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         tp->t_rxtcur = RACK_REXMTVAL(tp);
6610         if (TCPS_HAVEESTABLISHED(tp->t_state)) {
6611                 tp->t_rxtcur += TICKS_2_USEC(tcp_rexmit_slop);
6612         }
6613         if (tp->t_rxtcur > rack_rto_max) {
6614                 tp->t_rxtcur = rack_rto_max;
6615         }
6616 }
6617
6618 static void
6619 rack_cc_conn_init(struct tcpcb *tp)
6620 {
6621         struct tcp_rack *rack;
6622         uint32_t srtt;
6623
6624         rack = (struct tcp_rack *)tp->t_fb_ptr;
6625         srtt = tp->t_srtt;
6626         cc_conn_init(tp);
6627         /*
6628          * Now convert to rack's internal format,
6629          * if required.
6630          */
6631         if ((srtt == 0) && (tp->t_srtt != 0))
6632                 rack_convert_rtts(tp);
6633         /*
6634          * We want a chance to stay in slowstart as
6635          * we create a connection. TCP spec says that
6636          * initially ssthresh is infinite. For our
6637          * purposes that is the snd_wnd.
6638          */
6639         if (tp->snd_ssthresh < tp->snd_wnd) {
6640                 tp->snd_ssthresh = tp->snd_wnd;
6641         }
6642         /*
6643          * We also want to assure a IW worth of
6644          * data can get inflight.
6645          */
6646         if (rc_init_window(rack) < tp->snd_cwnd)
6647                 tp->snd_cwnd = rc_init_window(rack);
6648 }
6649
6650 /*
6651  * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
6652  * we will setup to retransmit the lowest seq number outstanding.
6653  */
6654 static int
6655 rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6656 {
6657         int32_t rexmt;
6658         struct inpcb *inp;
6659         int32_t retval = 0;
6660         bool isipv6;
6661
6662         inp = tp->t_inpcb;
6663         if (tp->t_timers->tt_flags & TT_STOPPED) {
6664                 return (1);
6665         }
6666         if (ctf_progress_timeout_check(tp, false)) {
6667                 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
6668                 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6669                 tcp_set_inp_to_drop(inp, ETIMEDOUT);
6670                 return (1);
6671         }
6672         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
6673         rack->r_ctl.retran_during_recovery = 0;
6674         rack->r_ctl.dsack_byte_cnt = 0;
6675         if (IN_FASTRECOVERY(tp->t_flags))
6676                 tp->t_flags |= TF_WASFRECOVERY;
6677         else
6678                 tp->t_flags &= ~TF_WASFRECOVERY;
6679         if (IN_CONGRECOVERY(tp->t_flags))
6680                 tp->t_flags |= TF_WASCRECOVERY;
6681         else
6682                 tp->t_flags &= ~TF_WASCRECOVERY;
6683         if (TCPS_HAVEESTABLISHED(tp->t_state) &&
6684             (tp->snd_una == tp->snd_max)) {
6685                 /* Nothing outstanding .. nothing to do */
6686                 return (0);
6687         }
6688         /*
6689          * Rack can only run one timer  at a time, so we cannot
6690          * run a KEEPINIT (gating SYN sending) and a retransmit
6691          * timer for the SYN. So if we are in a front state and
6692          * have a KEEPINIT timer we need to check the first transmit
6693          * against now to see if we have exceeded the KEEPINIT time
6694          * (if one is set).
6695          */
6696         if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
6697             (TP_KEEPINIT(tp) != 0)) {
6698                 struct rack_sendmap *rsm;
6699
6700                 rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6701                 if (rsm) {
6702                         /* Ok we have something outstanding to test keepinit with */
6703                         if ((TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) &&
6704                             ((cts - (uint32_t)rsm->r_tim_lastsent[0]) >= TICKS_2_USEC(TP_KEEPINIT(tp)))) {
6705                                 /* We have exceeded the KEEPINIT time */
6706                                 tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
6707                                 goto drop_it;
6708                         }
6709                 }
6710         }
6711         /*
6712          * Retransmission timer went off.  Message has not been acked within
6713          * retransmit interval.  Back off to a longer retransmit interval
6714          * and retransmit one segment.
6715          */
6716         rack_remxt_tmr(tp);
6717         if ((rack->r_ctl.rc_resend == NULL) ||
6718             ((rack->r_ctl.rc_resend->r_flags & RACK_RWND_COLLAPSED) == 0)) {
6719                 /*
6720                  * If the rwnd collapsed on
6721                  * the one we are retransmitting
6722                  * it does not count against the
6723                  * rxt count.
6724                  */
6725                 tp->t_rxtshift++;
6726         }
6727         if (tp->t_rxtshift > TCP_MAXRXTSHIFT) {
6728                 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
6729 drop_it:
6730                 tp->t_rxtshift = TCP_MAXRXTSHIFT;
6731                 KMOD_TCPSTAT_INC(tcps_timeoutdrop);
6732                 retval = 1;
6733                 tcp_set_inp_to_drop(rack->rc_inp,
6734                     (tp->t_softerror ? (uint16_t) tp->t_softerror : ETIMEDOUT));
6735                 goto out;
6736         }
6737         if (tp->t_state == TCPS_SYN_SENT) {
6738                 /*
6739                  * If the SYN was retransmitted, indicate CWND to be limited
6740                  * to 1 segment in cc_conn_init().
6741                  */
6742                 tp->snd_cwnd = 1;
6743         } else if (tp->t_rxtshift == 1) {
6744                 /*
6745                  * first retransmit; record ssthresh and cwnd so they can be
6746                  * recovered if this turns out to be a "bad" retransmit. A
6747                  * retransmit is considered "bad" if an ACK for this segment
6748                  * is received within RTT/2 interval; the assumption here is
6749                  * that the ACK was already in flight.  See "On Estimating
6750                  * End-to-End Network Path Properties" by Allman and Paxson
6751                  * for more details.
6752                  */
6753                 tp->snd_cwnd_prev = tp->snd_cwnd;
6754                 tp->snd_ssthresh_prev = tp->snd_ssthresh;
6755                 tp->snd_recover_prev = tp->snd_recover;
6756                 tp->t_badrxtwin = ticks + (USEC_2_TICKS(tp->t_srtt)/2);
6757                 tp->t_flags |= TF_PREVVALID;
6758         } else if ((tp->t_flags & TF_RCVD_TSTMP) == 0)
6759                 tp->t_flags &= ~TF_PREVVALID;
6760         KMOD_TCPSTAT_INC(tcps_rexmttimeo);
6761         if ((tp->t_state == TCPS_SYN_SENT) ||
6762             (tp->t_state == TCPS_SYN_RECEIVED))
6763                 rexmt = RACK_INITIAL_RTO * tcp_backoff[tp->t_rxtshift];
6764         else
6765                 rexmt = max(rack_rto_min, (tp->t_srtt + (tp->t_rttvar << 2))) * tcp_backoff[tp->t_rxtshift];
6766
6767         RACK_TCPT_RANGESET(tp->t_rxtcur, rexmt,
6768            max(rack_rto_min, rexmt), rack_rto_max);
6769         /*
6770          * We enter the path for PLMTUD if connection is established or, if
6771          * connection is FIN_WAIT_1 status, reason for the last is that if
6772          * amount of data we send is very small, we could send it in couple
6773          * of packets and process straight to FIN. In that case we won't
6774          * catch ESTABLISHED state.
6775          */
6776 #ifdef INET6
6777         isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) ? true : false;
6778 #else
6779         isipv6 = false;
6780 #endif
6781         if (((V_tcp_pmtud_blackhole_detect == 1) ||
6782             (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
6783             (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
6784             ((tp->t_state == TCPS_ESTABLISHED) ||
6785             (tp->t_state == TCPS_FIN_WAIT_1))) {
6786                 /*
6787                  * Idea here is that at each stage of mtu probe (usually,
6788                  * 1448 -> 1188 -> 524) should be given 2 chances to recover
6789                  * before further clamping down. 'tp->t_rxtshift % 2 == 0'
6790                  * should take care of that.
6791                  */
6792                 if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
6793                     (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
6794                     (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
6795                     tp->t_rxtshift % 2 == 0)) {
6796                         /*
6797                          * Enter Path MTU Black-hole Detection mechanism: -
6798                          * Disable Path MTU Discovery (IP "DF" bit). -
6799                          * Reduce MTU to lower value than what we negotiated
6800                          * with peer.
6801                          */
6802                         if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
6803                                 /* Record that we may have found a black hole. */
6804                                 tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
6805                                 /* Keep track of previous MSS. */
6806                                 tp->t_pmtud_saved_maxseg = tp->t_maxseg;
6807                         }
6808
6809                         /*
6810                          * Reduce the MSS to blackhole value or to the
6811                          * default in an attempt to retransmit.
6812                          */
6813 #ifdef INET6
6814                         if (isipv6 &&
6815                             tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
6816                                 /* Use the sysctl tuneable blackhole MSS. */
6817                                 tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
6818                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
6819                         } else if (isipv6) {
6820                                 /* Use the default MSS. */
6821                                 tp->t_maxseg = V_tcp_v6mssdflt;
6822                                 /*
6823                                  * Disable Path MTU Discovery when we switch
6824                                  * to minmss.
6825                                  */
6826                                 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
6827                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
6828                         }
6829 #endif
6830 #if defined(INET6) && defined(INET)
6831                         else
6832 #endif
6833 #ifdef INET
6834                         if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
6835                                 /* Use the sysctl tuneable blackhole MSS. */
6836                                 tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
6837                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
6838                         } else {
6839                                 /* Use the default MSS. */
6840                                 tp->t_maxseg = V_tcp_mssdflt;
6841                                 /*
6842                                  * Disable Path MTU Discovery when we switch
6843                                  * to minmss.
6844                                  */
6845                                 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
6846                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
6847                         }
6848 #endif
6849                 } else {
6850                         /*
6851                          * If further retransmissions are still unsuccessful
6852                          * with a lowered MTU, maybe this isn't a blackhole
6853                          * and we restore the previous MSS and blackhole
6854                          * detection flags. The limit '6' is determined by
6855                          * giving each probe stage (1448, 1188, 524) 2
6856                          * chances to recover.
6857                          */
6858                         if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
6859                             (tp->t_rxtshift >= 6)) {
6860                                 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
6861                                 tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
6862                                 tp->t_maxseg = tp->t_pmtud_saved_maxseg;
6863                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed);
6864                         }
6865                 }
6866         }
6867         /*
6868          * Disable RFC1323 and SACK if we haven't got any response to
6869          * our third SYN to work-around some broken terminal servers
6870          * (most of which have hopefully been retired) that have bad VJ
6871          * header compression code which trashes TCP segments containing
6872          * unknown-to-them TCP options.
6873          */
6874         if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
6875             (tp->t_rxtshift == 3))
6876                 tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP|TF_SACK_PERMIT);
6877         /*
6878          * If we backed off this far, our srtt estimate is probably bogus.
6879          * Clobber it so we'll take the next rtt measurement as our srtt;
6880          * move the current srtt into rttvar to keep the current retransmit
6881          * times until then.
6882          */
6883         if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
6884 #ifdef INET6
6885                 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0)
6886                         in6_losing(tp->t_inpcb);
6887                 else
6888 #endif
6889                         in_losing(tp->t_inpcb);
6890                 tp->t_rttvar += tp->t_srtt;
6891                 tp->t_srtt = 0;
6892         }
6893         sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
6894         tp->snd_recover = tp->snd_max;
6895         tp->t_flags |= TF_ACKNOW;
6896         tp->t_rtttime = 0;
6897         rack_cong_signal(tp, CC_RTO, tp->snd_una);
6898 out:
6899         return (retval);
6900 }
6901
6902 static int
6903 rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling)
6904 {
6905         int32_t ret = 0;
6906         int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
6907
6908         if (timers == 0) {
6909                 return (0);
6910         }
6911         if (tp->t_state == TCPS_LISTEN) {
6912                 /* no timers on listen sockets */
6913                 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
6914                         return (0);
6915                 return (1);
6916         }
6917         if ((timers & PACE_TMR_RACK) &&
6918             rack->rc_on_min_to) {
6919                 /*
6920                  * For the rack timer when we
6921                  * are on a min-timeout (which means rrr_conf = 3)
6922                  * we don't want to check the timer. It may
6923                  * be going off for a pace and thats ok we
6924                  * want to send the retransmit (if its ready).
6925                  *
6926                  * If its on a normal rack timer (non-min) then
6927                  * we will check if its expired.
6928                  */
6929                 goto skip_time_check;
6930         }
6931         if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
6932                 uint32_t left;
6933
6934                 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
6935                         ret = -1;
6936                         rack_log_to_processing(rack, cts, ret, 0);
6937                         return (0);
6938                 }
6939                 if (hpts_calling == 0) {
6940                         /*
6941                          * A user send or queued mbuf (sack) has called us? We
6942                          * return 0 and let the pacing guards
6943                          * deal with it if they should or
6944                          * should not cause a send.
6945                          */
6946                         ret = -2;
6947                         rack_log_to_processing(rack, cts, ret, 0);
6948                         return (0);
6949                 }
6950                 /*
6951                  * Ok our timer went off early and we are not paced false
6952                  * alarm, go back to sleep.
6953                  */
6954                 ret = -3;
6955                 left = rack->r_ctl.rc_timer_exp - cts;
6956                 tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(left));
6957                 rack_log_to_processing(rack, cts, ret, left);
6958                 return (1);
6959         }
6960 skip_time_check:
6961         rack->rc_tmr_stopped = 0;
6962         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
6963         if (timers & PACE_TMR_DELACK) {
6964                 ret = rack_timeout_delack(tp, rack, cts);
6965         } else if (timers & PACE_TMR_RACK) {
6966                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
6967                 rack->r_fast_output = 0;
6968                 ret = rack_timeout_rack(tp, rack, cts);
6969         } else if (timers & PACE_TMR_TLP) {
6970                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
6971                 ret = rack_timeout_tlp(tp, rack, cts);
6972         } else if (timers & PACE_TMR_RXT) {
6973                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
6974                 rack->r_fast_output = 0;
6975                 ret = rack_timeout_rxt(tp, rack, cts);
6976         } else if (timers & PACE_TMR_PERSIT) {
6977                 ret = rack_timeout_persist(tp, rack, cts);
6978         } else if (timers & PACE_TMR_KEEP) {
6979                 ret = rack_timeout_keepalive(tp, rack, cts);
6980         }
6981         rack_log_to_processing(rack, cts, ret, timers);
6982         return (ret);
6983 }
6984
6985 static void
6986 rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line)
6987 {
6988         struct timeval tv;
6989         uint32_t us_cts, flags_on_entry;
6990         uint8_t hpts_removed = 0;
6991
6992         flags_on_entry = rack->r_ctl.rc_hpts_flags;
6993         us_cts = tcp_get_usecs(&tv);
6994         if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
6995             ((TSTMP_GEQ(us_cts, rack->r_ctl.rc_last_output_to)) ||
6996              ((tp->snd_max - tp->snd_una) == 0))) {
6997                 tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
6998                 hpts_removed = 1;
6999                 /* If we were not delayed cancel out the flag. */
7000                 if ((tp->snd_max - tp->snd_una) == 0)
7001                         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
7002                 rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
7003         }
7004         if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
7005                 rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
7006                 if (rack->rc_inp->inp_in_hpts &&
7007                     ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) {
7008                         /*
7009                          * Canceling timer's when we have no output being
7010                          * paced. We also must remove ourselves from the
7011                          * hpts.
7012                          */
7013                         tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
7014                         hpts_removed = 1;
7015                 }
7016                 rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
7017         }
7018         if (hpts_removed == 0)
7019                 rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
7020 }
7021
7022 static void
7023 rack_timer_stop(struct tcpcb *tp, uint32_t timer_type)
7024 {
7025         return;
7026 }
7027
7028 static int
7029 rack_stopall(struct tcpcb *tp)
7030 {
7031         struct tcp_rack *rack;
7032         rack = (struct tcp_rack *)tp->t_fb_ptr;
7033         rack->t_timers_stopped = 1;
7034         return (0);
7035 }
7036
7037 static void
7038 rack_timer_activate(struct tcpcb *tp, uint32_t timer_type, uint32_t delta)
7039 {
7040         return;
7041 }
7042
7043 static int
7044 rack_timer_active(struct tcpcb *tp, uint32_t timer_type)
7045 {
7046         return (0);
7047 }
7048
7049 static void
7050 rack_stop_all_timers(struct tcpcb *tp)
7051 {
7052         struct tcp_rack *rack;
7053
7054         /*
7055          * Assure no timers are running.
7056          */
7057         if (tcp_timer_active(tp, TT_PERSIST)) {
7058                 /* We enter in persists, set the flag appropriately */
7059                 rack = (struct tcp_rack *)tp->t_fb_ptr;
7060                 rack->rc_in_persist = 1;
7061         }
7062         tcp_timer_suspend(tp, TT_PERSIST);
7063         tcp_timer_suspend(tp, TT_REXMT);
7064         tcp_timer_suspend(tp, TT_KEEP);
7065         tcp_timer_suspend(tp, TT_DELACK);
7066 }
7067
7068 static void
7069 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
7070     struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag)
7071 {
7072         int32_t idx;
7073         uint16_t stripped_flags;
7074
7075         rsm->r_rtr_cnt++;
7076         rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7077         rsm->r_dupack = 0;
7078         if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) {
7079                 rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS;
7080                 rsm->r_flags |= RACK_OVERMAX;
7081         }
7082         if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & RACK_TLP) == 0)) {
7083                 rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
7084                 rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
7085         }
7086         idx = rsm->r_rtr_cnt - 1;
7087         rsm->r_tim_lastsent[idx] = ts;
7088         stripped_flags = rsm->r_flags & ~(RACK_SENT_SP|RACK_SENT_FP);
7089         if (rsm->r_flags & RACK_ACKED) {
7090                 /* Problably MTU discovery messing with us */
7091                 rsm->r_flags &= ~RACK_ACKED;
7092                 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
7093         }
7094         if (rsm->r_in_tmap) {
7095                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7096                 rsm->r_in_tmap = 0;
7097         }
7098         TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7099         rsm->r_in_tmap = 1;
7100         if (rsm->r_flags & RACK_SACK_PASSED) {
7101                 /* We have retransmitted due to the SACK pass */
7102                 rsm->r_flags &= ~RACK_SACK_PASSED;
7103                 rsm->r_flags |= RACK_WAS_SACKPASS;
7104         }
7105 }
7106
7107 static uint32_t
7108 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
7109     struct rack_sendmap *rsm, uint64_t ts, int32_t *lenp, uint16_t add_flag)
7110 {
7111         /*
7112          * We (re-)transmitted starting at rsm->r_start for some length
7113          * (possibly less than r_end.
7114          */
7115         struct rack_sendmap *nrsm, *insret;
7116         uint32_t c_end;
7117         int32_t len;
7118
7119         len = *lenp;
7120         c_end = rsm->r_start + len;
7121         if (SEQ_GEQ(c_end, rsm->r_end)) {
7122                 /*
7123                  * We retransmitted the whole piece or more than the whole
7124                  * slopping into the next rsm.
7125                  */
7126                 rack_update_rsm(tp, rack, rsm, ts, add_flag);
7127                 if (c_end == rsm->r_end) {
7128                         *lenp = 0;
7129                         return (0);
7130                 } else {
7131                         int32_t act_len;
7132
7133                         /* Hangs over the end return whats left */
7134                         act_len = rsm->r_end - rsm->r_start;
7135                         *lenp = (len - act_len);
7136                         return (rsm->r_end);
7137                 }
7138                 /* We don't get out of this block. */
7139         }
7140         /*
7141          * Here we retransmitted less than the whole thing which means we
7142          * have to split this into what was transmitted and what was not.
7143          */
7144         nrsm = rack_alloc_full_limit(rack);
7145         if (nrsm == NULL) {
7146                 /*
7147                  * We can't get memory, so lets not proceed.
7148                  */
7149                 *lenp = 0;
7150                 return (0);
7151         }
7152         /*
7153          * So here we are going to take the original rsm and make it what we
7154          * retransmitted. nrsm will be the tail portion we did not
7155          * retransmit. For example say the chunk was 1, 11 (10 bytes). And
7156          * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
7157          * 1, 6 and the new piece will be 6, 11.
7158          */
7159         rack_clone_rsm(rack, nrsm, rsm, c_end);
7160         nrsm->r_dupack = 0;
7161         rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
7162         insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7163 #ifdef INVARIANTS
7164         if (insret != NULL) {
7165                 panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7166                       nrsm, insret, rack, rsm);
7167         }
7168 #endif
7169         if (rsm->r_in_tmap) {
7170                 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7171                 nrsm->r_in_tmap = 1;
7172         }
7173         rsm->r_flags &= (~RACK_HAS_FIN);
7174         rack_update_rsm(tp, rack, rsm, ts, add_flag);
7175         /* Log a split of rsm into rsm and nrsm */
7176         rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7177         *lenp = 0;
7178         return (0);
7179 }
7180
7181 static void
7182 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
7183                 uint32_t seq_out, uint8_t th_flags, int32_t err, uint64_t cts,
7184                 struct rack_sendmap *hintrsm, uint16_t add_flag, struct mbuf *s_mb, uint32_t s_moff)
7185 {
7186         struct tcp_rack *rack;
7187         struct rack_sendmap *rsm, *nrsm, *insret, fe;
7188         register uint32_t snd_max, snd_una;
7189
7190         /*
7191          * Add to the RACK log of packets in flight or retransmitted. If
7192          * there is a TS option we will use the TS echoed, if not we will
7193          * grab a TS.
7194          *
7195          * Retransmissions will increment the count and move the ts to its
7196          * proper place. Note that if options do not include TS's then we
7197          * won't be able to effectively use the ACK for an RTT on a retran.
7198          *
7199          * Notes about r_start and r_end. Lets consider a send starting at
7200          * sequence 1 for 10 bytes. In such an example the r_start would be
7201          * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
7202          * This means that r_end is actually the first sequence for the next
7203          * slot (11).
7204          *
7205          */
7206         /*
7207          * If err is set what do we do XXXrrs? should we not add the thing?
7208          * -- i.e. return if err != 0 or should we pretend we sent it? --
7209          * i.e. proceed with add ** do this for now.
7210          */
7211         INP_WLOCK_ASSERT(tp->t_inpcb);
7212         if (err)
7213                 /*
7214                  * We don't log errors -- we could but snd_max does not
7215                  * advance in this case either.
7216                  */
7217                 return;
7218
7219         if (th_flags & TH_RST) {
7220                 /*
7221                  * We don't log resets and we return immediately from
7222                  * sending
7223                  */
7224                 return;
7225         }
7226         rack = (struct tcp_rack *)tp->t_fb_ptr;
7227         snd_una = tp->snd_una;
7228         snd_max = tp->snd_max;
7229         if (th_flags & (TH_SYN | TH_FIN)) {
7230                 /*
7231                  * The call to rack_log_output is made before bumping
7232                  * snd_max. This means we can record one extra byte on a SYN
7233                  * or FIN if seq_out is adding more on and a FIN is present
7234                  * (and we are not resending).
7235                  */
7236                 if ((th_flags & TH_SYN) && (seq_out == tp->iss))
7237                         len++;
7238                 if (th_flags & TH_FIN)
7239                         len++;
7240                 if (SEQ_LT(snd_max, tp->snd_nxt)) {
7241                         /*
7242                          * The add/update as not been done for the FIN/SYN
7243                          * yet.
7244                          */
7245                         snd_max = tp->snd_nxt;
7246                 }
7247         }
7248         if (SEQ_LEQ((seq_out + len), snd_una)) {
7249                 /* Are sending an old segment to induce an ack (keep-alive)? */
7250                 return;
7251         }
7252         if (SEQ_LT(seq_out, snd_una)) {
7253                 /* huh? should we panic? */
7254                 uint32_t end;
7255
7256                 end = seq_out + len;
7257                 seq_out = snd_una;
7258                 if (SEQ_GEQ(end, seq_out))
7259                         len = end - seq_out;
7260                 else
7261                         len = 0;
7262         }
7263         if (len == 0) {
7264                 /* We don't log zero window probes */
7265                 return;
7266         }
7267         rack->r_ctl.rc_time_last_sent = cts;
7268         if (IN_FASTRECOVERY(tp->t_flags)) {
7269                 rack->r_ctl.rc_prr_out += len;
7270         }
7271         /* First question is it a retransmission or new? */
7272         if (seq_out == snd_max) {
7273                 /* Its new */
7274 again:
7275                 rsm = rack_alloc(rack);
7276                 if (rsm == NULL) {
7277                         /*
7278                          * Hmm out of memory and the tcb got destroyed while
7279                          * we tried to wait.
7280                          */
7281                         return;
7282                 }
7283                 if (th_flags & TH_FIN) {
7284                         rsm->r_flags = RACK_HAS_FIN|add_flag;
7285                 } else {
7286                         rsm->r_flags = add_flag;
7287                 }
7288                 rsm->r_tim_lastsent[0] = cts;
7289                 rsm->r_rtr_cnt = 1;
7290                 rsm->r_rtr_bytes = 0;
7291                 if (th_flags & TH_SYN) {
7292                         /* The data space is one beyond snd_una */
7293                         rsm->r_flags |= RACK_HAS_SYN;
7294                 }
7295                 rsm->r_start = seq_out;
7296                 rsm->r_end = rsm->r_start + len;
7297                 rsm->r_dupack = 0;
7298                 /*
7299                  * save off the mbuf location that
7300                  * sndmbuf_noadv returned (which is
7301                  * where we started copying from)..
7302                  */
7303                 rsm->m = s_mb;
7304                 rsm->soff = s_moff;
7305                 /* rsm->m will be NULL if RACK_HAS_SYN or RACK_HAS_FIN is set */
7306                 if (rsm->m) {
7307                         if (rsm->m->m_len <= rsm->soff) {
7308                                 /*
7309                                  * XXXrrs Question, will this happen?
7310                                  *
7311                                  * If sbsndptr is set at the correct place
7312                                  * then s_moff should always be somewhere
7313                                  * within rsm->m. But if the sbsndptr was
7314                                  * off then that won't be true. If it occurs
7315                                  * we need to walkout to the correct location.
7316                                  */
7317                                 struct mbuf *lm;
7318
7319                                 lm = rsm->m;
7320                                 while (lm->m_len <= rsm->soff) {
7321                                         rsm->soff -= lm->m_len;
7322                                         lm = lm->m_next;
7323                                         KASSERT(lm != NULL, ("%s rack:%p lm goes null orig_off:%u origmb:%p rsm->soff:%u",
7324                                                              __func__, rack, s_moff, s_mb, rsm->soff));
7325                                 }
7326                                 rsm->m = lm;
7327                                 counter_u64_add(rack_sbsndptr_wrong, 1);
7328                         } else
7329                                 counter_u64_add(rack_sbsndptr_right, 1);
7330                         rsm->orig_m_len = rsm->m->m_len;
7331                 } else
7332                         rsm->orig_m_len = 0;
7333                 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7334                 /* Log a new rsm */
7335                 rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_NEW, 0, __LINE__);
7336                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7337 #ifdef INVARIANTS
7338                 if (insret != NULL) {
7339                         panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7340                               nrsm, insret, rack, rsm);
7341                 }
7342 #endif
7343                 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7344                 rsm->r_in_tmap = 1;
7345                 /*
7346                  * Special case detection, is there just a single
7347                  * packet outstanding when we are not in recovery?
7348                  *
7349                  * If this is true mark it so.
7350                  */
7351                 if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
7352                     (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) == ctf_fixed_maxseg(tp))) {
7353                         struct rack_sendmap *prsm;
7354
7355                         prsm = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7356                         if (prsm)
7357                                 prsm->r_one_out_nr = 1;
7358                 }
7359                 return;
7360         }
7361         /*
7362          * If we reach here its a retransmission and we need to find it.
7363          */
7364         memset(&fe, 0, sizeof(fe));
7365 more:
7366         if (hintrsm && (hintrsm->r_start == seq_out)) {
7367                 rsm = hintrsm;
7368                 hintrsm = NULL;
7369         } else {
7370                 /* No hints sorry */
7371                 rsm = NULL;
7372         }
7373         if ((rsm) && (rsm->r_start == seq_out)) {
7374                 seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag);
7375                 if (len == 0) {
7376                         return;
7377                 } else {
7378                         goto more;
7379                 }
7380         }
7381         /* Ok it was not the last pointer go through it the hard way. */
7382 refind:
7383         fe.r_start = seq_out;
7384         rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
7385         if (rsm) {
7386                 if (rsm->r_start == seq_out) {
7387                         seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag);
7388                         if (len == 0) {
7389                                 return;
7390                         } else {
7391                                 goto refind;
7392                         }
7393                 }
7394                 if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
7395                         /* Transmitted within this piece */
7396                         /*
7397                          * Ok we must split off the front and then let the
7398                          * update do the rest
7399                          */
7400                         nrsm = rack_alloc_full_limit(rack);
7401                         if (nrsm == NULL) {
7402                                 rack_update_rsm(tp, rack, rsm, cts, add_flag);
7403                                 return;
7404                         }
7405                         /*
7406                          * copy rsm to nrsm and then trim the front of rsm
7407                          * to not include this part.
7408                          */
7409                         rack_clone_rsm(rack, nrsm, rsm, seq_out);
7410                         insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7411                         rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7412 #ifdef INVARIANTS
7413                         if (insret != NULL) {
7414                                 panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7415                                       nrsm, insret, rack, rsm);
7416                         }
7417 #endif
7418                         if (rsm->r_in_tmap) {
7419                                 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7420                                 nrsm->r_in_tmap = 1;
7421                         }
7422                         rsm->r_flags &= (~RACK_HAS_FIN);
7423                         seq_out = rack_update_entry(tp, rack, nrsm, cts, &len, add_flag);
7424                         if (len == 0) {
7425                                 return;
7426                         } else if (len > 0)
7427                                 goto refind;
7428                 }
7429         }
7430         /*
7431          * Hmm not found in map did they retransmit both old and on into the
7432          * new?
7433          */
7434         if (seq_out == tp->snd_max) {
7435                 goto again;
7436         } else if (SEQ_LT(seq_out, tp->snd_max)) {
7437 #ifdef INVARIANTS
7438                 printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
7439                        seq_out, len, tp->snd_una, tp->snd_max);
7440                 printf("Starting Dump of all rack entries\n");
7441                 RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
7442                         printf("rsm:%p start:%u end:%u\n",
7443                                rsm, rsm->r_start, rsm->r_end);
7444                 }
7445                 printf("Dump complete\n");
7446                 panic("seq_out not found rack:%p tp:%p",
7447                       rack, tp);
7448 #endif
7449         } else {
7450 #ifdef INVARIANTS
7451                 /*
7452                  * Hmm beyond sndmax? (only if we are using the new rtt-pack
7453                  * flag)
7454                  */
7455                 panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
7456                       seq_out, len, tp->snd_max, tp);
7457 #endif
7458         }
7459 }
7460
7461 /*
7462  * Record one of the RTT updates from an ack into
7463  * our sample structure.
7464  */
7465
7466 static void
7467 tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt, uint32_t len, uint32_t us_rtt,
7468                     int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt)
7469 {
7470         if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7471             (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) {
7472                 rack->r_ctl.rack_rs.rs_rtt_lowest = rtt;
7473         }
7474         if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7475             (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) {
7476                 rack->r_ctl.rack_rs.rs_rtt_highest = rtt;
7477         }
7478         if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
7479             if (us_rtt < rack->r_ctl.rc_gp_lowrtt)
7480                 rack->r_ctl.rc_gp_lowrtt = us_rtt;
7481             if (rack->rc_tp->snd_wnd > rack->r_ctl.rc_gp_high_rwnd)
7482                     rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
7483         }
7484         if ((confidence == 1) &&
7485             ((rsm == NULL) ||
7486              (rsm->r_just_ret) ||
7487              (rsm->r_one_out_nr &&
7488               len < (ctf_fixed_maxseg(rack->rc_tp) * 2)))) {
7489                 /*
7490                  * If the rsm had a just return
7491                  * hit it then we can't trust the
7492                  * rtt measurement for buffer deterimination
7493                  * Note that a confidence of 2, indicates
7494                  * SACK'd which overrides the r_just_ret or
7495                  * the r_one_out_nr. If it was a CUM-ACK and
7496                  * we had only two outstanding, but get an
7497                  * ack for only 1. Then that also lowers our
7498                  * confidence.
7499                  */
7500                 confidence = 0;
7501         }
7502         if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7503             (rack->r_ctl.rack_rs.rs_us_rtt > us_rtt)) {
7504                 if (rack->r_ctl.rack_rs.confidence == 0) {
7505                         /*
7506                          * We take anything with no current confidence
7507                          * saved.
7508                          */
7509                         rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
7510                         rack->r_ctl.rack_rs.confidence = confidence;
7511                         rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
7512                 } else if (confidence || rack->r_ctl.rack_rs.confidence) {
7513                         /*
7514                          * Once we have a confident number,
7515                          * we can update it with a smaller
7516                          * value since this confident number
7517                          * may include the DSACK time until
7518                          * the next segment (the second one) arrived.
7519                          */
7520                         rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
7521                         rack->r_ctl.rack_rs.confidence = confidence;
7522                         rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
7523                 }
7524         }
7525         rack_log_rtt_upd(rack->rc_tp, rack, us_rtt, len, rsm, confidence);
7526         rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID;
7527         rack->r_ctl.rack_rs.rs_rtt_tot += rtt;
7528         rack->r_ctl.rack_rs.rs_rtt_cnt++;
7529 }
7530
7531 /*
7532  * Collect new round-trip time estimate
7533  * and update averages and current timeout.
7534  */
7535 static void
7536 tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp)
7537 {
7538         int32_t delta;
7539         uint32_t o_srtt, o_var;
7540         int32_t hrtt_up = 0;
7541         int32_t rtt;
7542
7543         if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY)
7544                 /* No valid sample */
7545                 return;
7546         if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) {
7547                 /* We are to use the lowest RTT seen in a single ack */
7548                 rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
7549         } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) {
7550                 /* We are to use the highest RTT seen in a single ack */
7551                 rtt = rack->r_ctl.rack_rs.rs_rtt_highest;
7552         } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) {
7553                 /* We are to use the average RTT seen in a single ack */
7554                 rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot /
7555                                 (uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt);
7556         } else {
7557 #ifdef INVARIANTS
7558                 panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method);
7559 #endif
7560                 return;
7561         }
7562         if (rtt == 0)
7563                 rtt = 1;
7564         if (rack->rc_gp_rtt_set == 0) {
7565                 /*
7566                  * With no RTT we have to accept
7567                  * even one we are not confident of.
7568                  */
7569                 rack->r_ctl.rc_gp_srtt = rack->r_ctl.rack_rs.rs_us_rtt;
7570                 rack->rc_gp_rtt_set = 1;
7571         } else if (rack->r_ctl.rack_rs.confidence) {
7572                 /* update the running gp srtt */
7573                 rack->r_ctl.rc_gp_srtt -= (rack->r_ctl.rc_gp_srtt/8);
7574                 rack->r_ctl.rc_gp_srtt += rack->r_ctl.rack_rs.rs_us_rtt / 8;
7575         }
7576         if (rack->r_ctl.rack_rs.confidence) {
7577                 /*
7578                  * record the low and high for highly buffered path computation,
7579                  * we only do this if we are confident (not a retransmission).
7580                  */
7581                 if (rack->r_ctl.rc_highest_us_rtt < rack->r_ctl.rack_rs.rs_us_rtt) {
7582                         rack->r_ctl.rc_highest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7583                         hrtt_up = 1;
7584                 }
7585                 if (rack->rc_highly_buffered == 0) {
7586                         /*
7587                          * Currently once we declare a path has
7588                          * highly buffered there is no going
7589                          * back, which may be a problem...
7590                          */
7591                         if ((rack->r_ctl.rc_highest_us_rtt / rack->r_ctl.rc_lowest_us_rtt) > rack_hbp_thresh) {
7592                                 rack_log_rtt_shrinks(rack, rack->r_ctl.rack_rs.rs_us_rtt,
7593                                                      rack->r_ctl.rc_highest_us_rtt,
7594                                                      rack->r_ctl.rc_lowest_us_rtt,
7595                                                      RACK_RTTS_SEEHBP);
7596                                 rack->rc_highly_buffered = 1;
7597                         }
7598                 }
7599         }
7600         if ((rack->r_ctl.rack_rs.confidence) ||
7601             (rack->r_ctl.rack_rs.rs_us_rtrcnt == 1)) {
7602                 /*
7603                  * If we are highly confident of it <or> it was
7604                  * never retransmitted we accept it as the last us_rtt.
