
16305
16306                         sendwin = min(tp->snd_wnd, tp->snd_cwnd);
16307                         flight = ctf_flight_size(tp, rack->r_ctl.rc_out_at_rto);
16308                         if (flight >= sendwin) {
16309                                 so = inp->inp_socket;
16310                                 sb = &so->so_snd;
16311                                 goto just_return_nolock;
16312                         }
16313                         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
16314                         KASSERT(rsm != NULL, ("rsm is NULL rack:%p r_must_retran set", rack));
16315                         if (rsm == NULL) {
16316                                 /* TSNH */
16317                                 rack->r_must_retran = 0;
16318                                 rack->r_ctl.rc_out_at_rto = 0;
16319                                 rack->r_must_retran = 0;
16320                                 so = inp->inp_socket;
16321                                 sb = &so->so_snd;
16322                                 goto just_return_nolock;
16323                         }
16324                         sack_rxmit = 1;
16325                         len = rsm->r_end - rsm->r_start;
16326                         sendalot = 0;
16327                         sb_offset = rsm->r_start - tp->snd_una;
16328                         if (len >= segsiz)
16329                                 len = segsiz;
16330                 } else {
16331                         /* We must be done if there is nothing outstanding */
16332                         rack->r_must_retran = 0;
16333                         rack->r_ctl.rc_out_at_rto = 0;
16334                 }
16335         }
16336         /*
16337          * Enforce a connection sendmap count limit if set
16338          * as long as we are not retransmiting.
16339          */
16340         if ((rsm == NULL) &&
16341             (rack->do_detection == 0) &&
16342             (V_tcp_map_entries_limit > 0) &&
16343             (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
16344                 counter_u64_add(rack_to_alloc_limited, 1);
16345                 if (!rack->alloc_limit_reported) {
16346                         rack->alloc_limit_reported = 1;
16347                         counter_u64_add(rack_alloc_limited_conns, 1);
16348                 }
16349                 so = inp->inp_socket;
16350                 sb = &so->so_snd;
16351                 goto just_return_nolock;
16352         }
16353         if (rsm && (rsm->r_flags & RACK_HAS_FIN)) {
16354                 /* we are retransmitting the fin */
16355                 len--;
16356                 if (len) {
16357                         /*
16358                          * When retransmitting data do *not* include the
16359                          * FIN. This could happen from a TLP probe.
16360                          */
16361                         flags &= ~TH_FIN;
16362                 }
16363         }
16364 #ifdef INVARIANTS
16365         /* For debugging */
16366         rack->r_ctl.rc_rsm_at_retran = rsm;
16367 #endif
16368         if (rsm && rack->r_fsb_inited && rack_use_rsm_rfo &&
16369             ((rsm->r_flags & RACK_HAS_FIN) == 0)) {
16370                 int ret;
16371
16372                 ret = rack_fast_rsm_output(tp, rack, rsm, ts_val, cts, ms_cts, &tv, len);
16373                 if (ret == 0)
16374                         return (0);
16375         }
16376         so = inp->inp_socket;
16377         sb = &so->so_snd;
16378         if (do_a_prefetch == 0) {
16379                 kern_prefetch(sb, &do_a_prefetch);
16380                 do_a_prefetch = 1;
16381         }
16382 #ifdef NETFLIX_SHARED_CWND
16383         if ((tp->t_flags2 & TF2_TCP_SCWND_ALLOWED) &&
16384             rack->rack_enable_scwnd) {
16385                 /* We are doing cwnd sharing */
16386                 if (rack->gp_ready &&
16387                     (rack->rack_attempted_scwnd == 0) &&
16388                     (rack->r_ctl.rc_scw == NULL) &&
16389                     tp->t_lib) {
16390                         /* The pcbid is in, lets make an attempt */
16391                         counter_u64_add(rack_try_scwnd, 1);
16392                         rack->rack_attempted_scwnd = 1;
16393                         rack->r_ctl.rc_scw = tcp_shared_cwnd_alloc(tp,
16394                                                                    &rack->r_ctl.rc_scw_index,
16395                                                                    segsiz);
16396                 }
16397                 if (rack->r_ctl.rc_scw &&
16398                     (rack->rack_scwnd_is_idle == 1) &&
16399                     sbavail(&so->so_snd)) {
16400                         /* we are no longer out of data */
16401                         tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
16402                         rack->rack_scwnd_is_idle = 0;
16403                 }
16404                 if (rack->r_ctl.rc_scw) {
16405                         /* First lets update and get the cwnd */
16406                         rack->r_ctl.cwnd_to_use = cwnd_to_use = tcp_shared_cwnd_update(rack->r_ctl.rc_scw,
16407                                                                     rack->r_ctl.rc_scw_index,
16408                                                                     tp->snd_cwnd, tp->snd_wnd, segsiz);
16409                 }
16410         }
16411 #endif
16412         /*
16413          * Get standard flags, and add SYN or FIN if requested by 'hidden'
16414          * state flags.
16415          */
16416         if (tp->t_flags & TF_NEEDFIN)
16417                 flags |= TH_FIN;
16418         if (tp->t_flags & TF_NEEDSYN)
16419                 flags |= TH_SYN;
16420         if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
16421                 void *end_rsm;
16422                 end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
16423                 if (end_rsm)
16424                         kern_prefetch(end_rsm, &prefetch_rsm);
16425                 prefetch_rsm = 1;
16426         }
16427         SOCKBUF_LOCK(sb);
16428         /*
16429          * If snd_nxt == snd_max and we have transmitted a FIN, the
16430          * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a
16431          * negative length.  This can also occur when TCP opens up its
16432          * congestion window while receiving additional duplicate acks after
16433          * fast-retransmit because TCP will reset snd_nxt to snd_max after
16434          * the fast-retransmit.
16435          *
16436          * In the normal retransmit-FIN-only case, however, snd_nxt will be
16437          * set to snd_una, the sb_offset will be 0, and the length may wind
16438          * up 0.
16439          *
16440          * If sack_rxmit is true we are retransmitting from the scoreboard
16441          * in which case len is already set.
16442          */
16443         if ((sack_rxmit == 0) &&
16444             (TCPS_HAVEESTABLISHED(tp->t_state) || IS_FASTOPEN(tp->t_flags))) {
16445                 uint32_t avail;
16446
16447                 avail = sbavail(sb);
16448                 if (SEQ_GT(tp->snd_nxt, tp->snd_una) && avail)
16449                         sb_offset = tp->snd_nxt - tp->snd_una;
16450                 else
16451                         sb_offset = 0;
16452                 if ((IN_FASTRECOVERY(tp->t_flags) == 0) || rack->rack_no_prr) {
16453                         if (rack->r_ctl.rc_tlp_new_data) {
16454                                 /* TLP is forcing out new data */
16455                                 if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) {
16456                                         rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset);
16457                                 }
16458                                 if ((rack->r_ctl.rc_tlp_new_data + sb_offset) > tp->snd_wnd) {
16459                                         if (tp->snd_wnd > sb_offset)
16460                                                 len = tp->snd_wnd - sb_offset;
16461                                         else
16462                                                 len = 0;
16463                                 } else {
16464                                         len = rack->r_ctl.rc_tlp_new_data;
16465                                 }
16466                                 rack->r_ctl.rc_tlp_new_data = 0;
16467                                 doing_tlp = 1;
16468                         }  else {
16469                                 len = rack_what_can_we_send(tp, rack, cwnd_to_use, avail, sb_offset);
16470                         }
16471                         if ((rack->r_ctl.crte == NULL) && IN_FASTRECOVERY(tp->t_flags) && (len > segsiz)) {
16472                                 /*
16473                                  * For prr=off, we need to send only 1 MSS
16474                                  * at a time. We do this because another sack could
16475                                  * be arriving that causes us to send retransmits and
16476                                  * we don't want to be on a long pace due to a larger send
16477                                  * that keeps us from sending out the retransmit.
16478                                  */
16479                                 len = segsiz;
16480                         }
16481                 } else {
16482                         uint32_t outstanding;
16483                         /*
16484                          * We are inside of a Fast recovery episode, this
16485                          * is caused by a SACK or 3 dup acks. At this point
16486                          * we have sent all the retransmissions and we rely
16487                          * on PRR to dictate what we will send in the form of
16488                          * new data.
16489                          */
16490
16491                         outstanding = tp->snd_max - tp->snd_una;
16492                         if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) {
16493                                 if (tp->snd_wnd > outstanding) {
16494                                         len = tp->snd_wnd - outstanding;
16495                                         /* Check to see if we have the data */
16496                                         if ((sb_offset + len) > avail) {
16497                                                 /* It does not all fit */
16498                                                 if (avail > sb_offset)
16499                                                         len = avail - sb_offset;
16500                                                 else
16501                                                         len = 0;
16502                                         }
16503                                 } else {
16504                                         len = 0;
16505                                 }
16506                         } else if (avail > sb_offset) {
16507                                 len = avail - sb_offset;
16508                         } else {
16509                                 len = 0;
16510                         }
16511                         if (len > 0) {
16512                                 if (len > rack->r_ctl.rc_prr_sndcnt) {
16513                                         len = rack->r_ctl.rc_prr_sndcnt;
16514                                 }
16515                                 if (len > 0) {
16516                                         sub_from_prr = 1;
16517                                         counter_u64_add(rack_rtm_prr_newdata, 1);
16518                                 }
16519                         }
16520                         if (len > segsiz) {
16521                                 /*
16522                                  * We should never send more than a MSS when
16523                                  * retransmitting or sending new data in prr
16524                                  * mode unless the override flag is on. Most
16525                                  * likely the PRR algorithm is not going to
16526                                  * let us send a lot as well :-)
16527                                  */
16528                                 if (rack->r_ctl.rc_prr_sendalot == 0) {
16529                                         len = segsiz;
16530                                 }
16531                         } else if (len < segsiz) {
16532                                 /*
16533                                  * Do we send any? The idea here is if the
16534                                  * send empty's the socket buffer we want to
16535                                  * do it. However if not then lets just wait
16536                                  * for our prr_sndcnt to get bigger.
16537                                  */
16538                                 long leftinsb;
16539
16540                                 leftinsb = sbavail(sb) - sb_offset;
16541                                 if (leftinsb > len) {
16542                                         /* This send does not empty the sb */
16543                                         len = 0;
16544                                 }
16545                         }
16546                 }
16547         } else if (!TCPS_HAVEESTABLISHED(tp->t_state)) {
16548                 /*
16549                  * If you have not established
16550                  * and are not doing FAST OPEN
16551                  * no data please.
16552                  */
16553                 if ((sack_rxmit == 0) &&
16554                     (!IS_FASTOPEN(tp->t_flags))){
16555                         len = 0;
16556                         sb_offset = 0;
16557                 }
16558         }
16559         if (prefetch_so_done == 0) {
16560                 kern_prefetch(so, &prefetch_so_done);
16561                 prefetch_so_done = 1;
16562         }
16563         /*
16564          * Lop off SYN bit if it has already been sent.  However, if this is
16565          * SYN-SENT state and if segment contains data and if we don't know
16566          * that foreign host supports TAO, suppress sending segment.
16567          */
16568         if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una) &&
16569             ((sack_rxmit == 0) && (tp->t_rxtshift == 0))) {
16570                 /*
16571                  * When sending additional segments following a TFO SYN|ACK,
16572                  * do not include the SYN bit.
16573                  */
16574                 if (IS_FASTOPEN(tp->t_flags) &&
16575                     (tp->t_state == TCPS_SYN_RECEIVED))
16576                         flags &= ~TH_SYN;
16577         }
16578         /*
16579          * Be careful not to send data and/or FIN on SYN segments. This
16580          * measure is needed to prevent interoperability problems with not
16581          * fully conformant TCP implementations.
16582          */
16583         if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
16584                 len = 0;
16585                 flags &= ~TH_FIN;
16586         }
16587         /*
16588          * On TFO sockets, ensure no data is sent in the following cases:
16589          *
16590          *  - When retransmitting SYN|ACK on a passively-created socket
16591          *
16592          *  - When retransmitting SYN on an actively created socket
16593          *
16594          *  - When sending a zero-length cookie (cookie request) on an
16595          *    actively created socket
16596          *
16597          *  - When the socket is in the CLOSED state (RST is being sent)
16598          */
16599         if (IS_FASTOPEN(tp->t_flags) &&
16600             (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
16601              ((tp->t_state == TCPS_SYN_SENT) &&
16602               (tp->t_tfo_client_cookie_len == 0)) ||
16603              (flags & TH_RST))) {
16604                 sack_rxmit = 0;
16605                 len = 0;
16606         }
16607         /* Without fast-open there should never be data sent on a SYN */
16608         if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags))) {
16609                 tp->snd_nxt = tp->iss;
16610                 len = 0;
16611         }
16612         if ((len > segsiz) && (tcp_dsack_block_exists(tp))) {
16613                 /* We only send 1 MSS if we have a DSACK block */
16614                 add_flag |= RACK_SENT_W_DSACK;
16615                 len = segsiz;
16616         }
16617         orig_len = len;
16618         if (len <= 0) {
16619                 /*
16620                  * If FIN has been sent but not acked, but we haven't been
16621                  * called to retransmit, len will be < 0.  Otherwise, window
16622                  * shrank after we sent into it.  If window shrank to 0,
16623                  * cancel pending retransmit, pull snd_nxt back to (closed)
16624                  * window, and set the persist timer if it isn't already
16625                  * going.  If the window didn't close completely, just wait
16626                  * for an ACK.
16627                  *
16628                  * We also do a general check here to ensure that we will
16629                  * set the persist timer when we have data to send, but a
16630                  * 0-byte window. This makes sure the persist timer is set
16631                  * even if the packet hits one of the "goto send" lines
16632                  * below.
16633                  */
16634                 len = 0;
16635                 if ((tp->snd_wnd == 0) &&
16636                     (TCPS_HAVEESTABLISHED(tp->t_state)) &&
16637                     (tp->snd_una == tp->snd_max) &&
16638                     (sb_offset < (int)sbavail(sb))) {
16639                         rack_enter_persist(tp, rack, cts);
16640                 }
16641         } else if ((rsm == NULL) &&
16642                    (doing_tlp == 0) &&
16643                    (len < pace_max_seg)) {
16644                 /*
16645                  * We are not sending a maximum sized segment for
16646                  * some reason. Should we not send anything (think
16647                  * sws or persists)?
16648                  */
16649                 if ((tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
16650                     (TCPS_HAVEESTABLISHED(tp->t_state)) &&
16651                     (len < minseg) &&
16652                     (len < (int)(sbavail(sb) - sb_offset))) {
16653                         /*
16654                          * Here the rwnd is less than
16655                          * the minimum pacing size, this is not a retransmit,
16656                          * we are established and
16657                          * the send is not the last in the socket buffer
16658                          * we send nothing, and we may enter persists
16659                          * if nothing is outstanding.
16660                          */
16661                         len = 0;
16662                         if (tp->snd_max == tp->snd_una) {
16663                                 /*
16664                                  * Nothing out we can
16665                                  * go into persists.
16666                                  */
16667                                 rack_enter_persist(tp, rack, cts);
16668                         }
16669                      } else if ((cwnd_to_use >= max(minseg, (segsiz * 4))) &&
16670                            (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
16671                            (len < (int)(sbavail(sb) - sb_offset)) &&
16672                            (len < minseg)) {
16673                         /*
16674                          * Here we are not retransmitting, and
16675                          * the cwnd is not so small that we could
16676                          * not send at least a min size (rxt timer
16677                          * not having gone off), We have 2 segments or
16678                          * more already in flight, its not the tail end
16679                          * of the socket buffer  and the cwnd is blocking
16680                          * us from sending out a minimum pacing segment size.
16681                          * Lets not send anything.
16682                          */
16683                         len = 0;
16684                 } else if (((tp->snd_wnd - ctf_outstanding(tp)) <
16685                             min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
16686                            (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
16687                            (len < (int)(sbavail(sb) - sb_offset)) &&
16688                            (TCPS_HAVEESTABLISHED(tp->t_state))) {
16689                         /*
16690                          * Here we have a send window but we have
16691                          * filled it up and we can't send another pacing segment.
16692                          * We also have in flight more than 2 segments
16693                          * and we are not completing the sb i.e. we allow
16694                          * the last bytes of the sb to go out even if
16695                          * its not a full pacing segment.
16696                          */
16697                         len = 0;
16698                 } else if ((rack->r_ctl.crte != NULL) &&
16699                            (tp->snd_wnd >= (pace_max_seg * max(1, rack_hw_rwnd_factor))) &&
16700                            (cwnd_to_use >= (pace_max_seg + (4 * segsiz))) &&
16701                            (ctf_flight_size(tp, rack->r_ctl.rc_sacked) >= (2 * segsiz)) &&
16702                            (len < (int)(sbavail(sb) - sb_offset))) {
16703                         /*
16704                          * Here we are doing hardware pacing, this is not a TLP,
16705                          * we are not sending a pace max segment size, there is rwnd
16706                          * room to send at least N pace_max_seg, the cwnd is greater
16707                          * than or equal to a full pacing segments plus 4 mss and we have 2 or
16708                          * more segments in flight and its not the tail of the socket buffer.
16709                          *
16710                          * We don't want to send instead we need to get more ack's in to
16711                          * allow us to send a full pacing segment. Normally, if we are pacing
16712                          * about the right speed, we should have finished our pacing
16713                          * send as most of the acks have come back if we are at the
16714                          * right rate. This is a bit fuzzy since return path delay
16715                          * can delay the acks, which is why we want to make sure we
16716                          * have cwnd space to have a bit more than a max pace segments in flight.
16717                          *
16718                          * If we have not gotten our acks back we are pacing at too high a
16719                          * rate delaying will not hurt and will bring our GP estimate down by
16720                          * injecting the delay. If we don't do this we will send
16721                          * 2 MSS out in response to the acks being clocked in which
16722                          * defeats the point of hw-pacing (i.e. to help us get
16723                          * larger TSO's out).
16724                          */
16725                         len = 0;
16726
16727                 }
16728
16729         }
16730         /* len will be >= 0 after this point. */
16731         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
16732         rack_sndbuf_autoscale(rack);
16733         /*
16734          * Decide if we can use TCP Segmentation Offloading (if supported by
16735          * hardware).
16736          *
16737          * TSO may only be used if we are in a pure bulk sending state.  The
16738          * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
16739          * options prevent using TSO.  With TSO the TCP header is the same
16740          * (except for the sequence number) for all generated packets.  This
16741          * makes it impossible to transmit any options which vary per
16742          * generated segment or packet.
16743          *
16744          * IPv4 handling has a clear separation of ip options and ip header
16745          * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
16746          * the right thing below to provide length of just ip options and thus
16747          * checking for ipoptlen is enough to decide if ip options are present.
16748          */
16749         ipoptlen = 0;
16750 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
16751         /*
16752          * Pre-calculate here as we save another lookup into the darknesses
16753          * of IPsec that way and can actually decide if TSO is ok.
16754          */
16755 #ifdef INET6
16756         if (isipv6 && IPSEC_ENABLED(ipv6))
16757                 ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb);
16758 #ifdef INET
16759         else
16760 #endif
16761 #endif                          /* INET6 */
16762 #ifdef INET
16763                 if (IPSEC_ENABLED(ipv4))
16764                         ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb);
16765 #endif                          /* INET */
16766 #endif
16767
16768 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
16769         ipoptlen += ipsec_optlen;
16770 #endif
16771         if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > segsiz &&
16772             (tp->t_port == 0) &&
16773             ((tp->t_flags & TF_SIGNATURE) == 0) &&
16774             tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
16775             ipoptlen == 0)
16776                 tso = 1;
16777         {
16778                 uint32_t outstanding;
16779
16780                 outstanding = tp->snd_max - tp->snd_una;
16781                 if (tp->t_flags & TF_SENTFIN) {
16782                         /*
16783                          * If we sent a fin, snd_max is 1 higher than
16784                          * snd_una
16785                          */
16786                         outstanding--;
16787                 }
16788                 if (sack_rxmit) {
16789                         if ((rsm->r_flags & RACK_HAS_FIN) == 0)
16790                                 flags &= ~TH_FIN;
16791                 } else {
16792                         if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
16793                                    sbused(sb)))
16794                                 flags &= ~TH_FIN;
16795                 }
16796         }
16797         recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
16798             (long)TCP_MAXWIN << tp->rcv_scale);
16799
16800         /*
16801          * Sender silly window avoidance.   We transmit under the following
16802          * conditions when len is non-zero:
16803          *
16804          * - We have a full segment (or more with TSO) - This is the last
16805          * buffer in a write()/send() and we are either idle or running
16806          * NODELAY - we've timed out (e.g. persist timer) - we have more
16807          * then 1/2 the maximum send window's worth of data (receiver may be
16808          * limited the window size) - we need to retransmit
16809          */
16810         if (len) {
16811                 if (len >= segsiz) {
16812                         goto send;
16813                 }
16814                 /*
16815                  * NOTE! on localhost connections an 'ack' from the remote
16816                  * end may occur synchronously with the output and cause us
16817                  * to flush a buffer queued with moretocome.  XXX
16818                  *
16819                  */
16820                 if (!(tp->t_flags & TF_MORETOCOME) &&   /* normal case */
16821                     (idle || (tp->t_flags & TF_NODELAY)) &&
16822                     ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
16823                     (tp->t_flags & TF_NOPUSH) == 0) {
16824                         pass = 2;
16825                         goto send;
16826                 }
16827                 if ((tp->snd_una == tp->snd_max) && len) {      /* Nothing outstanding */
16828                         pass = 22;
16829                         goto send;
16830                 }
16831                 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) {
16832                         pass = 4;
16833                         goto send;
16834                 }
16835                 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { /* retransmit case */
16836                         pass = 5;
16837                         goto send;
16838                 }
16839                 if (sack_rxmit) {
16840                         pass = 6;
16841                         goto send;
16842                 }
16843                 if (((tp->snd_wnd - ctf_outstanding(tp)) < segsiz) &&
16844                     (ctf_outstanding(tp) < (segsiz * 2))) {
16845                         /*
16846                          * We have less than two MSS outstanding (delayed ack)
16847                          * and our rwnd will not let us send a full sized
16848                          * MSS. Lets go ahead and let this small segment
16849                          * out because we want to try to have at least two
16850                          * packets inflight to not be caught by delayed ack.
16851                          */
16852                         pass = 12;
16853                         goto send;
16854                 }
16855         }
16856         /*
16857          * Sending of standalone window updates.
16858          *
16859          * Window updates are important when we close our window due to a
16860          * full socket buffer and are opening it again after the application
16861          * reads data from it.  Once the window has opened again and the
16862          * remote end starts to send again the ACK clock takes over and
16863          * provides the most current window information.
16864          *
16865          * We must avoid the silly window syndrome whereas every read from
16866          * the receive buffer, no matter how small, causes a window update
16867          * to be sent.  We also should avoid sending a flurry of window
16868          * updates when the socket buffer had queued a lot of data and the
16869          * application is doing small reads.
16870          *
16871          * Prevent a flurry of pointless window updates by only sending an
16872          * update when we can increase the advertized window by more than
16873          * 1/4th of the socket buffer capacity.  When the buffer is getting
16874          * full or is very small be more aggressive and send an update
16875          * whenever we can increase by two mss sized segments. In all other
16876          * situations the ACK's to new incoming data will carry further
16877          * window increases.
16878          *
16879          * Don't send an independent window update if a delayed ACK is
16880          * pending (it will get piggy-backed on it) or the remote side
16881          * already has done a half-close and won't send more data.  Skip
16882          * this if the connection is in T/TCP half-open state.
16883          */
16884         if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
16885             !(tp->t_flags & TF_DELACK) &&
16886             !TCPS_HAVERCVDFIN(tp->t_state)) {
16887                 /*
16888                  * "adv" is the amount we could increase the window, taking
16889                  * into account that we are limited by TCP_MAXWIN <<
16890                  * tp->rcv_scale.
16891                  */
16892                 int32_t adv;
16893                 int oldwin;
16894
16895                 adv = recwin;
16896                 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
16897                         oldwin = (tp->rcv_adv - tp->rcv_nxt);
16898                         if (adv > oldwin)
16899                             adv -= oldwin;
16900                         else {
16901                                 /* We can't increase the window */
16902                                 adv = 0;
16903                         }
16904                 } else
16905                         oldwin = 0;
16906
16907                 /*
16908                  * If the new window size ends up being the same as or less
16909                  * than the old size when it is scaled, then don't force
16910                  * a window update.
16911                  */
16912                 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
16913                         goto dontupdate;
16914
16915                 if (adv >= (int32_t)(2 * segsiz) &&
16916                     (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
16917                      recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) ||
16918                      so->so_rcv.sb_hiwat <= 8 * segsiz)) {
16919                         pass = 7;
16920                         goto send;
16921                 }
16922                 if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) {
16923                         pass = 23;
16924                         goto send;
16925                 }
16926         }
16927 dontupdate:
16928
16929         /*
16930          * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
16931          * is also a catch-all for the retransmit timer timeout case.
16932          */
16933         if (tp->t_flags & TF_ACKNOW) {
16934                 pass = 8;
16935                 goto send;
16936         }
16937         if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) {
16938                 pass = 9;
16939                 goto send;
16940         }
16941         /*
16942          * If our state indicates that FIN should be sent and we have not
16943          * yet done so, then we need to send.
16944          */
16945         if ((flags & TH_FIN) &&
16946             (tp->snd_nxt == tp->snd_una)) {
16947                 pass = 11;
16948                 goto send;
16949         }
16950         /*
16951          * No reason to send a segment, just return.
16952          */
16953 just_return:
16954         SOCKBUF_UNLOCK(sb);
16955 just_return_nolock:
16956         {
16957                 int app_limited = CTF_JR_SENT_DATA;
16958
16959                 if (tot_len_this_send > 0) {
16960                         /* Make sure snd_nxt is up to max */
16961                         rack->r_ctl.fsb.recwin = recwin;
16962                         slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, NULL, segsiz);
16963                         if ((error == 0) &&
16964                             rack_use_rfo &&
16965                             ((flags & (TH_SYN|TH_FIN)) == 0) &&
16966                             (ipoptlen == 0) &&
16967                             (tp->snd_nxt == tp->snd_max) &&
16968                             (tp->rcv_numsacks == 0) &&
16969                             rack->r_fsb_inited &&
16970                             TCPS_HAVEESTABLISHED(tp->t_state) &&
16971                             (rack->r_must_retran == 0) &&
16972                             ((tp->t_flags & TF_NEEDFIN) == 0) &&
16973                             (len > 0) && (orig_len > 0) &&
16974                             (orig_len > len) &&
16975                             ((orig_len - len) >= segsiz) &&
16976                             ((optlen == 0) ||
16977                              ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
16978                                 /* We can send at least one more MSS using our fsb */
16979
16980                                 rack->r_fast_output = 1;
16981                                 rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
16982                                 rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
16983                                 rack->r_ctl.fsb.tcp_flags = flags;
16984                                 rack->r_ctl.fsb.left_to_send = orig_len - len;
16985                                 KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
16986                                         ("rack:%p left_to_send:%u sbavail:%u out:%u",
16987                                         rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
16988                                          (tp->snd_max - tp->snd_una)));
16989                                 if (rack->r_ctl.fsb.left_to_send < segsiz)
16990                                         rack->r_fast_output = 0;
16991                                 else {
16992                                         if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
16993                                                 rack->r_ctl.fsb.rfo_apply_push = 1;
16994                                         else
16995                                                 rack->r_ctl.fsb.rfo_apply_push = 0;
16996                                 }
16997                         } else
16998                                 rack->r_fast_output = 0;
16999
17000
17001                         rack_log_fsb(rack, tp, so, flags,
17002                                      ipoptlen, orig_len, len, 0,
17003                                      1, optlen, __LINE__, 1);
17004                         if (SEQ_GT(tp->snd_max, tp->snd_nxt))
17005                                 tp->snd_nxt = tp->snd_max;
17006                 } else {
17007                         int end_window = 0;
17008                         uint32_t seq = tp->gput_ack;
17009
17010                         rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
17011                         if (rsm) {
17012                                 /*
17013                                  * Mark the last sent that we just-returned (hinting
17014                                  * that delayed ack may play a role in any rtt measurement).
17015                                  */
17016                                 rsm->r_just_ret = 1;
17017                         }
17018                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1);
17019                         rack->r_ctl.rc_agg_delayed = 0;
17020                         rack->r_early = 0;
17021                         rack->r_late = 0;
17022                         rack->r_ctl.rc_agg_early = 0;
17023                         if ((ctf_outstanding(tp) +
17024                              min(max(segsiz, (rack->r_ctl.rc_high_rwnd/2)),
17025                                  minseg)) >= tp->snd_wnd) {
17026                                 /* We are limited by the rwnd */
17027                                 app_limited = CTF_JR_RWND_LIMITED;
17028                                 if (IN_FASTRECOVERY(tp->t_flags))
17029                                     rack->r_ctl.rc_prr_sndcnt = 0;
17030                         } else if (ctf_outstanding(tp) >= sbavail(sb)) {
17031                                 /* We are limited by whats available -- app limited */
17032                                 app_limited = CTF_JR_APP_LIMITED;
17033                                 if (IN_FASTRECOVERY(tp->t_flags))
17034                                     rack->r_ctl.rc_prr_sndcnt = 0;
17035                         } else if ((idle == 0) &&
17036                                    ((tp->t_flags & TF_NODELAY) == 0) &&
17037                                    ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
17038                                    (len < segsiz)) {
17039                                 /*
17040                                  * No delay is not on and the
17041                                  * user is sending less than 1MSS. This
17042                                  * brings out SWS avoidance so we
17043                                  * don't send. Another app-limited case.
17044                                  */
17045                                 app_limited = CTF_JR_APP_LIMITED;
17046                         } else if (tp->t_flags & TF_NOPUSH) {
17047                                 /*
17048                                  * The user has requested no push of
17049                                  * the last segment and we are
17050                                  * at the last segment. Another app
17051                                  * limited case.
17052                                  */
17053                                 app_limited = CTF_JR_APP_LIMITED;
17054                         } else if ((ctf_outstanding(tp) + minseg) > cwnd_to_use) {
17055                                 /* Its the cwnd */
17056                                 app_limited = CTF_JR_CWND_LIMITED;
17057                         } else if (IN_FASTRECOVERY(tp->t_flags) &&
17058                                    (rack->rack_no_prr == 0) &&
17059                                    (rack->r_ctl.rc_prr_sndcnt < segsiz)) {
17060                                 app_limited = CTF_JR_PRR;
17061                         } else {
17062                                 /* Now why here are we not sending? */
17063 #ifdef NOW
17064 #ifdef INVARIANTS
17065                                 panic("rack:%p hit JR_ASSESSING case cwnd_to_use:%u?", rack, cwnd_to_use);
17066 #endif
17067 #endif
17068                                 app_limited = CTF_JR_ASSESSING;
17069                         }
17070                         /*
17071                          * App limited in some fashion, for our pacing GP
17072                          * measurements we don't want any gap (even cwnd).
17073                          * Close  down the measurement window.
17074                          */
17075                         if (rack_cwnd_block_ends_measure &&
17076                             ((app_limited == CTF_JR_CWND_LIMITED) ||
17077                              (app_limited == CTF_JR_PRR))) {
17078                                 /*
17079                                  * The reason we are not sending is
17080                                  * the cwnd (or prr). We have been configured
17081                                  * to end the measurement window in
17082                                  * this case.
17083                                  */
17084                                 end_window = 1;
17085                         } else if (rack_rwnd_block_ends_measure &&
17086                                    (app_limited == CTF_JR_RWND_LIMITED)) {
17087                                 /*
17088                                  * We are rwnd limited and have been
17089                                  * configured to end the measurement
17090                                  * window in this case.
17091                                  */
17092                                 end_window = 1;
17093                         } else if (app_limited == CTF_JR_APP_LIMITED) {
17094                                 /*
17095                                  * A true application limited period, we have
17096                                  * ran out of data.
17097                                  */
17098                                 end_window = 1;
17099                         } else if (app_limited == CTF_JR_ASSESSING) {
17100                                 /*
17101                                  * In the assessing case we hit the end of
17102                                  * the if/else and had no known reason
17103                                  * This will panic us under invariants..
17104                                  *
17105                                  * If we get this out in logs we need to
17106                                  * investagate which reason we missed.
17107                                  */
17108                                 end_window = 1;
17109                         }
17110                         if (end_window) {
17111                                 uint8_t log = 0;
17112
17113                                 if ((tp->t_flags & TF_GPUTINPROG) &&
17114                                     SEQ_GT(tp->gput_ack, tp->snd_max)) {
17115                                         /* Mark the last packet has app limited */
17116                                         tp->gput_ack = tp->snd_max;
17117                                         log = 1;
17118                                 }
17119                                 rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
17120                                 if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
17121                                         if (rack->r_ctl.rc_app_limited_cnt == 0)
17122                                                 rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
17123                                         else {
17124                                                 /*
17125                                                  * Go out to the end app limited and mark
17126                                                  * this new one as next and move the end_appl up
17127                                                  * to this guy.
17128                                                  */
17129                                                 if (rack->r_ctl.rc_end_appl)
17130                                                         rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
17131                                                 rack->r_ctl.rc_end_appl = rsm;
17132                                         }
17133                                         rsm->r_flags |= RACK_APP_LIMITED;
17134                                         rack->r_ctl.rc_app_limited_cnt++;
17135                                 }
17136                                 if (log)
17137                                         rack_log_pacing_delay_calc(rack,
17138                                                                    rack->r_ctl.rc_app_limited_cnt, seq,
17139                                                                    tp->gput_ack, 0, 0, 4, __LINE__, NULL);
17140                         }
17141                 }
17142                 if (slot) {
17143                         /* set the rack tcb into the slot N */
17144                         counter_u64_add(rack_paced_segments, 1);
17145                 } else if (tot_len_this_send) {
17146                         counter_u64_add(rack_unpaced_segments, 1);
17147                 }
17148                 /* Check if we need to go into persists or not */
17149                 if ((tp->snd_max == tp->snd_una) &&
17150                     TCPS_HAVEESTABLISHED(tp->t_state) &&
17151                     sbavail(sb) &&
17152                     (sbavail(sb) > tp->snd_wnd) &&
17153                     (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg))) {
17154                         /* Yes lets make sure to move to persist before timer-start */
17155                         rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
17156                 }
17157                 rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, sup_rack);
17158                 rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling, app_limited, cwnd_to_use);
17159         }
17160 #ifdef NETFLIX_SHARED_CWND
17161         if ((sbavail(sb) == 0) &&
17162             rack->r_ctl.rc_scw) {
17163                 tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
17164                 rack->rack_scwnd_is_idle = 1;
17165         }
17166 #endif
17167 #ifdef TCP_ACCOUNTING
17168         if (tot_len_this_send > 0) {
17169                 crtsc = get_cyclecount();
17170                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17171                         tp->tcp_cnt_counters[SND_OUT_DATA]++;
17172                 }
17173                 counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], 1);
17174                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17175                         tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
17176                 }
17177                 counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
17178                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17179                         tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) / segsiz);
17180                 }
17181                 counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len_this_send + segsiz - 1) / segsiz));
17182         } else {
17183                 crtsc = get_cyclecount();
17184                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17185                         tp->tcp_cnt_counters[SND_LIMITED]++;
17186                 }
17187                 counter_u64_add(tcp_cnt_counters[SND_LIMITED], 1);
17188                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17189                         tp->tcp_proc_time[SND_LIMITED] += (crtsc - ts_val);
17190                 }
17191                 counter_u64_add(tcp_proc_time[SND_LIMITED], (crtsc - ts_val));
17192         }
17193         sched_unpin();
17194 #endif
17195         return (0);
17196
17197 send:
17198         if (rsm || sack_rxmit)
17199                 counter_u64_add(rack_nfto_resend, 1);
17200         else
17201                 counter_u64_add(rack_non_fto_send, 1);
17202         if ((flags & TH_FIN) &&
17203             sbavail(sb)) {
17204                 /*
17205                  * We do not transmit a FIN
17206                  * with data outstanding. We
17207                  * need to make it so all data
17208                  * is acked first.