7605                  */
7606                 rack->r_ctl.rc_last_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7607                 /* The lowest rtt can be set if its was not retransmited */
7608                 if (rack->r_ctl.rc_lowest_us_rtt > rack->r_ctl.rack_rs.rs_us_rtt) {
7609                         rack->r_ctl.rc_lowest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7610                         if (rack->r_ctl.rc_lowest_us_rtt == 0)
7611                                 rack->r_ctl.rc_lowest_us_rtt = 1;
7612                 }
7613         }
7614         o_srtt = tp->t_srtt;
7615         o_var = tp->t_rttvar;
7616         rack = (struct tcp_rack *)tp->t_fb_ptr;
7617         if (tp->t_srtt != 0) {
7618                 /*
7619                  * We keep a simple srtt in microseconds, like our rtt
7620                  * measurement. We don't need to do any tricks with shifting
7621                  * etc. Instead we just add in 1/8th of the new measurement
7622                  * and subtract out 1/8 of the old srtt. We do the same with
7623                  * the variance after finding the absolute value of the
7624                  * difference between this sample and the current srtt.
7625                  */
7626                 delta = tp->t_srtt - rtt;
7627                 /* Take off 1/8th of the current sRTT */
7628                 tp->t_srtt -= (tp->t_srtt >> 3);
7629                 /* Add in 1/8th of the new RTT just measured */
7630                 tp->t_srtt += (rtt >> 3);
7631                 if (tp->t_srtt <= 0)
7632                         tp->t_srtt = 1;
7633                 /* Now lets make the absolute value of the variance */
7634                 if (delta < 0)
7635                         delta = -delta;
7636                 /* Subtract out 1/8th */
7637                 tp->t_rttvar -= (tp->t_rttvar >> 3);
7638                 /* Add in 1/8th of the new variance we just saw */
7639                 tp->t_rttvar += (delta >> 3);
7640                 if (tp->t_rttvar <= 0)
7641                         tp->t_rttvar = 1;
7642                 if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar)
7643                         tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
7644         } else {
7645                 /*
7646                  * No rtt measurement yet - use the unsmoothed rtt. Set the
7647                  * variance to half the rtt (so our first retransmit happens
7648                  * at 3*rtt).
7649                  */
7650                 tp->t_srtt = rtt;
7651                 tp->t_rttvar = rtt >> 1;
7652                 tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
7653         }
7654         rack->rc_srtt_measure_made = 1;
7655         KMOD_TCPSTAT_INC(tcps_rttupdated);
7656         tp->t_rttupdated++;
7657 #ifdef STATS
7658         if (rack_stats_gets_ms_rtt == 0) {
7659                 /* Send in the microsecond rtt used for rxt timeout purposes */
7660                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt));
7661         } else if (rack_stats_gets_ms_rtt == 1) {
7662                 /* Send in the millisecond rtt used for rxt timeout purposes */
7663                 int32_t ms_rtt;
7664
7665                 /* Round up */
7666                 ms_rtt = (rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
7667                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
7668         } else if (rack_stats_gets_ms_rtt == 2) {
7669                 /* Send in the millisecond rtt has close to the path RTT as we can get  */
7670                 int32_t ms_rtt;
7671
7672                 /* Round up */
7673                 ms_rtt = (rack->r_ctl.rack_rs.rs_us_rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
7674                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
7675         }  else {
7676                 /* Send in the microsecond rtt has close to the path RTT as we can get  */
7677                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
7678         }
7679
7680 #endif
7681         /*
7682          * the retransmit should happen at rtt + 4 * rttvar. Because of the
7683          * way we do the smoothing, srtt and rttvar will each average +1/2
7684          * tick of bias.  When we compute the retransmit timer, we want 1/2
7685          * tick of rounding and 1 extra tick because of +-1/2 tick
7686          * uncertainty in the firing of the timer.  The bias will give us
7687          * exactly the 1.5 tick we need.  But, because the bias is
7688          * statistical, we have to test that we don't drop below the minimum
7689          * feasible timer (which is 2 ticks).
7690          */
7691         tp->t_rxtshift = 0;
7692         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
7693                       max(rack_rto_min, rtt + 2), rack_rto_max);
7694         rack_log_rtt_sample(rack, rtt);
7695         tp->t_softerror = 0;
7696 }
7697
7698
7699 static void
7700 rack_apply_updated_usrtt(struct tcp_rack *rack, uint32_t us_rtt, uint32_t us_cts)
7701 {
7702         /*
7703          * Apply to filter the inbound us-rtt at us_cts.
7704          */
7705         uint32_t old_rtt;
7706
7707         old_rtt = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
7708         apply_filter_min_small(&rack->r_ctl.rc_gp_min_rtt,
7709                                us_rtt, us_cts);
7710         if (rack->r_ctl.last_pacing_time &&
7711             rack->rc_gp_dyn_mul &&
7712             (rack->r_ctl.last_pacing_time > us_rtt))
7713                 rack->pacing_longer_than_rtt = 1;
7714         else
7715                 rack->pacing_longer_than_rtt = 0;
7716         if (old_rtt > us_rtt) {
7717                 /* We just hit a new lower rtt time */
7718                 rack_log_rtt_shrinks(rack,  us_cts,  old_rtt,
7719                                      __LINE__, RACK_RTTS_NEWRTT);
7720                 /*
7721                  * Only count it if its lower than what we saw within our
7722                  * calculated range.
7723                  */
7724                 if ((old_rtt - us_rtt) > rack_min_rtt_movement) {
7725                         if (rack_probertt_lower_within &&
7726                             rack->rc_gp_dyn_mul &&
7727                             (rack->use_fixed_rate == 0) &&
7728                             (rack->rc_always_pace)) {
7729                                 /*
7730                                  * We are seeing a new lower rtt very close
7731                                  * to the time that we would have entered probe-rtt.
7732                                  * This is probably due to the fact that a peer flow
7733                                  * has entered probe-rtt. Lets go in now too.
7734                                  */
7735                                 uint32_t val;
7736
7737                                 val = rack_probertt_lower_within * rack_time_between_probertt;
7738                                 val /= 100;
7739                                 if ((rack->in_probe_rtt == 0)  &&
7740                                     ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= (rack_time_between_probertt - val))) {
7741                                         rack_enter_probertt(rack, us_cts);
7742                                 }
7743                         }
7744                         rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
7745                 }
7746         }
7747 }
7748
7749 static int
7750 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
7751     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack)
7752 {
7753         int32_t i, all;
7754         uint32_t t, len_acked;
7755
7756         if ((rsm->r_flags & RACK_ACKED) ||
7757             (rsm->r_flags & RACK_WAS_ACKED))
7758                 /* Already done */
7759                 return (0);
7760         if (rsm->r_no_rtt_allowed) {
7761                 /* Not allowed */
7762                 return (0);
7763         }
7764         if (ack_type == CUM_ACKED) {
7765                 if (SEQ_GT(th_ack, rsm->r_end)) {
7766                         len_acked = rsm->r_end - rsm->r_start;
7767                         all = 1;
7768                 } else {
7769                         len_acked = th_ack - rsm->r_start;
7770                         all = 0;
7771                 }
7772         } else {
7773                 len_acked = rsm->r_end - rsm->r_start;
7774                 all = 0;
7775         }
7776         if (rsm->r_rtr_cnt == 1) {
7777                 uint32_t us_rtt;
7778
7779                 t = cts - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7780                 if ((int)t <= 0)
7781                         t = 1;
7782                 if (!tp->t_rttlow || tp->t_rttlow > t)
7783                         tp->t_rttlow = t;
7784                 if (!rack->r_ctl.rc_rack_min_rtt ||
7785                     SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7786                         rack->r_ctl.rc_rack_min_rtt = t;
7787                         if (rack->r_ctl.rc_rack_min_rtt == 0) {
7788                                 rack->r_ctl.rc_rack_min_rtt = 1;
7789                         }
7790                 }
7791                 if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]))
7792                         us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
7793                 else
7794                         us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
7795                 if (us_rtt == 0)
7796                         us_rtt = 1;
7797                 rack_apply_updated_usrtt(rack, us_rtt, tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
7798                 if (ack_type == SACKED) {
7799                         rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 1);
7800                         tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt, 2 , rsm, rsm->r_rtr_cnt);
7801                 } else {
7802                         /*
7803                          * We need to setup what our confidence
7804                          * is in this ack.
7805                          *
7806                          * If the rsm was app limited and it is
7807                          * less than a mss in length (the end
7808                          * of the send) then we have a gap. If we
7809                          * were app limited but say we were sending
7810                          * multiple MSS's then we are more confident
7811                          * int it.
7812                          *
7813                          * When we are not app-limited then we see if
7814                          * the rsm is being included in the current
7815                          * measurement, we tell this by the app_limited_needs_set
7816                          * flag.
7817                          *
7818                          * Note that being cwnd blocked is not applimited
7819                          * as well as the pacing delay between packets which
7820                          * are sending only 1 or 2 MSS's also will show up
7821                          * in the RTT. We probably need to examine this algorithm
7822                          * a bit more and enhance it to account for the delay
7823                          * between rsm's. We could do that by saving off the
7824                          * pacing delay of each rsm (in an rsm) and then
7825                          * factoring that in somehow though for now I am
7826                          * not sure how :)
7827                          */
7828                         int calc_conf = 0;
7829
7830                         if (rsm->r_flags & RACK_APP_LIMITED) {
7831                                 if (all && (len_acked <= ctf_fixed_maxseg(tp)))
7832                                         calc_conf = 0;
7833                                 else
7834                                         calc_conf = 1;
7835                         } else if (rack->app_limited_needs_set == 0) {
7836                                 calc_conf = 1;
7837                         } else {
7838                                 calc_conf = 0;
7839                         }
7840                         rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 2);
7841                         tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt,
7842                                             calc_conf, rsm, rsm->r_rtr_cnt);
7843                 }
7844                 if ((rsm->r_flags & RACK_TLP) &&
7845                     (!IN_FASTRECOVERY(tp->t_flags))) {
7846                         /* Segment was a TLP and our retrans matched */
7847                         if (rack->r_ctl.rc_tlp_cwnd_reduce) {
7848                                 rack->r_ctl.rc_rsm_start = tp->snd_max;
7849                                 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
7850                                 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
7851                                 rack_cong_signal(tp, CC_NDUPACK, tp->snd_una);
7852                         }
7853                 }
7854                 if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
7855                         /* New more recent rack_tmit_time */
7856                         rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7857                         rack->rc_rack_rtt = t;
7858                 }
7859                 return (1);
7860         }
7861         /*
7862          * We clear the soft/rxtshift since we got an ack.
7863          * There is no assurance we will call the commit() function
7864          * so we need to clear these to avoid incorrect handling.
7865          */
7866         tp->t_rxtshift = 0;
7867         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
7868                       rack_rto_min, rack_rto_max);
7869         tp->t_softerror = 0;
7870         if (to && (to->to_flags & TOF_TS) &&
7871             (ack_type == CUM_ACKED) &&
7872             (to->to_tsecr) &&
7873             ((rsm->r_flags & RACK_OVERMAX) == 0)) {
7874                 /*
7875                  * Now which timestamp does it match? In this block the ACK
7876                  * must be coming from a previous transmission.
7877                  */
7878                 for (i = 0; i < rsm->r_rtr_cnt; i++) {
7879                         if (rack_ts_to_msec(rsm->r_tim_lastsent[i]) == to->to_tsecr) {
7880                                 t = cts - (uint32_t)rsm->r_tim_lastsent[i];
7881                                 if ((int)t <= 0)
7882                                         t = 1;
7883                                 if ((i + 1) < rsm->r_rtr_cnt) {
7884                                         /*
7885                                          * The peer ack'd from our previous
7886                                          * transmission. We have a spurious
7887                                          * retransmission and thus we dont
7888                                          * want to update our rack_rtt.
7889                                          */
7890                                         return (0);
7891                                 }
7892                                 if (!tp->t_rttlow || tp->t_rttlow > t)
7893                                         tp->t_rttlow = t;
7894                                 if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7895                                         rack->r_ctl.rc_rack_min_rtt = t;
7896                                         if (rack->r_ctl.rc_rack_min_rtt == 0) {
7897                                                 rack->r_ctl.rc_rack_min_rtt = 1;
7898                                         }
7899                                 }
7900                                 if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
7901                                            (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
7902                                         /* New more recent rack_tmit_time */
7903                                         rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7904                                         rack->rc_rack_rtt = t;
7905                                 }
7906                                 rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[i], cts, 3);
7907                                 tcp_rack_xmit_timer(rack, t + 1, len_acked, t, 0, rsm,
7908                                                     rsm->r_rtr_cnt);
7909                                 return (1);
7910                         }
7911                 }
7912                 goto ts_not_found;
7913         } else {
7914                 /*
7915                  * Ok its a SACK block that we retransmitted. or a windows
7916                  * machine without timestamps. We can tell nothing from the
7917                  * time-stamp since its not there or the time the peer last
7918                  * recieved a segment that moved forward its cum-ack point.
7919                  */
7920 ts_not_found:
7921                 i = rsm->r_rtr_cnt - 1;
7922                 t = cts - (uint32_t)rsm->r_tim_lastsent[i];
7923                 if ((int)t <= 0)
7924                         t = 1;
7925                 if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7926                         /*
7927                          * We retransmitted and the ack came back in less
7928                          * than the smallest rtt we have observed. We most
7929                          * likely did an improper retransmit as outlined in
7930                          * 6.2 Step 2 point 2 in the rack-draft so we
7931                          * don't want to update our rack_rtt. We in
7932                          * theory (in future) might want to think about reverting our
7933                          * cwnd state but we won't for now.
7934                          */
7935                         return (0);
7936                 } else if (rack->r_ctl.rc_rack_min_rtt) {
7937                         /*
7938                          * We retransmitted it and the retransmit did the
7939                          * job.
7940                          */
7941                         if (!rack->r_ctl.rc_rack_min_rtt ||
7942                             SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7943                                 rack->r_ctl.rc_rack_min_rtt = t;
7944                                 if (rack->r_ctl.rc_rack_min_rtt == 0) {
7945                                         rack->r_ctl.rc_rack_min_rtt = 1;
7946                                 }
7947                         }
7948                         if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[i])) {
7949                                 /* New more recent rack_tmit_time */
7950                                 rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[i];
7951                                 rack->rc_rack_rtt = t;
7952                         }
7953                         return (1);
7954                 }
7955         }
7956         return (0);
7957 }
7958
7959 /*
7960  * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
7961  */
7962 static void
7963 rack_log_sack_passed(struct tcpcb *tp,
7964     struct tcp_rack *rack, struct rack_sendmap *rsm)
7965 {
7966         struct rack_sendmap *nrsm;
7967
7968         nrsm = rsm;
7969         TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap,
7970             rack_head, r_tnext) {
7971                 if (nrsm == rsm) {
7972                         /* Skip orginal segment he is acked */
7973                         continue;
7974                 }
7975                 if (nrsm->r_flags & RACK_ACKED) {
7976                         /*
7977                          * Skip ack'd segments, though we
7978                          * should not see these, since tmap
7979                          * should not have ack'd segments.
7980                          */
7981                         continue;
7982                 }
7983                 if (nrsm->r_flags & RACK_SACK_PASSED) {
7984                         /*
7985                          * We found one that is already marked
7986                          * passed, we have been here before and
7987                          * so all others below this are marked.
7988                          */
7989                         break;
7990                 }
7991                 nrsm->r_flags |= RACK_SACK_PASSED;
7992                 nrsm->r_flags &= ~RACK_WAS_SACKPASS;
7993         }
7994 }
7995
7996 static void
7997 rack_need_set_test(struct tcpcb *tp,
7998                    struct tcp_rack *rack,
7999                    struct rack_sendmap *rsm,
8000                    tcp_seq th_ack,
8001                    int line,
8002                    int use_which)
8003 {
8004
8005         if ((tp->t_flags & TF_GPUTINPROG) &&
8006             SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
8007                 /*
8008                  * We were app limited, and this ack
8009                  * butts up or goes beyond the point where we want
8010                  * to start our next measurement. We need
8011                  * to record the new gput_ts as here and
8012                  * possibly update the start sequence.
8013                  */
8014                 uint32_t seq, ts;
8015
8016                 if (rsm->r_rtr_cnt > 1) {
8017                         /*
8018                          * This is a retransmit, can we
8019                          * really make any assessment at this
8020                          * point?  We are not really sure of
8021                          * the timestamp, is it this or the
8022                          * previous transmission?
8023                          *
8024                          * Lets wait for something better that
8025                          * is not retransmitted.
8026                          */
8027                         return;
8028                 }
8029                 seq = tp->gput_seq;
8030                 ts = tp->gput_ts;
8031                 rack->app_limited_needs_set = 0;
8032                 tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
8033                 /* Do we start at a new end? */
8034                 if ((use_which == RACK_USE_BEG) &&
8035                     SEQ_GEQ(rsm->r_start, tp->gput_seq)) {
8036                         /*
8037                          * When we get an ACK that just eats
8038                          * up some of the rsm, we set RACK_USE_BEG
8039                          * since whats at r_start (i.e. th_ack)
8040                          * is left unacked and thats where the
8041                          * measurement not starts.
8042                          */
8043                         tp->gput_seq = rsm->r_start;
8044                         rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8045                 }
8046                 if ((use_which == RACK_USE_END) &&
8047                     SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
8048                             /*
8049                              * We use the end when the cumack
8050                              * is moving forward and completely
8051                              * deleting the rsm passed so basically
8052                              * r_end holds th_ack.
8053                              *
8054                              * For SACK's we also want to use the end
8055                              * since this piece just got sacked and
8056                              * we want to target anything after that
8057                              * in our measurement.
8058                              */
8059                             tp->gput_seq = rsm->r_end;
8060                             rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8061                 }
8062                 if (use_which == RACK_USE_END_OR_THACK) {
8063                         /*
8064                          * special case for ack moving forward,
8065                          * not a sack, we need to move all the
8066                          * way up to where this ack cum-ack moves
8067                          * to.
8068                          */
8069                         if (SEQ_GT(th_ack, rsm->r_end))
8070                                 tp->gput_seq = th_ack;
8071                         else
8072                                 tp->gput_seq = rsm->r_end;
8073                         rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8074                 }
8075                 if (SEQ_GT(tp->gput_seq, tp->gput_ack)) {
8076                         /*
8077                          * We moved beyond this guy's range, re-calculate
8078                          * the new end point.
8079                          */
8080                         if (rack->rc_gp_filled == 0) {
8081                                 tp->gput_ack = tp->gput_seq + max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
8082                         } else {
8083                                 tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
8084                         }
8085                 }
8086                 /*
8087                  * We are moving the goal post, we may be able to clear the
8088                  * measure_saw_probe_rtt flag.
8089                  */
8090                 if ((rack->in_probe_rtt == 0) &&
8091                     (rack->measure_saw_probe_rtt) &&
8092                     (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
8093                         rack->measure_saw_probe_rtt = 0;
8094                 rack_log_pacing_delay_calc(rack, ts, tp->gput_ts,
8095                                            seq, tp->gput_seq, 0, 5, line, NULL);
8096                 if (rack->rc_gp_filled &&
8097                     ((tp->gput_ack - tp->gput_seq) <
8098                      max(rc_init_window(rack), (MIN_GP_WIN *
8099                                                 ctf_fixed_maxseg(tp))))) {
8100                         uint32_t ideal_amount;
8101
8102                         ideal_amount = rack_get_measure_window(tp, rack);
8103                         if (ideal_amount > sbavail(&tp->t_inpcb->inp_socket->so_snd)) {
8104                                 /*
8105                                  * There is no sense of continuing this measurement
8106                                  * because its too small to gain us anything we
8107                                  * trust. Skip it and that way we can start a new
8108                                  * measurement quicker.
8109                                  */
8110                                 tp->t_flags &= ~TF_GPUTINPROG;
8111                                 rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
8112                                                            0, 0, 0, 6, __LINE__, NULL);
8113                         } else {
8114                                 /*
8115                                  * Reset the window further out.
8116                                  */
8117                                 tp->gput_ack = tp->gput_seq + ideal_amount;
8118                         }
8119                 }
8120         }
8121 }
8122
8123 static uint32_t
8124 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack,
8125                    struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts, int *moved_two)
8126 {
8127         uint32_t start, end, changed = 0;
8128         struct rack_sendmap stack_map;
8129         struct rack_sendmap *rsm, *nrsm, fe, *insret, *prev, *next;
8130         int32_t used_ref = 1;
8131         int moved = 0;
8132
8133         start = sack->start;
8134         end = sack->end;
8135         rsm = *prsm;
8136         memset(&fe, 0, sizeof(fe));
8137 do_rest_ofb:
8138         if ((rsm == NULL) ||
8139             (SEQ_LT(end, rsm->r_start)) ||
8140             (SEQ_GEQ(start, rsm->r_end)) ||
8141             (SEQ_LT(start, rsm->r_start))) {
8142                 /*
8143                  * We are not in the right spot,
8144                  * find the correct spot in the tree.
8145                  */
8146                 used_ref = 0;
8147                 fe.r_start = start;
8148                 rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
8149                 moved++;
8150         }
8151         if (rsm == NULL) {
8152                 /* TSNH */
8153                 goto out;
8154         }
8155         /* Ok we have an ACK for some piece of this rsm */
8156         if (rsm->r_start != start) {
8157                 if ((rsm->r_flags & RACK_ACKED) == 0) {
8158                         /**
8159                          * Need to split this in two pieces the before and after,
8160                          * the before remains in the map, the after must be
8161                          * added. In other words we have:
8162                          * rsm        |--------------|
8163                          * sackblk        |------->
8164                          * rsm will become
8165                          *     rsm    |---|
8166                          * and nrsm will be  the sacked piece
8167                          *     nrsm       |----------|
8168                          *
8169                          * But before we start down that path lets
8170                          * see if the sack spans over on top of
8171                          * the next guy and it is already sacked.
8172                          */
8173                         next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8174                         if (next && (next->r_flags & RACK_ACKED) &&
8175                             SEQ_GEQ(end, next->r_start)) {
8176                                 /**
8177                                  * So the next one is already acked, and
8178                                  * we can thus by hookery use our stack_map
8179                                  * to reflect the piece being sacked and
8180                                  * then adjust the two tree entries moving
8181                                  * the start and ends around. So we start like:
8182                                  *  rsm     |------------|             (not-acked)
8183                                  *  next                 |-----------| (acked)
8184                                  *  sackblk        |-------->
8185                                  *  We want to end like so:
8186                                  *  rsm     |------|                   (not-acked)
8187                                  *  next           |-----------------| (acked)
8188                                  *  nrsm           |-----|
8189                                  * Where nrsm is a temporary stack piece we
8190                                  * use to update all the gizmos.
8191                                  */
8192                                 /* Copy up our fudge block */
8193                                 nrsm = &stack_map;
8194                                 memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
8195                                 /* Now adjust our tree blocks */
8196                                 rsm->r_end = start;
8197                                 next->r_start = start;
8198                                 /* Now we must adjust back where next->m is */
8199                                 rack_setup_offset_for_rsm(rsm, next);
8200
8201                                 /* We don't need to adjust rsm, it did not change */
8202                                 /* Clear out the dup ack count of the remainder */
8203                                 rsm->r_dupack = 0;
8204                                 rsm->r_just_ret = 0;
8205                                 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8206                                 /* Now lets make sure our fudge block is right */
8207                                 nrsm->r_start = start;
8208                                 /* Now lets update all the stats and such */
8209                                 rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
8210                                 if (rack->app_limited_needs_set)
8211                                         rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
8212                                 changed += (nrsm->r_end - nrsm->r_start);
8213                                 rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
8214                                 if (nrsm->r_flags & RACK_SACK_PASSED) {
8215                                         counter_u64_add(rack_reorder_seen, 1);
8216                                         rack->r_ctl.rc_reorder_ts = cts;
8217                                 }
8218                                 /*
8219                                  * Now we want to go up from rsm (the
8220                                  * one left un-acked) to the next one
8221                                  * in the tmap. We do this so when
8222                                  * we walk backwards we include marking
8223                                  * sack-passed on rsm (The one passed in
8224                                  * is skipped since it is generally called
8225                                  * on something sacked before removing it
8226                                  * from the tmap).
8227                                  */
8228                                 if (rsm->r_in_tmap) {
8229                                         nrsm = TAILQ_NEXT(rsm, r_tnext);
8230                                         /*
8231                                          * Now that we have the next
8232                                          * one walk backwards from there.
8233                                          */
8234                                         if (nrsm && nrsm->r_in_tmap)
8235                                                 rack_log_sack_passed(tp, rack, nrsm);
8236                                 }
8237                                 /* Now are we done? */
8238                                 if (SEQ_LT(end, next->r_end) ||
8239                                     (end == next->r_end)) {
8240                                         /* Done with block */
8241                                         goto out;
8242                                 }
8243                                 rack_log_map_chg(tp, rack, &stack_map, rsm, next, MAP_SACK_M1, end, __LINE__);
8244                                 counter_u64_add(rack_sack_used_next_merge, 1);
8245                                 /* Postion for the next block */
8246                                 start = next->r_end;
8247                                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, next);
8248                                 if (rsm == NULL)
8249                                         goto out;
8250                         } else {
8251                                 /**
8252                                  * We can't use any hookery here, so we
8253                                  * need to split the map. We enter like
8254                                  * so:
8255                                  *  rsm      |--------|
8256                                  *  sackblk       |----->
8257                                  * We will add the new block nrsm and
8258                                  * that will be the new portion, and then
8259                                  * fall through after reseting rsm. So we
8260                                  * split and look like this:
8261                                  *  rsm      |----|
8262                                  *  sackblk       |----->
8263                                  *  nrsm          |---|
8264                                  * We then fall through reseting
8265                                  * rsm to nrsm, so the next block
8266                                  * picks it up.
8267                                  */
8268                                 nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
8269                                 if (nrsm == NULL) {
8270                                         /*
8271                                          * failed XXXrrs what can we do but loose the sack
8272                                          * info?
8273                                          */
8274                                         goto out;
8275                                 }
8276                                 counter_u64_add(rack_sack_splits, 1);
8277                                 rack_clone_rsm(rack, nrsm, rsm, start);
8278                                 rsm->r_just_ret = 0;
8279                                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8280 #ifdef INVARIANTS
8281                                 if (insret != NULL) {
8282                                         panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
8283                                               nrsm, insret, rack, rsm);
8284                                 }
8285 #endif
8286                                 if (rsm->r_in_tmap) {
8287                                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8288                                         nrsm->r_in_tmap = 1;
8289                                 }
8290                                 rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M2, end, __LINE__);
8291                                 rsm->r_flags &= (~RACK_HAS_FIN);
8292                                 /* Position us to point to the new nrsm that starts the sack blk */
8293                                 rsm = nrsm;
8294                         }
8295                 } else {
8296                         /* Already sacked this piece */
8297                         counter_u64_add(rack_sack_skipped_acked, 1);
8298                         moved++;
8299                         if (end == rsm->r_end) {
8300                                 /* Done with block */
8301                                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8302                                 goto out;
8303                         } else if (SEQ_LT(end, rsm->r_end)) {
8304                                 /* A partial sack to a already sacked block */
8305                                 moved++;
8306                                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8307                                 goto out;
8308                         } else {
8309                                 /*
8310                                  * The end goes beyond this guy
8311                                  * repostion the start to the
8312                                  * next block.
8313                                  */
8314                                 start = rsm->r_end;
8315                                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8316                                 if (rsm == NULL)
8317                                         goto out;
8318                         }
8319                 }
8320         }
8321         if (SEQ_GEQ(end, rsm->r_end)) {
8322                 /**
8323                  * The end of this block is either beyond this guy or right
8324                  * at this guy. I.e.:
8325                  *  rsm ---                 |-----|
8326                  *  end                     |-----|
8327                  *  <or>
8328                  *  end                     |---------|
8329                  */
8330                 if ((rsm->r_flags & RACK_ACKED) == 0) {
8331                         rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
8332                         changed += (rsm->r_end - rsm->r_start);
8333                         rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
8334                         if (rsm->r_in_tmap) /* should be true */
8335                                 rack_log_sack_passed(tp, rack, rsm);
8336                         /* Is Reordering occuring? */
8337                         if (rsm->r_flags & RACK_SACK_PASSED) {
8338                                 rsm->r_flags &= ~RACK_SACK_PASSED;
8339                                 counter_u64_add(rack_reorder_seen, 1);
8340                                 rack->r_ctl.rc_reorder_ts = cts;
8341                         }
8342                         if (rack->app_limited_needs_set)
8343                                 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
8344                         rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8345                         rsm->r_flags |= RACK_ACKED;
8346                         rsm->r_flags &= ~RACK_TLP;
8347                         if (rsm->r_in_tmap) {
8348                                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8349                                 rsm->r_in_tmap = 0;
8350                         }
8351                         rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_SACK_M3, end, __LINE__);
8352                 } else {
8353                         counter_u64_add(rack_sack_skipped_acked, 1);
8354                         moved++;
8355                 }
8356                 if (end == rsm->r_end) {
8357                         /* This block only - done, setup for next */
8358                         goto out;
8359                 }
8360                 /*
8361                  * There is more not coverend by this rsm move on
8362                  * to the next block in the RB tree.
8363                  */
8364                 nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8365                 start = rsm->r_end;
8366                 rsm = nrsm;
8367                 if (rsm == NULL)
8368                         goto out;
8369                 goto do_rest_ofb;
8370         }
8371         /**
8372          * The end of this sack block is smaller than
8373          * our rsm i.e.:
8374          *  rsm ---                 |-----|
8375          *  end                     |--|
8376          */
8377         if ((rsm->r_flags & RACK_ACKED) == 0) {
8378                 prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8379                 if (prev && (prev->r_flags & RACK_ACKED)) {
8380                         /**
8381                          * Goal, we want the right remainder of rsm to shrink
8382                          * in place and span from (rsm->r_start = end) to rsm->r_end.
8383                          * We want to expand prev to go all the way
8384                          * to prev->r_end <- end.
8385                          * so in the tree we have before:
8386                          *   prev     |--------|         (acked)
8387                          *   rsm               |-------| (non-acked)
8388                          *   sackblk           |-|
8389                          * We churn it so we end up with
8390                          *   prev     |----------|       (acked)
8391                          *   rsm                 |-----| (non-acked)
8392                          *   nrsm              |-| (temporary)
8393                          */
8394                         nrsm = &stack_map;
8395                         memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
8396                         prev->r_end = end;
8397                         rsm->r_start = end;
8398                         /* Now adjust nrsm (stack copy) to be
8399                          * the one that is the small
8400                          * piece that was "sacked".
8401                          */
8402                         nrsm->r_end = end;
8403                         rsm->r_dupack = 0;
8404                         rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8405                         /*
8406                          * Now that the rsm has had its start moved forward
8407                          * lets go ahead and get its new place in the world.
8408                          */
8409                         rack_setup_offset_for_rsm(prev, rsm);
8410                         /*
8411                          * Now nrsm is our new little piece
8412                          * that is acked (which was merged
8413                          * to prev). Update the rtt and changed
8414                          * based on that. Also check for reordering.
8415                          */
8416                         rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
8417                         if (rack->app_limited_needs_set)
8418                                 rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
8419                         changed += (nrsm->r_end - nrsm->r_start);
8420                         rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
8421                         if (nrsm->r_flags & RACK_SACK_PASSED) {
8422                                 counter_u64_add(rack_reorder_seen, 1);
8423                                 rack->r_ctl.rc_reorder_ts = cts;
8424                         }
8425                         rack_log_map_chg(tp, rack, prev, &stack_map, rsm, MAP_SACK_M4, end, __LINE__);
8426                         rsm = prev;
8427                         counter_u64_add(rack_sack_used_prev_merge, 1);
8428                 } else {
8429                         /**
8430                          * This is the case where our previous
8431                          * block is not acked either, so we must
8432                          * split the block in two.
8433                          */
8434                         nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
8435                         if (nrsm == NULL) {
8436                                 /* failed rrs what can we do but loose the sack info? */
8437                                 goto out;
8438                         }
8439                         /**
8440                          * In this case nrsm becomes
8441                          * nrsm->r_start = end;
8442                          * nrsm->r_end = rsm->r_end;
8443                          * which is un-acked.
8444                          * <and>
8445                          * rsm->r_end = nrsm->r_start;
8446                          * i.e. the remaining un-acked
8447                          * piece is left on the left
8448                          * hand side.
8449                          *
8450                          * So we start like this
8451                          * rsm      |----------| (not acked)
8452                          * sackblk  |---|
8453                          * build it so we have
8454                          * rsm      |---|         (acked)
8455                          * nrsm         |------|  (not acked)
8456                          */
8457                         counter_u64_add(rack_sack_splits, 1);
8458                         rack_clone_rsm(rack, nrsm, rsm, end);
8459                         rsm->r_flags &= (~RACK_HAS_FIN);
8460                         rsm->r_just_ret = 0;
8461                         insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8462 #ifdef INVARIANTS
8463                         if (insret != NULL) {
8464                                 panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
8465                                       nrsm, insret, rack, rsm);
8466                         }
8467 #endif
8468                         if (rsm->r_in_tmap) {
8469                                 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8470                                 nrsm->r_in_tmap = 1;
8471                         }
8472                         nrsm->r_dupack = 0;
8473                         rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
8474                         rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
8475                         changed += (rsm->r_end - rsm->r_start);
8476                         rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
8477                         if (rsm->r_in_tmap) /* should be true */
8478                                 rack_log_sack_passed(tp, rack, rsm);
8479                         /* Is Reordering occuring? */
8480                         if (rsm->r_flags & RACK_SACK_PASSED) {
8481                                 rsm->r_flags &= ~RACK_SACK_PASSED;
8482                                 counter_u64_add(rack_reorder_seen, 1);
8483                                 rack->r_ctl.rc_reorder_ts = cts;
8484                         }
8485                         if (rack->app_limited_needs_set)
8486                                 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
8487                         rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8488                         rsm->r_flags |= RACK_ACKED;
8489                         rsm->r_flags &= ~RACK_TLP;
8490                         rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M5, end, __LINE__);
8491                         if (rsm->r_in_tmap) {
8492                                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8493                                 rsm->r_in_tmap = 0;
8494                         }
8495                 }
8496         } else if (start != end){
8497                 /*
8498                  * The block was already acked.
8499                  */
8500                 counter_u64_add(rack_sack_skipped_acked, 1);
8501                 moved++;
8502         }
8503 out:
8504         if (rsm && (rsm->r_flags & RACK_ACKED)) {
8505                 /*
8506                  * Now can we merge where we worked
8507                  * with either the previous or
8508                  * next block?
8509                  */
8510                 next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8511                 while (next) {
8512                     if (next->r_flags & RACK_ACKED) {
8513                         /* yep this and next can be merged */
8514                         rsm = rack_merge_rsm(rack, rsm, next);
8515                         next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8516                     } else
8517                             break;
8518                 }
8519                 /* Now what about the previous? */
8520                 prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8521                 while (prev) {
8522                     if (prev->r_flags & RACK_ACKED) {
8523                         /* yep the previous and this can be merged */
8524                         rsm = rack_merge_rsm(rack, prev, rsm);
8525                         prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8526                     } else
8527                             break;
8528                 }
8529         }
8530         if (used_ref == 0) {
8531                 counter_u64_add(rack_sack_proc_all, 1);
8532         } else {
8533                 counter_u64_add(rack_sack_proc_short, 1);
8534         }
8535         /* Save off the next one for quick reference. */
8536         if (rsm)
8537                 nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8538         else
8539                 nrsm = NULL;
8540         *prsm = rack->r_ctl.rc_sacklast = nrsm;
8541         /* Pass back the moved. */
8542         *moved_two = moved;
8543         return (changed);
8544 }
8545
8546 static void inline
8547 rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack)
8548 {
8549         struct rack_sendmap *tmap;
8550
8551         tmap = NULL;
8552         while (rsm && (rsm->r_flags & RACK_ACKED)) {
8553                 /* Its no longer sacked, mark it so */
8554                 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
8555 #ifdef INVARIANTS
8556                 if (rsm->r_in_tmap) {
8557                         panic("rack:%p rsm:%p flags:0x%x in tmap?",
8558                               rack, rsm, rsm->r_flags);
8559                 }
8560 #endif
8561                 rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS);
8562                 /* Rebuild it into our tmap */
8563                 if (tmap == NULL) {
8564                         TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8565                         tmap = rsm;
8566                 } else {
8567                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext);
8568                         tmap = rsm;
8569                 }
8570                 tmap->r_in_tmap = 1;
8571                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8572         }
8573         /*
8574          * Now lets possibly clear the sack filter so we start
8575          * recognizing sacks that cover this area.
8576          */
8577         sack_filter_clear(&rack->r_ctl.rack_sf, th_ack);
8578
8579 }
8580
8581 static void
8582 rack_do_decay(struct tcp_rack *rack)
8583 {
8584         struct timeval res;
8585
8586 #define timersub(tvp, uvp, vvp)                                         \
8587         do {                                                            \
8588                 (vvp)->tv_sec = (tvp)->tv_sec - (uvp)->tv_sec;          \
8589                 (vvp)->tv_usec = (tvp)->tv_usec - (uvp)->tv_usec;       \
8590                 if ((vvp)->tv_usec < 0) {                               \
8591                         (vvp)->tv_sec--;                                \
8592                         (vvp)->tv_usec += 1000000;                      \
8593                 }                                                       \
8594         } while (0)
8595
8596         timersub(&rack->r_ctl.act_rcv_time, &rack->r_ctl.rc_last_time_decay, &res);
8597 #undef timersub
8598
8599         rack->r_ctl.input_pkt++;
8600         if ((rack->rc_in_persist) ||
8601             (res.tv_sec >= 1) ||
8602             (rack->rc_tp->snd_max == rack->rc_tp->snd_una)) {
8603                 /*
8604                  * Check for decay of non-SAD,
8605                  * we want all SAD detection metrics to
8606                  * decay 1/4 per second (or more) passed.
8607                  */
8608                 uint32_t pkt_delta;
8609
8610                 pkt_delta = rack->r_ctl.input_pkt - rack->r_ctl.saved_input_pkt;
8611                 /* Update our saved tracking values */
8612                 rack->r_ctl.saved_input_pkt = rack->r_ctl.input_pkt;
8613                 rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
8614                 /* Now do we escape without decay? */
8615 #ifdef NETFLIX_EXP_DETECTION
8616                 if (rack->rc_in_persist ||
8617                     (rack->rc_tp->snd_max == rack->rc_tp->snd_una) ||
8618                     (pkt_delta < tcp_sad_low_pps)){
8619                         /*
8620                          * We don't decay idle connections
8621                          * or ones that have a low input pps.
8622                          */
8623                         return;
8624                 }
8625                 /* Decay the counters */
8626                 rack->r_ctl.ack_count = ctf_decay_count(rack->r_ctl.ack_count,
8627                                                         tcp_sad_decay_val);
8628                 rack->r_ctl.sack_count = ctf_decay_count(rack->r_ctl.sack_count,
8629                                                          tcp_sad_decay_val);
8630                 rack->r_ctl.sack_moved_extra = ctf_decay_count(rack->r_ctl.sack_moved_extra,
8631                                                                tcp_sad_decay_val);
8632                 rack->r_ctl.sack_noextra_move = ctf_decay_count(rack->r_ctl.sack_noextra_move,
8633                                                                 tcp_sad_decay_val);
8634 #endif
8635         }
8636 }
8637
8638 static void
8639 rack_process_to_cumack(struct tcpcb *tp, struct tcp_rack *rack, register uint32_t th_ack, uint32_t cts, struct tcpopt *to)
8640 {
8641         struct rack_sendmap *rsm, *rm;
8642
8643         /*
8644          * The ACK point is advancing to th_ack, we must drop off
8645          * the packets in the rack log and calculate any eligble
8646          * RTT's.
8647          */
8648         rack->r_wanted_output = 1;
8649 more:
8650         rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
8651         if (rsm == NULL) {
8652                 if ((th_ack - 1) == tp->iss) {
8653                         /*
8654                          * For the SYN incoming case we will not
8655                          * have called tcp_output for the sending of
8656                          * the SYN, so there will be no map. All
8657                          * other cases should probably be a panic.
8658                          */
8659                         return;
8660                 }
8661                 if (tp->t_flags & TF_SENTFIN) {
8662                         /* if we sent a FIN we often will not have map */
8663                         return;
8664                 }
8665 #ifdef INVARIANTS
8666                 panic("No rack map tp:%p for state:%d ack:%u rack:%p snd_una:%u snd_max:%u snd_nxt:%u\n",
8667                       tp,
8668                       tp->t_state, th_ack, rack,
8669                       tp->snd_una, tp->snd_max, tp->snd_nxt);
8670 #endif
8671                 return;
8672         }
8673         if (SEQ_LT(th_ack, rsm->r_start)) {
8674                 /* Huh map is missing this */
8675 #ifdef INVARIANTS
8676                 printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n",
8677                        rsm->r_start,
8678                        th_ack, tp->t_state, rack->r_state);
8679 #endif
8680                 return;
8681         }
8682         rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED, th_ack);
8683         /* Now do we consume the whole thing? */
8684         if (SEQ_GEQ(th_ack, rsm->r_end)) {
8685                 /* Its all consumed. */
8686                 uint32_t left;
8687                 uint8_t newly_acked;
8688
8689                 rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_FREE, rsm->r_end, __LINE__);
8690                 rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
8691                 rsm->r_rtr_bytes = 0;
8692                 /* Record the time of highest cumack sent */
8693                 rack->r_ctl.rc_gp_cumack_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8694                 rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8695 #ifdef INVARIANTS
8696                 if (rm != rsm) {
8697                         panic("removing head in rack:%p rsm:%p rm:%p",
8698                               rack, rsm, rm);
8699                 }
8700 #endif
8701                 if (rsm->r_in_tmap) {
8702                         TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8703                         rsm->r_in_tmap = 0;
8704                 }
8705                 newly_acked = 1;
8706                 if (rsm->r_flags & RACK_ACKED) {
8707                         /*
8708                          * It was acked on the scoreboard -- remove
8709                          * it from total
8710                          */
8711                         rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
8712                         newly_acked = 0;
8713                 } else if (rsm->r_flags & RACK_SACK_PASSED) {
8714                         /*
8715                          * There are segments ACKED on the
8716                          * scoreboard further up. We are seeing
8717                          * reordering.
8718                          */
8719                         rsm->r_flags &= ~RACK_SACK_PASSED;
8720                         counter_u64_add(rack_reorder_seen, 1);
8721                         rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8722                         rsm->r_flags |= RACK_ACKED;
8723                         rack->r_ctl.rc_reorder_ts = cts;
8724                         if (rack->r_ent_rec_ns) {
8725                                 /*
8726                                  * We have sent no more, and we saw an sack
8727                                  * then ack arrive.
8728                                  */
8729                                 rack->r_might_revert = 1;
8730                         }
8731                 }
8732                 if ((rsm->r_flags & RACK_TO_REXT) &&
8733                     (tp->t_flags & TF_RCVD_TSTMP) &&
8734                     (to->to_flags & TOF_TS) &&
8735                     (tp->t_flags & TF_PREVVALID)) {
8736                         /*
8737                          * We can use the timestamp to see
8738                          * if this retransmission was from the
8739                          * first transmit. If so we made a mistake.
8740                          */
8741                         tp->t_flags &= ~TF_PREVVALID;
8742                         if (to->to_tsecr == rack_ts_to_msec(rsm->r_tim_lastsent[0])) {
8743                                 /* The first transmit is what this ack is for */
8744                                 rack_cong_signal(tp, CC_RTO_ERR, th_ack);
8745                         }
8746                 }
8747                 left = th_ack - rsm->r_end;
8748                 if (rack->app_limited_needs_set && newly_acked)
8749                         rack_need_set_test(tp, rack, rsm, th_ack, __LINE__, RACK_USE_END_OR_THACK);
8750                 /* Free back to zone */
8751                 rack_free(rack, rsm);
8752                 if (left) {
8753                         goto more;
8754                 }
8755                 /* Check for reneging */
8756                 rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
8757                 if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) {
8758                         /*
8759                          * The peer has moved snd_una up to
8760                          * the edge of this send, i.e. one
8761                          * that it had previously acked. The only
8762                          * way that can be true if the peer threw
8763                          * away data (space issues) that it had
8764                          * previously sacked (else it would have
8765                          * given us snd_una up to (rsm->r_end).
8766                          * We need to undo the acked markings here.
8767                          *
8768                          * Note we have to look to make sure th_ack is
8769                          * our rsm->r_start in case we get an old ack
8770                          * where th_ack is behind snd_una.
8771                          */
8772                         rack_peer_reneges(rack, rsm, th_ack);
8773                 }
8774                 return;
8775         }
8776         if (rsm->r_flags & RACK_ACKED) {
8777                 /*
8778                  * It was acked on the scoreboard -- remove it from
8779                  * total for the part being cum-acked.
8780                  */
8781                 rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
8782         }
8783         /*
8784          * Clear the dup ack count for
8785          * the piece that remains.
8786          */
8787         rsm->r_dupack = 0;
8788         rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8789         if (rsm->r_rtr_bytes) {
8790                 /*
8791                  * It was retransmitted adjust the
8792                  * sack holes for what was acked.
8793                  */
8794                 int ack_am;
8795
8796                 ack_am = (th_ack - rsm->r_start);
8797                 if (ack_am >= rsm->r_rtr_bytes) {
8798                         rack->r_ctl.rc_holes_rxt -= ack_am;
8799                         rsm->r_rtr_bytes -= ack_am;
8800                 }
8801         }
8802         /*
8803          * Update where the piece starts and record
8804          * the time of send of highest cumack sent.
8805          */
8806         rack->r_ctl.rc_gp_cumack_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8807         rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_TRIM_HEAD, th_ack, __LINE__);
8808         /* Now we need to move our offset forward too */
8809         if (rsm->orig_m_len != rsm->m->m_len) {
8810                 /* Fix up the orig_m_len and possibly the mbuf offset */
8811                 rack_adjust_orig_mlen(rsm);
8812         }
8813         rsm->soff += (th_ack - rsm->r_start);
8814         rsm->r_start = th_ack;
8815         /* Now do we need to move the mbuf fwd too? */
8816         while (rsm->soff >= rsm->m->m_len) {
8817                 rsm->soff -= rsm->m->m_len;
8818                 rsm->m = rsm->m->m_next;
8819                 KASSERT((rsm->m != NULL),
8820                         (" nrsm:%p hit at soff:%u null m",
8821                          rsm, rsm->soff));
8822         }
8823         rsm->orig_m_len = rsm->m->m_len;
8824         if (rack->app_limited_needs_set)
8825                 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_BEG);
8826 }
8827
8828 static void
8829 rack_handle_might_revert(struct tcpcb *tp, struct tcp_rack *rack)
8830 {
8831         struct rack_sendmap *rsm;
8832         int sack_pass_fnd = 0;
8833
8834         if (rack->r_might_revert) {
8835                 /*
8836                  * Ok we have reordering, have not sent anything, we
8837                  * might want to revert the congestion state if nothing
8838                  * further has SACK_PASSED on it. Lets check.
8839                  *
8840                  * We also get here when we have DSACKs come in for
8841                  * all the data that we FR'd. Note that a rxt or tlp
8842                  * timer clears this from happening.
8843                  */
8844
8845                 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
8846                         if (rsm->r_flags & RACK_SACK_PASSED) {
8847                                 sack_pass_fnd = 1;
8848                                 break;
8849                         }
8850                 }
8851                 if (sack_pass_fnd == 0) {
8852                         /*
8853                          * We went into recovery
8854                          * incorrectly due to reordering!
8855                          */
8856                         int orig_cwnd;
8857
8858                         rack->r_ent_rec_ns = 0;
8859                         orig_cwnd = tp->snd_cwnd;
8860                         tp->snd_cwnd = rack->r_ctl.rc_cwnd_at_erec;
8861                         tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at_erec;
8862                         tp->snd_recover = tp->snd_una;
8863                         rack_log_to_prr(rack, 14, orig_cwnd);
8864                         EXIT_RECOVERY(tp->t_flags);
8865                 }
8866                 rack->r_might_revert = 0;
8867         }
8868 }
8869
8870 #ifdef NETFLIX_EXP_DETECTION
8871 static void
8872 rack_do_detection(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t bytes_this_ack, uint32_t segsiz)
8873 {
8874         if ((rack->do_detection || tcp_force_detection) &&
8875             tcp_sack_to_ack_thresh &&
8876             tcp_sack_to_move_thresh &&
8877             ((rack->r_ctl.rc_num_maps_alloced > tcp_map_minimum) || rack->sack_attack_disable)) {
8878                 /*
8879                  * We have thresholds set to find
8880                  * possible attackers and disable sack.
8881                  * Check them.
8882                  */
8883                 uint64_t ackratio, moveratio, movetotal;
8884
8885                 /* Log detecting */
8886                 rack_log_sad(rack, 1);
8887                 ackratio = (uint64_t)(rack->r_ctl.sack_count);
8888                 ackratio *= (uint64_t)(1000);
8889                 if (rack->r_ctl.ack_count)
8890                         ackratio /= (uint64_t)(rack->r_ctl.ack_count);
8891                 else {
8892                         /* We really should not hit here */
8893                         ackratio = 1000;
8894                 }
8895                 if ((rack->sack_attack_disable == 0) &&
8896                     (ackratio > rack_highest_sack_thresh_seen))
8897                         rack_highest_sack_thresh_seen = (uint32_t)ackratio;
8898                 movetotal = rack->r_ctl.sack_moved_extra;
8899                 movetotal += rack->r_ctl.sack_noextra_move;
8900                 moveratio = rack->r_ctl.sack_moved_extra;
8901                 moveratio *= (uint64_t)1000;
8902                 if (movetotal)
8903                         moveratio /= movetotal;
8904                 else {
8905                         /* No moves, thats pretty good */
8906                         moveratio = 0;
8907                 }
8908                 if ((rack->sack_attack_disable == 0) &&
8909                     (moveratio > rack_highest_move_thresh_seen))
8910                         rack_highest_move_thresh_seen = (uint32_t)moveratio;
8911                 if (rack->sack_attack_disable == 0) {
8912                         if ((ackratio > tcp_sack_to_ack_thresh) &&
8913                             (moveratio > tcp_sack_to_move_thresh)) {
8914                                 /* Disable sack processing */
8915                                 rack->sack_attack_disable = 1;
8916                                 if (rack->r_rep_attack == 0) {
8917                                         rack->r_rep_attack = 1;
8918                                         counter_u64_add(rack_sack_attacks_detected, 1);
8919                                 }
8920                                 if (tcp_attack_on_turns_on_logging) {
8921                                         /*
8922                                          * Turn on logging, used for debugging
8923                                          * false positives.
8924                                          */
8925                                         rack->rc_tp->t_logstate = tcp_attack_on_turns_on_logging;
8926                                 }
8927                                 /* Clamp the cwnd at flight size */
8928                                 rack->r_ctl.rc_saved_cwnd = rack->rc_tp->snd_cwnd;
8929                                 rack->rc_tp->snd_cwnd = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
8930                                 rack_log_sad(rack, 2);
8931                         }
8932                 } else {
8933                         /* We are sack-disabled check for false positives */
8934                         if ((ackratio <= tcp_restoral_thresh) ||
8935                             (rack->r_ctl.rc_num_maps_alloced  < tcp_map_minimum)) {
8936                                 rack->sack_attack_disable = 0;
8937                                 rack_log_sad(rack, 3);
8938                                 /* Restart counting */
8939                                 rack->r_ctl.sack_count = 0;
8940                                 rack->r_ctl.sack_moved_extra = 0;
8941                                 rack->r_ctl.sack_noextra_move = 1;
8942                                 rack->r_ctl.ack_count = max(1,
8943                                       (bytes_this_ack / segsiz));
8944
8945                                 if (rack->r_rep_reverse == 0) {
8946                                         rack->r_rep_reverse = 1;
8947                                         counter_u64_add(rack_sack_attacks_reversed, 1);
8948                                 }
8949                                 /* Restore the cwnd */
8950                                 if (rack->r_ctl.rc_saved_cwnd > rack->rc_tp->snd_cwnd)
8951                                         rack->rc_tp->snd_cwnd = rack->r_ctl.rc_saved_cwnd;
8952                         }
8953                 }
8954         }
8955 }
8956 #endif
8957
8958 static void
8959 rack_note_dsack(struct tcp_rack *rack, tcp_seq start, tcp_seq end)
8960 {
8961
8962         uint32_t am;
8963
8964         if (SEQ_GT(end, start))
8965                 am = end - start;
8966         else
8967                 am = 0;
8968         /*
8969          * We keep track of how many DSACK blocks we get
8970          * after a recovery incident.
8971          */
8972         rack->r_ctl.dsack_byte_cnt += am;
8973         if (!IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
8974             rack->r_ctl.retran_during_recovery &&
8975             (rack->r_ctl.dsack_byte_cnt >= rack->r_ctl.retran_during_recovery)) {
8976                 /*
8977                  * False recovery most likely culprit is reordering. If
8978                  * nothing else is missing we need to revert.
8979                  */
8980                 rack->r_might_revert = 1;
8981                 rack_handle_might_revert(rack->rc_tp, rack);
8982                 rack->r_might_revert = 0;
8983                 rack->r_ctl.retran_during_recovery = 0;
8984                 rack->r_ctl.dsack_byte_cnt = 0;
8985         }
8986 }
8987
8988 static void
8989 rack_update_prr(struct tcpcb *tp, struct tcp_rack *rack, uint32_t changed, tcp_seq th_ack)
8990 {
8991         /* Deal with changed and PRR here (in recovery only) */
8992         uint32_t pipe, snd_una;
8993
8994         rack->r_ctl.rc_prr_delivered += changed;
8995
8996         if (sbavail(&rack->rc_inp->inp_socket->so_snd) <= (tp->snd_max - tp->snd_una)) {
8997                 /*
8998                  * It is all outstanding, we are application limited
8999                  * and thus we don't need more room to send anything.
9000                  * Note we use tp->snd_una here and not th_ack because
9001                  * the data as yet not been cut from the sb.
9002                  */
9003                 rack->r_ctl.rc_prr_sndcnt = 0;
9004                 return;
9005         }
9006         /* Compute prr_sndcnt */
9007         if (SEQ_GT(tp->snd_una, th_ack)) {
9008                 snd_una = tp->snd_una;
9009         } else {
9010                 snd_una = th_ack;
9011         }
9012         pipe = ((tp->snd_max - snd_una) - rack->r_ctl.rc_sacked) + rack->r_ctl.rc_holes_rxt;
9013         if (pipe > tp->snd_ssthresh) {
9014                 long sndcnt;
9015
9016                 sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh;
9017                 if (rack->r_ctl.rc_prr_recovery_fs > 0)
9018                         sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs;
9019                 else {
9020                         rack->r_ctl.rc_prr_sndcnt = 0;
9021                         rack_log_to_prr(rack, 9, 0);
9022                         sndcnt = 0;
9023                 }
9024                 sndcnt++;
9025                 if (sndcnt > (long)rack->r_ctl.rc_prr_out)
9026                         sndcnt -= rack->r_ctl.rc_prr_out;
9027                 else
9028                         sndcnt = 0;
9029                 rack->r_ctl.rc_prr_sndcnt = sndcnt;
9030                 rack_log_to_prr(rack, 10, 0);
9031         } else {
9032                 uint32_t limit;
9033
9034                 if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out)
9035                         limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out);
9036                 else
9037                         limit = 0;
9038                 if (changed > limit)
9039                         limit = changed;
9040                 limit += ctf_fixed_maxseg(tp);
9041                 if (tp->snd_ssthresh > pipe) {
9042                         rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit);
9043                         rack_log_to_prr(rack, 11, 0);
9044                 } else {
9045                         rack->r_ctl.rc_prr_sndcnt = min(0, limit);
9046                         rack_log_to_prr(rack, 12, 0);
9047                 }
9048         }
9049 }
9050
9051 static void
9052 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, int entered_recovery, int dup_ack_struck)
9053 {
9054         uint32_t changed;
9055         struct tcp_rack *rack;
9056         struct rack_sendmap *rsm;
9057         struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
9058         register uint32_t th_ack;
9059         int32_t i, j, k, num_sack_blks = 0;
9060         uint32_t cts, acked, ack_point, sack_changed = 0;
9061         int loop_start = 0, moved_two = 0;
9062         uint32_t tsused;
9063
9064
9065         INP_WLOCK_ASSERT(tp->t_inpcb);
9066         if (th->th_flags & TH_RST) {
9067                 /* We don't log resets */
9068                 return;
9069         }
9070         rack = (struct tcp_rack *)tp->t_fb_ptr;
9071         cts = tcp_get_usecs(NULL);
9072         rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9073         changed = 0;
9074         th_ack = th->th_ack;
9075         if (rack->sack_attack_disable == 0)
9076                 rack_do_decay(rack);
9077         if (BYTES_THIS_ACK(tp, th) >= ctf_fixed_maxseg(rack->rc_tp)) {
9078                 /*
9079                  * You only get credit for
9080                  * MSS and greater (and you get extra
9081                  * credit for larger cum-ack moves).
9082                  */
9083                 int ac;
9084
9085                 ac = BYTES_THIS_ACK(tp, th) / ctf_fixed_maxseg(rack->rc_tp);
9086                 rack->r_ctl.ack_count += ac;
9087                 counter_u64_add(rack_ack_total, ac);
9088         }
9089         if (rack->r_ctl.ack_count > 0xfff00000) {
9090                 /*
9091                  * reduce the number to keep us under
9092                  * a uint32_t.
9093                  */
9094                 rack->r_ctl.ack_count /= 2;
9095                 rack->r_ctl.sack_count /= 2;
9096         }
9097         if (SEQ_GT(th_ack, tp->snd_una)) {
9098                 rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__);
9099                 tp->t_acktime = ticks;
9100         }
9101         if (rsm && SEQ_GT(th_ack, rsm->r_start))
9102                 changed = th_ack - rsm->r_start;
9103         if (changed) {
9104                 rack_process_to_cumack(tp, rack, th_ack, cts, to);
9105         }
9106         if ((to->to_flags & TOF_SACK) == 0) {
9107                 /* We are done nothing left and no sack. */
9108                 rack_handle_might_revert(tp, rack);
9109                 /*
9110                  * For cases where we struck a dup-ack
9111                  * with no SACK, add to the changes so
9112                  * PRR will work right.
9113                  */
9114                 if (dup_ack_struck && (changed == 0)) {
9115                         changed += ctf_fixed_maxseg(rack->rc_tp);
9116                 }
9117                 goto out;
9118         }
9119         /* Sack block processing */
9120         if (SEQ_GT(th_ack, tp->snd_una))
9121                 ack_point = th_ack;
9122         else
9123                 ack_point = tp->snd_una;
9124         for (i = 0; i < to->to_nsacks; i++) {
9125                 bcopy((to->to_sacks + i * TCPOLEN_SACK),
9126                       &sack, sizeof(sack));
9127                 sack.start = ntohl(sack.start);
9128                 sack.end = ntohl(sack.end);
9129                 if (SEQ_GT(sack.end, sack.start) &&
9130                     SEQ_GT(sack.start, ack_point) &&
9131                     SEQ_LT(sack.start, tp->snd_max) &&
9132                     SEQ_GT(sack.end, ack_point) &&
9133                     SEQ_LEQ(sack.end, tp->snd_max)) {
9134                         sack_blocks[num_sack_blks] = sack;
9135                         num_sack_blks++;
9136 #ifdef NETFLIX_STATS
9137                 } else if (SEQ_LEQ(sack.start, th_ack) &&
9138                            SEQ_LEQ(sack.end, th_ack)) {
9139                         /*
9140                          * Its a D-SACK block.
9141                          */
9142                         tcp_record_dsack(sack.start, sack.end);
9143 #endif
9144                         rack_note_dsack(rack, sack.start, sack.end);
9145                 }
9146         }
9147         /*
9148          * Sort the SACK blocks so we can update the rack scoreboard with
9149          * just one pass.
9150          */
9151         num_sack_blks = sack_filter_blks(&rack->r_ctl.rack_sf, sack_blocks,
9152                                          num_sack_blks, th->th_ack);
9153         ctf_log_sack_filter(rack->rc_tp, num_sack_blks, sack_blocks);
9154         if (num_sack_blks == 0) {
9155                 /* Nothing to sack (DSACKs?) */
9156                 goto out_with_totals;
9157         }
9158         if (num_sack_blks < 2) {
9159                 /* Only one, we don't need to sort */
9160                 goto do_sack_work;
9161         }
9162         /* Sort the sacks */
9163         for (i = 0; i < num_sack_blks; i++) {
9164                 for (j = i + 1; j < num_sack_blks; j++) {
9165                         if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
9166                                 sack = sack_blocks[i];
9167                                 sack_blocks[i] = sack_blocks[j];
9168                                 sack_blocks[j] = sack;
9169                         }
9170                 }
9171         }
9172         /*
9173          * Now are any of the sack block ends the same (yes some
9174          * implementations send these)?
9175          */
9176 again:
9177         if (num_sack_blks == 0)
9178                 goto out_with_totals;
9179         if (num_sack_blks > 1) {
9180                 for (i = 0; i < num_sack_blks; i++) {
9181                         for (j = i + 1; j < num_sack_blks; j++) {
9182                                 if (sack_blocks[i].end == sack_blocks[j].end) {
9183                                         /*
9184                                          * Ok these two have the same end we
9185                                          * want the smallest end and then
9186                                          * throw away the larger and start
9187                                          * again.
9188                                          */
9189                                         if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
9190                                                 /*
9191                                                  * The second block covers
9192                                                  * more area use that
9193                                                  */
9194                                                 sack_blocks[i].start = sack_blocks[j].start;
9195                                         }
9196                                         /*
9197                                          * Now collapse out the dup-sack and
9198                                          * lower the count
9199                                          */
9200                                         for (k = (j + 1); k < num_sack_blks; k++) {
9201                                                 sack_blocks[j].start = sack_blocks[k].start;
9202                                                 sack_blocks[j].end = sack_blocks[k].end;
9203                                                 j++;
9204                                         }
9205                                         num_sack_blks--;
9206                                         goto again;
9207                                 }
9208                         }
9209                 }
9210         }
9211 do_sack_work:
9212         /*
9213          * First lets look to see if
9214          * we have retransmitted and
9215          * can use the transmit next?
9216          */
9217         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
9218         if (rsm &&
9219             SEQ_GT(sack_blocks[0].end, rsm->r_start) &&
9220             SEQ_LT(sack_blocks[0].start, rsm->r_end)) {
9221                 /*
9222                  * We probably did the FR and the next
9223                  * SACK in continues as we would expect.
9224                  */
9225                 acked = rack_proc_sack_blk(tp, rack, &sack_blocks[0], to, &rsm, cts, &moved_two);
9226                 if (acked) {
9227                         rack->r_wanted_output = 1;
9228                         changed += acked;
9229                         sack_changed += acked;
9230                 }
9231                 if (num_sack_blks == 1) {
9232                         /*
9233                          * This is what we would expect from
9234                          * a normal implementation to happen
9235                          * after we have retransmitted the FR,
9236                          * i.e the sack-filter pushes down
9237                          * to 1 block and the next to be retransmitted
9238                          * is the sequence in the sack block (has more
9239                          * are acked). Count this as ACK'd data to boost
9240                          * up the chances of recovering any false positives.
9241                          */
9242                         rack->r_ctl.ack_count += (acked / ctf_fixed_maxseg(rack->rc_tp));
9243                         counter_u64_add(rack_ack_total, (acked / ctf_fixed_maxseg(rack->rc_tp)));
9244                         counter_u64_add(rack_express_sack, 1);
9245                         if (rack->r_ctl.ack_count > 0xfff00000) {
9246                                 /*
9247                                  * reduce the number to keep us under
9248                                  * a uint32_t.
9249                                  */
9250                                 rack->r_ctl.ack_count /= 2;
9251                                 rack->r_ctl.sack_count /= 2;
9252                         }
9253                         goto out_with_totals;
9254                 } else {
9255                         /*
9256                          * Start the loop through the
9257                          * rest of blocks, past the first block.
9258                          */
9259                         moved_two = 0;
9260                         loop_start = 1;
9261                 }
9262         }
9263         /* Its a sack of some sort */
9264         rack->r_ctl.sack_count++;
9265         if (rack->r_ctl.sack_count > 0xfff00000) {
9266                 /*
9267                  * reduce the number to keep us under
9268                  * a uint32_t.
9269                  */
9270                 rack->r_ctl.ack_count /= 2;
9271                 rack->r_ctl.sack_count /= 2;
9272         }
9273         counter_u64_add(rack_sack_total, 1);
9274         if (rack->sack_attack_disable) {
9275                 /* An attacker disablement is in place */
9276                 if (num_sack_blks > 1) {
9277                         rack->r_ctl.sack_count += (num_sack_blks - 1);
9278                         rack->r_ctl.sack_moved_extra++;
9279                         counter_u64_add(rack_move_some, 1);
9280                         if (rack->r_ctl.sack_moved_extra > 0xfff00000) {
9281                                 rack->r_ctl.sack_moved_extra /= 2;
9282                                 rack->r_ctl.sack_noextra_move /= 2;
9283                         }
9284                 }
9285                 goto out;
9286         }
9287         rsm = rack->r_ctl.rc_sacklast;
9288         for (i = loop_start; i < num_sack_blks; i++) {
9289                 acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts, &moved_two);
9290                 if (acked) {
9291                         rack->r_wanted_output = 1;
9292                         changed += acked;
9293                         sack_changed += acked;
9294                 }
9295                 if (moved_two) {
9296                         /*
9297                          * If we did not get a SACK for at least a MSS and
9298                          * had to move at all, or if we moved more than our
9299                          * threshold, it counts against the "extra" move.
9300                          */
9301                         rack->r_ctl.sack_moved_extra += moved_two;
9302                         counter_u64_add(rack_move_some, 1);
9303                 } else {
9304                         /*
9305                          * else we did not have to move
9306                          * any more than we would expect.
9307                          */
9308                         rack->r_ctl.sack_noextra_move++;
9309                         counter_u64_add(rack_move_none, 1);
9310                 }
9311                 if (moved_two && (acked < ctf_fixed_maxseg(rack->rc_tp))) {
9312                         /*
9313                          * If the SACK was not a full MSS then
9314                          * we add to sack_count the number of
9315                          * MSS's (or possibly more than
9316                          * a MSS if its a TSO send) we had to skip by.
9317                          */
9318                         rack->r_ctl.sack_count += moved_two;
9319                         counter_u64_add(rack_sack_total, moved_two);
9320                 }
9321                 /*
9322                  * Now we need to setup for the next
9323                  * round. First we make sure we won't
9324                  * exceed the size of our uint32_t on
9325                  * the various counts, and then clear out
9326                  * moved_two.
9327                  */
9328                 if ((rack->r_ctl.sack_moved_extra > 0xfff00000) ||
9329                     (rack->r_ctl.sack_noextra_move > 0xfff00000)) {
9330                         rack->r_ctl.sack_moved_extra /= 2;
9331                         rack->r_ctl.sack_noextra_move /= 2;
9332                 }
9333                 if (rack->r_ctl.sack_count > 0xfff00000) {
9334                         rack->r_ctl.ack_count /= 2;
9335                         rack->r_ctl.sack_count /= 2;
9336                 }
9337                 moved_two = 0;
9338         }
9339 out_with_totals:
9340         if (num_sack_blks > 1) {
9341                 /*
9342                  * You get an extra stroke if
9343                  * you have more than one sack-blk, this
9344                  * could be where we are skipping forward
9345                  * and the sack-filter is still working, or
9346                  * it could be an attacker constantly
9347                  * moving us.
9348                  */
9349                 rack->r_ctl.sack_moved_extra++;
9350                 counter_u64_add(rack_move_some, 1);
9351         }
9352 out:
9353 #ifdef NETFLIX_EXP_DETECTION
9354         rack_do_detection(tp, rack, BYTES_THIS_ACK(tp, th), ctf_fixed_maxseg(rack->rc_tp));
9355 #endif
9356         if (changed) {
9357                 /* Something changed cancel the rack timer */
9358                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
9359         }
9360         tsused = tcp_get_usecs(NULL);
9361         rsm = tcp_rack_output(tp, rack, tsused);
9362         if ((!IN_FASTRECOVERY(tp->t_flags)) &&
9363             rsm) {
9364                 /* Enter recovery */
9365                 rack->r_ctl.rc_rsm_start = rsm->r_start;
9366                 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
9367                 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
9368                 entered_recovery = 1;
9369                 rack_cong_signal(tp, CC_NDUPACK, tp->snd_una);
9370                 /*
9371                  * When we enter recovery we need to assure we send
9372                  * one packet.
9373                  */
9374                 if (rack->rack_no_prr == 0) {
9375                         rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
9376                         rack_log_to_prr(rack, 8, 0);
9377                 }
9378                 rack->r_timer_override = 1;
9379                 rack->r_early = 0;
9380                 rack->r_ctl.rc_agg_early = 0;
9381         } else if (IN_FASTRECOVERY(tp->t_flags) &&
9382                    rsm &&
9383                    (rack->r_rr_config == 3)) {
9384                 /*
9385                  * Assure we can output and we get no
9386                  * remembered pace time except the retransmit.
9387                  */
9388                 rack->r_timer_override = 1;
9389                 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
9390                 rack->r_ctl.rc_resend = rsm;
9391         }
9392         if (IN_FASTRECOVERY(tp->t_flags) &&
9393             (rack->rack_no_prr == 0) &&
9394             (entered_recovery == 0)) {
9395                 rack_update_prr(tp, rack, changed, th_ack);
9396                 if ((rsm && (rack->r_ctl.rc_prr_sndcnt >= ctf_fixed_maxseg(tp)) &&
9397                      ((rack->rc_inp->inp_in_hpts == 0) &&
9398                       ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)))) {
9399                         /*
9400                          * If you are pacing output you don't want
9401                          * to override.
9402                          */
9403                         rack->r_early = 0;
9404                         rack->r_ctl.rc_agg_early = 0;
9405                         rack->r_timer_override = 1;
9406                 }
9407         }
9408 }
9409
9410 static void
9411 rack_strike_dupack(struct tcp_rack *rack)
9412 {
9413         struct rack_sendmap *rsm;
9414
9415         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
9416         while (rsm && (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
9417                 rsm = TAILQ_NEXT(rsm, r_tnext);
9418         }
9419         if (rsm && (rsm->r_dupack < 0xff)) {
9420                 rsm->r_dupack++;
9421                 if (rsm->r_dupack >= DUP_ACK_THRESHOLD) {
9422                         struct timeval tv;
9423                         uint32_t cts;
9424                         /*
9425                          * Here we see if we need to retransmit. For
9426                          * a SACK type connection if enough time has passed
9427                          * we will get a return of the rsm. For a non-sack
9428                          * connection we will get the rsm returned if the
9429                          * dupack value is 3 or more.
9430                          */
9431                         cts = tcp_get_usecs(&tv);
9432                         rack->r_ctl.rc_resend = tcp_rack_output(rack->rc_tp, rack, cts);
9433                         if (rack->r_ctl.rc_resend != NULL) {
9434                                 if (!IN_FASTRECOVERY(rack->rc_tp->t_flags)) {
9435                                         rack_cong_signal(rack->rc_tp, CC_NDUPACK,
9436                                                          rack->rc_tp->snd_una);
9437                                 }
9438                                 rack->r_wanted_output = 1;
9439                                 rack->r_timer_override = 1;
9440                                 rack_log_retran_reason(rack, rsm, __LINE__, 1, 3);
9441                         }
9442                 } else {
9443                         rack_log_retran_reason(rack, rsm, __LINE__, 0, 3);
9444                 }
9445         }
9446 }
9447
9448 static void
9449 rack_check_bottom_drag(struct tcpcb *tp,
9450                        struct tcp_rack *rack,
9451                        struct socket *so, int32_t acked)
9452 {
9453         uint32_t segsiz, minseg;
9454
9455         segsiz = ctf_fixed_maxseg(tp);
9456         minseg = segsiz;
9457
9458         if (tp->snd_max == tp->snd_una) {
9459                 /*
9460                  * We are doing dynamic pacing and we are way
9461                  * under. Basically everything got acked while
9462                  * we were still waiting on the pacer to expire.
9463                  *
9464                  * This means we need to boost the b/w in
9465                  * addition to any earlier boosting of
9466                  * the multipler.
9467                  */
9468                 rack->rc_dragged_bottom = 1;
9469                 rack_validate_multipliers_at_or_above100(rack);
9470                 /*
9471                  * Lets use the segment bytes acked plus
9472                  * the lowest RTT seen as the basis to
9473                  * form a b/w estimate. This will be off
9474                  * due to the fact that the true estimate
9475                  * should be around 1/2 the time of the RTT
9476                  * but we can settle for that.