17209                  */
17210                 flags &= ~TH_FIN;
17211         }
17212         /* Enforce stack imposed max seg size if we have one */
17213         if (rack->r_ctl.rc_pace_max_segs &&
17214             (len > rack->r_ctl.rc_pace_max_segs)) {
17215                 mark = 1;
17216                 len = rack->r_ctl.rc_pace_max_segs;
17217         }
17218         SOCKBUF_LOCK_ASSERT(sb);
17219         if (len > 0) {
17220                 if (len >= segsiz)
17221                         tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
17222                 else
17223                         tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
17224         }
17225         /*
17226          * Before ESTABLISHED, force sending of initial options unless TCP
17227          * set not to do any options. NOTE: we assume that the IP/TCP header
17228          * plus TCP options always fit in a single mbuf, leaving room for a
17229          * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr)
17230          * + optlen <= MCLBYTES
17231          */
17232         optlen = 0;
17233 #ifdef INET6
17234         if (isipv6)
17235                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
17236         else
17237 #endif
17238                 hdrlen = sizeof(struct tcpiphdr);
17239
17240         /*
17241          * Compute options for segment. We only have to care about SYN and
17242          * established connection segments.  Options for SYN-ACK segments
17243          * are handled in TCP syncache.
17244          */
17245         to.to_flags = 0;
17246         if ((tp->t_flags & TF_NOOPT) == 0) {
17247                 /* Maximum segment size. */
17248                 if (flags & TH_SYN) {
17249                         tp->snd_nxt = tp->iss;
17250                         to.to_mss = tcp_mssopt(&inp->inp_inc);
17251                         if (tp->t_port)
17252                                 to.to_mss -= V_tcp_udp_tunneling_overhead;
17253                         to.to_flags |= TOF_MSS;
17254
17255                         /*
17256                          * On SYN or SYN|ACK transmits on TFO connections,
17257                          * only include the TFO option if it is not a
17258                          * retransmit, as the presence of the TFO option may
17259                          * have caused the original SYN or SYN|ACK to have
17260                          * been dropped by a middlebox.
17261                          */
17262                         if (IS_FASTOPEN(tp->t_flags) &&
17263                             (tp->t_rxtshift == 0)) {
17264                                 if (tp->t_state == TCPS_SYN_RECEIVED) {
17265                                         to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
17266                                         to.to_tfo_cookie =
17267                                                 (u_int8_t *)&tp->t_tfo_cookie.server;
17268                                         to.to_flags |= TOF_FASTOPEN;
17269                                         wanted_cookie = 1;
17270                                 } else if (tp->t_state == TCPS_SYN_SENT) {
17271                                         to.to_tfo_len =
17272                                                 tp->t_tfo_client_cookie_len;
17273                                         to.to_tfo_cookie =
17274                                                 tp->t_tfo_cookie.client;
17275                                         to.to_flags |= TOF_FASTOPEN;
17276                                         wanted_cookie = 1;
17277                                         /*
17278                                          * If we wind up having more data to
17279                                          * send with the SYN than can fit in
17280                                          * one segment, don't send any more
17281                                          * until the SYN|ACK comes back from
17282                                          * the other end.
17283                                          */
17284                                         sendalot = 0;
17285                                 }
17286                         }
17287                 }
17288                 /* Window scaling. */
17289                 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
17290                         to.to_wscale = tp->request_r_scale;
17291                         to.to_flags |= TOF_SCALE;
17292                 }
17293                 /* Timestamps. */
17294                 if ((tp->t_flags & TF_RCVD_TSTMP) ||
17295                     ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
17296                         to.to_tsval = ms_cts + tp->ts_offset;
17297                         to.to_tsecr = tp->ts_recent;
17298                         to.to_flags |= TOF_TS;
17299                 }
17300                 /* Set receive buffer autosizing timestamp. */
17301                 if (tp->rfbuf_ts == 0 &&
17302                     (so->so_rcv.sb_flags & SB_AUTOSIZE))
17303                         tp->rfbuf_ts = tcp_ts_getticks();
17304                 /* Selective ACK's. */
17305                 if (tp->t_flags & TF_SACK_PERMIT) {
17306                         if (flags & TH_SYN)
17307                                 to.to_flags |= TOF_SACKPERM;
17308                         else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
17309                                  tp->rcv_numsacks > 0) {
17310                                 to.to_flags |= TOF_SACK;
17311                                 to.to_nsacks = tp->rcv_numsacks;
17312                                 to.to_sacks = (u_char *)tp->sackblks;
17313                         }
17314                 }
17315 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
17316                 /* TCP-MD5 (RFC2385). */
17317                 if (tp->t_flags & TF_SIGNATURE)
17318                         to.to_flags |= TOF_SIGNATURE;
17319 #endif                          /* TCP_SIGNATURE */
17320
17321                 /* Processing the options. */
17322                 hdrlen += optlen = tcp_addoptions(&to, opt);
17323                 /*
17324                  * If we wanted a TFO option to be added, but it was unable
17325                  * to fit, ensure no data is sent.
17326                  */
17327                 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
17328                     !(to.to_flags & TOF_FASTOPEN))
17329                         len = 0;
17330         }
17331         if (tp->t_port) {
17332                 if (V_tcp_udp_tunneling_port == 0) {
17333                         /* The port was removed?? */
17334                         SOCKBUF_UNLOCK(&so->so_snd);
17335 #ifdef TCP_ACCOUNTING
17336                         crtsc = get_cyclecount();
17337                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17338                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
17339                         }
17340                         counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
17341                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17342                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
17343                         }
17344                         counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
17345                         sched_unpin();
17346 #endif
17347                         return (EHOSTUNREACH);
17348                 }
17349                 hdrlen += sizeof(struct udphdr);
17350         }
17351 #ifdef INET6
17352         if (isipv6)
17353                 ipoptlen = ip6_optlen(tp->t_inpcb);
17354         else
17355 #endif
17356                 if (tp->t_inpcb->inp_options)
17357                         ipoptlen = tp->t_inpcb->inp_options->m_len -
17358                                 offsetof(struct ipoption, ipopt_list);
17359                 else
17360                         ipoptlen = 0;
17361 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
17362         ipoptlen += ipsec_optlen;
17363 #endif
17364
17365         /*
17366          * Adjust data length if insertion of options will bump the packet
17367          * length beyond the t_maxseg length. Clear the FIN bit because we
17368          * cut off the tail of the segment.
17369          */
17370         if (len + optlen + ipoptlen > tp->t_maxseg) {
17371                 if (tso) {
17372                         uint32_t if_hw_tsomax;
17373                         uint32_t moff;
17374                         int32_t max_len;
17375
17376                         /* extract TSO information */
17377                         if_hw_tsomax = tp->t_tsomax;
17378                         if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
17379                         if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
17380                         KASSERT(ipoptlen == 0,
17381                                 ("%s: TSO can't do IP options", __func__));
17382
17383                         /*
17384                          * Check if we should limit by maximum payload
17385                          * length:
17386                          */
17387                         if (if_hw_tsomax != 0) {
17388                                 /* compute maximum TSO length */
17389                                 max_len = (if_hw_tsomax - hdrlen -
17390                                            max_linkhdr);
17391                                 if (max_len <= 0) {
17392                                         len = 0;
17393                                 } else if (len > max_len) {
17394                                         sendalot = 1;
17395                                         len = max_len;
17396                                         mark = 2;
17397                                 }
17398                         }
17399                         /*
17400                          * Prevent the last segment from being fractional
17401                          * unless the send sockbuf can be emptied:
17402                          */
17403                         max_len = (tp->t_maxseg - optlen);
17404                         if ((sb_offset + len) < sbavail(sb)) {
17405                                 moff = len % (u_int)max_len;
17406                                 if (moff != 0) {
17407                                         mark = 3;
17408                                         len -= moff;
17409                                 }
17410                         }
17411                         /*
17412                          * In case there are too many small fragments don't
17413                          * use TSO:
17414                          */
17415                         if (len <= segsiz) {
17416                                 mark = 4;
17417                                 tso = 0;
17418                         }
17419                         /*
17420                          * Send the FIN in a separate segment after the bulk
17421                          * sending is done. We don't trust the TSO
17422                          * implementations to clear the FIN flag on all but
17423                          * the last segment.
17424                          */
17425                         if (tp->t_flags & TF_NEEDFIN) {
17426                                 sendalot = 4;
17427                         }
17428                 } else {
17429                         mark = 5;
17430                         if (optlen + ipoptlen >= tp->t_maxseg) {
17431                                 /*
17432                                  * Since we don't have enough space to put
17433                                  * the IP header chain and the TCP header in
17434                                  * one packet as required by RFC 7112, don't
17435                                  * send it. Also ensure that at least one
17436                                  * byte of the payload can be put into the
17437                                  * TCP segment.
17438                                  */
17439                                 SOCKBUF_UNLOCK(&so->so_snd);
17440                                 error = EMSGSIZE;
17441                                 sack_rxmit = 0;
17442                                 goto out;
17443                         }
17444                         len = tp->t_maxseg - optlen - ipoptlen;
17445                         sendalot = 5;
17446                 }
17447         } else {
17448                 tso = 0;
17449                 mark = 6;
17450         }
17451         KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
17452                 ("%s: len > IP_MAXPACKET", __func__));
17453 #ifdef DIAGNOSTIC
17454 #ifdef INET6
17455         if (max_linkhdr + hdrlen > MCLBYTES)
17456 #else
17457                 if (max_linkhdr + hdrlen > MHLEN)
17458 #endif
17459                         panic("tcphdr too big");
17460 #endif
17461
17462         /*
17463          * This KASSERT is here to catch edge cases at a well defined place.
17464          * Before, those had triggered (random) panic conditions further
17465          * down.
17466          */
17467         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
17468         if ((len == 0) &&
17469             (flags & TH_FIN) &&
17470             (sbused(sb))) {
17471                 /*
17472                  * We have outstanding data, don't send a fin by itself!.
17473                  */
17474                 goto just_return;
17475         }
17476         /*
17477          * Grab a header mbuf, attaching a copy of data to be transmitted,
17478          * and initialize the header from the template for sends on this
17479          * connection.
17480          */
17481         hw_tls = (sb->sb_flags & SB_TLS_IFNET) != 0;
17482         if (len) {
17483                 uint32_t max_val;
17484                 uint32_t moff;
17485
17486                 if (rack->r_ctl.rc_pace_max_segs)
17487                         max_val = rack->r_ctl.rc_pace_max_segs;
17488                 else if (rack->rc_user_set_max_segs)
17489                         max_val = rack->rc_user_set_max_segs * segsiz;
17490                 else
17491                         max_val = len;
17492                 /*
17493                  * We allow a limit on sending with hptsi.
17494                  */
17495                 if (len > max_val) {
17496                         mark = 7;
17497                         len = max_val;
17498                 }
17499 #ifdef INET6
17500                 if (MHLEN < hdrlen + max_linkhdr)
17501                         m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
17502                 else
17503 #endif
17504                         m = m_gethdr(M_NOWAIT, MT_DATA);
17505
17506                 if (m == NULL) {
17507                         SOCKBUF_UNLOCK(sb);
17508                         error = ENOBUFS;
17509                         sack_rxmit = 0;
17510                         goto out;
17511                 }
17512                 m->m_data += max_linkhdr;
17513                 m->m_len = hdrlen;
17514
17515                 /*
17516                  * Start the m_copy functions from the closest mbuf to the
17517                  * sb_offset in the socket buffer chain.
17518                  */
17519                 mb = sbsndptr_noadv(sb, sb_offset, &moff);
17520                 s_mb = mb;
17521                 s_moff = moff;
17522                 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
17523                         m_copydata(mb, moff, (int)len,
17524                                    mtod(m, caddr_t)+hdrlen);
17525                         if (SEQ_LT(tp->snd_nxt, tp->snd_max))
17526                                 sbsndptr_adv(sb, mb, len);
17527                         m->m_len += len;
17528                 } else {
17529                         struct sockbuf *msb;
17530
17531                         if (SEQ_LT(tp->snd_nxt, tp->snd_max))
17532                                 msb = NULL;
17533                         else
17534                                 msb = sb;
17535                         m->m_next = tcp_m_copym(
17536                                 mb, moff, &len,
17537                                 if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb,
17538                                 ((rsm == NULL) ? hw_tls : 0)
17539 #ifdef NETFLIX_COPY_ARGS
17540                                 , &filled_all
17541 #endif
17542                                 );
17543                         if (len <= (tp->t_maxseg - optlen)) {
17544                                 /*
17545                                  * Must have ran out of mbufs for the copy
17546                                  * shorten it to no longer need tso. Lets
17547                                  * not put on sendalot since we are low on
17548                                  * mbufs.
17549                                  */
17550                                 tso = 0;
17551                         }
17552                         if (m->m_next == NULL) {
17553                                 SOCKBUF_UNLOCK(sb);
17554                                 (void)m_free(m);
17555                                 error = ENOBUFS;
17556                                 sack_rxmit = 0;
17557                                 goto out;
17558                         }
17559                 }
17560                 if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
17561                         if (rsm && (rsm->r_flags & RACK_TLP)) {
17562                                 /*
17563                                  * TLP should not count in retran count, but
17564                                  * in its own bin
17565                                  */
17566                                 counter_u64_add(rack_tlp_retran, 1);
17567                                 counter_u64_add(rack_tlp_retran_bytes, len);
17568                         } else {
17569                                 tp->t_sndrexmitpack++;
17570                                 KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
17571                                 KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
17572                         }
17573 #ifdef STATS
17574                         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
17575                                                  len);
17576 #endif
17577                 } else {
17578                         KMOD_TCPSTAT_INC(tcps_sndpack);
17579                         KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
17580 #ifdef STATS
17581                         stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
17582                                                  len);
17583 #endif
17584                 }
17585                 /*
17586                  * If we're sending everything we've got, set PUSH. (This
17587                  * will keep happy those implementations which only give
17588                  * data to the user when a buffer fills or a PUSH comes in.)
17589                  */
17590                 if (sb_offset + len == sbused(sb) &&
17591                     sbused(sb) &&
17592                     !(flags & TH_SYN)) {
17593                         flags |= TH_PUSH;
17594                         add_flag |= RACK_HAD_PUSH;
17595                 }
17596
17597                 SOCKBUF_UNLOCK(sb);
17598         } else {
17599                 SOCKBUF_UNLOCK(sb);
17600                 if (tp->t_flags & TF_ACKNOW)
17601                         KMOD_TCPSTAT_INC(tcps_sndacks);
17602                 else if (flags & (TH_SYN | TH_FIN | TH_RST))
17603                         KMOD_TCPSTAT_INC(tcps_sndctrl);
17604                 else
17605                         KMOD_TCPSTAT_INC(tcps_sndwinup);
17606
17607                 m = m_gethdr(M_NOWAIT, MT_DATA);
17608                 if (m == NULL) {
17609                         error = ENOBUFS;
17610                         sack_rxmit = 0;
17611                         goto out;
17612                 }
17613 #ifdef INET6
17614                 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
17615                     MHLEN >= hdrlen) {
17616                         M_ALIGN(m, hdrlen);
17617                 } else
17618 #endif
17619                         m->m_data += max_linkhdr;
17620                 m->m_len = hdrlen;
17621         }
17622         SOCKBUF_UNLOCK_ASSERT(sb);
17623         m->m_pkthdr.rcvif = (struct ifnet *)0;
17624 #ifdef MAC
17625         mac_inpcb_create_mbuf(inp, m);
17626 #endif
17627         if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) &&  rack->r_fsb_inited) {
17628 #ifdef INET6
17629                 if (isipv6)
17630                         ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
17631                 else
17632 #endif                          /* INET6 */
17633                         ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
17634                 th = rack->r_ctl.fsb.th;
17635                 udp = rack->r_ctl.fsb.udp;
17636                 if (udp) {
17637                         if (isipv6)
17638                                 ulen = hdrlen + len - sizeof(struct ip6_hdr);
17639                         else
17640                                 ulen = hdrlen + len - sizeof(struct ip);
17641                         udp->uh_ulen = htons(ulen);
17642                 }
17643         } else {
17644 #ifdef INET6
17645                 if (isipv6) {
17646                         ip6 = mtod(m, struct ip6_hdr *);
17647                         if (tp->t_port) {
17648                                 udp = (struct udphdr *)((caddr_t)ip6 + ipoptlen + sizeof(struct ip6_hdr));
17649                                 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
17650                                 udp->uh_dport = tp->t_port;
17651                                 ulen = hdrlen + len - sizeof(struct ip6_hdr);
17652                                 udp->uh_ulen = htons(ulen);
17653                                 th = (struct tcphdr *)(udp + 1);
17654                         } else
17655                                 th = (struct tcphdr *)(ip6 + 1);
17656                         tcpip_fillheaders(inp, tp->t_port, ip6, th);
17657                 } else
17658 #endif                          /* INET6 */
17659                 {
17660                         ip = mtod(m, struct ip *);
17661 #ifdef TCPDEBUG
17662                         ipov = (struct ipovly *)ip;
17663 #endif
17664                         if (tp->t_port) {
17665                                 udp = (struct udphdr *)((caddr_t)ip + ipoptlen + sizeof(struct ip));
17666                                 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
17667                                 udp->uh_dport = tp->t_port;
17668                                 ulen = hdrlen + len - sizeof(struct ip);
17669                                 udp->uh_ulen = htons(ulen);
17670                                 th = (struct tcphdr *)(udp + 1);
17671                         } else
17672                                 th = (struct tcphdr *)(ip + 1);
17673                         tcpip_fillheaders(inp, tp->t_port, ip, th);
17674                 }
17675         }
17676         /*
17677          * Fill in fields, remembering maximum advertised window for use in
17678          * delaying messages about window sizes. If resending a FIN, be sure
17679          * not to use a new sequence number.
17680          */
17681         if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
17682             tp->snd_nxt == tp->snd_max)
17683                 tp->snd_nxt--;
17684         /*
17685          * If we are starting a connection, send ECN setup SYN packet. If we
17686          * are on a retransmit, we may resend those bits a number of times
17687          * as per RFC 3168.