9477                  */
9478                 if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_VALID) &&
9479                     acked) {
9480                         uint64_t bw, calc_bw, rtt;
9481
9482                         rtt = rack->r_ctl.rack_rs.rs_us_rtt;
9483                         if (rtt == 0) {
9484                                 /* no us sample is there a ms one? */
9485                                 if (rack->r_ctl.rack_rs.rs_rtt_lowest) {
9486                                         rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
9487                                 } else {
9488                                         goto no_measurement;
9489                                 }
9490                         }
9491                         bw = acked;
9492                         calc_bw = bw * 1000000;
9493                         calc_bw /= rtt;
9494                         if (rack->r_ctl.last_max_bw &&
9495                             (rack->r_ctl.last_max_bw < calc_bw)) {
9496                                 /*
9497                                  * If we have a last calculated max bw
9498                                  * enforce it.
9499                                  */
9500                                 calc_bw = rack->r_ctl.last_max_bw;
9501                         }
9502                         /* now plop it in */
9503                         if (rack->rc_gp_filled == 0) {
9504                                 if (calc_bw > ONE_POINT_TWO_MEG) {
9505                                         /*
9506                                          * If we have no measurement
9507                                          * don't let us set in more than
9508                                          * 1.2Mbps. If we are still too
9509                                          * low after pacing with this we
9510                                          * will hopefully have a max b/w
9511                                          * available to sanity check things.
9512                                          */
9513                                         calc_bw = ONE_POINT_TWO_MEG;
9514                                 }
9515                                 rack->r_ctl.rc_rtt_diff = 0;
9516                                 rack->r_ctl.gp_bw = calc_bw;
9517                                 rack->rc_gp_filled = 1;
9518                                 if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
9519                                         rack->r_ctl.num_measurements = RACK_REQ_AVG;
9520                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
9521                         } else if (calc_bw > rack->r_ctl.gp_bw) {
9522                                 rack->r_ctl.rc_rtt_diff = 0;
9523                                 if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
9524                                         rack->r_ctl.num_measurements = RACK_REQ_AVG;
9525                                 rack->r_ctl.gp_bw = calc_bw;
9526                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
9527                         } else
9528                                 rack_increase_bw_mul(rack, -1, 0, 0, 1);
9529                         if ((rack->gp_ready == 0) &&
9530                             (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
9531                                 /* We have enough measurements now */
9532                                 rack->gp_ready = 1;
9533                                 rack_set_cc_pacing(rack);
9534                                 if (rack->defer_options)
9535                                         rack_apply_deferred_options(rack);
9536                         }
9537                         /*
9538                          * For acks over 1mss we do a extra boost to simulate
9539                          * where we would get 2 acks (we want 110 for the mul).
9540                          */
9541                         if (acked > segsiz)
9542                                 rack_increase_bw_mul(rack, -1, 0, 0, 1);
9543                 } else {
9544                         /*
9545                          * zero rtt possibly?, settle for just an old increase.
9546                          */
9547 no_measurement:
9548                         rack_increase_bw_mul(rack, -1, 0, 0, 1);
9549                 }
9550         } else if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
9551                    (sbavail(&so->so_snd) > max((segsiz * (4 + rack_req_segs)),
9552                                                minseg)) &&
9553                    (rack->r_ctl.cwnd_to_use > max((segsiz * (rack_req_segs + 2)), minseg)) &&
9554                    (tp->snd_wnd > max((segsiz * (rack_req_segs + 2)), minseg)) &&
9555                    (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) <=
9556                     (segsiz * rack_req_segs))) {
9557                 /*
9558                  * We are doing dynamic GP pacing and
9559                  * we have everything except 1MSS or less
9560                  * bytes left out. We are still pacing away.
9561                  * And there is data that could be sent, This
9562                  * means we are inserting delayed ack time in
9563                  * our measurements because we are pacing too slow.
9564                  */
9565                 rack_validate_multipliers_at_or_above100(rack);
9566                 rack->rc_dragged_bottom = 1;
9567                 rack_increase_bw_mul(rack, -1, 0, 0, 1);
9568         }
9569 }
9570
9571
9572
9573 static void
9574 rack_gain_for_fastoutput(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t acked_amount)
9575 {
9576         /*
9577          * The fast output path is enabled and we
9578          * have moved the cumack forward. Lets see if
9579          * we can expand forward the fast path length by
9580          * that amount. What we would ideally like to
9581          * do is increase the number of bytes in the
9582          * fast path block (left_to_send) by the
9583          * acked amount. However we have to gate that
9584          * by two factors:
9585          * 1) The amount outstanding and the rwnd of the peer
9586          *    (i.e. we don't want to exceed the rwnd of the peer).
9587          *    <and>
9588          * 2) The amount of data left in the socket buffer (i.e.
9589          *    we can't send beyond what is in the buffer).
9590          *
9591          * Note that this does not take into account any increase
9592          * in the cwnd. We will only extend the fast path by
9593          * what was acked.
9594          */
9595         uint32_t new_total, gating_val;
9596
9597         new_total = acked_amount + rack->r_ctl.fsb.left_to_send;
9598         gating_val = min((sbavail(&so->so_snd) - (tp->snd_max - tp->snd_una)),
9599                          (tp->snd_wnd - (tp->snd_max - tp->snd_una)));
9600         if (new_total <= gating_val) {
9601                 /* We can increase left_to_send by the acked amount */
9602                 counter_u64_add(rack_extended_rfo, 1);
9603                 rack->r_ctl.fsb.left_to_send = new_total;
9604                 KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(&rack->rc_inp->inp_socket->so_snd) - (tp->snd_max - tp->snd_una))),
9605                         ("rack:%p left_to_send:%u sbavail:%u out:%u",
9606                          rack, rack->r_ctl.fsb.left_to_send,
9607                          sbavail(&rack->rc_inp->inp_socket->so_snd),
9608                          (tp->snd_max - tp->snd_una)));
9609
9610         }
9611 }
9612
9613 static void
9614 rack_adjust_sendmap(struct tcp_rack *rack, struct sockbuf *sb, tcp_seq snd_una)
9615 {
9616         /*
9617          * Here any sendmap entry that points to the
9618          * beginning mbuf must be adjusted to the correct
9619          * offset. This must be called with:
9620          * 1) The socket buffer locked
9621          * 2) snd_una adjusted to its new postion.
9622          *
9623          * Note that (2) implies rack_ack_received has also
9624          * been called.
9625          *
9626          * We grab the first mbuf in the socket buffer and
9627          * then go through the front of the sendmap, recalculating
9628          * the stored offset for any sendmap entry that has
9629          * that mbuf. We must use the sb functions to do this
9630          * since its possible an add was done has well as
9631          * the subtraction we may have just completed. This should
9632          * not be a penalty though, since we just referenced the sb
9633          * to go in and trim off the mbufs that we freed (of course
9634          * there will be a penalty for the sendmap references though).
9635          */
9636         struct mbuf *m;
9637         struct rack_sendmap *rsm;
9638
9639         SOCKBUF_LOCK_ASSERT(sb);
9640         m = sb->sb_mb;
9641         rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9642         if ((rsm == NULL) || (m == NULL)) {
9643                 /* Nothing outstanding */
9644                 return;
9645         }
9646         while (rsm->m == m) {
9647                 /* one to adjust */
9648 #ifdef INVARIANTS
9649                 struct mbuf *tm;
9650                 uint32_t soff;
9651
9652                 tm = sbsndmbuf(sb, (rsm->r_start - snd_una), &soff);
9653                 if (rsm->orig_m_len != m->m_len) {
9654                         rack_adjust_orig_mlen(rsm);
9655                 }
9656                 if (rsm->soff != soff) {
9657                         /*
9658                          * This is not a fatal error, we anticipate it
9659                          * might happen (the else code), so we count it here
9660                          * so that under invariant we can see that it really
9661                          * does happen.
9662                          */
9663                         counter_u64_add(rack_adjust_map_bw, 1);
9664                 }
9665                 rsm->m = tm;
9666                 rsm->soff = soff;
9667                 rsm->orig_m_len = rsm->m->m_len;
9668 #else
9669                 rsm->m = sbsndmbuf(sb, (rsm->r_start - snd_una), &rsm->soff);
9670                 rsm->orig_m_len = rsm->m->m_len;
9671 #endif
9672                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
9673                               rsm);
9674                 if (rsm == NULL)
9675                         break;
9676         }
9677 }
9678
9679 /*
9680  * Return value of 1, we do not need to call rack_process_data().
9681  * return value of 0, rack_process_data can be called.
9682  * For ret_val if its 0 the TCP is locked, if its non-zero
9683  * its unlocked and probably unsafe to touch the TCB.
9684  */
9685 static int
9686 rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
9687     struct tcpcb *tp, struct tcpopt *to,
9688     uint32_t tiwin, int32_t tlen,
9689     int32_t * ofia, int32_t thflags, int32_t *ret_val)
9690 {
9691         int32_t ourfinisacked = 0;
9692         int32_t nsegs, acked_amount;
9693         int32_t acked;
9694         struct mbuf *mfree;
9695         struct tcp_rack *rack;
9696         int32_t under_pacing = 0;
9697         int32_t recovery = 0;
9698
9699         rack = (struct tcp_rack *)tp->t_fb_ptr;
9700         if (SEQ_GT(th->th_ack, tp->snd_max)) {
9701                 __ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val,
9702                                       &rack->r_ctl.challenge_ack_ts,
9703                                       &rack->r_ctl.challenge_ack_cnt);
9704                 rack->r_wanted_output = 1;
9705                 return (1);
9706         }
9707         if (rack->gp_ready &&
9708             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
9709                 under_pacing = 1;
9710         }
9711         if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
9712                 int in_rec, dup_ack_struck = 0;
9713
9714                 in_rec = IN_FASTRECOVERY(tp->t_flags);
9715                 if (rack->rc_in_persist) {
9716                         tp->t_rxtshift = 0;
9717                         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
9718                                       rack_rto_min, rack_rto_max);
9719                 }
9720                 if ((th->th_ack == tp->snd_una) && (tiwin == tp->snd_wnd)) {
9721                         rack_strike_dupack(rack);
9722                         dup_ack_struck = 1;
9723                 }
9724                 rack_log_ack(tp, to, th, ((in_rec == 0) && IN_FASTRECOVERY(tp->t_flags)), dup_ack_struck);
9725         }
9726         if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
9727                 /*
9728                  * Old ack, behind (or duplicate to) the last one rcv'd
9729                  * Note: We mark reordering is occuring if its
9730                  * less than and we have not closed our window.
9731                  */
9732                 if (SEQ_LT(th->th_ack, tp->snd_una) && (sbspace(&so->so_rcv) > ctf_fixed_maxseg(tp))) {
9733                         counter_u64_add(rack_reorder_seen, 1);
9734                         rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
9735                 }
9736                 return (0);
9737         }
9738         /*
9739          * If we reach this point, ACK is not a duplicate, i.e., it ACKs
9740          * something we sent.
9741          */
9742         if (tp->t_flags & TF_NEEDSYN) {
9743                 /*
9744                  * T/TCP: Connection was half-synchronized, and our SYN has
9745                  * been ACK'd (so connection is now fully synchronized).  Go
9746                  * to non-starred state, increment snd_una for ACK of SYN,
9747                  * and check if we can do window scaling.
9748                  */
9749                 tp->t_flags &= ~TF_NEEDSYN;
9750                 tp->snd_una++;
9751                 /* Do window scaling? */
9752                 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
9753                     (TF_RCVD_SCALE | TF_REQ_SCALE)) {
9754                         tp->rcv_scale = tp->request_r_scale;
9755                         /* Send window already scaled. */
9756                 }
9757         }
9758         nsegs = max(1, m->m_pkthdr.lro_nsegs);
9759         INP_WLOCK_ASSERT(tp->t_inpcb);
9760
9761         acked = BYTES_THIS_ACK(tp, th);
9762         KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
9763         KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
9764         /*
9765          * If we just performed our first retransmit, and the ACK arrives
9766          * within our recovery window, then it was a mistake to do the
9767          * retransmit in the first place.  Recover our original cwnd and
9768          * ssthresh, and proceed to transmit where we left off.
9769          */
9770         if ((tp->t_flags & TF_PREVVALID) &&
9771             ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
9772                 tp->t_flags &= ~TF_PREVVALID;
9773                 if (tp->t_rxtshift == 1 &&
9774                     (int)(ticks - tp->t_badrxtwin) < 0)
9775                         rack_cong_signal(tp, CC_RTO_ERR, th->th_ack);
9776         }
9777         if (acked) {
9778                 /* assure we are not backed off */
9779                 tp->t_rxtshift = 0;
9780                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
9781                               rack_rto_min, rack_rto_max);
9782                 rack->rc_tlp_in_progress = 0;
9783                 rack->r_ctl.rc_tlp_cnt_out = 0;
9784                 /*
9785                  * If it is the RXT timer we want to
9786                  * stop it, so we can restart a TLP.
9787                  */
9788                 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
9789                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
9790 #ifdef NETFLIX_HTTP_LOGGING
9791                 tcp_http_check_for_comp(rack->rc_tp, th->th_ack);
9792 #endif
9793         }
9794         /*
9795          * If we have a timestamp reply, update smoothed round trip time. If
9796          * no timestamp is present but transmit timer is running and timed
9797          * sequence number was acked, update smoothed round trip time. Since
9798          * we now have an rtt measurement, cancel the timer backoff (cf.,
9799          * Phil Karn's retransmit alg.). Recompute the initial retransmit
9800          * timer.
9801          *
9802          * Some boxes send broken timestamp replies during the SYN+ACK
9803          * phase, ignore timestamps of 0 or we could calculate a huge RTT
9804          * and blow up the retransmit timer.
9805          */
9806         /*
9807          * If all outstanding data is acked, stop retransmit timer and
9808          * remember to restart (more output or persist). If there is more
9809          * data to be acked, restart retransmit timer, using current
9810          * (possibly backed-off) value.
9811          */
9812         if (acked == 0) {
9813                 if (ofia)
9814                         *ofia = ourfinisacked;
9815                 return (0);
9816         }
9817         if (IN_RECOVERY(tp->t_flags)) {
9818                 if (SEQ_LT(th->th_ack, tp->snd_recover) &&
9819                     (SEQ_LT(th->th_ack, tp->snd_max))) {
9820                         tcp_rack_partialack(tp);
9821                 } else {
9822                         rack_post_recovery(tp, th->th_ack);
9823                         recovery = 1;
9824                 }
9825         }
9826         /*
9827          * Let the congestion control algorithm update congestion control
9828          * related information. This typically means increasing the
9829          * congestion window.
9830          */
9831         rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, recovery);
9832         SOCKBUF_LOCK(&so->so_snd);
9833         acked_amount = min(acked, (int)sbavail(&so->so_snd));
9834         tp->snd_wnd -= acked_amount;
9835         mfree = sbcut_locked(&so->so_snd, acked_amount);
9836         if ((sbused(&so->so_snd) == 0) &&
9837             (acked > acked_amount) &&
9838             (tp->t_state >= TCPS_FIN_WAIT_1) &&
9839             (tp->t_flags & TF_SENTFIN)) {
9840                 /*
9841                  * We must be sure our fin
9842                  * was sent and acked (we can be
9843                  * in FIN_WAIT_1 without having
9844                  * sent the fin).
9845                  */
9846                 ourfinisacked = 1;
9847         }
9848         tp->snd_una = th->th_ack;
9849         if (acked_amount && sbavail(&so->so_snd))
9850                 rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
9851         rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
9852         SOCKBUF_UNLOCK(&so->so_snd);
9853         tp->t_flags |= TF_WAKESOW;
9854         m_freem(mfree);
9855         if (SEQ_GT(tp->snd_una, tp->snd_recover))
9856                 tp->snd_recover = tp->snd_una;
9857
9858         if (SEQ_LT(tp->snd_nxt, tp->snd_una)) {
9859                 tp->snd_nxt = tp->snd_una;
9860         }
9861         if (under_pacing &&
9862             (rack->use_fixed_rate == 0) &&
9863             (rack->in_probe_rtt == 0) &&
9864             rack->rc_gp_dyn_mul &&
9865             rack->rc_always_pace) {
9866                 /* Check if we are dragging bottom */
9867                 rack_check_bottom_drag(tp, rack, so, acked);
9868         }
9869         if (tp->snd_una == tp->snd_max) {
9870                 /* Nothing left outstanding */
9871                 tp->t_flags &= ~TF_PREVVALID;
9872                 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
9873                 rack->r_ctl.retran_during_recovery = 0;
9874                 rack->r_ctl.dsack_byte_cnt = 0;
9875                 if (rack->r_ctl.rc_went_idle_time == 0)
9876                         rack->r_ctl.rc_went_idle_time = 1;
9877                 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
9878                 if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
9879                         tp->t_acktime = 0;
9880                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
9881                 /* Set need output so persist might get set */
9882                 rack->r_wanted_output = 1;
9883                 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
9884                 if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
9885                     (sbavail(&so->so_snd) == 0) &&
9886                     (tp->t_flags2 & TF2_DROP_AF_DATA)) {
9887                         /*
9888                          * The socket was gone and the
9889                          * peer sent data (now or in the past), time to
9890                          * reset him.
9891                          */
9892                         *ret_val = 1;
9893                         /* tcp_close will kill the inp pre-log the Reset */
9894                         tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
9895                         tp = tcp_close(tp);
9896                         ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen);
9897                         return (1);
9898                 }
9899         }
9900         if (ofia)
9901                 *ofia = ourfinisacked;
9902         return (0);
9903 }
9904
9905 static void
9906 rack_collapsed_window(struct tcp_rack *rack)
9907 {
9908         /*
9909          * Now we must walk the
9910          * send map and divide the
9911          * ones left stranded. These
9912          * guys can't cause us to abort
9913          * the connection and are really
9914          * "unsent". However if a buggy
9915          * client actually did keep some
9916          * of the data i.e. collapsed the win
9917          * and refused to ack and then opened
9918          * the win and acked that data. We would
9919          * get into an ack war, the simplier
9920          * method then of just pretending we
9921          * did not send those segments something
9922          * won't work.
9923          */
9924         struct rack_sendmap *rsm, *nrsm, fe, *insret;
9925         tcp_seq max_seq;
9926
9927         max_seq = rack->rc_tp->snd_una + rack->rc_tp->snd_wnd;
9928         memset(&fe, 0, sizeof(fe));
9929         fe.r_start = max_seq;
9930         /* Find the first seq past or at maxseq */
9931         rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
9932         if (rsm == NULL) {
9933                 /* Nothing to do strange */
9934                 rack->rc_has_collapsed = 0;
9935                 return;
9936         }
9937         /*
9938          * Now do we need to split at
9939          * the collapse point?
9940          */
9941         if (SEQ_GT(max_seq, rsm->r_start)) {
9942                 nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
9943                 if (nrsm == NULL) {
9944                         /* We can't get a rsm, mark all? */
9945                         nrsm = rsm;
9946                         goto no_split;
9947                 }
9948                 /* Clone it */
9949                 rack_clone_rsm(rack, nrsm, rsm, max_seq);
9950                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
9951 #ifdef INVARIANTS
9952                 if (insret != NULL) {
9953                         panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
9954                               nrsm, insret, rack, rsm);
9955                 }
9956 #endif
9957                 rack_log_map_chg(rack->rc_tp, rack, NULL, rsm, nrsm, MAP_SPLIT, max_seq, __LINE__);
9958                 if (rsm->r_in_tmap) {
9959                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
9960                         nrsm->r_in_tmap = 1;
9961                 }
9962                 /*
9963                  * Set in the new RSM as the
9964                  * collapsed starting point
9965                  */
9966                 rsm = nrsm;
9967         }
9968 no_split:
9969         counter_u64_add(rack_collapsed_win, 1);
9970         RB_FOREACH_FROM(nrsm, rack_rb_tree_head, rsm) {
9971                 nrsm->r_flags |= RACK_RWND_COLLAPSED;
9972         }
9973         rack->rc_has_collapsed = 1;
9974 }
9975
9976 static void
9977 rack_un_collapse_window(struct tcp_rack *rack)
9978 {
9979         struct rack_sendmap *rsm;
9980
9981         RB_FOREACH_REVERSE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
9982                 if (rsm->r_flags & RACK_RWND_COLLAPSED)
9983                         rsm->r_flags &= ~RACK_RWND_COLLAPSED;
9984                 else
9985                         break;
9986         }
9987         rack->rc_has_collapsed = 0;
9988 }
9989
9990 static void
9991 rack_handle_delayed_ack(struct tcpcb *tp, struct tcp_rack *rack,
9992                         int32_t tlen, int32_t tfo_syn)
9993 {
9994         if (DELAY_ACK(tp, tlen) || tfo_syn) {
9995                 if (rack->rc_dack_mode &&
9996                     (tlen > 500) &&
9997                     (rack->rc_dack_toggle == 1)) {
9998                         goto no_delayed_ack;
9999                 }
10000                 rack_timer_cancel(tp, rack,
10001                                   rack->r_ctl.rc_rcvtime, __LINE__);
10002                 tp->t_flags |= TF_DELACK;
10003         } else {
10004 no_delayed_ack:
10005                 rack->r_wanted_output = 1;
10006                 tp->t_flags |= TF_ACKNOW;
10007                 if (rack->rc_dack_mode) {
10008                         if (tp->t_flags & TF_DELACK)
10009                                 rack->rc_dack_toggle = 1;
10010                         else
10011                                 rack->rc_dack_toggle = 0;
10012                 }
10013         }
10014 }
10015
10016 static void
10017 rack_validate_fo_sendwin_up(struct tcpcb *tp, struct tcp_rack *rack)
10018 {
10019         /*
10020          * If fast output is in progress, lets validate that
10021          * the new window did not shrink on us and make it
10022          * so fast output should end.
10023          */
10024         if (rack->r_fast_output) {
10025                 uint32_t out;
10026
10027                 /*
10028                  * Calculate what we will send if left as is
10029                  * and compare that to our send window.
10030                  */
10031                 out = ctf_outstanding(tp);
10032                 if ((out + rack->r_ctl.fsb.left_to_send) > tp->snd_wnd) {
10033                         /* ok we have an issue */
10034                         if (out >= tp->snd_wnd) {
10035                                 /* Turn off fast output the window is met or collapsed */
10036                                 rack->r_fast_output = 0;
10037                         } else {
10038                                 /* we have some room left */
10039                                 rack->r_ctl.fsb.left_to_send = tp->snd_wnd - out;
10040                                 if (rack->r_ctl.fsb.left_to_send < ctf_fixed_maxseg(tp)) {
10041                                         /* If not at least 1 full segment never mind */
10042                                         rack->r_fast_output = 0;
10043                                 }
10044                         }
10045                 }
10046         }
10047 }
10048
10049 /*
10050  * Return value of 1, the TCB is unlocked and most
10051  * likely gone, return value of 0, the TCP is still
10052  * locked.
10053  */
10054 static int
10055 rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
10056     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
10057     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
10058 {
10059         /*
10060          * Update window information. Don't look at window if no ACK: TAC's
10061          * send garbage on first SYN.
10062          */
10063         int32_t nsegs;
10064         int32_t tfo_syn;
10065         struct tcp_rack *rack;
10066
10067         rack = (struct tcp_rack *)tp->t_fb_ptr;
10068         INP_WLOCK_ASSERT(tp->t_inpcb);
10069         nsegs = max(1, m->m_pkthdr.lro_nsegs);
10070         if ((thflags & TH_ACK) &&
10071             (SEQ_LT(tp->snd_wl1, th->th_seq) ||
10072             (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
10073             (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
10074                 /* keep track of pure window updates */
10075                 if (tlen == 0 &&
10076                     tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
10077                         KMOD_TCPSTAT_INC(tcps_rcvwinupd);
10078                 tp->snd_wnd = tiwin;
10079                 rack_validate_fo_sendwin_up(tp, rack);
10080                 tp->snd_wl1 = th->th_seq;
10081                 tp->snd_wl2 = th->th_ack;
10082                 if (tp->snd_wnd > tp->max_sndwnd)
10083                         tp->max_sndwnd = tp->snd_wnd;
10084                 rack->r_wanted_output = 1;
10085         } else if (thflags & TH_ACK) {
10086                 if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
10087                         tp->snd_wnd = tiwin;
10088                         rack_validate_fo_sendwin_up(tp, rack);
10089                         tp->snd_wl1 = th->th_seq;
10090                         tp->snd_wl2 = th->th_ack;
10091                 }
10092         }
10093         if (tp->snd_wnd < ctf_outstanding(tp))
10094                 /* The peer collapsed the window */
10095                 rack_collapsed_window(rack);
10096         else if (rack->rc_has_collapsed)
10097                 rack_un_collapse_window(rack);
10098         /* Was persist timer active and now we have window space? */
10099         if ((rack->rc_in_persist != 0) &&
10100             (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
10101                                 rack->r_ctl.rc_pace_min_segs))) {
10102                 rack_exit_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10103                 tp->snd_nxt = tp->snd_max;
10104                 /* Make sure we output to start the timer */
10105                 rack->r_wanted_output = 1;
10106         }
10107         /* Do we enter persists? */
10108         if ((rack->rc_in_persist == 0) &&
10109             (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
10110             TCPS_HAVEESTABLISHED(tp->t_state) &&
10111             (tp->snd_max == tp->snd_una) &&
10112             sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
10113             (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
10114                 /*
10115                  * Here the rwnd is less than
10116                  * the pacing size, we are established,
10117                  * nothing is outstanding, and there is
10118                  * data to send. Enter persists.
10119                  */
10120                 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10121         }
10122         if (tp->t_flags2 & TF2_DROP_AF_DATA) {
10123                 m_freem(m);
10124                 return (0);
10125         }
10126         /*
10127          * don't process the URG bit, ignore them drag
10128          * along the up.
10129          */
10130         tp->rcv_up = tp->rcv_nxt;
10131         INP_WLOCK_ASSERT(tp->t_inpcb);
10132
10133         /*
10134          * Process the segment text, merging it into the TCP sequencing
10135          * queue, and arranging for acknowledgment of receipt if necessary.
10136          * This process logically involves adjusting tp->rcv_wnd as data is
10137          * presented to the user (this happens in tcp_usrreq.c, case
10138          * PRU_RCVD).  If a FIN has already been received on this connection
10139          * then we just ignore the text.
10140          */
10141         tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
10142                    IS_FASTOPEN(tp->t_flags));
10143         if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) &&
10144             TCPS_HAVERCVDFIN(tp->t_state) == 0) {
10145                 tcp_seq save_start = th->th_seq;
10146                 tcp_seq save_rnxt  = tp->rcv_nxt;
10147                 int     save_tlen  = tlen;
10148
10149                 m_adj(m, drop_hdrlen);  /* delayed header drop */
10150                 /*
10151                  * Insert segment which includes th into TCP reassembly
10152                  * queue with control block tp.  Set thflags to whether
10153                  * reassembly now includes a segment with FIN.  This handles
10154                  * the common case inline (segment is the next to be
10155                  * received on an established connection, and the queue is
10156                  * empty), avoiding linkage into and removal from the queue
10157                  * and repetition of various conversions. Set DELACK for
10158                  * segments received in order, but ack immediately when
10159                  * segments are out of order (so fast retransmit can work).
10160                  */
10161                 if (th->th_seq == tp->rcv_nxt &&
10162                     SEGQ_EMPTY(tp) &&
10163                     (TCPS_HAVEESTABLISHED(tp->t_state) ||
10164                     tfo_syn)) {
10165 #ifdef NETFLIX_SB_LIMITS
10166                         u_int mcnt, appended;
10167
10168                         if (so->so_rcv.sb_shlim) {
10169                                 mcnt = m_memcnt(m);
10170                                 appended = 0;
10171                                 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
10172                                     CFO_NOSLEEP, NULL) == false) {
10173                                         counter_u64_add(tcp_sb_shlim_fails, 1);
10174                                         m_freem(m);
10175                                         return (0);
10176                                 }
10177                         }
10178 #endif
10179                         rack_handle_delayed_ack(tp, rack, tlen, tfo_syn);
10180                         tp->rcv_nxt += tlen;
10181                         if (tlen &&
10182                             ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
10183                             (tp->t_fbyte_in == 0)) {
10184                                 tp->t_fbyte_in = ticks;
10185                                 if (tp->t_fbyte_in == 0)
10186                                         tp->t_fbyte_in = 1;
10187                                 if (tp->t_fbyte_out && tp->t_fbyte_in)
10188                                         tp->t_flags2 |= TF2_FBYTES_COMPLETE;
10189                         }
10190                         thflags = th->th_flags & TH_FIN;
10191                         KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
10192                         KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
10193                         SOCKBUF_LOCK(&so->so_rcv);
10194                         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10195                                 m_freem(m);
10196                         } else
10197 #ifdef NETFLIX_SB_LIMITS
10198                                 appended =
10199 #endif
10200                                         sbappendstream_locked(&so->so_rcv, m, 0);
10201
10202                         rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
10203                         SOCKBUF_UNLOCK(&so->so_rcv);
10204                         tp->t_flags |= TF_WAKESOR;
10205 #ifdef NETFLIX_SB_LIMITS
10206                         if (so->so_rcv.sb_shlim && appended != mcnt)
10207                                 counter_fo_release(so->so_rcv.sb_shlim,
10208                                     mcnt - appended);
10209 #endif
10210                 } else {
10211                         /*
10212                          * XXX: Due to the header drop above "th" is
10213                          * theoretically invalid by now.  Fortunately
10214                          * m_adj() doesn't actually frees any mbufs when
10215                          * trimming from the head.
10216                          */
10217                         tcp_seq temp = save_start;
10218
10219                         thflags = tcp_reass(tp, th, &temp, &tlen, m);
10220                         tp->t_flags |= TF_ACKNOW;
10221                 }
10222                 if ((tp->t_flags & TF_SACK_PERMIT) &&
10223                     (save_tlen > 0) &&
10224                     TCPS_HAVEESTABLISHED(tp->t_state)) {
10225                         if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) {
10226                                 /*
10227                                  * DSACK actually handled in the fastpath
10228                                  * above.
10229                                  */
10230                                 RACK_OPTS_INC(tcp_sack_path_1);
10231                                 tcp_update_sack_list(tp, save_start,
10232                                     save_start + save_tlen);
10233                         } else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) {
10234                                 if ((tp->rcv_numsacks >= 1) &&
10235                                     (tp->sackblks[0].end == save_start)) {
10236                                         /*
10237                                          * Partial overlap, recorded at todrop
10238                                          * above.
10239                                          */
10240                                         RACK_OPTS_INC(tcp_sack_path_2a);
10241                                         tcp_update_sack_list(tp,
10242                                             tp->sackblks[0].start,
10243                                             tp->sackblks[0].end);
10244                                 } else {
10245                                         RACK_OPTS_INC(tcp_sack_path_2b);
10246                                         tcp_update_dsack_list(tp, save_start,
10247                                             save_start + save_tlen);
10248                                 }
10249                         } else if (tlen >= save_tlen) {
10250                                 /* Update of sackblks. */
10251                                 RACK_OPTS_INC(tcp_sack_path_3);
10252                                 tcp_update_dsack_list(tp, save_start,
10253                                     save_start + save_tlen);
10254                         } else if (tlen > 0) {
10255                                 RACK_OPTS_INC(tcp_sack_path_4);
10256                                 tcp_update_dsack_list(tp, save_start,
10257                                     save_start + tlen);
10258                         }
10259                 }
10260         } else {
10261                 m_freem(m);
10262                 thflags &= ~TH_FIN;
10263         }
10264
10265         /*
10266          * If FIN is received ACK the FIN and let the user know that the
10267          * connection is closing.
10268          */
10269         if (thflags & TH_FIN) {
10270                 if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
10271                         socantrcvmore(so);
10272                         /* The socket upcall is handled by socantrcvmore. */
10273                         tp->t_flags &= ~TF_WAKESOR;
10274                         /*
10275                          * If connection is half-synchronized (ie NEEDSYN
10276                          * flag on) then delay ACK, so it may be piggybacked
10277                          * when SYN is sent. Otherwise, since we received a
10278                          * FIN then no more input can be expected, send ACK
10279                          * now.
10280                          */
10281                         if (tp->t_flags & TF_NEEDSYN) {
10282                                 rack_timer_cancel(tp, rack,
10283                                     rack->r_ctl.rc_rcvtime, __LINE__);
10284                                 tp->t_flags |= TF_DELACK;
10285                         } else {
10286                                 tp->t_flags |= TF_ACKNOW;
10287                         }
10288                         tp->rcv_nxt++;
10289                 }
10290                 switch (tp->t_state) {
10291                         /*
10292                          * In SYN_RECEIVED and ESTABLISHED STATES enter the
10293                          * CLOSE_WAIT state.
10294                          */
10295                 case TCPS_SYN_RECEIVED:
10296                         tp->t_starttime = ticks;
10297                         /* FALLTHROUGH */
10298                 case TCPS_ESTABLISHED:
10299                         rack_timer_cancel(tp, rack,
10300                             rack->r_ctl.rc_rcvtime, __LINE__);
10301                         tcp_state_change(tp, TCPS_CLOSE_WAIT);
10302                         break;
10303
10304                         /*
10305                          * If still in FIN_WAIT_1 STATE FIN has not been
10306                          * acked so enter the CLOSING state.
10307                          */
10308                 case TCPS_FIN_WAIT_1:
10309                         rack_timer_cancel(tp, rack,
10310                             rack->r_ctl.rc_rcvtime, __LINE__);
10311                         tcp_state_change(tp, TCPS_CLOSING);
10312                         break;
10313
10314                         /*
10315                          * In FIN_WAIT_2 state enter the TIME_WAIT state,
10316                          * starting the time-wait timer, turning off the
10317                          * other standard timers.
10318                          */
10319                 case TCPS_FIN_WAIT_2:
10320                         rack_timer_cancel(tp, rack,
10321                             rack->r_ctl.rc_rcvtime, __LINE__);
10322                         tcp_twstart(tp);
10323                         return (1);
10324                 }
10325         }
10326         /*
10327          * Return any desired output.
10328          */
10329         if ((tp->t_flags & TF_ACKNOW) ||
10330             (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) {
10331                 rack->r_wanted_output = 1;
10332         }
10333         INP_WLOCK_ASSERT(tp->t_inpcb);
10334         return (0);
10335 }
10336
10337 /*
10338  * Here nothing is really faster, its just that we
10339  * have broken out the fast-data path also just like
10340  * the fast-ack.
10341  */
10342 static int
10343 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
10344     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10345     uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos)
10346 {
10347         int32_t nsegs;
10348         int32_t newsize = 0;    /* automatic sockbuf scaling */
10349         struct tcp_rack *rack;
10350 #ifdef NETFLIX_SB_LIMITS
10351         u_int mcnt, appended;
10352 #endif
10353 #ifdef TCPDEBUG
10354         /*
10355          * The size of tcp_saveipgen must be the size of the max ip header,
10356          * now IPv6.
10357          */
10358         u_char tcp_saveipgen[IP6_HDR_LEN];
10359         struct tcphdr tcp_savetcp;
10360         short ostate = 0;
10361
10362 #endif
10363         /*
10364          * If last ACK falls within this segment's sequence numbers, record
10365          * the timestamp. NOTE that the test is modified according to the
10366          * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
10367          */
10368         if (__predict_false(th->th_seq != tp->rcv_nxt)) {
10369                 return (0);
10370         }
10371         if (__predict_false(tp->snd_nxt != tp->snd_max)) {
10372                 return (0);
10373         }
10374         if (tiwin && tiwin != tp->snd_wnd) {
10375                 return (0);
10376         }
10377         if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
10378                 return (0);
10379         }
10380         if (__predict_false((to->to_flags & TOF_TS) &&
10381             (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
10382                 return (0);
10383         }
10384         if (__predict_false((th->th_ack != tp->snd_una))) {
10385                 return (0);
10386         }
10387         if (__predict_false(tlen > sbspace(&so->so_rcv))) {
10388                 return (0);
10389         }
10390         if ((to->to_flags & TOF_TS) != 0 &&
10391             SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
10392                 tp->ts_recent_age = tcp_ts_getticks();
10393                 tp->ts_recent = to->to_tsval;
10394         }
10395         rack = (struct tcp_rack *)tp->t_fb_ptr;
10396         /*
10397          * This is a pure, in-sequence data packet with nothing on the
10398          * reassembly queue and we have enough buffer space to take it.
10399          */
10400         nsegs = max(1, m->m_pkthdr.lro_nsegs);
10401
10402 #ifdef NETFLIX_SB_LIMITS
10403         if (so->so_rcv.sb_shlim) {
10404                 mcnt = m_memcnt(m);
10405                 appended = 0;
10406                 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
10407                     CFO_NOSLEEP, NULL) == false) {
10408                         counter_u64_add(tcp_sb_shlim_fails, 1);
10409                         m_freem(m);
10410                         return (1);
10411                 }
10412         }
10413 #endif
10414         /* Clean receiver SACK report if present */
10415         if (tp->rcv_numsacks)
10416                 tcp_clean_sackreport(tp);
10417         KMOD_TCPSTAT_INC(tcps_preddat);
10418         tp->rcv_nxt += tlen;
10419         if (tlen &&
10420             ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
10421             (tp->t_fbyte_in == 0)) {
10422                 tp->t_fbyte_in = ticks;
10423                 if (tp->t_fbyte_in == 0)
10424                         tp->t_fbyte_in = 1;
10425                 if (tp->t_fbyte_out && tp->t_fbyte_in)
10426                         tp->t_flags2 |= TF2_FBYTES_COMPLETE;
10427         }
10428         /*
10429          * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
10430          */
10431         tp->snd_wl1 = th->th_seq;
10432         /*
10433          * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
10434          */
10435         tp->rcv_up = tp->rcv_nxt;
10436         KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
10437         KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
10438 #ifdef TCPDEBUG
10439         if (so->so_options & SO_DEBUG)
10440                 tcp_trace(TA_INPUT, ostate, tp,
10441                     (void *)tcp_saveipgen, &tcp_savetcp, 0);
10442 #endif
10443         newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
10444
10445         /* Add data to socket buffer. */
10446         SOCKBUF_LOCK(&so->so_rcv);
10447         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10448                 m_freem(m);
10449         } else {
10450                 /*
10451                  * Set new socket buffer size. Give up when limit is
10452                  * reached.
10453                  */
10454                 if (newsize)
10455                         if (!sbreserve_locked(&so->so_rcv,
10456                             newsize, so, NULL))
10457                                 so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
10458                 m_adj(m, drop_hdrlen);  /* delayed header drop */
10459 #ifdef NETFLIX_SB_LIMITS
10460                 appended =
10461 #endif
10462                         sbappendstream_locked(&so->so_rcv, m, 0);
10463                 ctf_calc_rwin(so, tp);
10464         }
10465         rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
10466         SOCKBUF_UNLOCK(&so->so_rcv);
10467         tp->t_flags |= TF_WAKESOR;
10468 #ifdef NETFLIX_SB_LIMITS
10469         if (so->so_rcv.sb_shlim && mcnt != appended)
10470                 counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended);
10471 #endif
10472         rack_handle_delayed_ack(tp, rack, tlen, 0);
10473         if (tp->snd_una == tp->snd_max)
10474                 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
10475         return (1);
10476 }
10477
10478 /*
10479  * This subfunction is used to try to highly optimize the
10480  * fast path. We again allow window updates that are
10481  * in sequence to remain in the fast-path. We also add
10482  * in the __predict's to attempt to help the compiler.
10483  * Note that if we return a 0, then we can *not* process
10484  * it and the caller should push the packet into the
10485  * slow-path.
10486  */
10487 static int
10488 rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
10489     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10490     uint32_t tiwin, int32_t nxt_pkt, uint32_t cts)
10491 {
10492         int32_t acked;
10493         int32_t nsegs;
10494 #ifdef TCPDEBUG
10495         /*
10496          * The size of tcp_saveipgen must be the size of the max ip header,
10497          * now IPv6.
10498          */
10499         u_char tcp_saveipgen[IP6_HDR_LEN];
10500         struct tcphdr tcp_savetcp;
10501         short ostate = 0;
10502 #endif
10503         int32_t under_pacing = 0;
10504         struct tcp_rack *rack;
10505
10506         if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
10507                 /* Old ack, behind (or duplicate to) the last one rcv'd */
10508                 return (0);
10509         }
10510         if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
10511                 /* Above what we have sent? */
10512                 return (0);
10513         }
10514         if (__predict_false(tp->snd_nxt != tp->snd_max)) {
10515                 /* We are retransmitting */
10516                 return (0);
10517         }
10518         if (__predict_false(tiwin == 0)) {
10519                 /* zero window */
10520                 return (0);
10521         }
10522         if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
10523                 /* We need a SYN or a FIN, unlikely.. */
10524                 return (0);
10525         }
10526         if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
10527                 /* Timestamp is behind .. old ack with seq wrap? */
10528                 return (0);
10529         }
10530         if (__predict_false(IN_RECOVERY(tp->t_flags))) {
10531                 /* Still recovering */
10532                 return (0);
10533         }
10534         rack = (struct tcp_rack *)tp->t_fb_ptr;
10535         if (rack->r_ctl.rc_sacked) {
10536                 /* We have sack holes on our scoreboard */
10537                 return (0);
10538         }
10539         /* Ok if we reach here, we can process a fast-ack */
10540         if (rack->gp_ready &&
10541             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
10542                 under_pacing = 1;
10543         }
10544         nsegs = max(1, m->m_pkthdr.lro_nsegs);
10545         rack_log_ack(tp, to, th, 0, 0);
10546         /* Did the window get updated? */
10547         if (tiwin != tp->snd_wnd) {
10548                 tp->snd_wnd = tiwin;
10549                 rack_validate_fo_sendwin_up(tp, rack);
10550                 tp->snd_wl1 = th->th_seq;
10551                 if (tp->snd_wnd > tp->max_sndwnd)
10552                         tp->max_sndwnd = tp->snd_wnd;
10553         }
10554         /* Do we exit persists? */
10555         if ((rack->rc_in_persist != 0) &&
10556             (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
10557                                rack->r_ctl.rc_pace_min_segs))) {
10558                 rack_exit_persist(tp, rack, cts);
10559         }
10560         /* Do we enter persists? */
10561         if ((rack->rc_in_persist == 0) &&
10562             (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
10563             TCPS_HAVEESTABLISHED(tp->t_state) &&
10564             (tp->snd_max == tp->snd_una) &&
10565             sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
10566             (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
10567                 /*
10568                  * Here the rwnd is less than
10569                  * the pacing size, we are established,
10570                  * nothing is outstanding, and there is
10571                  * data to send. Enter persists.
10572                  */
10573                 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10574         }
10575         /*
10576          * If last ACK falls within this segment's sequence numbers, record
10577          * the timestamp. NOTE that the test is modified according to the
10578          * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
10579          */
10580         if ((to->to_flags & TOF_TS) != 0 &&
10581             SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
10582                 tp->ts_recent_age = tcp_ts_getticks();
10583                 tp->ts_recent = to->to_tsval;
10584         }
10585         /*
10586          * This is a pure ack for outstanding data.
10587          */
10588         KMOD_TCPSTAT_INC(tcps_predack);
10589
10590         /*
10591          * "bad retransmit" recovery.
10592          */
10593         if ((tp->t_flags & TF_PREVVALID) &&
10594             ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
10595                 tp->t_flags &= ~TF_PREVVALID;
10596                 if (tp->t_rxtshift == 1 &&
10597                     (int)(ticks - tp->t_badrxtwin) < 0)
10598                         rack_cong_signal(tp, CC_RTO_ERR, th->th_ack);
10599         }
10600         /*
10601          * Recalculate the transmit timer / rtt.
10602          *
10603          * Some boxes send broken timestamp replies during the SYN+ACK
10604          * phase, ignore timestamps of 0 or we could calculate a huge RTT
10605          * and blow up the retransmit timer.
10606          */
10607         acked = BYTES_THIS_ACK(tp, th);
10608
10609 #ifdef TCP_HHOOK
10610         /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
10611         hhook_run_tcp_est_in(tp, th, to);
10612 #endif
10613         KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
10614         KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
10615         if (acked) {
10616                 struct mbuf *mfree;
10617
10618                 rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, 0);
10619                 SOCKBUF_LOCK(&so->so_snd);
10620                 mfree = sbcut_locked(&so->so_snd, acked);
10621                 tp->snd_una = th->th_ack;
10622                 /* Note we want to hold the sb lock through the sendmap adjust */
10623                 rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
10624                 /* Wake up the socket if we have room to write more */
10625                 rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
10626                 SOCKBUF_UNLOCK(&so->so_snd);
10627                 tp->t_flags |= TF_WAKESOW;
10628                 m_freem(mfree);
10629                 tp->t_rxtshift = 0;
10630                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
10631                               rack_rto_min, rack_rto_max);
10632                 rack->rc_tlp_in_progress = 0;
10633                 rack->r_ctl.rc_tlp_cnt_out = 0;
10634                 /*
10635                  * If it is the RXT timer we want to
10636                  * stop it, so we can restart a TLP.
10637                  */
10638                 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
10639                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
10640 #ifdef NETFLIX_HTTP_LOGGING
10641                 tcp_http_check_for_comp(rack->rc_tp, th->th_ack);
10642 #endif
10643         }
10644         /*
10645          * Let the congestion control algorithm update congestion control
10646          * related information. This typically means increasing the
10647          * congestion window.
10648          */
10649         if (tp->snd_wnd < ctf_outstanding(tp)) {
10650                 /* The peer collapsed the window */
10651                 rack_collapsed_window(rack);
10652         } else if (rack->rc_has_collapsed)
10653                 rack_un_collapse_window(rack);
10654
10655         /*
10656          * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
10657          */
10658         tp->snd_wl2 = th->th_ack;
10659         tp->t_dupacks = 0;
10660         m_freem(m);
10661         /* ND6_HINT(tp);         *//* Some progress has been made. */
10662
10663         /*
10664          * If all outstanding data are acked, stop retransmit timer,
10665          * otherwise restart timer using current (possibly backed-off)
10666          * value. If process is waiting for space, wakeup/selwakeup/signal.
10667          * If data are ready to send, let tcp_output decide between more
10668          * output or persist.
10669          */
10670 #ifdef TCPDEBUG
10671         if (so->so_options & SO_DEBUG)
10672                 tcp_trace(TA_INPUT, ostate, tp,
10673                     (void *)tcp_saveipgen,
10674                     &tcp_savetcp, 0);
10675 #endif
10676         if (under_pacing &&
10677             (rack->use_fixed_rate == 0) &&
10678             (rack->in_probe_rtt == 0) &&
10679             rack->rc_gp_dyn_mul &&
10680             rack->rc_always_pace) {
10681                 /* Check if we are dragging bottom */
10682                 rack_check_bottom_drag(tp, rack, so, acked);
10683         }
10684         if (tp->snd_una == tp->snd_max) {
10685                 tp->t_flags &= ~TF_PREVVALID;
10686                 rack->r_ctl.retran_during_recovery = 0;
10687                 rack->r_ctl.dsack_byte_cnt = 0;
10688                 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
10689                 if (rack->r_ctl.rc_went_idle_time == 0)
10690                         rack->r_ctl.rc_went_idle_time = 1;
10691                 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
10692                 if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
10693                         tp->t_acktime = 0;
10694                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
10695         }
10696         if (acked && rack->r_fast_output)
10697                 rack_gain_for_fastoutput(rack, tp, so, (uint32_t)acked);
10698         if (sbavail(&so->so_snd)) {
10699                 rack->r_wanted_output = 1;
10700         }
10701         return (1);
10702 }
10703
10704 /*
10705  * Return value of 1, the TCB is unlocked and most
10706  * likely gone, return value of 0, the TCP is still
10707  * locked.
10708  */
10709 static int
10710 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
10711     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10712     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
10713 {
10714         int32_t ret_val = 0;
10715         int32_t todrop;
10716         int32_t ourfinisacked = 0;
10717         struct tcp_rack *rack;
10718
10719         ctf_calc_rwin(so, tp);
10720         /*
10721          * If the state is SYN_SENT: if seg contains an ACK, but not for our
10722          * SYN, drop the input. if seg contains a RST, then drop the
10723          * connection. if seg does not contain SYN, then drop it. Otherwise
10724          * this is an acceptable SYN segment initialize tp->rcv_nxt and
10725          * tp->irs if seg contains ack then advance tp->snd_una if seg
10726          * contains an ECE and ECN support is enabled, the stream is ECN
10727          * capable. if SYN has been acked change to ESTABLISHED else
10728          * SYN_RCVD state arrange for segment to be acked (eventually)
10729          * continue processing rest of data/controls.
10730          */
10731         if ((thflags & TH_ACK) &&
10732             (SEQ_LEQ(th->th_ack, tp->iss) ||
10733             SEQ_GT(th->th_ack, tp->snd_max))) {
10734                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
10735                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
10736                 return (1);
10737         }
10738         if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
10739                 TCP_PROBE5(connect__refused, NULL, tp,
10740                     mtod(m, const char *), tp, th);
10741                 tp = tcp_drop(tp, ECONNREFUSED);
10742                 ctf_do_drop(m, tp);
10743                 return (1);
10744         }
10745         if (thflags & TH_RST) {
10746                 ctf_do_drop(m, tp);
10747                 return (1);
10748         }
10749         if (!(thflags & TH_SYN)) {
10750                 ctf_do_drop(m, tp);
10751                 return (1);
10752         }
10753         tp->irs = th->th_seq;
10754         tcp_rcvseqinit(tp);
10755         rack = (struct tcp_rack *)tp->t_fb_ptr;
10756         if (thflags & TH_ACK) {
10757                 int tfo_partial = 0;
10758
10759                 KMOD_TCPSTAT_INC(tcps_connects);
10760                 soisconnected(so);
10761 #ifdef MAC
10762                 mac_socketpeer_set_from_mbuf(m, so);
10763 #endif
10764                 /* Do window scaling on this connection? */
10765                 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
10766                     (TF_RCVD_SCALE | TF_REQ_SCALE)) {
10767                         tp->rcv_scale = tp->request_r_scale;
10768                 }
10769                 tp->rcv_adv += min(tp->rcv_wnd,
10770                     TCP_MAXWIN << tp->rcv_scale);
10771                 /*
10772                  * If not all the data that was sent in the TFO SYN
10773                  * has been acked, resend the remainder right away.
10774                  */
10775                 if (IS_FASTOPEN(tp->t_flags) &&
10776                     (tp->snd_una != tp->snd_max)) {
10777                         tp->snd_nxt = th->th_ack;
10778                         tfo_partial = 1;
10779                 }
10780                 /*
10781                  * If there's data, delay ACK; if there's also a FIN ACKNOW
10782                  * will be turned on later.
10783                  */
10784                 if (DELAY_ACK(tp, tlen) && tlen != 0 && !tfo_partial) {
10785                         rack_timer_cancel(tp, rack,
10786                                           rack->r_ctl.rc_rcvtime, __LINE__);
10787                         tp->t_flags |= TF_DELACK;
10788                 } else {
10789                         rack->r_wanted_output = 1;
10790                         tp->t_flags |= TF_ACKNOW;
10791                         rack->rc_dack_toggle = 0;
10792                 }
10793                 if (((thflags & (TH_CWR | TH_ECE)) == TH_ECE) &&
10794                     (V_tcp_do_ecn == 1)) {
10795                         tp->t_flags2 |= TF2_ECN_PERMIT;
10796                         KMOD_TCPSTAT_INC(tcps_ecn_shs);
10797                 }
10798                 if (SEQ_GT(th->th_ack, tp->snd_una)) {
10799                         /*
10800                          * We advance snd_una for the
10801                          * fast open case. If th_ack is
10802                          * acknowledging data beyond
10803                          * snd_una we can't just call
10804                          * ack-processing since the
10805                          * data stream in our send-map
10806                          * will start at snd_una + 1 (one
10807                          * beyond the SYN). If its just
10808                          * equal we don't need to do that
10809                          * and there is no send_map.
10810                          */
10811                         tp->snd_una++;
10812                 }
10813                 /*
10814                  * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
10815                  * SYN_SENT  --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
10816                  */
10817                 tp->t_starttime = ticks;
10818                 if (tp->t_flags & TF_NEEDFIN) {
10819                         tcp_state_change(tp, TCPS_FIN_WAIT_1);
10820                         tp->t_flags &= ~TF_NEEDFIN;
10821                         thflags &= ~TH_SYN;
10822                 } else {
10823                         tcp_state_change(tp, TCPS_ESTABLISHED);
10824                         TCP_PROBE5(connect__established, NULL, tp,
10825                             mtod(m, const char *), tp, th);
10826                         rack_cc_conn_init(tp);
10827                 }
10828         } else {
10829                 /*
10830                  * Received initial SYN in SYN-SENT[*] state => simultaneous
10831                  * open.  If segment contains CC option and there is a
10832                  * cached CC, apply TAO test. If it succeeds, connection is *
10833                  * half-synchronized. Otherwise, do 3-way handshake:
10834                  * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
10835                  * there was no CC option, clear cached CC value.
10836                  */
10837                 tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN);
10838                 tcp_state_change(tp, TCPS_SYN_RECEIVED);
10839         }
10840         INP_WLOCK_ASSERT(tp->t_inpcb);
10841         /*
10842          * Advance th->th_seq to correspond to first data byte. If data,
10843          * trim to stay within window, dropping FIN if necessary.
10844          */
10845         th->th_seq++;
10846         if (tlen > tp->rcv_wnd) {
10847                 todrop = tlen - tp->rcv_wnd;
10848                 m_adj(m, -todrop);
10849                 tlen = tp->rcv_wnd;
10850                 thflags &= ~TH_FIN;
10851                 KMOD_TCPSTAT_INC(tcps_rcvpackafterwin);
10852                 KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
10853         }
10854         tp->snd_wl1 = th->th_seq - 1;
10855         tp->rcv_up = th->th_seq;
10856         /*
10857          * Client side of transaction: already sent SYN and data. If the
10858          * remote host used T/TCP to validate the SYN, our data will be
10859          * ACK'd; if so, enter normal data segment processing in the middle
10860          * of step 5, ack processing. Otherwise, goto step 6.
10861          */
10862         if (thflags & TH_ACK) {
10863                 /* For syn-sent we need to possibly update the rtt */
10864                 if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
10865                         uint32_t t, mcts;
10866
10867                         mcts = tcp_ts_getticks();
10868                         t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
10869                         if (!tp->t_rttlow || tp->t_rttlow > t)
10870                                 tp->t_rttlow = t;
10871                         rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 4);
10872                         tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
10873                         tcp_rack_xmit_timer_commit(rack, tp);
10874                 }
10875                 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val))
10876                         return (ret_val);
10877                 /* We may have changed to FIN_WAIT_1 above */
10878                 if (tp->t_state == TCPS_FIN_WAIT_1) {
10879                         /*
10880                          * In FIN_WAIT_1 STATE in addition to the processing
10881                          * for the ESTABLISHED state if our FIN is now
10882                          * acknowledged then enter FIN_WAIT_2.
10883                          */
10884                         if (ourfinisacked) {
10885                                 /*
10886                                  * If we can't receive any more data, then
10887                                  * closing user can proceed. Starting the
10888                                  * timer is contrary to the specification,
10889                                  * but if we don't get a FIN we'll hang
10890                                  * forever.
10891                                  *
10892                                  * XXXjl: we should release the tp also, and
10893                                  * use a compressed state.
10894                                  */
10895                                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10896                                         soisdisconnected(so);
10897                                         tcp_timer_activate(tp, TT_2MSL,
10898                                             (tcp_fast_finwait2_recycle ?
10899                                             tcp_finwait2_timeout :
10900                                             TP_MAXIDLE(tp)));
10901                                 }
10902                                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
10903                         }
10904                 }
10905         }
10906         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
10907            tiwin, thflags, nxt_pkt));
10908 }
10909
10910 /*
10911  * Return value of 1, the TCB is unlocked and most
10912  * likely gone, return value of 0, the TCP is still
10913  * locked.
10914  */
10915 static int
10916 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
10917     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10918     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
10919 {
10920         struct tcp_rack *rack;
10921         int32_t ret_val = 0;
10922         int32_t ourfinisacked = 0;
10923
10924         ctf_calc_rwin(so, tp);
10925         if ((thflags & TH_ACK) &&
10926             (SEQ_LEQ(th->th_ack, tp->snd_una) ||
10927             SEQ_GT(th->th_ack, tp->snd_max))) {
10928                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
10929                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
10930                 return (1);
10931         }
10932         rack = (struct tcp_rack *)tp->t_fb_ptr;
10933         if (IS_FASTOPEN(tp->t_flags)) {
10934                 /*
10935                  * When a TFO connection is in SYN_RECEIVED, the
10936                  * only valid packets are the initial SYN, a
10937                  * retransmit/copy of the initial SYN (possibly with
10938                  * a subset of the original data), a valid ACK, a
10939                  * FIN, or a RST.
10940                  */
10941                 if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
10942                         tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
10943                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
10944                         return (1);
10945                 } else if (thflags & TH_SYN) {
10946                         /* non-initial SYN is ignored */
10947                         if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
10948                             (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
10949                             (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
10950                                 ctf_do_drop(m, NULL);
10951                                 return (0);
10952                         }
10953                 } else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
10954                         ctf_do_drop(m, NULL);
10955                         return (0);
10956                 }
10957         }
10958         if ((thflags & TH_RST) ||
10959             (tp->t_fin_is_rst && (thflags & TH_FIN)))
10960                 return (ctf_process_rst(m, th, so, tp));
10961         /*
10962          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
10963          * it's less than ts_recent, drop it.
10964          */
10965         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
10966             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
10967                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
10968                         return (ret_val);
10969         }
10970         /*
10971          * In the SYN-RECEIVED state, validate that the packet belongs to
10972          * this connection before trimming the data to fit the receive
10973          * window.  Check the sequence number versus IRS since we know the
10974          * sequence numbers haven't wrapped.  This is a partial fix for the
10975          * "LAND" DoS attack.
10976          */
10977         if (SEQ_LT(th->th_seq, tp->irs)) {
10978                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
10979                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
10980                 return (1);
10981         }
10982         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
10983                               &rack->r_ctl.challenge_ack_ts,
10984                               &rack->r_ctl.challenge_ack_cnt)) {
10985                 return (ret_val);
10986         }
10987         /*
10988          * If last ACK falls within this segment's sequence numbers, record
10989          * its timestamp. NOTE: 1) That the test incorporates suggestions
10990          * from the latest proposal of the tcplw@cray.com list (Braden
10991          * 1993/04/26). 2) That updating only on newer timestamps interferes
10992          * with our earlier PAWS tests, so this check should be solely
10993          * predicated on the sequence space of this segment. 3) That we
10994          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
10995          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
10996          * SEG.Len, This modified check allows us to overcome RFC1323's
10997          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
10998          * p.869. In such cases, we can still calculate the RTT correctly
10999          * when RCV.NXT == Last.ACK.Sent.
11000          */
11001         if ((to->to_flags & TOF_TS) != 0 &&
11002             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11003             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11004             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11005                 tp->ts_recent_age = tcp_ts_getticks();
11006                 tp->ts_recent = to->to_tsval;
11007         }
11008         tp->snd_wnd = tiwin;
11009         rack_validate_fo_sendwin_up(tp, rack);
11010         /*
11011          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11012          * is on (half-synchronized state), then queue data for later
11013          * processing; else drop segment and return.
11014          */
11015         if ((thflags & TH_ACK) == 0) {
11016                 if (IS_FASTOPEN(tp->t_flags)) {
11017                         rack_cc_conn_init(tp);
11018                 }
11019                 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11020                     tiwin, thflags, nxt_pkt));
11021         }
11022         KMOD_TCPSTAT_INC(tcps_connects);
11023         soisconnected(so);
11024         /* Do window scaling? */
11025         if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
11026             (TF_RCVD_SCALE | TF_REQ_SCALE)) {
11027                 tp->rcv_scale = tp->request_r_scale;
11028         }
11029         /*
11030          * Make transitions: SYN-RECEIVED  -> ESTABLISHED SYN-RECEIVED* ->
11031          * FIN-WAIT-1
11032          */
11033         tp->t_starttime = ticks;
11034         if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) {
11035                 tcp_fastopen_decrement_counter(tp->t_tfo_pending);
11036                 tp->t_tfo_pending = NULL;
11037         }
11038         if (tp->t_flags & TF_NEEDFIN) {
11039                 tcp_state_change(tp, TCPS_FIN_WAIT_1);
11040                 tp->t_flags &= ~TF_NEEDFIN;
11041         } else {
11042                 tcp_state_change(tp, TCPS_ESTABLISHED);
11043                 TCP_PROBE5(accept__established, NULL, tp,
11044                     mtod(m, const char *), tp, th);
11045                 /*
11046                  * TFO connections call cc_conn_init() during SYN
11047                  * processing.  Calling it again here for such connections
11048                  * is not harmless as it would undo the snd_cwnd reduction
11049                  * that occurs when a TFO SYN|ACK is retransmitted.
11050                  */
11051                 if (!IS_FASTOPEN(tp->t_flags))
11052                         rack_cc_conn_init(tp);
11053         }
11054         /*
11055          * Account for the ACK of our SYN prior to
11056          * regular ACK processing below, except for
11057          * simultaneous SYN, which is handled later.
11058          */
11059         if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN))
11060                 tp->snd_una++;
11061         /*
11062          * If segment contains data or ACK, will call tcp_reass() later; if
11063          * not, do so now to pass queued data to user.
11064          */
11065         if (tlen == 0 && (thflags & TH_FIN) == 0)
11066                 (void) tcp_reass(tp, (struct tcphdr *)0, NULL, 0,
11067                     (struct mbuf *)0);
11068         tp->snd_wl1 = th->th_seq - 1;
11069         /* For syn-recv we need to possibly update the rtt */
11070         if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
11071                 uint32_t t, mcts;
11072
11073                 mcts = tcp_ts_getticks();
11074                 t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
11075                 if (!tp->t_rttlow || tp->t_rttlow > t)
11076                         tp->t_rttlow = t;
11077                 rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 5);
11078                 tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
11079                 tcp_rack_xmit_timer_commit(rack, tp);
11080         }
11081         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11082                 return (ret_val);
11083         }
11084         if (tp->t_state == TCPS_FIN_WAIT_1) {
11085                 /* We could have went to FIN_WAIT_1 (or EST) above */
11086                 /*
11087                  * In FIN_WAIT_1 STATE in addition to the processing for the
11088                  * ESTABLISHED state if our FIN is now acknowledged then
11089                  * enter FIN_WAIT_2.
11090                  */
11091                 if (ourfinisacked) {
11092                         /*
11093                          * If we can't receive any more data, then closing
11094                          * user can proceed. Starting the timer is contrary
11095                          * to the specification, but if we don't get a FIN
11096                          * we'll hang forever.
11097                          *
11098                          * XXXjl: we should release the tp also, and use a
11099                          * compressed state.
11100                          */
11101                         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11102                                 soisdisconnected(so);
11103                                 tcp_timer_activate(tp, TT_2MSL,
11104                                     (tcp_fast_finwait2_recycle ?
11105                                     tcp_finwait2_timeout :
11106                                     TP_MAXIDLE(tp)));
11107                         }
11108                         tcp_state_change(tp, TCPS_FIN_WAIT_2);
11109                 }
11110         }
11111         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11112             tiwin, thflags, nxt_pkt));
11113 }
11114
11115 /*
11116  * Return value of 1, the TCB is unlocked and most
11117  * likely gone, return value of 0, the TCP is still
11118  * locked.
11119  */
11120 static int
11121 rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
11122     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11123     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11124 {
11125         int32_t ret_val = 0;
11126         struct tcp_rack *rack;
11127
11128         /*
11129          * Header prediction: check for the two common cases of a
11130          * uni-directional data xfer.  If the packet has no control flags,
11131          * is in-sequence, the window didn't change and we're not
11132          * retransmitting, it's a candidate.  If the length is zero and the
11133          * ack moved forward, we're the sender side of the xfer.  Just free
11134          * the data acked & wake any higher level process that was blocked
11135          * waiting for space.  If the length is non-zero and the ack didn't
11136          * move, we're the receiver side.  If we're getting packets in-order
11137          * (the reassembly queue is empty), add the data toc The socket
11138          * buffer and note that we need a delayed ack. Make sure that the
11139          * hidden state-flags are also off. Since we check for
11140          * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
11141          */
11142         rack = (struct tcp_rack *)tp->t_fb_ptr;
11143         if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
11144             __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_ACK)) == TH_ACK) &&
11145             __predict_true(SEGQ_EMPTY(tp)) &&
11146             __predict_true(th->th_seq == tp->rcv_nxt)) {
11147                 if (tlen == 0) {
11148                         if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
11149                             tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) {
11150                                 return (0);
11151                         }
11152                 } else {
11153                         if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
11154                             tiwin, nxt_pkt, iptos)) {
11155                                 return (0);
11156                         }
11157                 }
11158         }
11159         ctf_calc_rwin(so, tp);
11160
11161         if ((thflags & TH_RST) ||
11162             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11163                 return (ctf_process_rst(m, th, so, tp));
11164
11165         /*
11166          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11167          * synchronized state.
11168          */
11169         if (thflags & TH_SYN) {
11170                 ctf_challenge_ack(m, th, tp, &ret_val);
11171                 return (ret_val);
11172         }
11173         /*
11174          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11175          * it's less than ts_recent, drop it.
11176          */
11177         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11178             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11179                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11180                         return (ret_val);
11181         }
11182         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11183                               &rack->r_ctl.challenge_ack_ts,
11184                               &rack->r_ctl.challenge_ack_cnt)) {
11185                 return (ret_val);
11186         }
11187         /*
11188          * If last ACK falls within this segment's sequence numbers, record
11189          * its timestamp. NOTE: 1) That the test incorporates suggestions
11190          * from the latest proposal of the tcplw@cray.com list (Braden
11191          * 1993/04/26). 2) That updating only on newer timestamps interferes
11192          * with our earlier PAWS tests, so this check should be solely
11193          * predicated on the sequence space of this segment. 3) That we
11194          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11195          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11196          * SEG.Len, This modified check allows us to overcome RFC1323's
11197          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11198          * p.869. In such cases, we can still calculate the RTT correctly
11199          * when RCV.NXT == Last.ACK.Sent.
11200          */
11201         if ((to->to_flags & TOF_TS) != 0 &&
11202             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11203             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11204             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11205                 tp->ts_recent_age = tcp_ts_getticks();
11206                 tp->ts_recent = to->to_tsval;
11207         }
11208         /*
11209          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11210          * is on (half-synchronized state), then queue data for later
11211          * processing; else drop segment and return.
11212          */
11213         if ((thflags & TH_ACK) == 0) {
11214                 if (tp->t_flags & TF_NEEDSYN) {
11215                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11216                             tiwin, thflags, nxt_pkt));
11217
11218                 } else if (tp->t_flags & TF_ACKNOW) {
11219                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11220                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11221                         return (ret_val);
11222                 } else {
11223                         ctf_do_drop(m, NULL);
11224                         return (0);
11225                 }
11226         }
11227         /*
11228          * Ack processing.
11229          */
11230         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
11231                 return (ret_val);
11232         }
11233         if (sbavail(&so->so_snd)) {
11234                 if (ctf_progress_timeout_check(tp, true)) {
11235                         rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
11236                         tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11237                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11238                         return (1);
11239                 }
11240         }
11241         /* State changes only happen in rack_process_data() */
11242         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11243             tiwin, thflags, nxt_pkt));
11244 }
11245
11246 /*
11247  * Return value of 1, the TCB is unlocked and most
11248  * likely gone, return value of 0, the TCP is still
11249  * locked.
11250  */
11251 static int
11252 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
11253     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11254     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11255 {
11256         int32_t ret_val = 0;
11257         struct tcp_rack *rack;
11258
11259         rack = (struct tcp_rack *)tp->t_fb_ptr;
11260         ctf_calc_rwin(so, tp);
11261         if ((thflags & TH_RST) ||
11262             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11263                 return (ctf_process_rst(m, th, so, tp));
11264         /*
11265          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11266          * synchronized state.
11267          */
11268         if (thflags & TH_SYN) {
11269                 ctf_challenge_ack(m, th, tp, &ret_val);
11270                 return (ret_val);
11271         }
11272         /*
11273          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11274          * it's less than ts_recent, drop it.
11275          */
11276         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11277             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11278                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11279                         return (ret_val);
11280         }
11281         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11282                               &rack->r_ctl.challenge_ack_ts,
11283                               &rack->r_ctl.challenge_ack_cnt)) {
11284                 return (ret_val);
11285         }
11286         /*
11287          * If last ACK falls within this segment's sequence numbers, record
11288          * its timestamp. NOTE: 1) That the test incorporates suggestions
11289          * from the latest proposal of the tcplw@cray.com list (Braden
11290          * 1993/04/26). 2) That updating only on newer timestamps interferes
11291          * with our earlier PAWS tests, so this check should be solely
11292          * predicated on the sequence space of this segment. 3) That we
11293          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11294          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11295          * SEG.Len, This modified check allows us to overcome RFC1323's
11296          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11297          * p.869. In such cases, we can still calculate the RTT correctly
11298          * when RCV.NXT == Last.ACK.Sent.
11299          */
11300         if ((to->to_flags & TOF_TS) != 0 &&
11301             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11302             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11303             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11304                 tp->ts_recent_age = tcp_ts_getticks();
11305                 tp->ts_recent = to->to_tsval;
11306         }
11307         /*
11308          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11309          * is on (half-synchronized state), then queue data for later
11310          * processing; else drop segment and return.
11311          */
11312         if ((thflags & TH_ACK) == 0) {
11313                 if (tp->t_flags & TF_NEEDSYN) {
11314                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11315                             tiwin, thflags, nxt_pkt));
11316
11317                 } else if (tp->t_flags & TF_ACKNOW) {
11318                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11319                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11320                         return (ret_val);
11321                 } else {
11322                         ctf_do_drop(m, NULL);
11323                         return (0);
11324                 }
11325         }
11326         /*
11327          * Ack processing.
11328          */
11329         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
11330                 return (ret_val);
11331         }
11332         if (sbavail(&so->so_snd)) {
11333                 if (ctf_progress_timeout_check(tp, true)) {
11334                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11335                                                 tp, tick, PROGRESS_DROP, __LINE__);
11336                         tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11337                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11338                         return (1);
11339                 }
11340         }
11341         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11342             tiwin, thflags, nxt_pkt));
11343 }
11344
11345 static int
11346 rack_check_data_after_close(struct mbuf *m,
11347     struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so)
11348 {
11349         struct tcp_rack *rack;
11350
11351         rack = (struct tcp_rack *)tp->t_fb_ptr;
11352         if (rack->rc_allow_data_af_clo == 0) {
11353         close_now:
11354                 tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
11355                 /* tcp_close will kill the inp pre-log the Reset */
11356                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
11357                 tp = tcp_close(tp);
11358                 KMOD_TCPSTAT_INC(tcps_rcvafterclose);
11359                 ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen));
11360                 return (1);
11361         }
11362         if (sbavail(&so->so_snd) == 0)
11363                 goto close_now;
11364         /* Ok we allow data that is ignored and a followup reset */
11365         tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
11366         tp->rcv_nxt = th->th_seq + *tlen;
11367         tp->t_flags2 |= TF2_DROP_AF_DATA;
11368         rack->r_wanted_output = 1;
11369         *tlen = 0;
11370         return (0);
11371 }
11372
11373 /*
11374  * Return value of 1, the TCB is unlocked and most
11375  * likely gone, return value of 0, the TCP is still
11376  * locked.
11377  */
11378 static int
11379 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
11380     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11381     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11382 {
11383         int32_t ret_val = 0;
11384         int32_t ourfinisacked = 0;
11385         struct tcp_rack *rack;
11386
11387         rack = (struct tcp_rack *)tp->t_fb_ptr;
11388         ctf_calc_rwin(so, tp);
11389
11390         if ((thflags & TH_RST) ||
11391             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11392                 return (ctf_process_rst(m, th, so, tp));
11393         /*
11394          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11395          * synchronized state.
11396          */
11397         if (thflags & TH_SYN) {
11398                 ctf_challenge_ack(m, th, tp, &ret_val);
11399                 return (ret_val);
11400         }
11401         /*
11402          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11403          * it's less than ts_recent, drop it.
11404          */
11405         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11406             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11407                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11408                         return (ret_val);
11409         }
11410         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11411                               &rack->r_ctl.challenge_ack_ts,
11412                               &rack->r_ctl.challenge_ack_cnt)) {
11413                 return (ret_val);
11414         }
11415         /*
11416          * If new data are received on a connection after the user processes
11417          * are gone, then RST the other end.
11418          */
11419         if ((so->so_state & SS_NOFDREF) && tlen) {
11420                 if (rack_check_data_after_close(m, tp, &tlen, th, so))
11421                         return (1);
11422         }
11423         /*
11424          * If last ACK falls within this segment's sequence numbers, record
11425          * its timestamp. NOTE: 1) That the test incorporates suggestions
11426          * from the latest proposal of the tcplw@cray.com list (Braden
11427          * 1993/04/26). 2) That updating only on newer timestamps interferes
11428          * with our earlier PAWS tests, so this check should be solely
11429          * predicated on the sequence space of this segment. 3) That we
11430          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11431          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11432          * SEG.Len, This modified check allows us to overcome RFC1323's
11433          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11434          * p.869. In such cases, we can still calculate the RTT correctly
11435          * when RCV.NXT == Last.ACK.Sent.