17688          */
17689         if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) {
17690                 if (tp->t_rxtshift >= 1) {
17691                         if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
17692                                 flags |= TH_ECE | TH_CWR;
17693                 } else
17694                         flags |= TH_ECE | TH_CWR;
17695         }
17696         /* Handle parallel SYN for ECN */
17697         if ((tp->t_state == TCPS_SYN_RECEIVED) &&
17698             (tp->t_flags2 & TF2_ECN_SND_ECE)) {
17699                 flags |= TH_ECE;
17700                 tp->t_flags2 &= ~TF2_ECN_SND_ECE;
17701         }
17702         if (TCPS_HAVEESTABLISHED(tp->t_state) &&
17703             (tp->t_flags2 & TF2_ECN_PERMIT)) {
17704                 /*
17705                  * If the peer has ECN, mark data packets with ECN capable
17706                  * transmission (ECT). Ignore pure ack packets,
17707                  * retransmissions.
17708                  */
17709                 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
17710                     (sack_rxmit == 0)) {
17711 #ifdef INET6
17712                         if (isipv6)
17713                                 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
17714                         else
17715 #endif
17716                                 ip->ip_tos |= IPTOS_ECN_ECT0;
17717                         KMOD_TCPSTAT_INC(tcps_ecn_ect0);
17718                         /*
17719                          * Reply with proper ECN notifications.
17720                          * Only set CWR on new data segments.
17721                          */
17722                         if (tp->t_flags2 & TF2_ECN_SND_CWR) {
17723                                 flags |= TH_CWR;
17724                                 tp->t_flags2 &= ~TF2_ECN_SND_CWR;
17725                         }
17726                 }
17727                 if (tp->t_flags2 & TF2_ECN_SND_ECE)
17728                         flags |= TH_ECE;
17729         }
17730         /*
17731          * If we are doing retransmissions, then snd_nxt will not reflect
17732          * the first unsent octet.  For ACK only packets, we do not want the
17733          * sequence number of the retransmitted packet, we want the sequence
17734          * number of the next unsent octet.  So, if there is no data (and no
17735          * SYN or FIN), use snd_max instead of snd_nxt when filling in
17736          * ti_seq.  But if we are in persist state, snd_max might reflect
17737          * one byte beyond the right edge of the window, so use snd_nxt in
17738          * that case, since we know we aren't doing a retransmission.
17739          * (retransmit and persist are mutually exclusive...)
17740          */
17741         if (sack_rxmit == 0) {
17742                 if (len || (flags & (TH_SYN | TH_FIN))) {
17743                         th->th_seq = htonl(tp->snd_nxt);
17744                         rack_seq = tp->snd_nxt;
17745                 } else {
17746                         th->th_seq = htonl(tp->snd_max);
17747                         rack_seq = tp->snd_max;
17748                 }
17749         } else {
17750                 th->th_seq = htonl(rsm->r_start);
17751                 rack_seq = rsm->r_start;
17752         }
17753         th->th_ack = htonl(tp->rcv_nxt);
17754         th->th_flags = flags;
17755         /*
17756          * Calculate receive window.  Don't shrink window, but avoid silly
17757          * window syndrome.
17758          * If a RST segment is sent, advertise a window of zero.
17759          */
17760         if (flags & TH_RST) {
17761                 recwin = 0;
17762         } else {
17763                 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
17764                     recwin < (long)segsiz) {
17765                         recwin = 0;
17766                 }
17767                 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
17768                     recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
17769                         recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
17770         }
17771
17772         /*
17773          * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
17774          * <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK> case is
17775          * handled in syncache.
17776          */
17777         if (flags & TH_SYN)
17778                 th->th_win = htons((u_short)
17779                                    (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
17780         else {
17781                 /* Avoid shrinking window with window scaling. */
17782                 recwin = roundup2(recwin, 1 << tp->rcv_scale);
17783                 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
17784         }
17785         /*
17786          * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
17787          * window.  This may cause the remote transmitter to stall.  This
17788          * flag tells soreceive() to disable delayed acknowledgements when
17789          * draining the buffer.  This can occur if the receiver is
17790          * attempting to read more data than can be buffered prior to
17791          * transmitting on the connection.
17792          */
17793         if (th->th_win == 0) {
17794                 tp->t_sndzerowin++;
17795                 tp->t_flags |= TF_RXWIN0SENT;
17796         } else
17797                 tp->t_flags &= ~TF_RXWIN0SENT;
17798         tp->snd_up = tp->snd_una;       /* drag it along, its deprecated */
17799         /* Now are we using fsb?, if so copy the template data to the mbuf */
17800         if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) && rack->r_fsb_inited) {
17801                 uint8_t *cpto;
17802
17803                 cpto = mtod(m, uint8_t *);
17804                 memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
17805                 /*
17806                  * We have just copied in:
17807                  * IP/IP6
17808                  * <optional udphdr>
17809                  * tcphdr (no options)
17810                  *
17811                  * We need to grab the correct pointers into the mbuf
17812                  * for both the tcp header, and possibly the udp header (if tunneling).
17813                  * We do this by using the offset in the copy buffer and adding it
17814                  * to the mbuf base pointer (cpto).
17815                  */
17816 #ifdef INET6
17817                 if (isipv6)
17818                         ip6 = mtod(m, struct ip6_hdr *);
17819                 else
17820 #endif                          /* INET6 */
17821                         ip = mtod(m, struct ip *);
17822                 th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
17823                 /* If we have a udp header lets set it into the mbuf as well */
17824                 if (udp)
17825                         udp = (struct udphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.udp - rack->r_ctl.fsb.tcp_ip_hdr));
17826         }
17827 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
17828         if (to.to_flags & TOF_SIGNATURE) {
17829                 /*
17830                  * Calculate MD5 signature and put it into the place
17831                  * determined before.
17832                  * NOTE: since TCP options buffer doesn't point into
17833                  * mbuf's data, calculate offset and use it.
17834                  */
17835                 if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
17836                                                        (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
17837                         /*
17838                          * Do not send segment if the calculation of MD5
17839                          * digest has failed.
17840                          */
17841                         goto out;
17842                 }
17843         }
17844 #endif
17845         if (optlen) {
17846                 bcopy(opt, th + 1, optlen);
17847                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
17848         }
17849         /*
17850          * Put TCP length in extended header, and then checksum extended
17851          * header and data.
17852          */
17853         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
17854 #ifdef INET6
17855         if (isipv6) {
17856                 /*
17857                  * ip6_plen is not need to be filled now, and will be filled
17858                  * in ip6_output.
17859                  */
17860                 if (tp->t_port) {
17861                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
17862                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
17863                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
17864                         th->th_sum = htons(0);
17865                         UDPSTAT_INC(udps_opackets);
17866                 } else {
17867                         m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
17868                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
17869                         th->th_sum = in6_cksum_pseudo(ip6,
17870                                                       sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
17871                                                       0);
17872                 }
17873         }
17874 #endif
17875 #if defined(INET6) && defined(INET)
17876         else
17877 #endif
17878 #ifdef INET
17879         {
17880                 if (tp->t_port) {
17881                         m->m_pkthdr.csum_flags = CSUM_UDP;
17882                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
17883                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
17884                                                 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
17885                         th->th_sum = htons(0);
17886                         UDPSTAT_INC(udps_opackets);
17887                 } else {
17888                         m->m_pkthdr.csum_flags = CSUM_TCP;
17889                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
17890                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
17891                                                ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
17892                                                                         IPPROTO_TCP + len + optlen));
17893                 }
17894                 /* IP version must be set here for ipv4/ipv6 checking later */
17895                 KASSERT(ip->ip_v == IPVERSION,
17896                         ("%s: IP version incorrect: %d", __func__, ip->ip_v));
17897         }
17898 #endif
17899         /*
17900          * Enable TSO and specify the size of the segments. The TCP pseudo
17901          * header checksum is always provided. XXX: Fixme: This is currently
17902          * not the case for IPv6.
17903          */
17904         if (tso) {
17905                 KASSERT(len > tp->t_maxseg - optlen,
17906                         ("%s: len <= tso_segsz", __func__));
17907                 m->m_pkthdr.csum_flags |= CSUM_TSO;
17908                 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
17909         }
17910         KASSERT(len + hdrlen == m_length(m, NULL),
17911                 ("%s: mbuf chain different than expected: %d + %u != %u",
17912                  __func__, len, hdrlen, m_length(m, NULL)));
17913
17914 #ifdef TCP_HHOOK
17915         /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
17916         hhook_run_tcp_est_out(tp, th, &to, len, tso);
17917 #endif
17918         /* We're getting ready to send; log now. */
17919         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
17920                 union tcp_log_stackspecific log;
17921
17922                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
17923                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
17924                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
17925                 if (rack->rack_no_prr)
17926                         log.u_bbr.flex1 = 0;
17927                 else
17928                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
17929                 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
17930                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
17931                 log.u_bbr.flex4 = orig_len;
17932                 if (filled_all)
17933                         log.u_bbr.flex5 = 0x80000000;
17934                 else
17935                         log.u_bbr.flex5 = 0;
17936                 /* Save off the early/late values */
17937                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
17938                 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
17939                 log.u_bbr.bw_inuse = rack_get_bw(rack);
17940                 if (rsm || sack_rxmit) {
17941                         if (doing_tlp)
17942                                 log.u_bbr.flex8 = 2;
17943                         else
17944                                 log.u_bbr.flex8 = 1;
17945                 } else {
17946                         log.u_bbr.flex8 = 0;
17947                 }
17948                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
17949                 log.u_bbr.flex7 = mark;
17950                 log.u_bbr.flex7 <<= 8;
17951                 log.u_bbr.flex7 |= pass;
17952                 log.u_bbr.pkts_out = tp->t_maxseg;
17953                 log.u_bbr.timeStamp = cts;
17954                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
17955                 log.u_bbr.lt_epoch = cwnd_to_use;
17956                 log.u_bbr.delivered = sendalot;
17957                 lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
17958                                      len, &log, false, NULL, NULL, 0, &tv);
17959         } else
17960                 lgb = NULL;
17961
17962         /*
17963          * Fill in IP length and desired time to live and send to IP level.
17964          * There should be a better way to handle ttl and tos; we could keep
17965          * them in the template, but need a way to checksum without them.
17966          */
17967         /*
17968          * m->m_pkthdr.len should have been set before cksum calcuration,
17969          * because in6_cksum() need it.
17970          */
17971 #ifdef INET6
17972         if (isipv6) {
17973                 /*
17974                  * we separately set hoplimit for every segment, since the
17975                  * user might want to change the value via setsockopt. Also,
17976                  * desired default hop limit might be changed via Neighbor
17977                  * Discovery.
17978                  */
17979                 rack->r_ctl.fsb.hoplimit = ip6->ip6_hlim = in6_selecthlim(inp, NULL);
17980
17981                 /*
17982                  * Set the packet size here for the benefit of DTrace
17983                  * probes. ip6_output() will set it properly; it's supposed
17984                  * to include the option header lengths as well.
17985                  */
17986                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
17987
17988                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
17989                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
17990                 else
17991                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
17992
17993                 if (tp->t_state == TCPS_SYN_SENT)
17994                         TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
17995
17996                 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
17997                 /* TODO: IPv6 IP6TOS_ECT bit on */
17998                 error = ip6_output(m,
17999 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18000                                    inp->in6p_outputopts,
18001 #else
18002                                    NULL,
18003 #endif
18004                                    &inp->inp_route6,
18005                                    ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0),
18006                                    NULL, NULL, inp);
18007
18008                 if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
18009                         mtu = inp->inp_route6.ro_nh->nh_mtu;
18010         }
18011 #endif                          /* INET6 */
18012 #if defined(INET) && defined(INET6)
18013         else
18014 #endif
18015 #ifdef INET
18016         {
18017                 ip->ip_len = htons(m->m_pkthdr.len);
18018 #ifdef INET6
18019                 if (inp->inp_vflag & INP_IPV6PROTO)
18020                         ip->ip_ttl = in6_selecthlim(inp, NULL);
18021 #endif                          /* INET6 */
18022                 rack->r_ctl.fsb.hoplimit = ip->ip_ttl;
18023                 /*
18024                  * If we do path MTU discovery, then we set DF on every
18025                  * packet. This might not be the best thing to do according
18026                  * to RFC3390 Section 2. However the tcp hostcache migitates
18027                  * the problem so it affects only the first tcp connection
18028                  * with a host.
18029                  *
18030                  * NB: Don't set DF on small MTU/MSS to have a safe
18031                  * fallback.
18032                  */
18033                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
18034                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18035                         if (tp->t_port == 0 || len < V_tcp_minmss) {
18036                                 ip->ip_off |= htons(IP_DF);
18037                         }
18038                 } else {
18039                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18040                 }
18041
18042                 if (tp->t_state == TCPS_SYN_SENT)
18043                         TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
18044
18045                 TCP_PROBE5(send, NULL, tp, ip, tp, th);
18046
18047                 error = ip_output(m,
18048 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18049                                   inp->inp_options,
18050 #else
18051                                   NULL,
18052 #endif
18053                                   &inp->inp_route,
18054                                   ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0,
18055                                   inp);
18056                 if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
18057                         mtu = inp->inp_route.ro_nh->nh_mtu;
18058         }
18059 #endif                          /* INET */
18060
18061 out:
18062         if (lgb) {
18063                 lgb->tlb_errno = error;
18064                 lgb = NULL;
18065         }
18066         /*
18067          * In transmit state, time the transmission and arrange for the
18068          * retransmit.  In persist state, just set snd_max.
18069          */
18070         if (error == 0) {
18071                 rack->forced_ack = 0;   /* If we send something zap the FA flag */
18072                 if (rsm && (doing_tlp == 0)) {
18073                         /* Set we retransmitted */
18074                         rack->rc_gp_saw_rec = 1;
18075                 } else {
18076                         if (cwnd_to_use > tp->snd_ssthresh) {
18077                                 /* Set we sent in CA */
18078                                 rack->rc_gp_saw_ca = 1;
18079                         } else {
18080                                 /* Set we sent in SS */
18081                                 rack->rc_gp_saw_ss = 1;
18082                         }
18083                 }
18084                 if (TCPS_HAVEESTABLISHED(tp->t_state) &&
18085                     (tp->t_flags & TF_SACK_PERMIT) &&
18086                     tp->rcv_numsacks > 0)
18087                         tcp_clean_dsack_blocks(tp);
18088                 tot_len_this_send += len;
18089                 if (len == 0)
18090                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1);
18091                 else if (len == 1) {
18092                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1);
18093                 } else if (len > 1) {
18094                         int idx;
18095
18096                         idx = (len / segsiz) + 3;
18097                         if (idx >= TCP_MSS_ACCT_ATIMER)
18098                                 counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
18099                         else
18100                                 counter_u64_add(rack_out_size[idx], 1);
18101                 }
18102         }
18103         if ((rack->rack_no_prr == 0) &&
18104             sub_from_prr &&
18105             (error == 0)) {
18106                 if (rack->r_ctl.rc_prr_sndcnt >= len)
18107                         rack->r_ctl.rc_prr_sndcnt -= len;
18108                 else
18109                         rack->r_ctl.rc_prr_sndcnt = 0;
18110         }
18111         sub_from_prr = 0;
18112         if (doing_tlp && (rsm == NULL)) {
18113                 /* New send doing a TLP */
18114                 add_flag |= RACK_TLP;
18115                 tp->t_sndtlppack++;
18116                 tp->t_sndtlpbyte += len;
18117         }
18118         rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error,
18119                         rack_to_usec_ts(&tv),
18120                         rsm, add_flag, s_mb, s_moff);
18121
18122
18123         if ((error == 0) &&
18124             (len > 0) &&
18125             (tp->snd_una == tp->snd_max))
18126                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
18127         {
18128                 tcp_seq startseq = tp->snd_nxt;
18129
18130                 /* Track our lost count */
18131                 if (rsm && (doing_tlp == 0))
18132                         rack->r_ctl.rc_loss_count += rsm->r_end - rsm->r_start;
18133                 /*
18134                  * Advance snd_nxt over sequence space of this segment.
18135                  */
18136                 if (error)
18137                         /* We don't log or do anything with errors */
18138                         goto nomore;
18139                 if (doing_tlp == 0) {
18140                         if (rsm == NULL) {
18141                                 /*
18142                                  * Not a retransmission of some
18143                                  * sort, new data is going out so
18144                                  * clear our TLP count and flag.
18145                                  */
18146                                 rack->rc_tlp_in_progress = 0;
18147                                 rack->r_ctl.rc_tlp_cnt_out = 0;
18148                         }
18149                 } else {
18150                         /*
18151                          * We have just sent a TLP, mark that it is true
18152                          * and make sure our in progress is set so we
18153                          * continue to check the count.
18154                          */
18155                         rack->rc_tlp_in_progress = 1;
18156                         rack->r_ctl.rc_tlp_cnt_out++;
18157                 }
18158                 if (flags & (TH_SYN | TH_FIN)) {
18159                         if (flags & TH_SYN)
18160                                 tp->snd_nxt++;
18161                         if (flags & TH_FIN) {
18162                                 tp->snd_nxt++;
18163                                 tp->t_flags |= TF_SENTFIN;
18164                         }
18165                 }
18166                 /* In the ENOBUFS case we do *not* update snd_max */
18167                 if (sack_rxmit)
18168                         goto nomore;
18169
18170                 tp->snd_nxt += len;
18171                 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
18172                         if (tp->snd_una == tp->snd_max) {
18173                                 /*
18174                                  * Update the time we just added data since
18175                                  * none was outstanding.
18176                                  */
18177                                 rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
18178                                 tp->t_acktime = ticks;
18179                         }
18180                         tp->snd_max = tp->snd_nxt;
18181                         /*
18182                          * Time this transmission if not a retransmission and
18183                          * not currently timing anything.
18184                          * This is only relevant in case of switching back to
18185                          * the base stack.