11436          */
11437         if ((to->to_flags & TOF_TS) != 0 &&
11438             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11439             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11440             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11441                 tp->ts_recent_age = tcp_ts_getticks();
11442                 tp->ts_recent = to->to_tsval;
11443         }
11444         /*
11445          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11446          * is on (half-synchronized state), then queue data for later
11447          * processing; else drop segment and return.
11448          */
11449         if ((thflags & TH_ACK) == 0) {
11450                 if (tp->t_flags & TF_NEEDSYN) {
11451                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11452                             tiwin, thflags, nxt_pkt));
11453                 } else if (tp->t_flags & TF_ACKNOW) {
11454                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11455                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11456                         return (ret_val);
11457                 } else {
11458                         ctf_do_drop(m, NULL);
11459                         return (0);
11460                 }
11461         }
11462         /*
11463          * Ack processing.
11464          */
11465         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11466                 return (ret_val);
11467         }
11468         if (ourfinisacked) {
11469                 /*
11470                  * If we can't receive any more data, then closing user can
11471                  * proceed. Starting the timer is contrary to the
11472                  * specification, but if we don't get a FIN we'll hang
11473                  * forever.
11474                  *
11475                  * XXXjl: we should release the tp also, and use a
11476                  * compressed state.
11477                  */
11478                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11479                         soisdisconnected(so);
11480                         tcp_timer_activate(tp, TT_2MSL,
11481                             (tcp_fast_finwait2_recycle ?
11482                             tcp_finwait2_timeout :
11483                             TP_MAXIDLE(tp)));
11484                 }
11485                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
11486         }
11487         if (sbavail(&so->so_snd)) {
11488                 if (ctf_progress_timeout_check(tp, true)) {
11489                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11490                                                 tp, tick, PROGRESS_DROP, __LINE__);
11491                         tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11492                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11493                         return (1);
11494                 }
11495         }
11496         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11497             tiwin, thflags, nxt_pkt));
11498 }
11499
11500 /*
11501  * Return value of 1, the TCB is unlocked and most
11502  * likely gone, return value of 0, the TCP is still
11503  * locked.
11504  */
11505 static int
11506 rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so,
11507     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11508     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11509 {
11510         int32_t ret_val = 0;
11511         int32_t ourfinisacked = 0;
11512         struct tcp_rack *rack;
11513
11514         rack = (struct tcp_rack *)tp->t_fb_ptr;
11515         ctf_calc_rwin(so, tp);
11516
11517         if ((thflags & TH_RST) ||
11518             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11519                 return (ctf_process_rst(m, th, so, tp));
11520         /*
11521          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11522          * synchronized state.
11523          */
11524         if (thflags & TH_SYN) {
11525                 ctf_challenge_ack(m, th, tp, &ret_val);
11526                 return (ret_val);
11527         }
11528         /*
11529          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11530          * it's less than ts_recent, drop it.
11531          */
11532         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11533             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11534                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11535                         return (ret_val);
11536         }
11537         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11538                               &rack->r_ctl.challenge_ack_ts,
11539                               &rack->r_ctl.challenge_ack_cnt)) {
11540                 return (ret_val);
11541         }
11542         /*
11543          * If new data are received on a connection after the user processes
11544          * are gone, then RST the other end.
11545          */
11546         if ((so->so_state & SS_NOFDREF) && tlen) {
11547                 if (rack_check_data_after_close(m, tp, &tlen, th, so))
11548                         return (1);
11549         }
11550         /*
11551          * If last ACK falls within this segment's sequence numbers, record
11552          * its timestamp. NOTE: 1) That the test incorporates suggestions
11553          * from the latest proposal of the tcplw@cray.com list (Braden
11554          * 1993/04/26). 2) That updating only on newer timestamps interferes
11555          * with our earlier PAWS tests, so this check should be solely
11556          * predicated on the sequence space of this segment. 3) That we
11557          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11558          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11559          * SEG.Len, This modified check allows us to overcome RFC1323's
11560          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11561          * p.869. In such cases, we can still calculate the RTT correctly
11562          * when RCV.NXT == Last.ACK.Sent.
11563          */
11564         if ((to->to_flags & TOF_TS) != 0 &&
11565             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11566             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11567             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11568                 tp->ts_recent_age = tcp_ts_getticks();
11569                 tp->ts_recent = to->to_tsval;
11570         }
11571         /*
11572          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11573          * is on (half-synchronized state), then queue data for later
11574          * processing; else drop segment and return.
11575          */
11576         if ((thflags & TH_ACK) == 0) {
11577                 if (tp->t_flags & TF_NEEDSYN) {
11578                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11579                             tiwin, thflags, nxt_pkt));
11580                 } else if (tp->t_flags & TF_ACKNOW) {
11581                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11582                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11583                         return (ret_val);
11584                 } else {
11585                         ctf_do_drop(m, NULL);
11586                         return (0);
11587                 }
11588         }
11589         /*
11590          * Ack processing.
11591          */
11592         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11593                 return (ret_val);
11594         }
11595         if (ourfinisacked) {
11596                 tcp_twstart(tp);
11597                 m_freem(m);
11598                 return (1);
11599         }
11600         if (sbavail(&so->so_snd)) {
11601                 if (ctf_progress_timeout_check(tp, true)) {
11602                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11603                                                 tp, tick, PROGRESS_DROP, __LINE__);
11604                         tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11605                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11606                         return (1);
11607                 }
11608         }
11609         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11610             tiwin, thflags, nxt_pkt));
11611 }
11612
11613 /*
11614  * Return value of 1, the TCB is unlocked and most
11615  * likely gone, return value of 0, the TCP is still
11616  * locked.
11617  */
11618 static int
11619 rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
11620     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11621     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11622 {
11623         int32_t ret_val = 0;
11624         int32_t ourfinisacked = 0;
11625         struct tcp_rack *rack;
11626
11627         rack = (struct tcp_rack *)tp->t_fb_ptr;
11628         ctf_calc_rwin(so, tp);
11629
11630         if ((thflags & TH_RST) ||
11631             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11632                 return (ctf_process_rst(m, th, so, tp));
11633         /*
11634          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11635          * synchronized state.
11636          */
11637         if (thflags & TH_SYN) {
11638                 ctf_challenge_ack(m, th, tp, &ret_val);
11639                 return (ret_val);
11640         }
11641         /*
11642          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11643          * it's less than ts_recent, drop it.
11644          */
11645         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11646             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11647                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11648                         return (ret_val);
11649         }
11650         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11651                               &rack->r_ctl.challenge_ack_ts,
11652                               &rack->r_ctl.challenge_ack_cnt)) {
11653                 return (ret_val);
11654         }
11655         /*
11656          * If new data are received on a connection after the user processes
11657          * are gone, then RST the other end.
11658          */
11659         if ((so->so_state & SS_NOFDREF) && tlen) {
11660                 if (rack_check_data_after_close(m, tp, &tlen, th, so))
11661                         return (1);
11662         }
11663         /*
11664          * If last ACK falls within this segment's sequence numbers, record
11665          * its timestamp. NOTE: 1) That the test incorporates suggestions
11666          * from the latest proposal of the tcplw@cray.com list (Braden
11667          * 1993/04/26). 2) That updating only on newer timestamps interferes
11668          * with our earlier PAWS tests, so this check should be solely
11669          * predicated on the sequence space of this segment. 3) That we
11670          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11671          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11672          * SEG.Len, This modified check allows us to overcome RFC1323's
11673          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11674          * p.869. In such cases, we can still calculate the RTT correctly
11675          * when RCV.NXT == Last.ACK.Sent.
11676          */
11677         if ((to->to_flags & TOF_TS) != 0 &&
11678             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11679             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11680             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11681                 tp->ts_recent_age = tcp_ts_getticks();
11682                 tp->ts_recent = to->to_tsval;
11683         }
11684         /*
11685          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11686          * is on (half-synchronized state), then queue data for later
11687          * processing; else drop segment and return.
11688          */
11689         if ((thflags & TH_ACK) == 0) {
11690                 if (tp->t_flags & TF_NEEDSYN) {
11691                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11692                             tiwin, thflags, nxt_pkt));
11693                 } else if (tp->t_flags & TF_ACKNOW) {
11694                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11695                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11696                         return (ret_val);
11697                 } else {
11698                         ctf_do_drop(m, NULL);
11699                         return (0);
11700                 }
11701         }
11702         /*
11703          * case TCPS_LAST_ACK: Ack processing.
11704          */
11705         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11706                 return (ret_val);
11707         }
11708         if (ourfinisacked) {
11709                 tp = tcp_close(tp);
11710                 ctf_do_drop(m, tp);
11711                 return (1);
11712         }
11713         if (sbavail(&so->so_snd)) {
11714                 if (ctf_progress_timeout_check(tp, true)) {
11715                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11716                                                 tp, tick, PROGRESS_DROP, __LINE__);
11717                         tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11718                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11719                         return (1);
11720                 }
11721         }
11722         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11723             tiwin, thflags, nxt_pkt));
11724 }
11725
11726 /*
11727  * Return value of 1, the TCB is unlocked and most
11728  * likely gone, return value of 0, the TCP is still
11729  * locked.
11730  */
11731 static int
11732 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
11733     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11734     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11735 {
11736         int32_t ret_val = 0;
11737         int32_t ourfinisacked = 0;
11738         struct tcp_rack *rack;
11739
11740         rack = (struct tcp_rack *)tp->t_fb_ptr;
11741         ctf_calc_rwin(so, tp);
11742
11743         /* Reset receive buffer auto scaling when not in bulk receive mode. */
11744         if ((thflags & TH_RST) ||
11745             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11746                 return (ctf_process_rst(m, th, so, tp));
11747         /*
11748          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11749          * synchronized state.
11750          */
11751         if (thflags & TH_SYN) {
11752                 ctf_challenge_ack(m, th, tp, &ret_val);
11753                 return (ret_val);
11754         }
11755         /*
11756          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11757          * it's less than ts_recent, drop it.
11758          */
11759         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11760             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11761                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11762                         return (ret_val);
11763         }
11764         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11765                               &rack->r_ctl.challenge_ack_ts,
11766                               &rack->r_ctl.challenge_ack_cnt)) {
11767                 return (ret_val);
11768         }
11769         /*
11770          * If new data are received on a connection after the user processes
11771          * are gone, then RST the other end.
11772          */
11773         if ((so->so_state & SS_NOFDREF) &&
11774             tlen) {
11775                 if (rack_check_data_after_close(m, tp, &tlen, th, so))
11776                         return (1);
11777         }
11778         /*
11779          * If last ACK falls within this segment's sequence numbers, record
11780          * its timestamp. NOTE: 1) That the test incorporates suggestions
11781          * from the latest proposal of the tcplw@cray.com list (Braden
11782          * 1993/04/26). 2) That updating only on newer timestamps interferes
11783          * with our earlier PAWS tests, so this check should be solely
11784          * predicated on the sequence space of this segment. 3) That we
11785          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11786          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11787          * SEG.Len, This modified check allows us to overcome RFC1323's
11788          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11789          * p.869. In such cases, we can still calculate the RTT correctly
11790          * when RCV.NXT == Last.ACK.Sent.
11791          */
11792         if ((to->to_flags & TOF_TS) != 0 &&
11793             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11794             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11795             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11796                 tp->ts_recent_age = tcp_ts_getticks();
11797                 tp->ts_recent = to->to_tsval;
11798         }
11799         /*
11800          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11801          * is on (half-synchronized state), then queue data for later
11802          * processing; else drop segment and return.
11803          */
11804         if ((thflags & TH_ACK) == 0) {
11805                 if (tp->t_flags & TF_NEEDSYN) {
11806                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11807                             tiwin, thflags, nxt_pkt));
11808                 } else if (tp->t_flags & TF_ACKNOW) {
11809                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11810                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11811                         return (ret_val);
11812                 } else {
11813                         ctf_do_drop(m, NULL);
11814                         return (0);
11815                 }
11816         }
11817         /*
11818          * Ack processing.
11819          */
11820         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11821                 return (ret_val);
11822         }
11823         if (sbavail(&so->so_snd)) {
11824                 if (ctf_progress_timeout_check(tp, true)) {
11825                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11826                                                 tp, tick, PROGRESS_DROP, __LINE__);
11827                         tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11828                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11829                         return (1);
11830                 }
11831         }
11832         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11833             tiwin, thflags, nxt_pkt));
11834 }
11835
11836 static void inline
11837 rack_clear_rate_sample(struct tcp_rack *rack)
11838 {
11839         rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY;
11840         rack->r_ctl.rack_rs.rs_rtt_cnt = 0;
11841         rack->r_ctl.rack_rs.rs_rtt_tot = 0;
11842 }
11843
11844 static void
11845 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override)
11846 {
11847         uint64_t bw_est, rate_wanted;
11848         int chged = 0;
11849         uint32_t user_max, orig_min, orig_max;
11850
11851         orig_min = rack->r_ctl.rc_pace_min_segs;
11852         orig_max = rack->r_ctl.rc_pace_max_segs;
11853         user_max = ctf_fixed_maxseg(tp) * rack->rc_user_set_max_segs;
11854         if (ctf_fixed_maxseg(tp) != rack->r_ctl.rc_pace_min_segs)
11855                 chged = 1;
11856         rack->r_ctl.rc_pace_min_segs = ctf_fixed_maxseg(tp);
11857         if (rack->use_fixed_rate || rack->rc_force_max_seg) {
11858                 if (user_max != rack->r_ctl.rc_pace_max_segs)
11859                         chged = 1;
11860         }
11861         if (rack->rc_force_max_seg) {
11862                 rack->r_ctl.rc_pace_max_segs = user_max;
11863         } else if (rack->use_fixed_rate) {
11864                 bw_est = rack_get_bw(rack);
11865                 if ((rack->r_ctl.crte == NULL) ||
11866                     (bw_est != rack->r_ctl.crte->rate)) {
11867                         rack->r_ctl.rc_pace_max_segs = user_max;
11868                 } else {
11869                         /* We are pacing right at the hardware rate */
11870                         uint32_t segsiz;
11871
11872                         segsiz = min(ctf_fixed_maxseg(tp),
11873                                      rack->r_ctl.rc_pace_min_segs);
11874                         rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(
11875                                                            tp, bw_est, segsiz, 0,
11876                                                            rack->r_ctl.crte, NULL);
11877                 }
11878         } else if (rack->rc_always_pace) {
11879                 if (rack->r_ctl.gp_bw ||
11880 #ifdef NETFLIX_PEAKRATE
11881                     rack->rc_tp->t_maxpeakrate ||
11882 #endif
11883                     rack->r_ctl.init_rate) {
11884                         /* We have a rate of some sort set */
11885                         uint32_t  orig;
11886
11887                         bw_est = rack_get_bw(rack);
11888                         orig = rack->r_ctl.rc_pace_max_segs;
11889                         if (fill_override)
11890                                 rate_wanted = *fill_override;
11891                         else
11892                                 rate_wanted = rack_get_output_bw(rack, bw_est, NULL, NULL);
11893                         if (rate_wanted) {
11894                                 /* We have something */
11895                                 rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack,
11896                                                                                    rate_wanted,
11897                                                                                    ctf_fixed_maxseg(rack->rc_tp));
11898                         } else
11899                                 rack->r_ctl.rc_pace_max_segs = rack->r_ctl.rc_pace_min_segs;
11900                         if (orig != rack->r_ctl.rc_pace_max_segs)
11901                                 chged = 1;
11902                 } else if ((rack->r_ctl.gp_bw == 0) &&
11903                            (rack->r_ctl.rc_pace_max_segs == 0)) {
11904                         /*
11905                          * If we have nothing limit us to bursting
11906                          * out IW sized pieces.
11907                          */
11908                         chged = 1;
11909                         rack->r_ctl.rc_pace_max_segs = rc_init_window(rack);
11910                 }
11911         }
11912         if (rack->r_ctl.rc_pace_max_segs > PACE_MAX_IP_BYTES) {
11913                 chged = 1;
11914                 rack->r_ctl.rc_pace_max_segs = PACE_MAX_IP_BYTES;
11915         }
11916         if (chged)
11917                 rack_log_type_pacing_sizes(tp, rack, orig_min, orig_max, line, 2);
11918 }
11919
11920
11921 static void
11922 rack_init_fsb_block(struct tcpcb *tp, struct tcp_rack *rack)
11923 {
11924 #ifdef INET6
11925         struct ip6_hdr *ip6 = NULL;
11926 #endif
11927 #ifdef INET
11928         struct ip *ip = NULL;
11929 #endif
11930         struct udphdr *udp = NULL;
11931
11932         /* Ok lets fill in the fast block, it can only be used with no IP options! */
11933 #ifdef INET6
11934         if (rack->r_is_v6) {
11935                 rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
11936                 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
11937                 if (tp->t_port) {
11938                         rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
11939                         udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
11940                         udp->uh_sport = htons(V_tcp_udp_tunneling_port);
11941                         udp->uh_dport = tp->t_port;
11942                         rack->r_ctl.fsb.udp = udp;
11943                         rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
11944                 } else
11945                 {
11946                         rack->r_ctl.fsb.th = (struct tcphdr *)(ip6 + 1);
11947                         rack->r_ctl.fsb.udp = NULL;
11948                 }
11949                 tcpip_fillheaders(rack->rc_inp,
11950                                   tp->t_port,
11951                                   ip6, rack->r_ctl.fsb.th);
11952         } else
11953 #endif                          /* INET6 */
11954         {
11955                 rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr);
11956                 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
11957                 if (tp->t_port) {
11958                         rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
11959                         udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
11960                         udp->uh_sport = htons(V_tcp_udp_tunneling_port);
11961                         udp->uh_dport = tp->t_port;
11962                         rack->r_ctl.fsb.udp = udp;
11963                         rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
11964                 } else
11965                 {
11966                         rack->r_ctl.fsb.udp = NULL;
11967                         rack->r_ctl.fsb.th = (struct tcphdr *)(ip + 1);
11968                 }
11969                 tcpip_fillheaders(rack->rc_inp,
11970                                   tp->t_port,
11971                                   ip, rack->r_ctl.fsb.th);
11972         }
11973         rack->r_fsb_inited = 1;
11974 }
11975
11976 static int
11977 rack_init_fsb(struct tcpcb *tp, struct tcp_rack *rack)
11978 {
11979         /*
11980          * Allocate the larger of spaces V6 if available else just
11981          * V4 and include udphdr (overbook)
11982          */
11983 #ifdef INET6
11984         rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + sizeof(struct udphdr);
11985 #else
11986         rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr) + sizeof(struct udphdr);
11987 #endif
11988         rack->r_ctl.fsb.tcp_ip_hdr = malloc(rack->r_ctl.fsb.tcp_ip_hdr_len,
11989                                             M_TCPFSB, M_NOWAIT|M_ZERO);
11990         if (rack->r_ctl.fsb.tcp_ip_hdr == NULL) {
11991                 return (ENOMEM);
11992         }
11993         rack->r_fsb_inited = 0;
11994         return (0);
11995 }
11996
11997 static int
11998 rack_init(struct tcpcb *tp)
11999 {
12000         struct tcp_rack *rack = NULL;
12001         struct rack_sendmap *insret;
12002         uint32_t iwin, snt, us_cts;
12003         int err;
12004
12005         tp->t_fb_ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT);
12006         if (tp->t_fb_ptr == NULL) {
12007                 /*
12008                  * We need to allocate memory but cant. The INP and INP_INFO
12009                  * locks and they are recusive (happens during setup. So a
12010                  * scheme to drop the locks fails :(
12011                  *
12012                  */
12013                 return (ENOMEM);
12014         }
12015         memset(tp->t_fb_ptr, 0, sizeof(struct tcp_rack));
12016
12017         rack = (struct tcp_rack *)tp->t_fb_ptr;
12018         RB_INIT(&rack->r_ctl.rc_mtree);
12019         TAILQ_INIT(&rack->r_ctl.rc_free);
12020         TAILQ_INIT(&rack->r_ctl.rc_tmap);
12021         rack->rc_tp = tp;
12022         rack->rc_inp = tp->t_inpcb;
12023         /* Set the flag */
12024         rack->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
12025         /* Probably not needed but lets be sure */
12026         rack_clear_rate_sample(rack);
12027         /*
12028          * Save off the default values, socket options will poke
12029          * at these if pacing is not on or we have not yet
12030          * reached where pacing is on (gp_ready/fixed enabled).
12031          * When they get set into the CC module (when gp_ready
12032          * is enabled or we enable fixed) then we will set these
12033          * values into the CC and place in here the old values
12034          * so we have a restoral. Then we will set the flag
12035          * rc_pacing_cc_set. That way whenever we turn off pacing
12036          * or switch off this stack, we will know to go restore
12037          * the saved values.
12038          */
12039         rack->r_ctl.rc_saved_beta.beta = V_newreno_beta_ecn;
12040         rack->r_ctl.rc_saved_beta.beta_ecn = V_newreno_beta_ecn;
12041         /* We want abe like behavior as well */
12042         rack->r_ctl.rc_saved_beta.newreno_flags = CC_NEWRENO_BETA_ECN;
12043         rack->r_ctl.rc_reorder_fade = rack_reorder_fade;
12044         rack->rc_allow_data_af_clo = rack_ignore_data_after_close;
12045         rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh;
12046         if (use_rack_rr)
12047                 rack->use_rack_rr = 1;
12048         if (V_tcp_delack_enabled)
12049                 tp->t_delayed_ack = 1;
12050         else
12051                 tp->t_delayed_ack = 0;
12052 #ifdef TCP_ACCOUNTING
12053         if (rack_tcp_accounting) {
12054                 tp->t_flags2 |= TF2_TCP_ACCOUNTING;
12055         }
12056 #endif
12057         if (rack_enable_shared_cwnd)
12058                 rack->rack_enable_scwnd = 1;
12059         rack->rc_user_set_max_segs = rack_hptsi_segments;
12060         rack->rc_force_max_seg = 0;
12061         if (rack_use_imac_dack)
12062                 rack->rc_dack_mode = 1;
12063         TAILQ_INIT(&rack->r_ctl.opt_list);
12064         rack->r_ctl.rc_reorder_shift = rack_reorder_thresh;
12065         rack->r_ctl.rc_pkt_delay = rack_pkt_delay;
12066         rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp;
12067         rack->r_ctl.rc_lowest_us_rtt = 0xffffffff;
12068         rack->r_ctl.rc_highest_us_rtt = 0;
12069         rack->r_ctl.bw_rate_cap = rack_bw_rate_cap;
12070         if (rack_use_cmp_acks)
12071                 rack->r_use_cmp_ack = 1;
12072         if (rack_disable_prr)
12073                 rack->rack_no_prr = 1;
12074         if (rack_gp_no_rec_chg)
12075                 rack->rc_gp_no_rec_chg = 1;
12076         if (rack_pace_every_seg && tcp_can_enable_pacing()) {
12077                 rack->rc_always_pace = 1;
12078                 if (rack->use_fixed_rate || rack->gp_ready)
12079                         rack_set_cc_pacing(rack);
12080         } else
12081                 rack->rc_always_pace = 0;
12082         if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack)
12083                 rack->r_mbuf_queue = 1;
12084         else
12085                 rack->r_mbuf_queue = 0;
12086         if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
12087                 tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
12088         else
12089                 tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
12090         rack_set_pace_segments(tp, rack, __LINE__, NULL);
12091         if (rack_limits_scwnd)
12092                 rack->r_limit_scw = 1;
12093         else
12094                 rack->r_limit_scw = 0;
12095         rack->rc_labc = V_tcp_abc_l_var;
12096         rack->r_ctl.rc_high_rwnd = tp->snd_wnd;
12097         rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
12098         rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method;
12099         rack->rack_tlp_threshold_use = rack_tlp_threshold_use;
12100         rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr;
12101         rack->r_ctl.rc_min_to = rack_min_to;
12102         microuptime(&rack->r_ctl.act_rcv_time);
12103         rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
12104         rack->r_running_late = 0;
12105         rack->r_running_early = 0;
12106         rack->rc_init_win = rack_default_init_window;
12107         rack->r_ctl.rack_per_of_gp_ss = rack_per_of_gp_ss;
12108         if (rack_hw_up_only)
12109                 rack->r_up_only = 1;
12110         if (rack_do_dyn_mul) {
12111                 /* When dynamic adjustment is on CA needs to start at 100% */
12112                 rack->rc_gp_dyn_mul = 1;
12113                 if (rack_do_dyn_mul >= 100)
12114                         rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
12115         } else
12116                 rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
12117         rack->r_ctl.rack_per_of_gp_rec = rack_per_of_gp_rec;
12118         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
12119         rack->r_ctl.rc_tlp_rxt_last_time = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time);
12120         setup_time_filter_small(&rack->r_ctl.rc_gp_min_rtt, FILTER_TYPE_MIN,
12121                                 rack_probertt_filter_life);
12122         us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
12123         rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
12124         rack->r_ctl.rc_time_of_last_probertt = us_cts;
12125         rack->r_ctl.challenge_ack_ts = tcp_ts_getticks();
12126         rack->r_ctl.rc_time_probertt_starts = 0;
12127         /* We require at least one measurement, even if the sysctl is 0 */
12128         if (rack_req_measurements)
12129                 rack->r_ctl.req_measurements = rack_req_measurements;
12130         else
12131                 rack->r_ctl.req_measurements = 1;
12132         if (rack_enable_hw_pacing)
12133                 rack->rack_hdw_pace_ena = 1;
12134         if (rack_hw_rate_caps)
12135                 rack->r_rack_hw_rate_caps = 1;
12136         /* Do we force on detection? */
12137 #ifdef NETFLIX_EXP_DETECTION
12138         if (tcp_force_detection)
12139                 rack->do_detection = 1;
12140         else
12141 #endif
12142                 rack->do_detection = 0;
12143         if (rack_non_rxt_use_cr)
12144                 rack->rack_rec_nonrxt_use_cr = 1;
12145         err = rack_init_fsb(tp, rack);
12146         if (err) {
12147                 uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12148                 tp->t_fb_ptr = NULL;
12149                 return (err);
12150         }
12151         if (tp->snd_una != tp->snd_max) {
12152                 /* Create a send map for the current outstanding data */
12153                 struct rack_sendmap *rsm;
12154
12155                 rsm = rack_alloc(rack);
12156                 if (rsm == NULL) {
12157                         uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12158                         tp->t_fb_ptr = NULL;
12159                         return (ENOMEM);
12160                 }
12161                 rsm->r_no_rtt_allowed = 1;
12162                 rsm->r_tim_lastsent[0] = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
12163                 rsm->r_rtr_cnt = 1;
12164                 rsm->r_rtr_bytes = 0;
12165                 if (tp->t_flags & TF_SENTFIN) {
12166                         rsm->r_end = tp->snd_max - 1;
12167                         rsm->r_flags |= RACK_HAS_FIN;
12168                 } else {
12169                         rsm->r_end = tp->snd_max;
12170                 }
12171                 if (tp->snd_una == tp->iss) {
12172                         /* The data space is one beyond snd_una */
12173                         rsm->r_flags |= RACK_HAS_SYN;
12174                         rsm->r_start = tp->iss;
12175                         rsm->r_end = rsm->r_start + (tp->snd_max - tp->snd_una);
12176                 } else
12177                         rsm->r_start = tp->snd_una;
12178                 rsm->r_dupack = 0;
12179                 if (rack->rc_inp->inp_socket->so_snd.sb_mb != NULL) {
12180                         rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd, 0, &rsm->soff);
12181                         rsm->orig_m_len = rsm->m->m_len;
12182                 } else {
12183                         /*
12184                          * This can happen if we have a stand-alone FIN or
12185                          *  SYN.
12186                          */
12187                         rsm->m = NULL;
12188                         rsm->orig_m_len = 0;
12189                         rsm->soff = 0;
12190                 }
12191                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12192 #ifdef INVARIANTS
12193                 if (insret != NULL) {
12194                         panic("Insert in rb tree fails ret:%p rack:%p rsm:%p",
12195                               insret, rack, rsm);
12196                 }
12197 #endif
12198                 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
12199                 rsm->r_in_tmap = 1;
12200         }
12201         /*
12202          * Timers in Rack are kept in microseconds so lets
12203          * convert any initial incoming variables
12204          * from ticks into usecs. Note that we
12205          * also change the values of t_srtt and t_rttvar, if
12206          * they are non-zero. They are kept with a 5
12207          * bit decimal so we have to carefully convert
12208          * these to get the full precision.
12209          */
12210         rack_convert_rtts(tp);
12211         tp->t_rttlow = TICKS_2_USEC(tp->t_rttlow);
12212         if (rack_def_profile)
12213                 rack_set_profile(rack, rack_def_profile);
12214         /* Cancel the GP measurement in progress */
12215         tp->t_flags &= ~TF_GPUTINPROG;
12216         if (SEQ_GT(tp->snd_max, tp->iss))
12217                 snt = tp->snd_max - tp->iss;
12218         else
12219                 snt = 0;
12220         iwin = rc_init_window(rack);
12221         if (snt < iwin) {
12222                 /* We are not past the initial window
12223                  * so we need to make sure cwnd is
12224                  * correct.
12225                  */
12226                 if (tp->snd_cwnd < iwin)
12227                         tp->snd_cwnd = iwin;
12228                 /*
12229                  * If we are within the initial window
12230                  * we want ssthresh to be unlimited. Setting
12231                  * it to the rwnd (which the default stack does
12232                  * and older racks) is not really a good idea
12233                  * since we want to be in SS and grow both the
12234                  * cwnd and the rwnd (via dynamic rwnd growth). If
12235                  * we set it to the rwnd then as the peer grows its
12236                  * rwnd we will be stuck in CA and never hit SS.
12237                  *
12238                  * Its far better to raise it up high (this takes the
12239                  * risk that there as been a loss already, probably
12240                  * we should have an indicator in all stacks of loss
12241                  * but we don't), but considering the normal use this
12242                  * is a risk worth taking. The consequences of not
12243                  * hitting SS are far worse than going one more time
12244                  * into it early on (before we have sent even a IW).
12245                  * It is highly unlikely that we will have had a loss
12246                  * before getting the IW out.
12247                  */
12248                 tp->snd_ssthresh = 0xffffffff;
12249         }
12250         rack_stop_all_timers(tp);
12251         /* Lets setup the fsb block */
12252         rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
12253         rack_log_rtt_shrinks(rack,  us_cts,  tp->t_rxtcur,
12254                              __LINE__, RACK_RTTS_INIT);
12255         return (0);
12256 }
12257
12258 static int
12259 rack_handoff_ok(struct tcpcb *tp)
12260 {
12261         if ((tp->t_state == TCPS_CLOSED) ||
12262             (tp->t_state == TCPS_LISTEN)) {
12263                 /* Sure no problem though it may not stick */
12264                 return (0);
12265         }
12266         if ((tp->t_state == TCPS_SYN_SENT) ||
12267             (tp->t_state == TCPS_SYN_RECEIVED)) {
12268                 /*
12269                  * We really don't know if you support sack,
12270                  * you have to get to ESTAB or beyond to tell.
12271                  */
12272                 return (EAGAIN);
12273         }
12274         if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) > 1)) {
12275                 /*
12276                  * Rack will only send a FIN after all data is acknowledged.
12277                  * So in this case we have more data outstanding. We can't
12278                  * switch stacks until either all data and only the FIN
12279                  * is left (in which case rack_init() now knows how
12280                  * to deal with that) <or> all is acknowledged and we
12281                  * are only left with incoming data, though why you
12282                  * would want to switch to rack after all data is acknowledged
12283                  * I have no idea (rrs)!
12284                  */
12285                 return (EAGAIN);
12286         }
12287         if ((tp->t_flags & TF_SACK_PERMIT) || rack_sack_not_required){
12288                 return (0);
12289         }
12290         /*
12291          * If we reach here we don't do SACK on this connection so we can
12292          * never do rack.
12293          */
12294         return (EINVAL);
12295 }
12296
12297
12298 static void
12299 rack_fini(struct tcpcb *tp, int32_t tcb_is_purged)
12300 {
12301         int ack_cmp = 0;
12302
12303         if (tp->t_fb_ptr) {
12304                 struct tcp_rack *rack;
12305                 struct rack_sendmap *rsm, *nrsm, *rm;
12306
12307                 rack = (struct tcp_rack *)tp->t_fb_ptr;
12308                 if (tp->t_in_pkt) {
12309                         /*
12310                          * Since we are switching we need to process any
12311                          * inbound packets in case a compressed ack is
12312                          * in queue or the new stack does not support
12313                          * mbuf queuing. These packets in theory should
12314                          * have been handled by the old stack anyway.
12315                          */
12316                         if ((rack->rc_inp->inp_flags & (INP_DROPPED|INP_TIMEWAIT)) ||
12317                             (rack->rc_inp->inp_flags2 & INP_FREED)) {
12318                                 /* Kill all the packets */
12319                                 struct mbuf *save, *m;
12320
12321                                 m = tp->t_in_pkt;
12322                                 tp->t_in_pkt = NULL;
12323                                 tp->t_tail_pkt = NULL;
12324                                 while (m) {
12325                                         save = m->m_nextpkt;
12326                                         m->m_nextpkt = NULL;
12327                                         m_freem(m);
12328                                         m = save;
12329                                 }
12330                         } else {
12331                                 /* Process all the packets */
12332                                 ctf_do_queued_segments(rack->rc_inp->inp_socket, rack->rc_tp, 0);
12333                         }
12334                         if ((tp->t_inpcb) &&
12335                             (tp->t_inpcb->inp_flags2 & INP_MBUF_ACKCMP))
12336                                 ack_cmp = 1;
12337                         if (ack_cmp) {
12338                                 /* Total if we used large or small (if ack-cmp was used). */
12339                                 if (rack->rc_inp->inp_flags2 & INP_MBUF_L_ACKS)
12340                                         counter_u64_add(rack_large_ackcmp, 1);
12341                                 else
12342                                         counter_u64_add(rack_small_ackcmp, 1);
12343                         }
12344                 }
12345                 tp->t_flags &= ~TF_FORCEDATA;
12346 #ifdef NETFLIX_SHARED_CWND
12347                 if (rack->r_ctl.rc_scw) {
12348                         uint32_t limit;
12349
12350                         if (rack->r_limit_scw)
12351                                 limit = max(1, rack->r_ctl.rc_lowest_us_rtt);
12352                         else
12353                                 limit = 0;
12354                         tcp_shared_cwnd_free_full(tp, rack->r_ctl.rc_scw,
12355                                                   rack->r_ctl.rc_scw_index,
12356                                                   limit);
12357                         rack->r_ctl.rc_scw = NULL;
12358                 }
12359 #endif
12360                 if (rack->r_ctl.fsb.tcp_ip_hdr) {
12361                         free(rack->r_ctl.fsb.tcp_ip_hdr, M_TCPFSB);
12362                         rack->r_ctl.fsb.tcp_ip_hdr = NULL;
12363                         rack->r_ctl.fsb.th = NULL;
12364                 }
12365                 /* Convert back to ticks, with  */
12366                 if (tp->t_srtt > 1) {
12367                         uint32_t val, frac;
12368
12369                         val = USEC_2_TICKS(tp->t_srtt);
12370                         frac = tp->t_srtt % (HPTS_USEC_IN_SEC / hz);
12371                         tp->t_srtt = val << TCP_RTT_SHIFT;
12372                         /*
12373                          * frac is the fractional part here is left
12374                          * over from converting to hz and shifting.
12375                          * We need to convert this to the 5 bit
12376                          * remainder.
12377                          */
12378                         if (frac) {
12379                                 if (hz == 1000) {
12380                                         frac = (((uint64_t)frac *  (uint64_t)TCP_RTT_SCALE) / (uint64_t)HPTS_USEC_IN_MSEC);
12381                                 } else {
12382                                         frac = (((uint64_t)frac * (uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE) /(uint64_t)HPTS_USEC_IN_SEC);
12383                                 }
12384                                 tp->t_srtt += frac;
12385                         }
12386                 }
12387                 if (tp->t_rttvar) {
12388                         uint32_t val, frac;
12389
12390                         val = USEC_2_TICKS(tp->t_rttvar);
12391                         frac = tp->t_srtt % (HPTS_USEC_IN_SEC / hz);
12392                         tp->t_rttvar = val <<  TCP_RTTVAR_SHIFT;
12393                         /*
12394                          * frac is the fractional part here is left
12395                          * over from converting to hz and shifting.
12396                          * We need to convert this to the 5 bit
12397                          * remainder.