18186                          */
18187                         if (tp->t_rtttime == 0) {
18188                                 tp->t_rtttime = ticks;
18189                                 tp->t_rtseq = startseq;
18190                                 KMOD_TCPSTAT_INC(tcps_segstimed);
18191                         }
18192                         if (len &&
18193                             ((tp->t_flags & TF_GPUTINPROG) == 0))
18194                                 rack_start_gp_measurement(tp, rack, startseq, sb_offset);
18195                 }
18196                 /*
18197                  * If we are doing FO we need to update the mbuf position and subtract
18198                  * this happens when the peer sends us duplicate information and
18199                  * we thus want to send a DSACK.
18200                  *
18201                  * XXXRRS: This brings to mind a ?, when we send a DSACK block is TSO
18202                  * turned off? If not then we are going to echo multiple DSACK blocks
18203                  * out (with the TSO), which we should not be doing.
18204                  */
18205                 if (rack->r_fast_output && len) {
18206                         if (rack->r_ctl.fsb.left_to_send > len)
18207                                 rack->r_ctl.fsb.left_to_send -= len;
18208                         else
18209                                 rack->r_ctl.fsb.left_to_send = 0;
18210                         if (rack->r_ctl.fsb.left_to_send < segsiz)
18211                                 rack->r_fast_output = 0;
18212                         if (rack->r_fast_output) {
18213                                 rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
18214                                 rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
18215                         }
18216                 }
18217         }
18218 nomore:
18219         if (error) {
18220                 rack->r_ctl.rc_agg_delayed = 0;
18221                 rack->r_early = 0;
18222                 rack->r_late = 0;
18223                 rack->r_ctl.rc_agg_early = 0;
18224                 SOCKBUF_UNLOCK_ASSERT(sb);      /* Check gotos. */
18225                 /*
18226                  * Failures do not advance the seq counter above. For the
18227                  * case of ENOBUFS we will fall out and retry in 1ms with
18228                  * the hpts. Everything else will just have to retransmit
18229                  * with the timer.
18230                  *
18231                  * In any case, we do not want to loop around for another
18232                  * send without a good reason.
18233                  */
18234                 sendalot = 0;
18235                 switch (error) {
18236                 case EPERM:
18237                         tp->t_softerror = error;
18238 #ifdef TCP_ACCOUNTING
18239                         crtsc = get_cyclecount();
18240                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18241                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
18242                         }
18243                         counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
18244                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18245                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
18246                         }
18247                         counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
18248                         sched_unpin();
18249 #endif
18250                         return (error);
18251                 case ENOBUFS:
18252                         /*
18253                          * Pace us right away to retry in a some
18254                          * time
18255                          */
18256                         slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
18257                         if (rack->rc_enobuf < 0x7f)
18258                                 rack->rc_enobuf++;
18259                         if (slot < (10 * HPTS_USEC_IN_MSEC))
18260                                 slot = 10 * HPTS_USEC_IN_MSEC;
18261                         if (rack->r_ctl.crte != NULL) {
18262                                 counter_u64_add(rack_saw_enobuf_hw, 1);
18263                                 tcp_rl_log_enobuf(rack->r_ctl.crte);
18264                         }
18265                         counter_u64_add(rack_saw_enobuf, 1);
18266                         goto enobufs;
18267                 case EMSGSIZE:
18268                         /*
18269                          * For some reason the interface we used initially
18270                          * to send segments changed to another or lowered
18271                          * its MTU. If TSO was active we either got an
18272                          * interface without TSO capabilits or TSO was
18273                          * turned off. If we obtained mtu from ip_output()
18274                          * then update it and try again.
18275                          */
18276                         if (tso)
18277                                 tp->t_flags &= ~TF_TSO;
18278                         if (mtu != 0) {
18279                                 tcp_mss_update(tp, -1, mtu, NULL, NULL);
18280                                 goto again;
18281                         }
18282                         slot = 10 * HPTS_USEC_IN_MSEC;
18283                         rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
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 ENETUNREACH:
18298                         counter_u64_add(rack_saw_enetunreach, 1);
18299                 case EHOSTDOWN:
18300                 case EHOSTUNREACH:
18301                 case ENETDOWN:
18302                         if (TCPS_HAVERCVDSYN(tp->t_state)) {
18303                                 tp->t_softerror = error;
18304                         }
18305                         /* FALLTHROUGH */
18306                 default:
18307                         slot = 10 * HPTS_USEC_IN_MSEC;
18308                         rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
18309 #ifdef TCP_ACCOUNTING
18310                         crtsc = get_cyclecount();
18311                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18312                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
18313                         }
18314                         counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
18315                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18316                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
18317                         }
18318                         counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
18319                         sched_unpin();
18320 #endif
18321                         return (error);
18322                 }
18323         } else {
18324                 rack->rc_enobuf = 0;
18325                 if (IN_FASTRECOVERY(tp->t_flags) && rsm)
18326                         rack->r_ctl.retran_during_recovery += len;
18327         }
18328         KMOD_TCPSTAT_INC(tcps_sndtotal);
18329
18330         /*
18331          * Data sent (as far as we can tell). If this advertises a larger
18332          * window than any other segment, then remember the size of the
18333          * advertised window. Any pending ACK has now been sent.
18334          */
18335         if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
18336                 tp->rcv_adv = tp->rcv_nxt + recwin;
18337
18338         tp->last_ack_sent = tp->rcv_nxt;
18339         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
18340 enobufs:
18341         if (sendalot) {
18342                 /* Do we need to turn off sendalot? */
18343                 if (rack->r_ctl.rc_pace_max_segs &&
18344                     (tot_len_this_send >= rack->r_ctl.rc_pace_max_segs)) {
18345                         /* We hit our max. */
18346                         sendalot = 0;
18347                 } else if ((rack->rc_user_set_max_segs) &&
18348                            (tot_len_this_send >= (rack->rc_user_set_max_segs * segsiz))) {
18349                         /* We hit the user defined max */
18350                         sendalot = 0;
18351                 }
18352         }
18353         if ((error == 0) && (flags & TH_FIN))
18354                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN);
18355         if (flags & TH_RST) {
18356                 /*
18357                  * We don't send again after sending a RST.
18358                  */
18359                 slot = 0;
18360                 sendalot = 0;
18361                 if (error == 0)
18362                         tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
18363         } else if ((slot == 0) && (sendalot == 0) && tot_len_this_send) {
18364                 /*
18365                  * Get our pacing rate, if an error
18366                  * occurred in sending (ENOBUF) we would
18367                  * hit the else if with slot preset. Other
18368                  * errors return.
18369                  */
18370                 slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, rsm, segsiz);
18371         }
18372         if (rsm &&
18373             (rsm->r_flags & RACK_HAS_SYN) == 0 &&
18374             rack->use_rack_rr) {
18375                 /* Its a retransmit and we use the rack cheat? */
18376                 if ((slot == 0) ||
18377                     (rack->rc_always_pace == 0) ||
18378                     (rack->r_rr_config == 1)) {
18379                         /*
18380                          * We have no pacing set or we
18381                          * are using old-style rack or
18382                          * we are overriden to use the old 1ms pacing.
18383                          */
18384                         slot = rack->r_ctl.rc_min_to;
18385                 }
18386         }
18387         /* We have sent clear the flag */
18388         rack->r_ent_rec_ns = 0;
18389         if (rack->r_must_retran) {
18390                 if (rsm) {
18391                         rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
18392                         if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
18393                                 /*
18394                                  * We have retransmitted all.
18395                                  */
18396                                 rack->r_must_retran = 0;
18397                                 rack->r_ctl.rc_out_at_rto = 0;
18398                         }
18399                 } else if (SEQ_GEQ(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
18400                         /*
18401                          * Sending new data will also kill
18402                          * the loop.
18403                          */
18404                         rack->r_must_retran = 0;
18405                         rack->r_ctl.rc_out_at_rto = 0;
18406                 }
18407         }
18408         rack->r_ctl.fsb.recwin = recwin;
18409         if ((tp->t_flags & (TF_WASCRECOVERY|TF_WASFRECOVERY)) &&
18410             SEQ_GT(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
18411                 /*
18412                  * We hit an RTO and now have past snd_max at the RTO
18413                  * clear all the WAS flags.
18414                  */
18415                 tp->t_flags &= ~(TF_WASCRECOVERY|TF_WASFRECOVERY);
18416         }
18417         if (slot) {
18418                 /* set the rack tcb into the slot N */
18419                 counter_u64_add(rack_paced_segments, 1);
18420                 if ((error == 0) &&
18421                     rack_use_rfo &&
18422                     ((flags & (TH_SYN|TH_FIN)) == 0) &&
18423                     (rsm == NULL) &&
18424                     (tp->snd_nxt == tp->snd_max) &&
18425                     (ipoptlen == 0) &&
18426                     (tp->rcv_numsacks == 0) &&
18427                     rack->r_fsb_inited &&
18428                     TCPS_HAVEESTABLISHED(tp->t_state) &&
18429                     (rack->r_must_retran == 0) &&
18430                     ((tp->t_flags & TF_NEEDFIN) == 0) &&
18431                     (len > 0) && (orig_len > 0) &&
18432                     (orig_len > len) &&
18433                     ((orig_len - len) >= segsiz) &&
18434                     ((optlen == 0) ||
18435                      ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
18436                         /* We can send at least one more MSS using our fsb */
18437
18438                         rack->r_fast_output = 1;
18439                         rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
18440                         rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
18441                         rack->r_ctl.fsb.tcp_flags = flags;
18442                         rack->r_ctl.fsb.left_to_send = orig_len - len;
18443                         KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
18444                                 ("rack:%p left_to_send:%u sbavail:%u out:%u",
18445                                  rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
18446                                  (tp->snd_max - tp->snd_una)));
18447                         if (rack->r_ctl.fsb.left_to_send < segsiz)
18448                                 rack->r_fast_output = 0;
18449                         else {
18450                                 if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
18451                                         rack->r_ctl.fsb.rfo_apply_push = 1;
18452                                 else
18453                                         rack->r_ctl.fsb.rfo_apply_push = 0;
18454                         }
18455                 } else
18456                         rack->r_fast_output = 0;
18457                 rack_log_fsb(rack, tp, so, flags,
18458                              ipoptlen, orig_len, len, error,
18459                              (rsm == NULL), optlen, __LINE__, 2);
18460         } else if (sendalot) {
18461                 int ret;
18462
18463                 if (len)
18464                         counter_u64_add(rack_unpaced_segments, 1);
18465                 sack_rxmit = 0;
18466                 if ((error == 0) &&
18467                     rack_use_rfo &&
18468                     ((flags & (TH_SYN|TH_FIN)) == 0) &&
18469                     (rsm == NULL) &&
18470                     (ipoptlen == 0) &&
18471                     (tp->rcv_numsacks == 0) &&
18472                     (tp->snd_nxt == tp->snd_max) &&
18473                     (rack->r_must_retran == 0) &&
18474                     rack->r_fsb_inited &&
18475                     TCPS_HAVEESTABLISHED(tp->t_state) &&
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 use fast_output for more */
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                         }
18496                         if (rack->r_fast_output) {
18497                                 if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
18498                                         rack->r_ctl.fsb.rfo_apply_push = 1;
18499                                 else
18500                                         rack->r_ctl.fsb.rfo_apply_push = 0;
18501                                 rack_log_fsb(rack, tp, so, flags,
18502                                              ipoptlen, orig_len, len, error,
18503                                              (rsm == NULL), optlen, __LINE__, 3);
18504                                 error = 0;
18505                                 ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
18506                                 if (ret >= 0)
18507                                         return (ret);
18508                                 else if (error)
18509                                         goto nomore;
18510
18511                         }
18512                 }
18513                 goto again;
18514         } else if (len) {
18515                 counter_u64_add(rack_unpaced_segments, 1);
18516         }
18517         /* Assure when we leave that snd_nxt will point to top */
18518         if (SEQ_GT(tp->snd_max, tp->snd_nxt))
18519                 tp->snd_nxt = tp->snd_max;
18520         rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, 0);
18521 #ifdef TCP_ACCOUNTING
18522         crtsc = get_cyclecount() - ts_val;
18523         if (tot_len_this_send) {
18524                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18525                         tp->tcp_cnt_counters[SND_OUT_DATA]++;
18526                 }
18527                 counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], 1);
18528                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18529                         tp->tcp_proc_time[SND_OUT_DATA] += crtsc;
18530                 }
18531                 counter_u64_add(tcp_proc_time[SND_OUT_DATA], crtsc);
18532                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18533                         tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) /segsiz);
18534                 }
18535                 counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len_this_send + segsiz - 1) /segsiz));
18536         } else {
18537                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18538                         tp->tcp_cnt_counters[SND_OUT_ACK]++;
18539                 }
18540                 counter_u64_add(tcp_cnt_counters[SND_OUT_ACK], 1);
18541                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18542                         tp->tcp_proc_time[SND_OUT_ACK] += crtsc;
18543                 }
18544                 counter_u64_add(tcp_proc_time[SND_OUT_ACK], crtsc);
18545         }
18546         sched_unpin();
18547 #endif
18548         if (error == ENOBUFS)
18549                 error = 0;
18550         return (error);
18551 }
18552
18553 static void
18554 rack_update_seg(struct tcp_rack *rack)
18555 {
18556         uint32_t orig_val;
18557
18558         orig_val = rack->r_ctl.rc_pace_max_segs;
18559         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
18560         if (orig_val != rack->r_ctl.rc_pace_max_segs)
18561                 rack_log_pacing_delay_calc(rack, 0, 0, orig_val, 0, 0, 15, __LINE__, NULL);
18562 }
18563
18564 static void
18565 rack_mtu_change(struct tcpcb *tp)
18566 {
18567         /*
18568          * The MSS may have changed
18569          */
18570         struct tcp_rack *rack;
18571
18572         rack = (struct tcp_rack *)tp->t_fb_ptr;
18573         if (rack->r_ctl.rc_pace_min_segs != ctf_fixed_maxseg(tp)) {
18574                 /*
18575                  * The MTU has changed we need to resend everything
18576                  * since all we have sent is lost. We first fix
18577                  * up the mtu though.
18578                  */
18579                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
18580                 /* We treat this like a full retransmit timeout without the cwnd adjustment */
18581                 rack_remxt_tmr(tp);
18582                 rack->r_fast_output = 0;
18583                 rack->r_ctl.rc_out_at_rto = ctf_flight_size(tp,
18584                                                 rack->r_ctl.rc_sacked);
18585                 rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
18586                 rack->r_must_retran = 1;
18587
18588         }
18589         sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
18590         /* We don't use snd_nxt to retransmit */
18591         tp->snd_nxt = tp->snd_max;
18592 }
18593
18594 static int
18595 rack_set_profile(struct tcp_rack *rack, int prof)
18596 {
18597         int err = EINVAL;
18598         if (prof == 1) {
18599                 /* pace_always=1 */
18600                 if (rack->rc_always_pace == 0) {
18601                         if (tcp_can_enable_pacing() == 0)
18602                                 return (EBUSY);
18603                 }
18604                 rack->rc_always_pace = 1;
18605                 if (rack->use_fixed_rate || rack->gp_ready)
18606                         rack_set_cc_pacing(rack);
18607                 rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
18608                 rack->rack_attempt_hdwr_pace = 0;
18609                 /* cmpack=1 */
18610                 if (rack_use_cmp_acks)
18611                         rack->r_use_cmp_ack = 1;
18612                 if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
18613                     rack->r_use_cmp_ack)
18614                         rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
18615                 /* scwnd=1 */
18616                 rack->rack_enable_scwnd = 1;
18617                 /* dynamic=100 */
18618                 rack->rc_gp_dyn_mul = 1;
18619                 /* gp_inc_ca */
18620                 rack->r_ctl.rack_per_of_gp_ca = 100;
18621                 /* rrr_conf=3 */
18622                 rack->r_rr_config = 3;
18623                 /* npush=2 */
18624                 rack->r_ctl.rc_no_push_at_mrtt = 2;
18625                 /* fillcw=1 */
18626                 rack->rc_pace_to_cwnd = 1;
18627                 rack->rc_pace_fill_if_rttin_range = 0;
18628                 rack->rtt_limit_mul = 0;
18629                 /* noprr=1 */
18630                 rack->rack_no_prr = 1;
18631                 /* lscwnd=1 */
18632                 rack->r_limit_scw = 1;
18633                 /* gp_inc_rec */
18634                 rack->r_ctl.rack_per_of_gp_rec = 90;
18635                 err = 0;
18636
18637         } else if (prof == 3) {
18638                 /* Same as profile one execept fill_cw becomes 2 (less aggressive set) */
18639                 /* pace_always=1 */
18640                 if (rack->rc_always_pace == 0) {
18641                         if (tcp_can_enable_pacing() == 0)
18642                                 return (EBUSY);
18643                 }
18644                 rack->rc_always_pace = 1;
18645                 if (rack->use_fixed_rate || rack->gp_ready)
18646                         rack_set_cc_pacing(rack);
18647                 rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
18648                 rack->rack_attempt_hdwr_pace = 0;
18649                 /* cmpack=1 */
18650                 if (rack_use_cmp_acks)
18651                         rack->r_use_cmp_ack = 1;
18652                 if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
18653                     rack->r_use_cmp_ack)
18654                         rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
18655                 /* scwnd=1 */
18656                 rack->rack_enable_scwnd = 1;
18657                 /* dynamic=100 */
18658                 rack->rc_gp_dyn_mul = 1;
18659                 /* gp_inc_ca */
18660                 rack->r_ctl.rack_per_of_gp_ca = 100;
18661                 /* rrr_conf=3 */
18662                 rack->r_rr_config = 3;
18663                 /* npush=2 */
18664                 rack->r_ctl.rc_no_push_at_mrtt = 2;
18665                 /* fillcw=2 */
18666                 rack->rc_pace_to_cwnd = 1;
18667                 rack->r_fill_less_agg = 1;
18668                 rack->rc_pace_fill_if_rttin_range = 0;
18669                 rack->rtt_limit_mul = 0;
18670                 /* noprr=1 */
18671                 rack->rack_no_prr = 1;
18672                 /* lscwnd=1 */
18673                 rack->r_limit_scw = 1;
18674                 /* gp_inc_rec */
18675                 rack->r_ctl.rack_per_of_gp_rec = 90;
18676                 err = 0;
18677
18678
18679         } else if (prof == 2) {
18680                 /* cmpack=1 */
18681                 if (rack->rc_always_pace == 0) {
18682                         if (tcp_can_enable_pacing() == 0)
18683                                 return (EBUSY);
18684                 }
18685                 rack->rc_always_pace = 1;
18686                 if (rack->use_fixed_rate || rack->gp_ready)
18687                         rack_set_cc_pacing(rack);
18688                 rack->r_use_cmp_ack = 1;
18689                 if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
18690                         rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
18691                 /* pace_always=1 */
18692                 rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
18693                 /* scwnd=1 */
18694                 rack->rack_enable_scwnd = 1;
18695                 /* dynamic=100 */
18696                 rack->rc_gp_dyn_mul = 1;
18697                 rack->r_ctl.rack_per_of_gp_ca = 100;
18698                 /* rrr_conf=3 */
18699                 rack->r_rr_config = 3;
18700                 /* npush=2 */
18701                 rack->r_ctl.rc_no_push_at_mrtt = 2;
18702                 /* fillcw=1 */
18703                 rack->rc_pace_to_cwnd = 1;
18704                 rack->rc_pace_fill_if_rttin_range = 0;
18705                 rack->rtt_limit_mul = 0;
18706                 /* noprr=1 */
18707                 rack->rack_no_prr = 1;
18708                 /* lscwnd=0 */
18709                 rack->r_limit_scw = 0;
18710                 err = 0;
18711         } else if (prof == 0) {
18712                 /* This changes things back to the default settings */
18713                 err = 0;
18714                 if (rack->rc_always_pace) {
18715                         tcp_decrement_paced_conn();
18716                         rack_undo_cc_pacing(rack);
18717                         rack->rc_always_pace = 0;
18718                 }
18719                 if (rack_pace_every_seg && tcp_can_enable_pacing()) {
18720                         rack->rc_always_pace = 1;
18721                         if (rack->use_fixed_rate || rack->gp_ready)
18722                                 rack_set_cc_pacing(rack);
18723                 } else
18724                         rack->rc_always_pace = 0;
18725                 if (rack_use_cmp_acks)
18726                         rack->r_use_cmp_ack = 1;
18727                 else
18728                         rack->r_use_cmp_ack = 0;
18729                 if (rack_disable_prr)
18730                         rack->rack_no_prr = 1;
18731                 else
18732                         rack->rack_no_prr = 0;
18733                 if (rack_gp_no_rec_chg)
18734                         rack->rc_gp_no_rec_chg = 1;
18735                 else
18736                         rack->rc_gp_no_rec_chg = 0;
18737                 if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack) {
18738                         rack->r_mbuf_queue = 1;
18739                         if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
18740                                 rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
18741                         rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
18742                 } else {
18743                         rack->r_mbuf_queue = 0;
18744                         rack->rc_inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
18745                 }
18746                 if (rack_enable_shared_cwnd)
18747                         rack->rack_enable_scwnd = 1;
18748                 else
18749                         rack->rack_enable_scwnd = 0;
18750                 if (rack_do_dyn_mul) {
18751                         /* When dynamic adjustment is on CA needs to start at 100% */
18752                         rack->rc_gp_dyn_mul = 1;
18753                         if (rack_do_dyn_mul >= 100)
18754                                 rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
18755                 } else {
18756                         rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
18757                         rack->rc_gp_dyn_mul = 0;
18758                 }
18759                 rack->r_rr_config = 0;
18760                 rack->r_ctl.rc_no_push_at_mrtt = 0;
18761                 rack->rc_pace_to_cwnd = 0;
18762                 rack->rc_pace_fill_if_rttin_range = 0;
18763                 rack->rtt_limit_mul = 0;
18764
18765                 if (rack_enable_hw_pacing)
18766                         rack->rack_hdw_pace_ena = 1;
18767                 else
18768                         rack->rack_hdw_pace_ena = 0;
18769                 if (rack_disable_prr)
18770                         rack->rack_no_prr = 1;
18771                 else
18772                         rack->rack_no_prr = 0;
18773                 if (rack_limits_scwnd)
18774                         rack->r_limit_scw  = 1;
18775                 else
18776                         rack->r_limit_scw  = 0;
18777                 err = 0;
18778         }
18779         return (err);
18780 }
18781
18782 static int
18783 rack_add_deferred_option(struct tcp_rack *rack, int sopt_name, uint64_t loptval)
18784 {
18785         struct deferred_opt_list *dol;
18786
18787         dol = malloc(sizeof(struct deferred_opt_list),
18788                      M_TCPFSB, M_NOWAIT|M_ZERO);
18789         if (dol == NULL) {
18790                 /*
18791                  * No space yikes -- fail out..
18792                  */
18793                 return (0);
18794         }
18795         dol->optname = sopt_name;
18796         dol->optval = loptval;
18797         TAILQ_INSERT_TAIL(&rack->r_ctl.opt_list, dol, next);
18798         return (1);
18799 }
18800
18801 static int
18802 rack_process_option(struct tcpcb *tp, struct tcp_rack *rack, int sopt_name,
18803                     uint32_t optval, uint64_t loptval)
18804 {
18805         struct epoch_tracker et;
18806         struct sockopt sopt;
18807         struct cc_newreno_opts opt;
18808         uint64_t val;
18809         int error = 0;
18810         uint16_t ca, ss;
18811
18812         switch (sopt_name) {
18813
18814         case TCP_RACK_PACING_BETA:
18815                 RACK_OPTS_INC(tcp_rack_beta);
18816                 if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
18817                         /* This only works for newreno. */
18818                         error = EINVAL;
18819                         break;
18820                 }
18821                 if (rack->rc_pacing_cc_set) {
18822                         /*
18823                          * Set them into the real CC module
18824                          * whats in the rack pcb is the old values
18825                          * to be used on restoral/
18826                          */
18827                         sopt.sopt_dir = SOPT_SET;
18828                         opt.name = CC_NEWRENO_BETA;
18829                         opt.val = optval;
18830                         if (CC_ALGO(tp)->ctl_output != NULL)
18831                                 error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
18832                         else {
18833                                 error = ENOENT;
18834                                 break;
18835                         }
18836                 } else {
18837                         /*
18838                          * Not pacing yet so set it into our local
18839                          * rack pcb storage.
18840                          */
18841                         rack->r_ctl.rc_saved_beta.beta = optval;
18842                 }
18843                 break;
18844         case TCP_RACK_PACING_BETA_ECN:
18845                 RACK_OPTS_INC(tcp_rack_beta_ecn);
18846                 if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
18847                         /* This only works for newreno. */
18848                         error = EINVAL;
18849                         break;
18850                 }
18851                 if (rack->rc_pacing_cc_set) {
18852                         /*
18853                          * Set them into the real CC module
18854                          * whats in the rack pcb is the old values
18855                          * to be used on restoral/
18856                          */
18857                         sopt.sopt_dir = SOPT_SET;
18858                         opt.name = CC_NEWRENO_BETA_ECN;
18859                         opt.val = optval;
18860                         if (CC_ALGO(tp)->ctl_output != NULL)
18861                                 error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
18862                         else
18863                                 error = ENOENT;
18864                 } else {
18865                         /*
18866                          * Not pacing yet so set it into our local
18867                          * rack pcb storage.
18868                          */
18869                         rack->r_ctl.rc_saved_beta.beta_ecn = optval;
18870                         rack->r_ctl.rc_saved_beta.newreno_flags = CC_NEWRENO_BETA_ECN;
18871                 }
18872                 break;
18873         case TCP_DEFER_OPTIONS:
18874                 RACK_OPTS_INC(tcp_defer_opt);
18875                 if (optval) {
18876                         if (rack->gp_ready) {
18877                                 /* Too late */
18878                                 error = EINVAL;
18879                                 break;
18880                         }
18881                         rack->defer_options = 1;
18882                 } else
18883                         rack->defer_options = 0;
18884                 break;
18885         case TCP_RACK_MEASURE_CNT:
18886                 RACK_OPTS_INC(tcp_rack_measure_cnt);
18887                 if (optval && (optval <= 0xff)) {
18888                         rack->r_ctl.req_measurements = optval;
18889                 } else
18890                         error = EINVAL;
18891                 break;
18892         case TCP_REC_ABC_VAL:
18893                 RACK_OPTS_INC(tcp_rec_abc_val);
18894                 if (optval > 0)
18895                         rack->r_use_labc_for_rec = 1;
18896                 else
18897                         rack->r_use_labc_for_rec = 0;
18898                 break;
18899         case TCP_RACK_ABC_VAL:
18900                 RACK_OPTS_INC(tcp_rack_abc_val);
18901                 if ((optval > 0) && (optval < 255))
18902                         rack->rc_labc = optval;
18903                 else
18904                         error = EINVAL;
18905                 break;
18906         case TCP_HDWR_UP_ONLY:
18907                 RACK_OPTS_INC(tcp_pacing_up_only);
18908                 if (optval)
18909                         rack->r_up_only = 1;
18910                 else
18911                         rack->r_up_only = 0;
18912                 break;
18913         case TCP_PACING_RATE_CAP:
18914                 RACK_OPTS_INC(tcp_pacing_rate_cap);
18915                 rack->r_ctl.bw_rate_cap = loptval;
18916                 break;
18917         case TCP_RACK_PROFILE:
18918                 RACK_OPTS_INC(tcp_profile);
18919                 error = rack_set_profile(rack, optval);
18920                 break;
18921         case TCP_USE_CMP_ACKS:
18922                 RACK_OPTS_INC(tcp_use_cmp_acks);
18923                 if ((optval == 0) && (rack->rc_inp->inp_flags2 & INP_MBUF_ACKCMP)) {
18924                         /* You can't turn it off once its on! */
18925                         error = EINVAL;
18926                 } else if ((optval == 1) && (rack->r_use_cmp_ack == 0)) {
18927                         rack->r_use_cmp_ack = 1;
18928                         rack->r_mbuf_queue = 1;
18929                         tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
18930                 }
18931                 if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
18932                         rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
18933                 break;
18934         case TCP_SHARED_CWND_TIME_LIMIT:
18935                 RACK_OPTS_INC(tcp_lscwnd);
18936                 if (optval)
18937                         rack->r_limit_scw = 1;
18938                 else
18939                         rack->r_limit_scw = 0;
18940                 break;
18941         case TCP_RACK_PACE_TO_FILL:
18942                 RACK_OPTS_INC(tcp_fillcw);
18943                 if (optval == 0)
18944                         rack->rc_pace_to_cwnd = 0;
18945                 else {
18946                         rack->rc_pace_to_cwnd = 1;
18947                         if (optval > 1)
18948                                 rack->r_fill_less_agg = 1;
18949                 }
18950                 if ((optval >= rack_gp_rtt_maxmul) &&
18951                     rack_gp_rtt_maxmul &&
18952                     (optval < 0xf)) {
18953                         rack->rc_pace_fill_if_rttin_range = 1;
18954                         rack->rtt_limit_mul = optval;
18955                 } else {
18956                         rack->rc_pace_fill_if_rttin_range = 0;
18957                         rack->rtt_limit_mul = 0;
18958                 }
18959                 break;
18960         case TCP_RACK_NO_PUSH_AT_MAX:
18961                 RACK_OPTS_INC(tcp_npush);
18962                 if (optval == 0)
18963                         rack->r_ctl.rc_no_push_at_mrtt = 0;
18964                 else if (optval < 0xff)
18965                         rack->r_ctl.rc_no_push_at_mrtt = optval;
18966                 else
18967                         error = EINVAL;
18968                 break;
18969         case TCP_SHARED_CWND_ENABLE:
18970                 RACK_OPTS_INC(tcp_rack_scwnd);
18971                 if (optval == 0)
18972                         rack->rack_enable_scwnd = 0;
18973                 else
18974                         rack->rack_enable_scwnd = 1;
18975                 break;
18976         case TCP_RACK_MBUF_QUEUE:
18977                 /* Now do we use the LRO mbuf-queue feature */
18978                 RACK_OPTS_INC(tcp_rack_mbufq);
18979                 if (optval || rack->r_use_cmp_ack)
18980                         rack->r_mbuf_queue = 1;
18981                 else
18982                         rack->r_mbuf_queue = 0;
18983                 if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
18984                         tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
18985                 else
18986                         tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
18987                 break;
18988         case TCP_RACK_NONRXT_CFG_RATE:
18989                 RACK_OPTS_INC(tcp_rack_cfg_rate);
18990                 if (optval == 0)
18991                         rack->rack_rec_nonrxt_use_cr = 0;
18992                 else
18993                         rack->rack_rec_nonrxt_use_cr = 1;
18994                 break;
18995         case TCP_NO_PRR:
18996                 RACK_OPTS_INC(tcp_rack_noprr);
18997                 if (optval == 0)
18998                         rack->rack_no_prr = 0;
18999                 else if (optval == 1)
19000                         rack->rack_no_prr = 1;
19001                 else if (optval == 2)
19002                         rack->no_prr_addback = 1;
19003                 else
19004                         error = EINVAL;
19005                 break;
19006         case TCP_TIMELY_DYN_ADJ:
19007                 RACK_OPTS_INC(tcp_timely_dyn);
19008                 if (optval == 0)
19009                         rack->rc_gp_dyn_mul = 0;
19010                 else {
19011                         rack->rc_gp_dyn_mul = 1;
19012                         if (optval >= 100) {
19013                                 /*
19014                                  * If the user sets something 100 or more
19015                                  * its the gp_ca value.
19016                                  */
19017                                 rack->r_ctl.rack_per_of_gp_ca  = optval;
19018                         }
19019                 }
19020                 break;
19021         case TCP_RACK_DO_DETECTION:
19022                 RACK_OPTS_INC(tcp_rack_do_detection);
19023                 if (optval == 0)
19024                         rack->do_detection = 0;
19025                 else
19026                         rack->do_detection = 1;
19027                 break;
19028         case TCP_RACK_TLP_USE:
19029                 if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) {
19030                         error = EINVAL;
19031                         break;
19032                 }
19033                 RACK_OPTS_INC(tcp_tlp_use);
19034                 rack->rack_tlp_threshold_use = optval;
19035                 break;
19036         case TCP_RACK_TLP_REDUCE:
19037                 /* RACK TLP cwnd reduction (bool) */
19038                 RACK_OPTS_INC(tcp_rack_tlp_reduce);
19039                 rack->r_ctl.rc_tlp_cwnd_reduce = optval;
19040                 break;
19041         /*  Pacing related ones */
19042         case TCP_RACK_PACE_ALWAYS:
19043                 /*
19044                  * zero is old rack method, 1 is new
19045                  * method using a pacing rate.
19046                  */
19047                 RACK_OPTS_INC(tcp_rack_pace_always);
19048                 if (optval > 0) {
19049                         if (rack->rc_always_pace) {
19050                                 error = EALREADY;
19051                                 break;
19052                         } else if (tcp_can_enable_pacing()) {
19053                                 rack->rc_always_pace = 1;
19054                                 if (rack->use_fixed_rate || rack->gp_ready)
19055                                         rack_set_cc_pacing(rack);
19056                         }
19057                         else {
19058                                 error = ENOSPC;
19059                                 break;
19060                         }
19061                 } else {
19062                         if (rack->rc_always_pace) {
19063                                 tcp_decrement_paced_conn();
19064                                 rack->rc_always_pace = 0;
19065                                 rack_undo_cc_pacing(rack);
19066                         }
19067                 }
19068                 if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
19069                         tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19070                 else
19071                         tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19072                 /* A rate may be set irate or other, if so set seg size */
19073                 rack_update_seg(rack);
19074                 break;
19075         case TCP_BBR_RACK_INIT_RATE:
19076                 RACK_OPTS_INC(tcp_initial_rate);
19077                 val = optval;
19078                 /* Change from kbits per second to bytes per second */
19079                 val *= 1000;
19080                 val /= 8;
19081                 rack->r_ctl.init_rate = val;
19082                 if (rack->rc_init_win != rack_default_init_window) {
19083                         uint32_t win, snt;
19084
19085                         /*
19086                          * Options don't always get applied
19087                          * in the order you think. So in order
19088                          * to assure we update a cwnd we need
19089                          * to check and see if we are still
19090                          * where we should raise the cwnd.
19091                          */
19092                         win = rc_init_window(rack);
19093                         if (SEQ_GT(tp->snd_max, tp->iss))
19094                                 snt = tp->snd_max - tp->iss;
19095                         else
19096                                 snt = 0;
19097                         if ((snt < win) &&
19098                             (tp->snd_cwnd < win))
19099                                 tp->snd_cwnd = win;
19100                 }
19101                 if (rack->rc_always_pace)
19102                         rack_update_seg(rack);
19103                 break;
19104         case TCP_BBR_IWINTSO:
19105                 RACK_OPTS_INC(tcp_initial_win);
19106                 if (optval && (optval <= 0xff)) {
19107                         uint32_t win, snt;
19108
19109                         rack->rc_init_win = optval;
19110                         win = rc_init_window(rack);
19111                         if (SEQ_GT(tp->snd_max, tp->iss))
19112                                 snt = tp->snd_max - tp->iss;
19113                         else
19114                                 snt = 0;
19115                         if ((snt < win) &&
19116                             (tp->t_srtt |
19117 #ifdef NETFLIX_PEAKRATE
19118                              tp->t_maxpeakrate |
19119 #endif
19120                              rack->r_ctl.init_rate)) {
19121                                 /*
19122                                  * We are not past the initial window
19123                                  * and we have some bases for pacing,
19124                                  * so we need to possibly adjust up
19125                                  * the cwnd. Note even if we don't set
19126                                  * the cwnd, its still ok to raise the rc_init_win
19127                                  * which can be used coming out of idle when we
19128                                  * would have a rate.