12398                          */
12399                         if (frac) {
12400                                 if (hz == 1000) {
12401                                         frac = (((uint64_t)frac *  (uint64_t)TCP_RTT_SCALE) / (uint64_t)HPTS_USEC_IN_MSEC);
12402                                 } else {
12403                                         frac = (((uint64_t)frac * (uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE) /(uint64_t)HPTS_USEC_IN_SEC);
12404                                 }
12405                                 tp->t_rttvar += frac;
12406                         }
12407                 }
12408                 tp->t_rxtcur = USEC_2_TICKS(tp->t_rxtcur);
12409                 tp->t_rttlow = USEC_2_TICKS(tp->t_rttlow);
12410                 if (rack->rc_always_pace) {
12411                         tcp_decrement_paced_conn();
12412                         rack_undo_cc_pacing(rack);
12413                         rack->rc_always_pace = 0;
12414                 }
12415                 /* Clean up any options if they were not applied */
12416                 while (!TAILQ_EMPTY(&rack->r_ctl.opt_list)) {
12417                         struct deferred_opt_list *dol;
12418
12419                         dol = TAILQ_FIRST(&rack->r_ctl.opt_list);
12420                         TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
12421                         free(dol, M_TCPDO);
12422                 }
12423                 /* rack does not use force data but other stacks may clear it */
12424                 if (rack->r_ctl.crte != NULL) {
12425                         tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
12426                         rack->rack_hdrw_pacing = 0;
12427                         rack->r_ctl.crte = NULL;
12428                 }
12429 #ifdef TCP_BLACKBOX
12430                 tcp_log_flowend(tp);
12431 #endif
12432                 RB_FOREACH_SAFE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm) {
12433                         rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12434 #ifdef INVARIANTS
12435                         if (rm != rsm) {
12436                                 panic("At fini, rack:%p rsm:%p rm:%p",
12437                                       rack, rsm, rm);
12438                         }
12439 #endif
12440                         uma_zfree(rack_zone, rsm);
12441                 }
12442                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
12443                 while (rsm) {
12444                         TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
12445                         uma_zfree(rack_zone, rsm);
12446                         rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
12447                 }
12448                 rack->rc_free_cnt = 0;
12449                 uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12450                 tp->t_fb_ptr = NULL;
12451         }
12452         if (tp->t_inpcb) {
12453                 tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
12454                 tp->t_inpcb->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
12455                 tp->t_inpcb->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
12456                 tp->t_inpcb->inp_flags2 &= ~INP_MBUF_ACKCMP;
12457                 /* Cancel the GP measurement in progress */
12458                 tp->t_flags &= ~TF_GPUTINPROG;
12459                 tp->t_inpcb->inp_flags2 &= ~INP_MBUF_L_ACKS;
12460         }
12461         /* Make sure snd_nxt is correctly set */
12462         tp->snd_nxt = tp->snd_max;
12463 }
12464
12465 static void
12466 rack_set_state(struct tcpcb *tp, struct tcp_rack *rack)
12467 {
12468         switch (tp->t_state) {
12469         case TCPS_SYN_SENT:
12470                 rack->r_state = TCPS_SYN_SENT;
12471                 rack->r_substate = rack_do_syn_sent;
12472                 break;
12473         case TCPS_SYN_RECEIVED:
12474                 rack->r_state = TCPS_SYN_RECEIVED;
12475                 rack->r_substate = rack_do_syn_recv;
12476                 break;
12477         case TCPS_ESTABLISHED:
12478                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12479                 rack->r_state = TCPS_ESTABLISHED;
12480                 rack->r_substate = rack_do_established;
12481                 break;
12482         case TCPS_CLOSE_WAIT:
12483                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12484                 rack->r_state = TCPS_CLOSE_WAIT;
12485                 rack->r_substate = rack_do_close_wait;
12486                 break;
12487         case TCPS_FIN_WAIT_1:
12488                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12489                 rack->r_state = TCPS_FIN_WAIT_1;
12490                 rack->r_substate = rack_do_fin_wait_1;
12491                 break;
12492         case TCPS_CLOSING:
12493                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12494                 rack->r_state = TCPS_CLOSING;
12495                 rack->r_substate = rack_do_closing;
12496                 break;
12497         case TCPS_LAST_ACK:
12498                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12499                 rack->r_state = TCPS_LAST_ACK;
12500                 rack->r_substate = rack_do_lastack;
12501                 break;
12502         case TCPS_FIN_WAIT_2:
12503                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12504                 rack->r_state = TCPS_FIN_WAIT_2;
12505                 rack->r_substate = rack_do_fin_wait_2;
12506                 break;
12507         case TCPS_LISTEN:
12508         case TCPS_CLOSED:
12509         case TCPS_TIME_WAIT:
12510         default:
12511                 break;
12512         };
12513         if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
12514                 rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
12515
12516 }
12517
12518 static void
12519 rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb)
12520 {
12521         /*
12522          * We received an ack, and then did not
12523          * call send or were bounced out due to the
12524          * hpts was running. Now a timer is up as well, is
12525          * it the right timer?
12526          */
12527         struct rack_sendmap *rsm;
12528         int tmr_up;
12529
12530         tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
12531         if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT))
12532                 return;
12533         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
12534         if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) &&
12535             (tmr_up == PACE_TMR_RXT)) {
12536                 /* Should be an RXT */
12537                 return;
12538         }
12539         if (rsm == NULL) {
12540                 /* Nothing outstanding? */
12541                 if (tp->t_flags & TF_DELACK) {
12542                         if (tmr_up == PACE_TMR_DELACK)
12543                                 /* We are supposed to have delayed ack up and we do */
12544                                 return;
12545                 } else if (sbavail(&tp->t_inpcb->inp_socket->so_snd) && (tmr_up == PACE_TMR_RXT)) {
12546                         /*
12547                          * if we hit enobufs then we would expect the possiblity
12548                          * of nothing outstanding and the RXT up (and the hptsi timer).
12549                          */
12550                         return;
12551                 } else if (((V_tcp_always_keepalive ||
12552                              rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
12553                             (tp->t_state <= TCPS_CLOSING)) &&
12554                            (tmr_up == PACE_TMR_KEEP) &&
12555                            (tp->snd_max == tp->snd_una)) {
12556                         /* We should have keep alive up and we do */
12557                         return;
12558                 }
12559         }
12560         if (SEQ_GT(tp->snd_max, tp->snd_una) &&
12561                    ((tmr_up == PACE_TMR_TLP) ||
12562                     (tmr_up == PACE_TMR_RACK) ||
12563                     (tmr_up == PACE_TMR_RXT))) {
12564                 /*
12565                  * Either a Rack, TLP or RXT is fine if  we
12566                  * have outstanding data.
12567                  */
12568                 return;
12569         } else if (tmr_up == PACE_TMR_DELACK) {
12570                 /*
12571                  * If the delayed ack was going to go off
12572                  * before the rtx/tlp/rack timer were going to
12573                  * expire, then that would be the timer in control.
12574                  * Note we don't check the time here trusting the
12575                  * code is correct.
12576                  */
12577                 return;
12578         }
12579         /*
12580          * Ok the timer originally started is not what we want now.
12581          * We will force the hpts to be stopped if any, and restart
12582          * with the slot set to what was in the saved slot.
12583          */
12584         if (rack->rc_inp->inp_in_hpts) {
12585                 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
12586                         uint32_t us_cts;
12587
12588                         us_cts = tcp_get_usecs(NULL);
12589                         if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
12590                                 rack->r_early = 1;
12591                                 rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
12592                         }
12593                         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
12594                 }
12595                 tcp_hpts_remove(tp->t_inpcb, HPTS_REMOVE_OUTPUT);
12596         }
12597         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
12598         rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
12599 }
12600
12601
12602 static void
12603 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)
12604 {
12605         tp->snd_wnd = tiwin;
12606         rack_validate_fo_sendwin_up(tp, rack);
12607         tp->snd_wl1 = seq;
12608         tp->snd_wl2 = ack;
12609         if (tp->snd_wnd > tp->max_sndwnd)
12610                 tp->max_sndwnd = tp->snd_wnd;
12611         if (tp->snd_wnd < (tp->snd_max - high_seq)) {
12612                 /* The peer collapsed the window */
12613                 rack_collapsed_window(rack);
12614         } else if (rack->rc_has_collapsed)
12615                 rack_un_collapse_window(rack);
12616         /* Do we exit persists? */
12617         if ((rack->rc_in_persist != 0) &&
12618             (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
12619                                 rack->r_ctl.rc_pace_min_segs))) {
12620                 rack_exit_persist(tp, rack, cts);
12621         }
12622         /* Do we enter persists? */
12623         if ((rack->rc_in_persist == 0) &&
12624             (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
12625             TCPS_HAVEESTABLISHED(tp->t_state) &&
12626             (tp->snd_max == tp->snd_una) &&
12627             sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
12628             (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
12629                 /*
12630                  * Here the rwnd is less than
12631                  * the pacing size, we are established,
12632                  * nothing is outstanding, and there is
12633                  * data to send. Enter persists.
12634                  */
12635                 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
12636         }
12637 }
12638
12639 static void
12640 rack_log_input_packet(struct tcpcb *tp, struct tcp_rack *rack, struct tcp_ackent *ae, int ackval, uint32_t high_seq)
12641 {
12642
12643         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
12644                 union tcp_log_stackspecific log;
12645                 struct timeval ltv;
12646                 char tcp_hdr_buf[60];
12647                 struct tcphdr *th;
12648                 struct timespec ts;
12649                 uint32_t orig_snd_una;
12650                 uint8_t xx = 0;
12651
12652 #ifdef NETFLIX_HTTP_LOGGING
12653                 struct http_sendfile_track *http_req;
12654
12655                 if (SEQ_GT(ae->ack, tp->snd_una)) {
12656                         http_req = tcp_http_find_req_for_seq(tp, (ae->ack-1));
12657                 } else {
12658                         http_req = tcp_http_find_req_for_seq(tp, ae->ack);
12659                 }
12660 #endif
12661                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
12662                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
12663                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
12664                 if (rack->rack_no_prr == 0)
12665                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
12666                 else
12667                         log.u_bbr.flex1 = 0;
12668                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
12669                 log.u_bbr.use_lt_bw <<= 1;
12670                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
12671                 log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
12672                 log.u_bbr.inflight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
12673                 log.u_bbr.pkts_out = tp->t_maxseg;
12674                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
12675                 log.u_bbr.flex7 = 1;
12676                 log.u_bbr.lost = ae->flags;
12677                 log.u_bbr.cwnd_gain = ackval;
12678                 log.u_bbr.pacing_gain = 0x2;
12679                 if (ae->flags & TSTMP_HDWR) {
12680                         /* Record the hardware timestamp if present */
12681                         log.u_bbr.flex3 = M_TSTMP;
12682                         ts.tv_sec = ae->timestamp / 1000000000;
12683                         ts.tv_nsec = ae->timestamp % 1000000000;
12684                         ltv.tv_sec = ts.tv_sec;
12685                         ltv.tv_usec = ts.tv_nsec / 1000;
12686                         log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
12687                 } else if (ae->flags & TSTMP_LRO) {
12688                         /* Record the LRO the arrival timestamp */
12689                         log.u_bbr.flex3 = M_TSTMP_LRO;
12690                         ts.tv_sec = ae->timestamp / 1000000000;
12691                         ts.tv_nsec = ae->timestamp % 1000000000;
12692                         ltv.tv_sec = ts.tv_sec;
12693                         ltv.tv_usec = ts.tv_nsec / 1000;
12694                         log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
12695                 }
12696                 log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
12697                 /* Log the rcv time */
12698                 log.u_bbr.delRate = ae->timestamp;
12699 #ifdef NETFLIX_HTTP_LOGGING
12700                 log.u_bbr.applimited = tp->t_http_closed;
12701                 log.u_bbr.applimited <<= 8;
12702                 log.u_bbr.applimited |= tp->t_http_open;
12703                 log.u_bbr.applimited <<= 8;
12704                 log.u_bbr.applimited |= tp->t_http_req;
12705                 if (http_req) {
12706                         /* Copy out any client req info */
12707                         /* seconds */
12708                         log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC);
12709                         /* useconds */
12710                         log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC);
12711                         log.u_bbr.rttProp = http_req->timestamp;
12712                         log.u_bbr.cur_del_rate = http_req->start;
12713                         if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) {
12714                                 log.u_bbr.flex8 |= 1;
12715                         } else {
12716                                 log.u_bbr.flex8 |= 2;
12717                                 log.u_bbr.bw_inuse = http_req->end;
12718                         }
12719                         log.u_bbr.flex6 = http_req->start_seq;
12720                         if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) {
12721                                 log.u_bbr.flex8 |= 4;
12722                                 log.u_bbr.epoch = http_req->end_seq;
12723                         }
12724                 }
12725 #endif
12726                 memset(tcp_hdr_buf, 0, sizeof(tcp_hdr_buf));
12727                 th = (struct tcphdr *)tcp_hdr_buf;
12728                 th->th_seq = ae->seq;
12729                 th->th_ack = ae->ack;
12730                 th->th_win = ae->win;
12731                 /* Now fill in the ports */
12732                 th->th_sport = tp->t_inpcb->inp_fport;
12733                 th->th_dport = tp->t_inpcb->inp_lport;
12734                 th->th_flags = ae->flags & 0xff;
12735                 /* Now do we have a timestamp option? */
12736                 if (ae->flags & HAS_TSTMP) {
12737                         u_char *cp;
12738                         uint32_t val;
12739
12740                         th->th_off = ((sizeof(struct tcphdr) + TCPOLEN_TSTAMP_APPA) >> 2);
12741                         cp = (u_char *)(th + 1);
12742                         *cp = TCPOPT_NOP;
12743                         cp++;
12744                         *cp = TCPOPT_NOP;
12745                         cp++;
12746                         *cp = TCPOPT_TIMESTAMP;
12747                         cp++;
12748                         *cp = TCPOLEN_TIMESTAMP;
12749                         cp++;
12750                         val = htonl(ae->ts_value);
12751                         bcopy((char *)&val,
12752                               (char *)cp, sizeof(uint32_t));
12753                         val = htonl(ae->ts_echo);
12754                         bcopy((char *)&val,
12755                               (char *)(cp + 4), sizeof(uint32_t));
12756                 } else
12757                         th->th_off = (sizeof(struct tcphdr) >> 2);
12758
12759                 /*
12760                  * For sane logging we need to play a little trick.
12761                  * If the ack were fully processed we would have moved
12762                  * snd_una to high_seq, but since compressed acks are
12763                  * processed in two phases, at this point (logging) snd_una
12764                  * won't be advanced. So we would see multiple acks showing
12765                  * the advancement. We can prevent that by "pretending" that
12766                  * snd_una was advanced and then un-advancing it so that the
12767                  * logging code has the right value for tlb_snd_una.
12768                  */
12769                 if (tp->snd_una != high_seq) {
12770                         orig_snd_una = tp->snd_una;
12771                         tp->snd_una = high_seq;
12772                         xx = 1;
12773                 } else
12774                         xx = 0;
12775                 TCP_LOG_EVENTP(tp, th,
12776                                &tp->t_inpcb->inp_socket->so_rcv,
12777                                &tp->t_inpcb->inp_socket->so_snd, TCP_LOG_IN, 0,
12778                                0, &log, true, &ltv);
12779                 if (xx) {
12780                         tp->snd_una = orig_snd_una;
12781                 }
12782         }
12783
12784 }
12785
12786 static int
12787 rack_do_compressed_ack_processing(struct tcpcb *tp, struct socket *so, struct mbuf *m, int nxt_pkt, struct timeval *tv)
12788 {
12789         /*
12790          * Handle a "special" compressed ack mbuf. Each incoming
12791          * ack has only four possible dispositions:
12792          *
12793          * A) It moves the cum-ack forward
12794          * B) It is behind the cum-ack.
12795          * C) It is a window-update ack.
12796          * D) It is a dup-ack.
12797          *
12798          * Note that we can have between 1 -> TCP_COMP_ACK_ENTRIES
12799          * in the incoming mbuf. We also need to still pay attention
12800          * to nxt_pkt since there may be another packet after this
12801          * one.
12802          */
12803 #ifdef TCP_ACCOUNTING
12804         uint64_t ts_val;
12805         uint64_t rdstc;
12806 #endif
12807         int segsiz;
12808         struct timespec ts;
12809         struct tcp_rack *rack;
12810         struct tcp_ackent *ae;
12811         uint32_t tiwin, us_cts, cts, acked, acked_amount, high_seq, win_seq, the_win, win_upd_ack;
12812         int cnt, i, did_out, ourfinisacked = 0;
12813         int win_up_req = 0;
12814         struct tcpopt to_holder, *to = NULL;
12815         int nsegs = 0;
12816         int under_pacing = 1;
12817         int recovery = 0;
12818         int idx;
12819 #ifdef TCP_ACCOUNTING
12820         sched_pin();
12821 #endif
12822         rack = (struct tcp_rack *)tp->t_fb_ptr;
12823         if (rack->gp_ready &&
12824             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT))
12825                 under_pacing = 0;
12826         else
12827                 under_pacing = 1;
12828
12829         if (rack->r_state != tp->t_state)
12830                 rack_set_state(tp, rack);
12831         to = &to_holder;
12832         to->to_flags = 0;
12833         KASSERT((m->m_len >= sizeof(struct tcp_ackent)),
12834                 ("tp:%p m_cmpack:%p with invalid len:%u", tp, m, m->m_len));
12835         cnt = m->m_len / sizeof(struct tcp_ackent);
12836         idx = cnt / 5;
12837         if (idx >= MAX_NUM_OF_CNTS)
12838                 idx = MAX_NUM_OF_CNTS - 1;
12839         counter_u64_add(rack_proc_comp_ack[idx], 1);
12840         counter_u64_add(rack_multi_single_eq, cnt);
12841         high_seq = tp->snd_una;
12842         the_win = tp->snd_wnd;
12843         win_seq = tp->snd_wl1;
12844         win_upd_ack = tp->snd_wl2;
12845         cts = us_cts = tcp_tv_to_usectick(tv);
12846         segsiz = ctf_fixed_maxseg(tp);
12847         if ((rack->rc_gp_dyn_mul) &&
12848             (rack->use_fixed_rate == 0) &&
12849             (rack->rc_always_pace)) {
12850                 /* Check in on probertt */
12851                 rack_check_probe_rtt(rack, us_cts);
12852         }
12853         for (i = 0; i < cnt; i++) {
12854 #ifdef TCP_ACCOUNTING
12855                 ts_val = get_cyclecount();
12856 #endif
12857                 rack_clear_rate_sample(rack);
12858                 ae = ((mtod(m, struct tcp_ackent *)) + i);
12859                 /* Setup the window */
12860                 tiwin = ae->win << tp->snd_scale;
12861                 /* figure out the type of ack */
12862                 if (SEQ_LT(ae->ack, high_seq)) {
12863                         /* Case B*/
12864                         ae->ack_val_set = ACK_BEHIND;
12865                 } else if (SEQ_GT(ae->ack, high_seq)) {
12866                         /* Case A */
12867                         ae->ack_val_set = ACK_CUMACK;
12868                 } else if (tiwin == the_win) {
12869                         /* Case D */
12870                         ae->ack_val_set = ACK_DUPACK;
12871                 } else {
12872                         /* Case C */
12873                         ae->ack_val_set = ACK_RWND;
12874                 }
12875                 rack_log_input_packet(tp, rack, ae, ae->ack_val_set, high_seq);
12876                 /* Validate timestamp */
12877                 if (ae->flags & HAS_TSTMP) {
12878                         /* Setup for a timestamp */
12879                         to->to_flags = TOF_TS;
12880                         ae->ts_echo -= tp->ts_offset;
12881                         to->to_tsecr = ae->ts_echo;
12882                         to->to_tsval = ae->ts_value;
12883                         /*
12884                          * If echoed timestamp is later than the current time, fall back to
12885                          * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
12886                          * were used when this connection was established.
12887                          */
12888                         if (TSTMP_GT(ae->ts_echo, cts))
12889                                 ae->ts_echo = 0;
12890                         if (tp->ts_recent &&
12891                             TSTMP_LT(ae->ts_value, tp->ts_recent)) {
12892                                 if (ctf_ts_check_ac(tp, (ae->flags & 0xff))) {
12893 #ifdef TCP_ACCOUNTING
12894                                         rdstc = get_cyclecount();
12895                                         if (rdstc > ts_val) {
12896                                                 counter_u64_add(tcp_proc_time[ae->ack_val_set] ,
12897                                                                 (rdstc - ts_val));
12898                                                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
12899                                                         tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
12900                                                 }
12901                                         }
12902 #endif
12903                                         continue;
12904                                 }
12905                         }
12906                         if (SEQ_LEQ(ae->seq, tp->last_ack_sent) &&
12907                             SEQ_LEQ(tp->last_ack_sent, ae->seq)) {
12908                                 tp->ts_recent_age = tcp_ts_getticks();
12909                                 tp->ts_recent = ae->ts_value;
12910                         }
12911                 } else {
12912                         /* Setup for a no options */
12913                         to->to_flags = 0;
12914                 }
12915                 /* Update the rcv time and perform idle reduction possibly */
12916                 if  (tp->t_idle_reduce &&
12917                      (tp->snd_max == tp->snd_una) &&
12918                      ((ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
12919                         counter_u64_add(rack_input_idle_reduces, 1);
12920                         rack_cc_after_idle(rack, tp);
12921                 }
12922                 tp->t_rcvtime = ticks;
12923                 /* Now what about ECN? */
12924                 if (tp->t_flags2 & TF2_ECN_PERMIT) {
12925                         if (ae->flags & TH_CWR) {
12926                                 tp->t_flags2 &= ~TF2_ECN_SND_ECE;
12927                                 tp->t_flags |= TF_ACKNOW;
12928                         }
12929                         switch (ae->codepoint & IPTOS_ECN_MASK) {
12930                         case IPTOS_ECN_CE:
12931                                 tp->t_flags2 |= TF2_ECN_SND_ECE;
12932                                 KMOD_TCPSTAT_INC(tcps_ecn_ce);
12933                                 break;
12934                         case IPTOS_ECN_ECT0:
12935                                 KMOD_TCPSTAT_INC(tcps_ecn_ect0);
12936                                 break;
12937                         case IPTOS_ECN_ECT1:
12938                                 KMOD_TCPSTAT_INC(tcps_ecn_ect1);
12939                                 break;
12940                         }
12941
12942                         /* Process a packet differently from RFC3168. */
12943                         cc_ecnpkt_handler_flags(tp, ae->flags, ae->codepoint);
12944                         /* Congestion experienced. */
12945                         if (ae->flags & TH_ECE) {
12946                                 rack_cong_signal(tp,  CC_ECN, ae->ack);
12947                         }
12948                 }
12949 #ifdef TCP_ACCOUNTING
12950                 /* Count for the specific type of ack in */
12951                 counter_u64_add(tcp_cnt_counters[ae->ack_val_set], 1);
12952                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
12953                         tp->tcp_cnt_counters[ae->ack_val_set]++;
12954                 }
12955 #endif
12956                 /*
12957                  * Note how we could move up these in the determination
12958                  * above, but we don't so that way the timestamp checks (and ECN)
12959                  * is done first before we do any processing on the ACK.
12960                  * The non-compressed path through the code has this
12961                  * weakness (noted by @jtl) that it actually does some
12962                  * processing before verifying the timestamp information.
12963                  * We don't take that path here which is why we set
12964                  * the ack_val_set first, do the timestamp and ecn
12965                  * processing, and then look at what we have setup.
12966                  */
12967                 if (ae->ack_val_set == ACK_BEHIND) {
12968                         /*
12969                          * Case B flag reordering, if window is not closed
12970                          * or it could be a keep-alive or persists
12971                          */
12972                         if (SEQ_LT(ae->ack, tp->snd_una) && (sbspace(&so->so_rcv) > segsiz)) {
12973                                 counter_u64_add(rack_reorder_seen, 1);
12974                                 rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
12975                         }
12976                 } else if (ae->ack_val_set == ACK_DUPACK) {
12977                         /* Case D */
12978
12979                         rack_strike_dupack(rack);
12980                 } else if (ae->ack_val_set == ACK_RWND) {
12981                         /* Case C */
12982
12983                         win_up_req = 1;
12984                         win_upd_ack = ae->ack;
12985                         win_seq = ae->seq;
12986                         the_win = tiwin;
12987                 } else {
12988                         /* Case A */
12989
12990                         if (SEQ_GT(ae->ack, tp->snd_max)) {
12991                                 /*
12992                                  * We just send an ack since the incoming
12993                                  * ack is beyond the largest seq we sent.
12994                                  */
12995                                 if ((tp->t_flags & TF_ACKNOW) == 0) {
12996                                         ctf_ack_war_checks(tp, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt);
12997                                         if (tp->t_flags && TF_ACKNOW)
12998                                                 rack->r_wanted_output = 1;
12999                                 }
13000                         } else {
13001                                 nsegs++;
13002                                 /* If the window changed setup to update */
13003                                 if (tiwin != tp->snd_wnd) {
13004                                         win_up_req = 1;
13005                                         win_upd_ack = ae->ack;
13006                                         win_seq = ae->seq;
13007                                         the_win = tiwin;
13008                                 }
13009 #ifdef TCP_ACCOUNTING
13010                                 /* Account for the acks */
13011                                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13012                                         tp->tcp_cnt_counters[CNT_OF_ACKS_IN] += (((ae->ack - high_seq) + segsiz - 1) / segsiz);
13013                                 }
13014                                 counter_u64_add(tcp_cnt_counters[CNT_OF_ACKS_IN],
13015                                                 (((ae->ack - high_seq) + segsiz - 1) / segsiz));
13016 #endif
13017                                 high_seq = ae->ack;
13018                                 /* Setup our act_rcv_time */
13019                                 if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
13020                                         ts.tv_sec = ae->timestamp / 1000000000;
13021                                         ts.tv_nsec = ae->timestamp % 1000000000;
13022                                         rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
13023                                         rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
13024                                 } else {
13025                                         rack->r_ctl.act_rcv_time = *tv;
13026                                 }
13027                                 rack_process_to_cumack(tp, rack, ae->ack, cts, to);
13028                         }
13029                 }
13030                 /* And lets be sure to commit the rtt measurements for this ack */
13031                 tcp_rack_xmit_timer_commit(rack, tp);
13032 #ifdef TCP_ACCOUNTING
13033                 rdstc = get_cyclecount();
13034                 if (rdstc > ts_val) {
13035                         counter_u64_add(tcp_proc_time[ae->ack_val_set] , (rdstc - ts_val));
13036                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13037                                 tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
13038                                 if (ae->ack_val_set == ACK_CUMACK)
13039                                         tp->tcp_proc_time[CYC_HANDLE_MAP] += (rdstc - ts_val);
13040                         }
13041                 }
13042 #endif
13043         }
13044 #ifdef TCP_ACCOUNTING
13045         ts_val = get_cyclecount();
13046 #endif
13047         acked_amount = acked = (high_seq - tp->snd_una);
13048         if (win_up_req) {
13049                 rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts, high_seq);
13050         }
13051         if (acked) {
13052                 if (rack->sack_attack_disable == 0)
13053                         rack_do_decay(rack);
13054                 if (acked >= segsiz) {
13055                         /*
13056                          * You only get credit for
13057                          * MSS and greater (and you get extra
13058                          * credit for larger cum-ack moves).
13059                          */
13060                         int ac;
13061
13062                         ac = acked / segsiz;
13063                         rack->r_ctl.ack_count += ac;
13064                         counter_u64_add(rack_ack_total, ac);
13065                 }
13066                 if (rack->r_ctl.ack_count > 0xfff00000) {
13067                         /*
13068                          * reduce the number to keep us under
13069                          * a uint32_t.
13070                          */
13071                         rack->r_ctl.ack_count /= 2;
13072                         rack->r_ctl.sack_count /= 2;
13073                 }
13074                 if (tp->t_flags & TF_NEEDSYN) {
13075                         /*
13076                          * T/TCP: Connection was half-synchronized, and our SYN has
13077                          * been ACK'd (so connection is now fully synchronized).  Go
13078                          * to non-starred state, increment snd_una for ACK of SYN,
13079                          * and check if we can do window scaling.
13080                          */
13081                         tp->t_flags &= ~TF_NEEDSYN;
13082                         tp->snd_una++;
13083                         acked_amount = acked = (high_seq - tp->snd_una);
13084                 }
13085                 if (acked > sbavail(&so->so_snd))
13086                         acked_amount = sbavail(&so->so_snd);
13087 #ifdef NETFLIX_EXP_DETECTION
13088                 /*
13089                  * We only care on a cum-ack move if we are in a sack-disabled
13090                  * state. We have already added in to the ack_count, and we never
13091                  * would disable on a cum-ack move, so we only care to do the
13092                  * detection if it may "undo" it, i.e. we were in disabled already.
13093                  */
13094                 if (rack->sack_attack_disable)
13095                         rack_do_detection(tp, rack, acked_amount, segsiz);
13096 #endif
13097                 if (IN_FASTRECOVERY(tp->t_flags) &&
13098                     (rack->rack_no_prr == 0))
13099                         rack_update_prr(tp, rack, acked_amount, high_seq);
13100                 if (IN_RECOVERY(tp->t_flags)) {
13101                         if (SEQ_LT(high_seq, tp->snd_recover) &&
13102                             (SEQ_LT(high_seq, tp->snd_max))) {
13103                                 tcp_rack_partialack(tp);
13104                         } else {
13105                                 rack_post_recovery(tp, high_seq);
13106                                 recovery = 1;
13107                         }
13108                 }
13109                 /* Handle the rack-log-ack part (sendmap) */
13110                 if ((sbused(&so->so_snd) == 0) &&
13111                     (acked > acked_amount) &&
13112                     (tp->t_state >= TCPS_FIN_WAIT_1) &&
13113                     (tp->t_flags & TF_SENTFIN)) {
13114                         /*
13115                          * We must be sure our fin
13116                          * was sent and acked (we can be
13117                          * in FIN_WAIT_1 without having
13118                          * sent the fin).
13119                          */
13120                         ourfinisacked = 1;
13121                         /*
13122                          * Lets make sure snd_una is updated
13123                          * since most likely acked_amount = 0 (it
13124                          * should be).
13125                          */
13126                         tp->snd_una = high_seq;
13127                 }
13128                 /* Did we make a RTO error? */
13129                 if ((tp->t_flags & TF_PREVVALID) &&
13130                     ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
13131                         tp->t_flags &= ~TF_PREVVALID;
13132                         if (tp->t_rxtshift == 1 &&
13133                             (int)(ticks - tp->t_badrxtwin) < 0)
13134                                 rack_cong_signal(tp, CC_RTO_ERR, high_seq);
13135                 }
13136                 /* Handle the data in the socket buffer */
13137                 KMOD_TCPSTAT_ADD(tcps_rcvackpack, 1);
13138                 KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
13139                 if (acked_amount > 0) {
13140                         struct mbuf *mfree;
13141
13142                         rack_ack_received(tp, rack, high_seq, nsegs, CC_ACK, recovery);
13143                         SOCKBUF_LOCK(&so->so_snd);
13144                         mfree = sbcut_locked(&so->so_snd, acked);
13145                         tp->snd_una = high_seq;
13146                         /* Note we want to hold the sb lock through the sendmap adjust */
13147                         rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
13148                         /* Wake up the socket if we have room to write more */
13149                         rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
13150                         SOCKBUF_UNLOCK(&so->so_snd);
13151                         tp->t_flags |= TF_WAKESOW;
13152                         m_freem(mfree);
13153                 }
13154                 /* update progress */
13155                 tp->t_acktime = ticks;
13156                 rack_log_progress_event(rack, tp, tp->t_acktime,
13157                                         PROGRESS_UPDATE, __LINE__);
13158                 /* Clear out shifts and such */
13159                 tp->t_rxtshift = 0;
13160                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
13161                                    rack_rto_min, rack_rto_max);
13162                 rack->rc_tlp_in_progress = 0;
13163                 rack->r_ctl.rc_tlp_cnt_out = 0;
13164                 /* Send recover and snd_nxt must be dragged along */
13165                 if (SEQ_GT(tp->snd_una, tp->snd_recover))
13166                         tp->snd_recover = tp->snd_una;
13167                 if (SEQ_LT(tp->snd_nxt, tp->snd_una))
13168                         tp->snd_nxt = tp->snd_una;
13169                 /*
13170                  * If the RXT timer is running we want to
13171                  * stop it, so we can restart a TLP (or new RXT).
13172                  */
13173                 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
13174                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13175 #ifdef NETFLIX_HTTP_LOGGING
13176                 tcp_http_check_for_comp(rack->rc_tp, high_seq);
13177 #endif
13178                 tp->snd_wl2 = high_seq;
13179                 tp->t_dupacks = 0;
13180                 if (under_pacing &&
13181                     (rack->use_fixed_rate == 0) &&
13182                     (rack->in_probe_rtt == 0) &&
13183                     rack->rc_gp_dyn_mul &&
13184                     rack->rc_always_pace) {
13185                         /* Check if we are dragging bottom */
13186                         rack_check_bottom_drag(tp, rack, so, acked);
13187                 }
13188                 if (tp->snd_una == tp->snd_max) {
13189                         tp->t_flags &= ~TF_PREVVALID;
13190                         rack->r_ctl.retran_during_recovery = 0;
13191                         rack->r_ctl.dsack_byte_cnt = 0;
13192                         rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
13193                         if (rack->r_ctl.rc_went_idle_time == 0)
13194                                 rack->r_ctl.rc_went_idle_time = 1;
13195                         rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
13196                         if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
13197                                 tp->t_acktime = 0;
13198                         /* Set so we might enter persists... */
13199                         rack->r_wanted_output = 1;
13200                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13201                         sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
13202                         if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
13203                             (sbavail(&so->so_snd) == 0) &&
13204                             (tp->t_flags2 & TF2_DROP_AF_DATA)) {
13205                                 /*
13206                                  * The socket was gone and the
13207                                  * peer sent data (not now in the past), time to
13208                                  * reset him.
13209                                  */
13210                                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13211                                 /* tcp_close will kill the inp pre-log the Reset */
13212                                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
13213 #ifdef TCP_ACCOUNTING
13214                                 rdstc = get_cyclecount();
13215                                 if (rdstc > ts_val) {
13216                                         counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13217                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13218                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13219                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13220                                         }
13221                                 }
13222 #endif
13223                                 m_freem(m);
13224                                 tp = tcp_close(tp);
13225                                 if (tp == NULL) {
13226 #ifdef TCP_ACCOUNTING
13227                                         sched_unpin();
13228 #endif
13229                                         return (1);
13230                                 }
13231                                 /*
13232                                  * We would normally do drop-with-reset which would
13233                                  * send back a reset. We can't since we don't have
13234                                  * all the needed bits. Instead lets arrange for
13235                                  * a call to tcp_output(). That way since we
13236                                  * are in the closed state we will generate a reset.
13237                                  *
13238                                  * Note if tcp_accounting is on we don't unpin since
13239                                  * we do that after the goto label.
13240                                  */
13241                                 goto send_out_a_rst;
13242                         }
13243                         if ((sbused(&so->so_snd) == 0) &&
13244                             (tp->t_state >= TCPS_FIN_WAIT_1) &&
13245                             (tp->t_flags & TF_SENTFIN)) {
13246                                 /*
13247                                  * If we can't receive any more data, then closing user can
13248                                  * proceed. Starting the timer is contrary to the
13249                                  * specification, but if we don't get a FIN we'll hang
13250                                  * forever.
13251                                  *
13252                                  */
13253                                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13254                                         soisdisconnected(so);
13255                                         tcp_timer_activate(tp, TT_2MSL,
13256                                                            (tcp_fast_finwait2_recycle ?
13257                                                             tcp_finwait2_timeout :
13258                                                             TP_MAXIDLE(tp)));
13259                                 }
13260                                 if (ourfinisacked == 0) {
13261                                         /*
13262                                          * We don't change to fin-wait-2 if we have our fin acked
13263                                          * which means we are probably in TCPS_CLOSING.
13264                                          */
13265                                         tcp_state_change(tp, TCPS_FIN_WAIT_2);
13266                                 }
13267                         }
13268                 }
13269                 /* Wake up the socket if we have room to write more */
13270                 if (sbavail(&so->so_snd)) {
13271                         rack->r_wanted_output = 1;
13272                         if (ctf_progress_timeout_check(tp, true)) {
13273                                 rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13274                                                         tp, tick, PROGRESS_DROP, __LINE__);
13275                                 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
13276                                 /*
13277                                  * We cheat here and don't send a RST, we should send one
13278                                  * when the pacer drops the connection.