19129                                  */
19130                                 if (tp->snd_cwnd < win)
19131                                         tp->snd_cwnd = win;
19132                         }
19133                         if (rack->rc_always_pace)
19134                                 rack_update_seg(rack);
19135                 } else
19136                         error = EINVAL;
19137                 break;
19138         case TCP_RACK_FORCE_MSEG:
19139                 RACK_OPTS_INC(tcp_rack_force_max_seg);
19140                 if (optval)
19141                         rack->rc_force_max_seg = 1;
19142                 else
19143                         rack->rc_force_max_seg = 0;
19144                 break;
19145         case TCP_RACK_PACE_MAX_SEG:
19146                 /* Max segments size in a pace in bytes */
19147                 RACK_OPTS_INC(tcp_rack_max_seg);
19148                 rack->rc_user_set_max_segs = optval;
19149                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
19150                 break;
19151         case TCP_RACK_PACE_RATE_REC:
19152                 /* Set the fixed pacing rate in Bytes per second ca */
19153                 RACK_OPTS_INC(tcp_rack_pace_rate_rec);
19154                 rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19155                 if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
19156                         rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19157                 if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
19158                         rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19159                 rack->use_fixed_rate = 1;
19160                 if (rack->rc_always_pace)
19161                         rack_set_cc_pacing(rack);
19162                 rack_log_pacing_delay_calc(rack,
19163                                            rack->r_ctl.rc_fixed_pacing_rate_ss,
19164                                            rack->r_ctl.rc_fixed_pacing_rate_ca,
19165                                            rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19166                                            __LINE__, NULL);
19167                 break;
19168
19169         case TCP_RACK_PACE_RATE_SS:
19170                 /* Set the fixed pacing rate in Bytes per second ca */
19171                 RACK_OPTS_INC(tcp_rack_pace_rate_ss);
19172                 rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19173                 if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
19174                         rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19175                 if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
19176                         rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19177                 rack->use_fixed_rate = 1;
19178                 if (rack->rc_always_pace)
19179                         rack_set_cc_pacing(rack);
19180                 rack_log_pacing_delay_calc(rack,
19181                                            rack->r_ctl.rc_fixed_pacing_rate_ss,
19182                                            rack->r_ctl.rc_fixed_pacing_rate_ca,
19183                                            rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19184                                            __LINE__, NULL);
19185                 break;
19186
19187         case TCP_RACK_PACE_RATE_CA:
19188                 /* Set the fixed pacing rate in Bytes per second ca */
19189                 RACK_OPTS_INC(tcp_rack_pace_rate_ca);
19190                 rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19191                 if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
19192                         rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19193                 if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
19194                         rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19195                 rack->use_fixed_rate = 1;
19196                 if (rack->rc_always_pace)
19197                         rack_set_cc_pacing(rack);
19198                 rack_log_pacing_delay_calc(rack,
19199                                            rack->r_ctl.rc_fixed_pacing_rate_ss,
19200                                            rack->r_ctl.rc_fixed_pacing_rate_ca,
19201                                            rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19202                                            __LINE__, NULL);
19203                 break;
19204         case TCP_RACK_GP_INCREASE_REC:
19205                 RACK_OPTS_INC(tcp_gp_inc_rec);
19206                 rack->r_ctl.rack_per_of_gp_rec = optval;
19207                 rack_log_pacing_delay_calc(rack,
19208                                            rack->r_ctl.rack_per_of_gp_ss,
19209                                            rack->r_ctl.rack_per_of_gp_ca,
19210                                            rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
19211                                            __LINE__, NULL);
19212                 break;
19213         case TCP_RACK_GP_INCREASE_CA:
19214                 RACK_OPTS_INC(tcp_gp_inc_ca);
19215                 ca = optval;
19216                 if (ca < 100) {
19217                         /*
19218                          * We don't allow any reduction
19219                          * over the GP b/w.
19220                          */
19221                         error = EINVAL;
19222                         break;
19223                 }
19224                 rack->r_ctl.rack_per_of_gp_ca = ca;
19225                 rack_log_pacing_delay_calc(rack,
19226                                            rack->r_ctl.rack_per_of_gp_ss,
19227                                            rack->r_ctl.rack_per_of_gp_ca,
19228                                            rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
19229                                            __LINE__, NULL);
19230                 break;
19231         case TCP_RACK_GP_INCREASE_SS:
19232                 RACK_OPTS_INC(tcp_gp_inc_ss);
19233                 ss = optval;
19234                 if (ss < 100) {
19235                         /*
19236                          * We don't allow any reduction
19237                          * over the GP b/w.
19238                          */
19239                         error = EINVAL;
19240                         break;
19241                 }
19242                 rack->r_ctl.rack_per_of_gp_ss = ss;
19243                 rack_log_pacing_delay_calc(rack,
19244                                            rack->r_ctl.rack_per_of_gp_ss,
19245                                            rack->r_ctl.rack_per_of_gp_ca,
19246                                            rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
19247                                            __LINE__, NULL);
19248                 break;
19249         case TCP_RACK_RR_CONF:
19250                 RACK_OPTS_INC(tcp_rack_rrr_no_conf_rate);
19251                 if (optval && optval <= 3)
19252                         rack->r_rr_config = optval;
19253                 else
19254                         rack->r_rr_config = 0;
19255                 break;
19256         case TCP_HDWR_RATE_CAP:
19257                 RACK_OPTS_INC(tcp_hdwr_rate_cap);
19258                 if (optval) {
19259                         if (rack->r_rack_hw_rate_caps == 0)
19260                                 rack->r_rack_hw_rate_caps = 1;
19261                         else
19262                                 error = EALREADY;
19263                 } else {
19264                         rack->r_rack_hw_rate_caps = 0;
19265                 }
19266                 break;
19267         case TCP_BBR_HDWR_PACE:
19268                 RACK_OPTS_INC(tcp_hdwr_pacing);
19269                 if (optval){
19270                         if (rack->rack_hdrw_pacing == 0) {
19271                                 rack->rack_hdw_pace_ena = 1;
19272                                 rack->rack_attempt_hdwr_pace = 0;
19273                         } else
19274                                 error = EALREADY;
19275                 } else {
19276                         rack->rack_hdw_pace_ena = 0;
19277 #ifdef RATELIMIT
19278                         if (rack->r_ctl.crte != NULL) {
19279                                 rack->rack_hdrw_pacing = 0;
19280                                 rack->rack_attempt_hdwr_pace = 0;
19281                                 tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
19282                                 rack->r_ctl.crte = NULL;
19283                         }
19284 #endif
19285                 }
19286                 break;
19287         /*  End Pacing related ones */
19288         case TCP_RACK_PRR_SENDALOT:
19289                 /* Allow PRR to send more than one seg */
19290                 RACK_OPTS_INC(tcp_rack_prr_sendalot);
19291                 rack->r_ctl.rc_prr_sendalot = optval;
19292                 break;
19293         case TCP_RACK_MIN_TO:
19294                 /* Minimum time between rack t-o's in ms */
19295                 RACK_OPTS_INC(tcp_rack_min_to);
19296                 rack->r_ctl.rc_min_to = optval;
19297                 break;
19298         case TCP_RACK_EARLY_SEG:
19299                 /* If early recovery max segments */
19300                 RACK_OPTS_INC(tcp_rack_early_seg);
19301                 rack->r_ctl.rc_early_recovery_segs = optval;
19302                 break;
19303         case TCP_RACK_REORD_THRESH:
19304                 /* RACK reorder threshold (shift amount) */
19305                 RACK_OPTS_INC(tcp_rack_reord_thresh);
19306                 if ((optval > 0) && (optval < 31))
19307                         rack->r_ctl.rc_reorder_shift = optval;
19308                 else
19309                         error = EINVAL;
19310                 break;
19311         case TCP_RACK_REORD_FADE:
19312                 /* Does reordering fade after ms time */
19313                 RACK_OPTS_INC(tcp_rack_reord_fade);
19314                 rack->r_ctl.rc_reorder_fade = optval;
19315                 break;
19316         case TCP_RACK_TLP_THRESH:
19317                 /* RACK TLP theshold i.e. srtt+(srtt/N) */
19318                 RACK_OPTS_INC(tcp_rack_tlp_thresh);
19319                 if (optval)
19320                         rack->r_ctl.rc_tlp_threshold = optval;
19321                 else
19322                         error = EINVAL;
19323                 break;
19324         case TCP_BBR_USE_RACK_RR:
19325                 RACK_OPTS_INC(tcp_rack_rr);
19326                 if (optval)
19327                         rack->use_rack_rr = 1;
19328                 else
19329                         rack->use_rack_rr = 0;
19330                 break;
19331         case TCP_FAST_RSM_HACK:
19332                 RACK_OPTS_INC(tcp_rack_fastrsm_hack);
19333                 if (optval)
19334                         rack->fast_rsm_hack = 1;
19335                 else
19336                         rack->fast_rsm_hack = 0;
19337                 break;
19338         case TCP_RACK_PKT_DELAY:
19339                 /* RACK added ms i.e. rack-rtt + reord + N */
19340                 RACK_OPTS_INC(tcp_rack_pkt_delay);
19341                 rack->r_ctl.rc_pkt_delay = optval;
19342                 break;
19343         case TCP_DELACK:
19344                 RACK_OPTS_INC(tcp_rack_delayed_ack);
19345                 if (optval == 0)
19346                         tp->t_delayed_ack = 0;
19347                 else
19348                         tp->t_delayed_ack = 1;
19349                 if (tp->t_flags & TF_DELACK) {
19350                         tp->t_flags &= ~TF_DELACK;
19351                         tp->t_flags |= TF_ACKNOW;
19352                         NET_EPOCH_ENTER(et);
19353                         rack_output(tp);
19354                         NET_EPOCH_EXIT(et);
19355                 }
19356                 break;
19357
19358         case TCP_BBR_RACK_RTT_USE:
19359                 RACK_OPTS_INC(tcp_rack_rtt_use);
19360                 if ((optval != USE_RTT_HIGH) &&
19361                     (optval != USE_RTT_LOW) &&
19362                     (optval != USE_RTT_AVG))
19363                         error = EINVAL;
19364                 else
19365                         rack->r_ctl.rc_rate_sample_method = optval;
19366                 break;
19367         case TCP_DATA_AFTER_CLOSE:
19368                 RACK_OPTS_INC(tcp_data_after_close);
19369                 if (optval)
19370                         rack->rc_allow_data_af_clo = 1;
19371                 else
19372                         rack->rc_allow_data_af_clo = 0;
19373                 break;
19374         default:
19375                 break;
19376         }
19377 #ifdef NETFLIX_STATS
19378         tcp_log_socket_option(tp, sopt_name, optval, error);
19379 #endif
19380         return (error);
19381 }
19382
19383
19384 static void
19385 rack_apply_deferred_options(struct tcp_rack *rack)
19386 {
19387         struct deferred_opt_list *dol, *sdol;
19388         uint32_t s_optval;
19389
19390         TAILQ_FOREACH_SAFE(dol, &rack->r_ctl.opt_list, next, sdol) {
19391                 TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
19392                 /* Disadvantage of deferal is you loose the error return */
19393                 s_optval = (uint32_t)dol->optval;
19394                 (void)rack_process_option(rack->rc_tp, rack, dol->optname, s_optval, dol->optval);
19395                 free(dol, M_TCPDO);
19396         }
19397 }
19398
19399 /*
19400  * rack_ctloutput() must drop the inpcb lock before performing copyin on
19401  * socket option arguments.  When it re-acquires the lock after the copy, it
19402  * has to revalidate that the connection is still valid for the socket
19403  * option.
19404  */
19405 static int
19406 rack_set_sockopt(struct socket *so, struct sockopt *sopt,
19407     struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack)
19408 {
19409         uint64_t loptval;
19410         int32_t error = 0, optval;
19411
19412         switch (sopt->sopt_name) {
19413         case TCP_RACK_TLP_REDUCE:               /*  URL:tlp_reduce */
19414         /*  Pacing related ones */
19415         case TCP_RACK_PACE_ALWAYS:              /*  URL:pace_always */
19416         case TCP_BBR_RACK_INIT_RATE:            /*  URL:irate */
19417         case TCP_BBR_IWINTSO:                   /*  URL:tso_iwin */
19418         case TCP_RACK_PACE_MAX_SEG:             /*  URL:pace_max_seg */
19419         case TCP_RACK_FORCE_MSEG:               /*  URL:force_max_seg */
19420         case TCP_RACK_PACE_RATE_CA:             /*  URL:pr_ca */
19421         case TCP_RACK_PACE_RATE_SS:             /*  URL:pr_ss*/
19422         case TCP_RACK_PACE_RATE_REC:            /*  URL:pr_rec */
19423         case TCP_RACK_GP_INCREASE_CA:           /*  URL:gp_inc_ca */
19424         case TCP_RACK_GP_INCREASE_SS:           /*  URL:gp_inc_ss */
19425         case TCP_RACK_GP_INCREASE_REC:          /*  URL:gp_inc_rec */
19426         case TCP_RACK_RR_CONF:                  /*  URL:rrr_conf */
19427         case TCP_BBR_HDWR_PACE:                 /*  URL:hdwrpace */
19428         case TCP_HDWR_RATE_CAP:                 /*  URL: hdwrcap boolean */
19429         case TCP_PACING_RATE_CAP:               /*  URL:cap-- used by side-channel */
19430         case TCP_HDWR_UP_ONLY:                  /*  URL:uponly -- hardware pacing  boolean */
19431        /* End pacing related */
19432         case TCP_FAST_RSM_HACK:                 /*  URL:frsm_hack */
19433         case TCP_DELACK:                        /*  URL:delack (in base TCP i.e. tcp_hints along with cc etc ) */
19434         case TCP_RACK_PRR_SENDALOT:             /*  URL:prr_sendalot */
19435         case TCP_RACK_MIN_TO:                   /*  URL:min_to */
19436         case TCP_RACK_EARLY_SEG:                /*  URL:early_seg */
19437         case TCP_RACK_REORD_THRESH:             /*  URL:reord_thresh */
19438         case TCP_RACK_REORD_FADE:               /*  URL:reord_fade */
19439         case TCP_RACK_TLP_THRESH:               /*  URL:tlp_thresh */
19440         case TCP_RACK_PKT_DELAY:                /*  URL:pkt_delay */
19441         case TCP_RACK_TLP_USE:                  /*  URL:tlp_use */
19442         case TCP_BBR_RACK_RTT_USE:              /*  URL:rttuse */
19443         case TCP_BBR_USE_RACK_RR:               /*  URL:rackrr */
19444         case TCP_RACK_DO_DETECTION:             /*  URL:detect */
19445         case TCP_NO_PRR:                        /*  URL:noprr */
19446         case TCP_TIMELY_DYN_ADJ:                /*  URL:dynamic */
19447         case TCP_DATA_AFTER_CLOSE:              /*  no URL */
19448         case TCP_RACK_NONRXT_CFG_RATE:          /*  URL:nonrxtcr */
19449         case TCP_SHARED_CWND_ENABLE:            /*  URL:scwnd */
19450         case TCP_RACK_MBUF_QUEUE:               /*  URL:mqueue */
19451         case TCP_RACK_NO_PUSH_AT_MAX:           /*  URL:npush */
19452         case TCP_RACK_PACE_TO_FILL:             /*  URL:fillcw */
19453         case TCP_SHARED_CWND_TIME_LIMIT:        /*  URL:lscwnd */
19454         case TCP_RACK_PROFILE:                  /*  URL:profile */
19455         case TCP_USE_CMP_ACKS:                  /*  URL:cmpack */
19456         case TCP_RACK_ABC_VAL:                  /*  URL:labc */
19457         case TCP_REC_ABC_VAL:                   /*  URL:reclabc */
19458         case TCP_RACK_MEASURE_CNT:              /*  URL:measurecnt */
19459         case TCP_DEFER_OPTIONS:                 /*  URL:defer */
19460         case TCP_RACK_PACING_BETA:              /*  URL:pacing_beta */
19461         case TCP_RACK_PACING_BETA_ECN:          /*  URL:pacing_beta_ecn */
19462                 break;
19463         default:
19464                 /* Filter off all unknown options to the base stack */
19465                 return (tcp_default_ctloutput(so, sopt, inp, tp));
19466                 break;
19467         }
19468         INP_WUNLOCK(inp);
19469         if (sopt->sopt_name == TCP_PACING_RATE_CAP) {
19470                 error = sooptcopyin(sopt, &loptval, sizeof(loptval), sizeof(loptval));
19471                 /*
19472                  * We truncate it down to 32 bits for the socket-option trace this
19473                  * means rates > 34Gbps won't show right, but thats probably ok.
19474                  */
19475                 optval = (uint32_t)loptval;
19476         } else {
19477                 error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
19478                 /* Save it in 64 bit form too */
19479                 loptval = optval;
19480         }
19481         if (error)
19482                 return (error);
19483         INP_WLOCK(inp);
19484         if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
19485                 INP_WUNLOCK(inp);
19486                 return (ECONNRESET);
19487         }
19488         if (rack->defer_options && (rack->gp_ready == 0) &&
19489             (sopt->sopt_name != TCP_DEFER_OPTIONS) &&
19490             (sopt->sopt_name != TCP_RACK_PACING_BETA) &&
19491             (sopt->sopt_name != TCP_RACK_PACING_BETA_ECN) &&
19492             (sopt->sopt_name != TCP_RACK_MEASURE_CNT)) {
19493                 /* Options are beind deferred */
19494                 if (rack_add_deferred_option(rack, sopt->sopt_name, loptval)) {
19495                         INP_WUNLOCK(inp);
19496                         return (0);
19497                 } else {
19498                         /* No memory to defer, fail */
19499                         INP_WUNLOCK(inp);
19500                         return (ENOMEM);
19501                 }
19502         }
19503         error = rack_process_option(tp, rack, sopt->sopt_name, optval, loptval);
19504         INP_WUNLOCK(inp);
19505         return (error);
19506 }
19507
19508 static void
19509 rack_fill_info(struct tcpcb *tp, struct tcp_info *ti)
19510 {
19511
19512         INP_WLOCK_ASSERT(tp->t_inpcb);
19513         bzero(ti, sizeof(*ti));
19514
19515         ti->tcpi_state = tp->t_state;
19516         if ((tp->t_flags & TF_REQ_TSTMP) && (tp->t_flags & TF_RCVD_TSTMP))
19517                 ti->tcpi_options |= TCPI_OPT_TIMESTAMPS;
19518         if (tp->t_flags & TF_SACK_PERMIT)
19519                 ti->tcpi_options |= TCPI_OPT_SACK;
19520         if ((tp->t_flags & TF_REQ_SCALE) && (tp->t_flags & TF_RCVD_SCALE)) {
19521                 ti->tcpi_options |= TCPI_OPT_WSCALE;
19522                 ti->tcpi_snd_wscale = tp->snd_scale;
19523                 ti->tcpi_rcv_wscale = tp->rcv_scale;
19524         }
19525         if (tp->t_flags2 & TF2_ECN_PERMIT)
19526                 ti->tcpi_options |= TCPI_OPT_ECN;
19527         if (tp->t_flags & TF_FASTOPEN)
19528                 ti->tcpi_options |= TCPI_OPT_TFO;
19529         /* still kept in ticks is t_rcvtime */
19530         ti->tcpi_last_data_recv = ((uint32_t)ticks - tp->t_rcvtime) * tick;
19531         /* Since we hold everything in precise useconds this is easy */
19532         ti->tcpi_rtt = tp->t_srtt;
19533         ti->tcpi_rttvar = tp->t_rttvar;
19534         ti->tcpi_rto = tp->t_rxtcur;
19535         ti->tcpi_snd_ssthresh = tp->snd_ssthresh;
19536         ti->tcpi_snd_cwnd = tp->snd_cwnd;
19537         /*
19538          * FreeBSD-specific extension fields for tcp_info.
19539          */
19540         ti->tcpi_rcv_space = tp->rcv_wnd;
19541         ti->tcpi_rcv_nxt = tp->rcv_nxt;
19542         ti->tcpi_snd_wnd = tp->snd_wnd;
19543         ti->tcpi_snd_bwnd = 0;          /* Unused, kept for compat. */
19544         ti->tcpi_snd_nxt = tp->snd_nxt;
19545         ti->tcpi_snd_mss = tp->t_maxseg;
19546         ti->tcpi_rcv_mss = tp->t_maxseg;
19547         ti->tcpi_snd_rexmitpack = tp->t_sndrexmitpack;
19548         ti->tcpi_rcv_ooopack = tp->t_rcvoopack;
19549         ti->tcpi_snd_zerowin = tp->t_sndzerowin;
19550 #ifdef NETFLIX_STATS
19551         ti->tcpi_total_tlp = tp->t_sndtlppack;
19552         ti->tcpi_total_tlp_bytes = tp->t_sndtlpbyte;
19553         memcpy(&ti->tcpi_rxsyninfo, &tp->t_rxsyninfo, sizeof(struct tcpsyninfo));
19554 #endif
19555 #ifdef TCP_OFFLOAD
19556         if (tp->t_flags & TF_TOE) {
19557                 ti->tcpi_options |= TCPI_OPT_TOE;
19558                 tcp_offload_tcp_info(tp, ti);
19559         }
19560 #endif
19561 }
19562
19563 static int
19564 rack_get_sockopt(struct socket *so, struct sockopt *sopt,
19565     struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack)
19566 {
19567         int32_t error, optval;
19568         uint64_t val, loptval;
19569         struct  tcp_info ti;
19570         /*
19571          * Because all our options are either boolean or an int, we can just
19572          * pull everything into optval and then unlock and copy. If we ever
19573          * add a option that is not a int, then this will have quite an
19574          * impact to this routine.
19575          */
19576         error = 0;
19577         switch (sopt->sopt_name) {
19578         case TCP_INFO:
19579                 /* First get the info filled */
19580                 rack_fill_info(tp, &ti);
19581                 /* Fix up the rtt related fields if needed */
19582                 INP_WUNLOCK(inp);
19583                 error = sooptcopyout(sopt, &ti, sizeof ti);
19584                 return (error);
19585         /*
19586          * Beta is the congestion control value for NewReno that influences how
19587          * much of a backoff happens when loss is detected. It is normally set
19588          * to 50 for 50% i.e. the cwnd is reduced to 50% of its previous value
19589          * when you exit recovery.
19590          */
19591         case TCP_RACK_PACING_BETA:
19592                 if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0)
19593                         error = EINVAL;
19594                 else if (rack->rc_pacing_cc_set == 0)
19595                         optval = rack->r_ctl.rc_saved_beta.beta;
19596                 else {
19597                         /*
19598                          * Reach out into the CC data and report back what
19599                          * I have previously set. Yeah it looks hackish but
19600                          * we don't want to report the saved values.
19601                          */
19602                         if (tp->ccv->cc_data)
19603                                 optval = ((struct newreno *)tp->ccv->cc_data)->beta;
19604                         else
19605                                 error = EINVAL;
19606                 }
19607                 break;
19608                 /*
19609                  * Beta_ecn is the congestion control value for NewReno that influences how
19610                  * much of a backoff happens when a ECN mark is detected. It is normally set
19611                  * to 80 for 80% i.e. the cwnd is reduced by 20% of its previous value when
19612                  * you exit recovery. Note that classic ECN has a beta of 50, it is only
19613                  * ABE Ecn that uses this "less" value, but we do too with pacing :)
19614                  */
19615
19616         case TCP_RACK_PACING_BETA_ECN:
19617                 if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0)
19618                         error = EINVAL;
19619                 else if (rack->rc_pacing_cc_set == 0)
19620                         optval = rack->r_ctl.rc_saved_beta.beta_ecn;
19621                 else {
19622                         /*
19623                          * Reach out into the CC data and report back what
19624                          * I have previously set. Yeah it looks hackish but
19625                          * we don't want to report the saved values.
19626                          */
19627                         if (tp->ccv->cc_data)
19628                                 optval = ((struct newreno *)tp->ccv->cc_data)->beta_ecn;
19629                         else
19630                                 error = EINVAL;
19631                 }
19632                 break;
19633         case TCP_FAST_RSM_HACK:
19634                 optval = rack->fast_rsm_hack;
19635                 break;
19636         case TCP_DEFER_OPTIONS:
19637                 optval = rack->defer_options;
19638                 break;
19639         case TCP_RACK_MEASURE_CNT:
19640                 optval = rack->r_ctl.req_measurements;
19641                 break;
19642         case TCP_REC_ABC_VAL:
19643                 optval = rack->r_use_labc_for_rec;
19644                 break;
19645         case TCP_RACK_ABC_VAL:
19646                 optval = rack->rc_labc;
19647                 break;
19648         case TCP_HDWR_UP_ONLY:
19649                 optval= rack->r_up_only;
19650                 break;
19651         case TCP_PACING_RATE_CAP:
19652                 loptval = rack->r_ctl.bw_rate_cap;
19653                 break;
19654         case TCP_RACK_PROFILE:
19655                 /* You cannot retrieve a profile, its write only */
19656                 error = EINVAL;
19657                 break;
19658         case TCP_USE_CMP_ACKS:
19659                 optval = rack->r_use_cmp_ack;
19660                 break;
19661         case TCP_RACK_PACE_TO_FILL:
19662                 optval = rack->rc_pace_to_cwnd;
19663                 if (optval && rack->r_fill_less_agg)
19664                         optval++;
19665                 break;
19666         case TCP_RACK_NO_PUSH_AT_MAX:
19667                 optval = rack->r_ctl.rc_no_push_at_mrtt;
19668                 break;
19669         case TCP_SHARED_CWND_ENABLE:
19670                 optval = rack->rack_enable_scwnd;
19671                 break;
19672         case TCP_RACK_NONRXT_CFG_RATE:
19673                 optval = rack->rack_rec_nonrxt_use_cr;
19674                 break;
19675         case TCP_NO_PRR:
19676                 if (rack->rack_no_prr  == 1)
19677                         optval = 1;
19678                 else if (rack->no_prr_addback == 1)
19679                         optval = 2;
19680                 else
19681                         optval = 0;
19682                 break;
19683         case TCP_RACK_DO_DETECTION:
19684                 optval = rack->do_detection;
19685                 break;
19686         case TCP_RACK_MBUF_QUEUE:
19687                 /* Now do we use the LRO mbuf-queue feature */
19688                 optval = rack->r_mbuf_queue;
19689                 break;
19690         case TCP_TIMELY_DYN_ADJ:
19691                 optval = rack->rc_gp_dyn_mul;
19692                 break;
19693         case TCP_BBR_IWINTSO:
19694                 optval = rack->rc_init_win;
19695                 break;
19696         case TCP_RACK_TLP_REDUCE:
19697                 /* RACK TLP cwnd reduction (bool) */
19698                 optval = rack->r_ctl.rc_tlp_cwnd_reduce;
19699                 break;
19700         case TCP_BBR_RACK_INIT_RATE:
19701                 val = rack->r_ctl.init_rate;
19702                 /* convert to kbits per sec */
19703                 val *= 8;
19704                 val /= 1000;
19705                 optval = (uint32_t)val;
19706                 break;
19707         case TCP_RACK_FORCE_MSEG:
19708                 optval = rack->rc_force_max_seg;
19709                 break;
19710         case TCP_RACK_PACE_MAX_SEG:
19711                 /* Max segments in a pace */
19712                 optval = rack->rc_user_set_max_segs;
19713                 break;
19714         case TCP_RACK_PACE_ALWAYS:
19715                 /* Use the always pace method */
19716                 optval = rack->rc_always_pace;
19717                 break;
19718         case TCP_RACK_PRR_SENDALOT:
19719                 /* Allow PRR to send more than one seg */
19720                 optval = rack->r_ctl.rc_prr_sendalot;
19721                 break;
19722         case TCP_RACK_MIN_TO:
19723                 /* Minimum time between rack t-o's in ms */
19724                 optval = rack->r_ctl.rc_min_to;
19725                 break;
19726         case TCP_RACK_EARLY_SEG:
19727                 /* If early recovery max segments */
19728                 optval = rack->r_ctl.rc_early_recovery_segs;
19729                 break;
19730         case TCP_RACK_REORD_THRESH:
19731                 /* RACK reorder threshold (shift amount) */
19732                 optval = rack->r_ctl.rc_reorder_shift;
19733                 break;
19734         case TCP_RACK_REORD_FADE:
19735                 /* Does reordering fade after ms time */
19736                 optval = rack->r_ctl.rc_reorder_fade;
19737                 break;
19738         case TCP_BBR_USE_RACK_RR:
19739                 /* Do we use the rack cheat for rxt */
19740                 optval = rack->use_rack_rr;
19741                 break;
19742         case TCP_RACK_RR_CONF:
19743                 optval = rack->r_rr_config;
19744                 break;
19745         case TCP_HDWR_RATE_CAP:
19746                 optval = rack->r_rack_hw_rate_caps;
19747                 break;
19748         case TCP_BBR_HDWR_PACE:
19749                 optval = rack->rack_hdw_pace_ena;
19750                 break;
19751         case TCP_RACK_TLP_THRESH:
19752                 /* RACK TLP theshold i.e. srtt+(srtt/N) */
19753                 optval = rack->r_ctl.rc_tlp_threshold;
19754                 break;
19755         case TCP_RACK_PKT_DELAY:
19756                 /* RACK added ms i.e. rack-rtt + reord + N */
19757                 optval = rack->r_ctl.rc_pkt_delay;
19758                 break;
19759         case TCP_RACK_TLP_USE:
19760                 optval = rack->rack_tlp_threshold_use;
19761                 break;
19762         case TCP_RACK_PACE_RATE_CA:
19763                 optval = rack->r_ctl.rc_fixed_pacing_rate_ca;
19764                 break;
19765         case TCP_RACK_PACE_RATE_SS:
19766                 optval = rack->r_ctl.rc_fixed_pacing_rate_ss;
19767                 break;
19768         case TCP_RACK_PACE_RATE_REC:
19769                 optval = rack->r_ctl.rc_fixed_pacing_rate_rec;
19770                 break;
19771         case TCP_RACK_GP_INCREASE_SS:
19772                 optval = rack->r_ctl.rack_per_of_gp_ca;
19773                 break;
19774         case TCP_RACK_GP_INCREASE_CA:
19775                 optval = rack->r_ctl.rack_per_of_gp_ss;
19776                 break;
19777         case TCP_BBR_RACK_RTT_USE:
19778                 optval = rack->r_ctl.rc_rate_sample_method;
19779                 break;
19780         case TCP_DELACK:
19781                 optval = tp->t_delayed_ack;
19782                 break;
19783         case TCP_DATA_AFTER_CLOSE:
19784                 optval = rack->rc_allow_data_af_clo;
19785                 break;
19786         case TCP_SHARED_CWND_TIME_LIMIT:
19787                 optval = rack->r_limit_scw;
19788                 break;
19789         default:
19790                 return (tcp_default_ctloutput(so, sopt, inp, tp));
19791                 break;
19792         }
19793         INP_WUNLOCK(inp);
19794         if (error == 0) {
19795                 if (TCP_PACING_RATE_CAP)
19796                         error = sooptcopyout(sopt, &loptval, sizeof loptval);
19797                 else
19798                         error = sooptcopyout(sopt, &optval, sizeof optval);
19799         }
19800         return (error);
19801 }
19802
19803 static int
19804 rack_ctloutput(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp)
19805 {
19806         int32_t error = EINVAL;
19807         struct tcp_rack *rack;
19808
19809         rack = (struct tcp_rack *)tp->t_fb_ptr;
19810         if (rack == NULL) {
19811                 /* Huh? */
19812                 goto out;
19813         }
19814         if (sopt->sopt_dir == SOPT_SET) {
19815                 return (rack_set_sockopt(so, sopt, inp, tp, rack));
19816         } else if (sopt->sopt_dir == SOPT_GET) {
19817                 return (rack_get_sockopt(so, sopt, inp, tp, rack));
19818         }
19819 out:
19820         INP_WUNLOCK(inp);
19821         return (error);
19822 }
19823
19824 static int
19825 rack_pru_options(struct tcpcb *tp, int flags)
19826 {
19827         if (flags & PRUS_OOB)
19828                 return (EOPNOTSUPP);
19829         return (0);
19830 }
19831
19832 static struct tcp_function_block __tcp_rack = {
19833         .tfb_tcp_block_name = __XSTRING(STACKNAME),
19834         .tfb_tcp_output = rack_output,
19835         .tfb_do_queued_segments = ctf_do_queued_segments,
19836         .tfb_do_segment_nounlock = rack_do_segment_nounlock,
19837         .tfb_tcp_do_segment = rack_do_segment,
19838         .tfb_tcp_ctloutput = rack_ctloutput,
19839         .tfb_tcp_fb_init = rack_init,
19840         .tfb_tcp_fb_fini = rack_fini,
19841         .tfb_tcp_timer_stop_all = rack_stopall,
19842         .tfb_tcp_timer_activate = rack_timer_activate,
19843         .tfb_tcp_timer_active = rack_timer_active,
19844         .tfb_tcp_timer_stop = rack_timer_stop,
19845         .tfb_tcp_rexmit_tmr = rack_remxt_tmr,
19846         .tfb_tcp_handoff_ok = rack_handoff_ok,
19847         .tfb_tcp_mtu_chg = rack_mtu_change,
19848         .tfb_pru_options = rack_pru_options,
19849
19850 };
19851
19852 static const char *rack_stack_names[] = {
19853         __XSTRING(STACKNAME),
19854 #ifdef STACKALIAS
19855         __XSTRING(STACKALIAS),
19856 #endif
19857 };
19858
19859 static int
19860 rack_ctor(void *mem, int32_t size, void *arg, int32_t how)
19861 {
19862         memset(mem, 0, size);
19863         return (0);
19864 }
19865
19866 static void
19867 rack_dtor(void *mem, int32_t size, void *arg)
19868 {
19869
19870 }
19871
19872 static bool rack_mod_inited = false;
19873
19874 static int
19875 tcp_addrack(module_t mod, int32_t type, void *data)
19876 {
19877         int32_t err = 0;
19878         int num_stacks;
19879
19880         switch (type) {
19881         case MOD_LOAD:
19882                 rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
19883                     sizeof(struct rack_sendmap),
19884                     rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0);
19885
19886                 rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
19887                     sizeof(struct tcp_rack),
19888                     rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
19889
19890                 sysctl_ctx_init(&rack_sysctl_ctx);
19891                 rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
19892                     SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
19893                     OID_AUTO,
19894 #ifdef STACKALIAS
19895                     __XSTRING(STACKALIAS),
19896 #else
19897                     __XSTRING(STACKNAME),
19898 #endif
19899                     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
19900                     "");
19901                 if (rack_sysctl_root == NULL) {
19902                         printf("Failed to add sysctl node\n");
19903                         err = EFAULT;
19904                         goto free_uma;
19905                 }
19906                 rack_init_sysctls();
19907                 num_stacks = nitems(rack_stack_names);
19908                 err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK,
19909                     rack_stack_names, &num_stacks);
19910                 if (err) {
19911                         printf("Failed to register %s stack name for "
19912                             "%s module\n", rack_stack_names[num_stacks],
19913                             __XSTRING(MODNAME));
19914                         sysctl_ctx_free(&rack_sysctl_ctx);
19915 free_uma:
19916                         uma_zdestroy(rack_zone);
19917                         uma_zdestroy(rack_pcb_zone);
19918                         rack_counter_destroy();
19919                         printf("Failed to register rack module -- err:%d\n", err);
19920                         return (err);
19921                 }
19922                 tcp_lro_reg_mbufq();
19923                 rack_mod_inited = true;
19924                 break;
19925         case MOD_QUIESCE:
19926                 err = deregister_tcp_functions(&__tcp_rack, true, false);
19927                 break;
19928         case MOD_UNLOAD:
19929                 err = deregister_tcp_functions(&__tcp_rack, false, true);
19930                 if (err == EBUSY)
19931                         break;
19932                 if (rack_mod_inited) {
19933                         uma_zdestroy(rack_zone);
19934                         uma_zdestroy(rack_pcb_zone);
19935                         sysctl_ctx_free(&rack_sysctl_ctx);
19936                         rack_counter_destroy();
19937                         rack_mod_inited = false;
19938                 }
19939                 tcp_lro_dereg_mbufq();
19940                 err = 0;
19941                 break;
19942         default:
19943                 return (EOPNOTSUPP);
19944         }
19945         return (err);
19946 }
19947
19948 static moduledata_t tcp_rack = {
19949         .name = __XSTRING(MODNAME),
19950         .evhand = tcp_addrack,
19951         .priv = 0
19952 };
19953
19954 MODULE_VERSION(MODNAME, 1);
19955 DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
19956 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);