13279                                  */
13280 #ifdef TCP_ACCOUNTING
13281                                 rdstc = get_cyclecount();
13282                                 if (rdstc > ts_val) {
13283                                         counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13284                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13285                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13286                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13287                                         }
13288                                 }
13289                                 sched_unpin();
13290 #endif
13291                                 INP_WUNLOCK(rack->rc_inp);
13292                                 m_freem(m);
13293                                 return (1);
13294                         }
13295                 }
13296                 if (ourfinisacked) {
13297                         switch(tp->t_state) {
13298                         case TCPS_CLOSING:
13299 #ifdef TCP_ACCOUNTING
13300                                 rdstc = get_cyclecount();
13301                                 if (rdstc > ts_val) {
13302                                         counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13303                                                         (rdstc - ts_val));
13304                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13305                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13306                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13307                                         }
13308                                 }
13309                                 sched_unpin();
13310 #endif
13311                                 tcp_twstart(tp);
13312                                 m_freem(m);
13313                                 return (1);
13314                                 break;
13315                         case TCPS_LAST_ACK:
13316 #ifdef TCP_ACCOUNTING
13317                                 rdstc = get_cyclecount();
13318                                 if (rdstc > ts_val) {
13319                                         counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13320                                                         (rdstc - ts_val));
13321                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13322                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13323                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13324                                         }
13325                                 }
13326                                 sched_unpin();
13327 #endif
13328                                 tp = tcp_close(tp);
13329                                 ctf_do_drop(m, tp);
13330                                 return (1);
13331                                 break;
13332                         case TCPS_FIN_WAIT_1:
13333 #ifdef TCP_ACCOUNTING
13334                                 rdstc = get_cyclecount();
13335                                 if (rdstc > ts_val) {
13336                                         counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13337                                                         (rdstc - ts_val));
13338                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13339                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13340                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13341                                         }
13342                                 }
13343 #endif
13344                                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13345                                         soisdisconnected(so);
13346                                         tcp_timer_activate(tp, TT_2MSL,
13347                                                            (tcp_fast_finwait2_recycle ?
13348                                                             tcp_finwait2_timeout :
13349                                                             TP_MAXIDLE(tp)));
13350                                 }
13351                                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
13352                                 break;
13353                         default:
13354                                 break;
13355                         }
13356                 }
13357                 if (rack->r_fast_output) {
13358                         /*
13359                          * We re doing fast output.. can we expand that?
13360                          */
13361                         rack_gain_for_fastoutput(rack, tp, so, acked_amount);
13362                 }
13363 #ifdef TCP_ACCOUNTING
13364                 rdstc = get_cyclecount();
13365                 if (rdstc > ts_val) {
13366                         counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13367                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13368                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13369                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13370                         }
13371                 }
13372
13373         } else if (win_up_req) {
13374                 rdstc = get_cyclecount();
13375                 if (rdstc > ts_val) {
13376                         counter_u64_add(tcp_proc_time[ACK_RWND] , (rdstc - ts_val));
13377                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13378                                 tp->tcp_proc_time[ACK_RWND] += (rdstc - ts_val);
13379                         }
13380                 }
13381 #endif
13382         }
13383         /* Now is there a next packet, if so we are done */
13384         m_freem(m);
13385         did_out = 0;
13386         if (nxt_pkt) {
13387 #ifdef TCP_ACCOUNTING
13388                 sched_unpin();
13389 #endif
13390                 rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 5, nsegs);
13391                 return (0);
13392         }
13393         rack_handle_might_revert(tp, rack);
13394         ctf_calc_rwin(so, tp);
13395         if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
13396         send_out_a_rst:
13397                 (void)tp->t_fb->tfb_tcp_output(tp);
13398                 did_out = 1;
13399         }
13400         rack_free_trim(rack);
13401 #ifdef TCP_ACCOUNTING
13402         sched_unpin();
13403 #endif
13404         rack_timer_audit(tp, rack, &so->so_snd);
13405         rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 6, nsegs);
13406         return (0);
13407 }
13408
13409
13410 static int
13411 rack_do_segment_nounlock(struct mbuf *m, struct tcphdr *th, struct socket *so,
13412     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos,
13413     int32_t nxt_pkt, struct timeval *tv)
13414 {
13415 #ifdef TCP_ACCOUNTING
13416         uint64_t ts_val;
13417 #endif
13418         int32_t thflags, retval, did_out = 0;
13419         int32_t way_out = 0;
13420         uint32_t cts;
13421         uint32_t tiwin;
13422         struct timespec ts;
13423         struct tcpopt to;
13424         struct tcp_rack *rack;
13425         struct rack_sendmap *rsm;
13426         int32_t prev_state = 0;
13427 #ifdef TCP_ACCOUNTING
13428         int ack_val_set = 0xf;
13429 #endif
13430         uint32_t us_cts;
13431         /*
13432          * tv passed from common code is from either M_TSTMP_LRO or
13433          * tcp_get_usecs() if no LRO m_pkthdr timestamp is present.
13434          */
13435         if (m->m_flags & M_ACKCMP) {
13436                 return (rack_do_compressed_ack_processing(tp, so, m, nxt_pkt, tv));
13437         }
13438         if (m->m_flags & M_ACKCMP) {
13439                 panic("Impossible reach m has ackcmp? m:%p tp:%p", m, tp);
13440         }
13441         counter_u64_add(rack_proc_non_comp_ack, 1);
13442         thflags = th->th_flags;
13443 #ifdef TCP_ACCOUNTING
13444         sched_pin();
13445         if (thflags & TH_ACK)
13446                 ts_val = get_cyclecount();
13447 #endif
13448         cts = tcp_tv_to_usectick(tv);
13449         rack = (struct tcp_rack *)tp->t_fb_ptr;
13450
13451         if ((m->m_flags & M_TSTMP) ||
13452             (m->m_flags & M_TSTMP_LRO)) {
13453                 mbuf_tstmp2timespec(m, &ts);
13454                 rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
13455                 rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
13456         } else
13457                 rack->r_ctl.act_rcv_time = *tv;
13458         kern_prefetch(rack, &prev_state);
13459         prev_state = 0;
13460         /*
13461          * Unscale the window into a 32-bit value. For the SYN_SENT state
13462          * the scale is zero.
13463          */
13464         tiwin = th->th_win << tp->snd_scale;
13465         /*
13466          * Parse options on any incoming segment.
13467          */
13468         memset(&to, 0, sizeof(to));
13469         tcp_dooptions(&to, (u_char *)(th + 1),
13470             (th->th_off << 2) - sizeof(struct tcphdr),
13471             (thflags & TH_SYN) ? TO_SYN : 0);
13472 #ifdef TCP_ACCOUNTING
13473         if (thflags & TH_ACK) {
13474                 /*
13475                  * We have a tradeoff here. We can either do what we are
13476                  * doing i.e. pinning to this CPU and then doing the accounting
13477                  * <or> we could do a critical enter, setup the rdtsc and cpu
13478                  * as in below, and then validate we are on the same CPU on
13479                  * exit. I have choosen to not do the critical enter since
13480                  * that often will gain you a context switch, and instead lock
13481                  * us (line above this if) to the same CPU with sched_pin(). This
13482                  * means we may be context switched out for a higher priority
13483                  * interupt but we won't be moved to another CPU.
13484                  *
13485                  * If this occurs (which it won't very often since we most likely
13486                  * are running this code in interupt context and only a higher
13487                  * priority will bump us ... clock?) we will falsely add in
13488                  * to the time the interupt processing time plus the ack processing
13489                  * time. This is ok since its a rare event.
13490                  */
13491                 ack_val_set = tcp_do_ack_accounting(tp, th, &to, tiwin,
13492                                                     ctf_fixed_maxseg(tp));
13493         }
13494 #endif
13495         NET_EPOCH_ASSERT();
13496         INP_WLOCK_ASSERT(tp->t_inpcb);
13497         KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
13498             __func__));
13499         KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
13500             __func__));
13501         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
13502                 union tcp_log_stackspecific log;
13503                 struct timeval ltv;
13504 #ifdef NETFLIX_HTTP_LOGGING
13505                 struct http_sendfile_track *http_req;
13506
13507                 if (SEQ_GT(th->th_ack, tp->snd_una)) {
13508                         http_req = tcp_http_find_req_for_seq(tp, (th->th_ack-1));
13509                 } else {
13510                         http_req = tcp_http_find_req_for_seq(tp, th->th_ack);
13511                 }
13512 #endif
13513                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
13514                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
13515                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
13516                 if (rack->rack_no_prr == 0)
13517                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
13518                 else
13519                         log.u_bbr.flex1 = 0;
13520                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
13521                 log.u_bbr.use_lt_bw <<= 1;
13522                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
13523                 log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
13524                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
13525                 log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
13526                 log.u_bbr.flex3 = m->m_flags;
13527                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
13528                 log.u_bbr.lost = thflags;
13529                 log.u_bbr.pacing_gain = 0x1;
13530 #ifdef TCP_ACCOUNTING
13531                 log.u_bbr.cwnd_gain = ack_val_set;
13532 #endif
13533                 log.u_bbr.flex7 = 2;
13534                 if (m->m_flags & M_TSTMP) {
13535                         /* Record the hardware timestamp if present */
13536                         mbuf_tstmp2timespec(m, &ts);
13537                         ltv.tv_sec = ts.tv_sec;
13538                         ltv.tv_usec = ts.tv_nsec / 1000;
13539                         log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
13540                 } else if (m->m_flags & M_TSTMP_LRO) {
13541                         /* Record the LRO the arrival timestamp */
13542                         mbuf_tstmp2timespec(m, &ts);
13543                         ltv.tv_sec = ts.tv_sec;
13544                         ltv.tv_usec = ts.tv_nsec / 1000;
13545                         log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
13546                 }
13547                 log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
13548                 /* Log the rcv time */
13549                 log.u_bbr.delRate = m->m_pkthdr.rcv_tstmp;
13550 #ifdef NETFLIX_HTTP_LOGGING
13551                 log.u_bbr.applimited = tp->t_http_closed;
13552                 log.u_bbr.applimited <<= 8;
13553                 log.u_bbr.applimited |= tp->t_http_open;
13554                 log.u_bbr.applimited <<= 8;
13555                 log.u_bbr.applimited |= tp->t_http_req;
13556                 if (http_req) {
13557                         /* Copy out any client req info */
13558                         /* seconds */
13559                         log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC);
13560                         /* useconds */
13561                         log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC);
13562                         log.u_bbr.rttProp = http_req->timestamp;
13563                         log.u_bbr.cur_del_rate = http_req->start;
13564                         if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) {
13565                                 log.u_bbr.flex8 |= 1;
13566                         } else {
13567                                 log.u_bbr.flex8 |= 2;
13568                                 log.u_bbr.bw_inuse = http_req->end;
13569                         }
13570                         log.u_bbr.flex6 = http_req->start_seq;
13571                         if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) {
13572                                 log.u_bbr.flex8 |= 4;
13573                                 log.u_bbr.epoch = http_req->end_seq;
13574                         }
13575                 }
13576 #endif
13577                 TCP_LOG_EVENTP(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0,
13578                     tlen, &log, true, &ltv);
13579         }
13580         if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) {
13581                 way_out = 4;
13582                 retval = 0;
13583                 goto done_with_input;
13584         }
13585         /*
13586          * If a segment with the ACK-bit set arrives in the SYN-SENT state
13587          * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9.
13588          */
13589         if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) &&
13590             (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) {
13591                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13592                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13593 #ifdef TCP_ACCOUNTING
13594                 sched_unpin();
13595 #endif
13596                 return (1);
13597         }
13598
13599         /*
13600          * Parse options on any incoming segment.
13601          */
13602         tcp_dooptions(&to, (u_char *)(th + 1),
13603             (th->th_off << 2) - sizeof(struct tcphdr),
13604             (thflags & TH_SYN) ? TO_SYN : 0);
13605
13606         /*
13607          * If timestamps were negotiated during SYN/ACK and a
13608          * segment without a timestamp is received, silently drop
13609          * the segment, unless it is a RST segment or missing timestamps are
13610          * tolerated.
13611          * See section 3.2 of RFC 7323.
13612          */
13613         if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) &&
13614             ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) {
13615                 way_out = 5;
13616                 retval = 0;
13617                 goto done_with_input;
13618         }
13619
13620         /*
13621          * Segment received on connection. Reset idle time and keep-alive
13622          * timer. XXX: This should be done after segment validation to
13623          * ignore broken/spoofed segs.
13624          */
13625         if  (tp->t_idle_reduce &&
13626              (tp->snd_max == tp->snd_una) &&
13627              ((ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
13628                 counter_u64_add(rack_input_idle_reduces, 1);
13629                 rack_cc_after_idle(rack, tp);
13630         }
13631         tp->t_rcvtime = ticks;
13632 #ifdef STATS
13633         stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
13634 #endif
13635         if (tiwin > rack->r_ctl.rc_high_rwnd)
13636                 rack->r_ctl.rc_high_rwnd = tiwin;
13637         /*
13638          * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move
13639          * this to occur after we've validated the segment.
13640          */
13641         if (tp->t_flags2 & TF2_ECN_PERMIT) {
13642                 if (thflags & TH_CWR) {
13643                         tp->t_flags2 &= ~TF2_ECN_SND_ECE;
13644                         tp->t_flags |= TF_ACKNOW;
13645                 }
13646                 switch (iptos & IPTOS_ECN_MASK) {
13647                 case IPTOS_ECN_CE:
13648                         tp->t_flags2 |= TF2_ECN_SND_ECE;
13649                         KMOD_TCPSTAT_INC(tcps_ecn_ce);
13650                         break;
13651                 case IPTOS_ECN_ECT0:
13652                         KMOD_TCPSTAT_INC(tcps_ecn_ect0);
13653                         break;
13654                 case IPTOS_ECN_ECT1:
13655                         KMOD_TCPSTAT_INC(tcps_ecn_ect1);
13656                         break;
13657                 }
13658
13659                 /* Process a packet differently from RFC3168. */
13660                 cc_ecnpkt_handler(tp, th, iptos);
13661
13662                 /* Congestion experienced. */
13663                 if (thflags & TH_ECE) {
13664                         rack_cong_signal(tp, CC_ECN, th->th_ack);
13665                 }
13666         }
13667
13668         /*
13669          * If echoed timestamp is later than the current time, fall back to
13670          * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
13671          * were used when this connection was established.
13672          */
13673         if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
13674                 to.to_tsecr -= tp->ts_offset;
13675                 if (TSTMP_GT(to.to_tsecr, cts))
13676                         to.to_tsecr = 0;
13677         }
13678
13679         /*
13680          * If its the first time in we need to take care of options and
13681          * verify we can do SACK for rack!
13682          */
13683         if (rack->r_state == 0) {
13684                 /* Should be init'd by rack_init() */
13685                 KASSERT(rack->rc_inp != NULL,
13686                     ("%s: rack->rc_inp unexpectedly NULL", __func__));
13687                 if (rack->rc_inp == NULL) {
13688                         rack->rc_inp = tp->t_inpcb;
13689                 }
13690
13691                 /*
13692                  * Process options only when we get SYN/ACK back. The SYN
13693                  * case for incoming connections is handled in tcp_syncache.
13694                  * According to RFC1323 the window field in a SYN (i.e., a
13695                  * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
13696                  * this is traditional behavior, may need to be cleaned up.
13697                  */
13698                 if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
13699                         /* Handle parallel SYN for ECN */
13700                         if (!(thflags & TH_ACK) &&
13701                             ((thflags & (TH_CWR | TH_ECE)) == (TH_CWR | TH_ECE)) &&
13702                             ((V_tcp_do_ecn == 1) || (V_tcp_do_ecn == 2))) {
13703                                 tp->t_flags2 |= TF2_ECN_PERMIT;
13704                                 tp->t_flags2 |= TF2_ECN_SND_ECE;
13705                                 TCPSTAT_INC(tcps_ecn_shs);
13706                         }
13707                         if ((to.to_flags & TOF_SCALE) &&
13708                             (tp->t_flags & TF_REQ_SCALE)) {
13709                                 tp->t_flags |= TF_RCVD_SCALE;
13710                                 tp->snd_scale = to.to_wscale;
13711                         } else
13712                                 tp->t_flags &= ~TF_REQ_SCALE;
13713                         /*
13714                          * Initial send window.  It will be updated with the
13715                          * next incoming segment to the scaled value.
13716                          */
13717                         tp->snd_wnd = th->th_win;
13718                         rack_validate_fo_sendwin_up(tp, rack);
13719                         if ((to.to_flags & TOF_TS) &&
13720                             (tp->t_flags & TF_REQ_TSTMP)) {
13721                                 tp->t_flags |= TF_RCVD_TSTMP;
13722                                 tp->ts_recent = to.to_tsval;
13723                                 tp->ts_recent_age = cts;
13724                         } else
13725                                 tp->t_flags &= ~TF_REQ_TSTMP;
13726                         if (to.to_flags & TOF_MSS) {
13727                                 tcp_mss(tp, to.to_mss);
13728                         }
13729                         if ((tp->t_flags & TF_SACK_PERMIT) &&
13730                             (to.to_flags & TOF_SACKPERM) == 0)
13731                                 tp->t_flags &= ~TF_SACK_PERMIT;
13732                         if (IS_FASTOPEN(tp->t_flags)) {
13733                                 if (to.to_flags & TOF_FASTOPEN) {
13734                                         uint16_t mss;
13735
13736                                         if (to.to_flags & TOF_MSS)
13737                                                 mss = to.to_mss;
13738                                         else
13739                                                 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0)
13740                                                         mss = TCP6_MSS;
13741                                                 else
13742                                                         mss = TCP_MSS;
13743                                         tcp_fastopen_update_cache(tp, mss,
13744                                             to.to_tfo_len, to.to_tfo_cookie);
13745                                 } else
13746                                         tcp_fastopen_disable_path(tp);
13747                         }
13748                 }
13749                 /*
13750                  * At this point we are at the initial call. Here we decide
13751                  * if we are doing RACK or not. We do this by seeing if
13752                  * TF_SACK_PERMIT is set and the sack-not-required is clear.
13753                  * The code now does do dup-ack counting so if you don't
13754                  * switch back you won't get rack & TLP, but you will still
13755                  * get this stack.
13756                  */
13757
13758                 if ((rack_sack_not_required == 0) &&
13759                     ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
13760                         tcp_switch_back_to_default(tp);
13761                         (*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen,
13762                             tlen, iptos);
13763 #ifdef TCP_ACCOUNTING
13764                         sched_unpin();
13765 #endif
13766                         return (1);
13767                 }
13768                 tcp_set_hpts(tp->t_inpcb);
13769                 sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack);
13770         }
13771         if (thflags & TH_FIN)
13772                 tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN);
13773         us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
13774         if ((rack->rc_gp_dyn_mul) &&
13775             (rack->use_fixed_rate == 0) &&
13776             (rack->rc_always_pace)) {
13777                 /* Check in on probertt */
13778                 rack_check_probe_rtt(rack, us_cts);
13779         }
13780         if (rack->forced_ack) {
13781                 uint32_t us_rtt;
13782
13783                 /*
13784                  * A persist or keep-alive was forced out, update our
13785                  * min rtt time. Note we do not worry about lost
13786                  * retransmissions since KEEP-ALIVES and persists
13787                  * are usually way long on times of sending (though
13788                  * if we were really paranoid or worried we could
13789                  * at least use timestamps if available to validate).
13790                  */
13791                 rack->forced_ack = 0;
13792                 us_rtt = us_cts - rack->r_ctl.forced_ack_ts;
13793                 if (us_rtt == 0)
13794                         us_rtt = 1;
13795                 rack_log_rtt_upd(tp, rack, us_rtt, 0, NULL, 3);
13796                 rack_apply_updated_usrtt(rack, us_rtt, us_cts);
13797         }
13798         /*
13799          * This is the one exception case where we set the rack state
13800          * always. All other times (timers etc) we must have a rack-state
13801          * set (so we assure we have done the checks above for SACK).
13802          */
13803         rack->r_ctl.rc_rcvtime = cts;
13804         if (rack->r_state != tp->t_state)
13805                 rack_set_state(tp, rack);
13806         if (SEQ_GT(th->th_ack, tp->snd_una) &&
13807             (rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree)) != NULL)
13808                 kern_prefetch(rsm, &prev_state);
13809         prev_state = rack->r_state;
13810         rack_clear_rate_sample(rack);
13811         retval = (*rack->r_substate) (m, th, so,
13812             tp, &to, drop_hdrlen,
13813             tlen, tiwin, thflags, nxt_pkt, iptos);
13814 #ifdef INVARIANTS
13815         if ((retval == 0) &&
13816             (tp->t_inpcb == NULL)) {
13817                 panic("retval:%d tp:%p t_inpcb:NULL state:%d",
13818                     retval, tp, prev_state);
13819         }
13820 #endif
13821         if (retval == 0) {
13822                 /*
13823                  * If retval is 1 the tcb is unlocked and most likely the tp
13824                  * is gone.
13825                  */
13826                 INP_WLOCK_ASSERT(tp->t_inpcb);
13827                 if ((rack->rc_gp_dyn_mul) &&
13828                     (rack->rc_always_pace) &&
13829                     (rack->use_fixed_rate == 0) &&
13830                     rack->in_probe_rtt &&
13831                     (rack->r_ctl.rc_time_probertt_starts == 0)) {
13832                         /*
13833                          * If we are going for target, lets recheck before
13834                          * we output.
13835                          */
13836                         rack_check_probe_rtt(rack, us_cts);
13837                 }
13838                 if (rack->set_pacing_done_a_iw == 0) {
13839                         /* How much has been acked? */
13840                         if ((tp->snd_una - tp->iss) > (ctf_fixed_maxseg(tp) * 10)) {
13841                                 /* We have enough to set in the pacing segment size */
13842                                 rack->set_pacing_done_a_iw = 1;
13843                                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
13844                         }
13845                 }
13846                 tcp_rack_xmit_timer_commit(rack, tp);
13847 #ifdef TCP_ACCOUNTING
13848                 /*
13849                  * If we set the ack_val_se to what ack processing we are doing
13850                  * we also want to track how many cycles we burned. Note
13851                  * the bits after tcp_output we let be "free". This is because
13852                  * we are also tracking the tcp_output times as well. Note the
13853                  * use of 0xf here since we only have 11 counter (0 - 0xa) and
13854                  * 0xf cannot be returned and is what we initialize it too to
13855                  * indicate we are not doing the tabulations.
13856                  */
13857                 if (ack_val_set != 0xf) {
13858                         uint64_t crtsc;
13859
13860                         crtsc = get_cyclecount();
13861                         counter_u64_add(tcp_proc_time[ack_val_set] , (crtsc - ts_val));
13862                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13863                                 tp->tcp_proc_time[ack_val_set] += (crtsc - ts_val);
13864                         }
13865                 }
13866 #endif
13867                 if (nxt_pkt == 0) {
13868                         if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
13869 do_output_now:
13870                                 did_out = 1;
13871                                 (void)tp->t_fb->tfb_tcp_output(tp);
13872                         }
13873                         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
13874                         rack_free_trim(rack);
13875                 }
13876                 if ((nxt_pkt == 0) &&
13877                     ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
13878                     (SEQ_GT(tp->snd_max, tp->snd_una) ||
13879                      (tp->t_flags & TF_DELACK) ||
13880                      ((V_tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
13881                       (tp->t_state <= TCPS_CLOSING)))) {
13882                         /* We could not send (probably in the hpts but stopped the timer earlier)? */
13883                         if ((tp->snd_max == tp->snd_una) &&
13884                             ((tp->t_flags & TF_DELACK) == 0) &&
13885                             (rack->rc_inp->inp_in_hpts) &&
13886                             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
13887                                 /* keep alive not needed if we are hptsi output yet */
13888                                 ;
13889                         } else {
13890                                 int late = 0;
13891                                 if (rack->rc_inp->inp_in_hpts) {
13892                                         if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
13893                                                 us_cts = tcp_get_usecs(NULL);
13894                                                 if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
13895                                                         rack->r_early = 1;
13896                                                         rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
13897                                                 } else
13898                                                         late = 1;
13899                                                 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
13900                                         }
13901                                         tcp_hpts_remove(tp->t_inpcb, HPTS_REMOVE_OUTPUT);
13902                                 }
13903                                 if (late && (did_out == 0)) {
13904                                         /*
13905                                          * We are late in the sending
13906                                          * and we did not call the output
13907                                          * (this probably should not happen).
13908                                          */
13909                                         goto do_output_now;
13910                                 }
13911                                 rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
13912                         }
13913                         way_out = 1;
13914                 } else if (nxt_pkt == 0) {
13915                         /* Do we have the correct timer running? */
13916                         rack_timer_audit(tp, rack, &so->so_snd);
13917                         way_out = 2;
13918                 }
13919         done_with_input:
13920                 rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out, max(1, m->m_pkthdr.lro_nsegs));
13921                 if (did_out)
13922                         rack->r_wanted_output = 0;
13923 #ifdef INVARIANTS
13924                 if (tp->t_inpcb == NULL) {
13925                         panic("OP:%d retval:%d tp:%p t_inpcb:NULL state:%d",
13926                               did_out,
13927                               retval, tp, prev_state);
13928                 }
13929 #endif
13930 #ifdef TCP_ACCOUNTING
13931         } else {
13932                 /*
13933                  * Track the time (see above).
13934                  */
13935                 if (ack_val_set != 0xf) {
13936                         uint64_t crtsc;
13937
13938                         crtsc = get_cyclecount();
13939                         counter_u64_add(tcp_proc_time[ack_val_set] , (crtsc - ts_val));
13940                         /*
13941                          * Note we *DO NOT* increment the per-tcb counters since
13942                          * in the else the TP may be gone!!
13943                          */
13944                 }
13945 #endif
13946         }
13947 #ifdef TCP_ACCOUNTING
13948         sched_unpin();
13949 #endif
13950         return (retval);
13951 }
13952
13953 void
13954 rack_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
13955     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
13956 {
13957         struct timeval tv;
13958
13959         /* First lets see if we have old packets */
13960         if (tp->t_in_pkt) {
13961                 if (ctf_do_queued_segments(so, tp, 1)) {
13962                         m_freem(m);
13963                         return;
13964                 }
13965         }
13966         if (m->m_flags & M_TSTMP_LRO) {
13967                 tv.tv_sec = m->m_pkthdr.rcv_tstmp /1000000000;
13968                 tv.tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000)/1000;
13969         } else {
13970                 /* Should not be should we kassert instead? */
13971                 tcp_get_usecs(&tv);
13972         }
13973         if (rack_do_segment_nounlock(m, th, so, tp,
13974                                      drop_hdrlen, tlen, iptos, 0, &tv) == 0) {
13975                 tcp_handle_wakeup(tp, so);
13976                 INP_WUNLOCK(tp->t_inpcb);
13977         }
13978 }
13979
13980 struct rack_sendmap *
13981 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused)
13982 {
13983         struct rack_sendmap *rsm = NULL;
13984         int32_t idx;
13985         uint32_t srtt = 0, thresh = 0, ts_low = 0;
13986
13987         /* Return the next guy to be re-transmitted */
13988         if (RB_EMPTY(&rack->r_ctl.rc_mtree)) {
13989                 return (NULL);
13990         }
13991         if (tp->t_flags & TF_SENTFIN) {
13992                 /* retran the end FIN? */
13993                 return (NULL);
13994         }
13995         /* ok lets look at this one */
13996         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
13997         if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) {
13998                 goto check_it;
13999         }
14000         rsm = rack_find_lowest_rsm(rack);
14001         if (rsm == NULL) {
14002                 return (NULL);
14003         }
14004 check_it:
14005         if (((rack->rc_tp->t_flags & TF_SACK_PERMIT) == 0) &&
14006             (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
14007                 /*
14008                  * No sack so we automatically do the 3 strikes and
14009                  * retransmit (no rack timer would be started).
14010                  */
14011
14012                 return (rsm);
14013         }
14014         if (rsm->r_flags & RACK_ACKED) {
14015                 return (NULL);
14016         }
14017         if (((rsm->r_flags & RACK_SACK_PASSED) == 0) &&
14018             (rsm->r_dupack < DUP_ACK_THRESHOLD)) {
14019                 /* Its not yet ready */
14020                 return (NULL);
14021         }
14022         srtt = rack_grab_rtt(tp, rack);
14023         idx = rsm->r_rtr_cnt - 1;
14024         ts_low = (uint32_t)rsm->r_tim_lastsent[idx];
14025         thresh = rack_calc_thresh_rack(rack, srtt, tsused);
14026         if ((tsused == ts_low) ||
14027             (TSTMP_LT(tsused, ts_low))) {
14028                 /* No time since sending */
14029                 return (NULL);
14030         }
14031         if ((tsused - ts_low) < thresh) {
14032                 /* It has not been long enough yet */
14033                 return (NULL);
14034         }
14035         if ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
14036             ((rsm->r_flags & RACK_SACK_PASSED) &&
14037              (rack->sack_attack_disable == 0))) {
14038                 /*
14039                  * We have passed the dup-ack threshold <or>
14040                  * a SACK has indicated this is missing.
14041                  * Note that if you are a declared attacker
14042                  * it is only the dup-ack threshold that
14043                  * will cause retransmits.
14044                  */
14045                 /* log retransmit reason */
14046                 rack_log_retran_reason(rack, rsm, (tsused - ts_low), thresh, 1);
14047                 rack->r_fast_output = 0;
14048                 return (rsm);
14049         }
14050         return (NULL);
14051 }
14052
14053 static void
14054 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
14055                            uint64_t bw_est, uint64_t bw, uint64_t len_time, int method,
14056                            int line, struct rack_sendmap *rsm)
14057 {
14058         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
14059                 union tcp_log_stackspecific log;
14060                 struct timeval tv;
14061
14062                 memset(&log, 0, sizeof(log));
14063                 log.u_bbr.flex1 = slot;
14064                 log.u_bbr.flex2 = len;
14065                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_min_segs;
14066                 log.u_bbr.flex4 = rack->r_ctl.rc_pace_max_segs;
14067                 log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ss;
14068                 log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_ca;
14069                 log.u_bbr.use_lt_bw = rack->rc_ack_can_sendout_data;
14070                 log.u_bbr.use_lt_bw <<= 1;
14071                 log.u_bbr.use_lt_bw |= rack->r_late;
14072                 log.u_bbr.use_lt_bw <<= 1;
14073                 log.u_bbr.use_lt_bw |= rack->r_early;
14074                 log.u_bbr.use_lt_bw <<= 1;
14075                 log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
14076                 log.u_bbr.use_lt_bw <<= 1;
14077                 log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
14078                 log.u_bbr.use_lt_bw <<= 1;
14079                 log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
14080                 log.u_bbr.use_lt_bw <<= 1;
14081                 log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
14082                 log.u_bbr.use_lt_bw <<= 1;
14083                 log.u_bbr.use_lt_bw |= rack->gp_ready;
14084                 log.u_bbr.pkt_epoch = line;
14085                 log.u_bbr.epoch = rack->r_ctl.rc_agg_delayed;
14086                 log.u_bbr.lt_epoch = rack->r_ctl.rc_agg_early;
14087                 log.u_bbr.applimited = rack->r_ctl.rack_per_of_gp_rec;
14088                 log.u_bbr.bw_inuse = bw_est;
14089                 log.u_bbr.delRate = bw;
14090                 if (rack->r_ctl.gp_bw == 0)
14091                         log.u_bbr.cur_del_rate = 0;
14092                 else
14093                         log.u_bbr.cur_del_rate = rack_get_bw(rack);
14094                 log.u_bbr.rttProp = len_time;
14095                 log.u_bbr.pkts_out = rack->r_ctl.rc_rack_min_rtt;
14096                 log.u_bbr.lost = rack->r_ctl.rc_probertt_sndmax_atexit;
14097                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
14098                 if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) {
14099                         /* We are in slow start */
14100                         log.u_bbr.flex7 = 1;
14101                 } else {
14102                         /* we are on congestion avoidance */
14103                         log.u_bbr.flex7 = 0;
14104                 }
14105                 log.u_bbr.flex8 = method;
14106                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14107                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14108                 log.u_bbr.cwnd_gain = rack->rc_gp_saw_rec;
14109                 log.u_bbr.cwnd_gain <<= 1;
14110                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
14111                 log.u_bbr.cwnd_gain <<= 1;
14112                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
14113                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
14114                     &rack->rc_inp->inp_socket->so_rcv,
14115                     &rack->rc_inp->inp_socket->so_snd,
14116                     BBR_LOG_HPTSI_CALC, 0,
14117                     0, &log, false, &tv);
14118         }
14119 }
14120
14121 static uint32_t
14122 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss)
14123 {
14124         uint32_t new_tso, user_max;
14125
14126         user_max = rack->rc_user_set_max_segs * mss;
14127         if (rack->rc_force_max_seg) {
14128                 return (user_max);
14129         }
14130         if (rack->use_fixed_rate &&
14131             ((rack->r_ctl.crte == NULL) ||
14132              (bw != rack->r_ctl.crte->rate))) {
14133                 /* Use the user mss since we are not exactly matched */
14134                 return (user_max);
14135         }
14136         new_tso = tcp_get_pacing_burst_size(rack->rc_tp, bw, mss, rack_pace_one_seg, rack->r_ctl.crte, NULL);
14137         if (new_tso > user_max)
14138                 new_tso = user_max;
14139         return (new_tso);
14140 }
14141
14142 static int32_t
14143 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)
14144 {
14145         uint64_t lentim, fill_bw;
14146
14147         /* Lets first see if we are full, if so continue with normal rate */
14148         rack->r_via_fill_cw = 0;
14149         if (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.cwnd_to_use)
14150                 return (slot);
14151         if ((ctf_outstanding(rack->rc_tp) + (segsiz-1)) > rack->rc_tp->snd_wnd)
14152                 return (slot);
14153         if (rack->r_ctl.rc_last_us_rtt == 0)
14154                 return (slot);
14155         if (rack->rc_pace_fill_if_rttin_range &&
14156             (rack->r_ctl.rc_last_us_rtt >=
14157              (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack->rtt_limit_mul))) {
14158                 /* The rtt is huge, N * smallest, lets not fill */
14159                 return (slot);
14160         }
14161         /*
14162          * first lets calculate the b/w based on the last us-rtt
14163          * and the sndwnd.
14164          */
14165         fill_bw = rack->r_ctl.cwnd_to_use;
14166         /* Take the rwnd if its smaller */
14167         if (fill_bw > rack->rc_tp->snd_wnd)
14168                 fill_bw = rack->rc_tp->snd_wnd;
14169         if (rack->r_fill_less_agg) {
14170                 /*
14171                  * Now take away the inflight (this will reduce our
14172                  * aggressiveness and yeah, if we get that much out in 1RTT
14173                  * we will have had acks come back and still be behind).
14174                  */
14175                 fill_bw -= ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14176         }
14177         /* Now lets make it into a b/w */
14178         fill_bw *= (uint64_t)HPTS_USEC_IN_SEC;
14179         fill_bw /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
14180         /* We are below the min b/w */
14181         if (non_paced)
14182                 *rate_wanted = fill_bw;
14183         if ((fill_bw < RACK_MIN_BW) || (fill_bw < *rate_wanted))
14184                 return (slot);
14185         if (rack->r_ctl.bw_rate_cap && (fill_bw > rack->r_ctl.bw_rate_cap))
14186                 fill_bw = rack->r_ctl.bw_rate_cap;
14187         rack->r_via_fill_cw = 1;
14188         if (rack->r_rack_hw_rate_caps &&
14189             (rack->r_ctl.crte != NULL)) {
14190                 uint64_t high_rate;
14191
14192                 high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
14193                 if (fill_bw > high_rate) {
14194                         /* We are capping bw at the highest rate table entry */
14195                         if (*rate_wanted > high_rate) {
14196                                 /* The original rate was also capped */
14197                                 rack->r_via_fill_cw = 0;
14198                         }
14199                         rack_log_hdwr_pacing(rack,
14200                                              fill_bw, high_rate, __LINE__,
14201                                              0, 3);
14202                         fill_bw = high_rate;
14203                         if (capped)
14204                                 *capped = 1;
14205                 }
14206         } else if ((rack->r_ctl.crte == NULL) &&
14207                    (rack->rack_hdrw_pacing == 0) &&
14208                    (rack->rack_hdw_pace_ena) &&
14209                    rack->r_rack_hw_rate_caps &&
14210                    (rack->rack_attempt_hdwr_pace == 0) &&
14211                    (rack->rc_inp->inp_route.ro_nh != NULL) &&
14212                    (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
14213                 /*
14214                  * Ok we may have a first attempt that is greater than our top rate
14215                  * lets check.
14216                  */
14217                 uint64_t high_rate;
14218
14219                 high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
14220                 if (high_rate) {
14221                         if (fill_bw > high_rate) {
14222                                 fill_bw = high_rate;
14223                                 if (capped)
14224                                         *capped = 1;
14225                         }
14226                 }
14227         }
14228         /*
14229          * Ok fill_bw holds our mythical b/w to fill the cwnd
14230          * in a rtt, what does that time wise equate too?
14231          */
14232         lentim = (uint64_t)(len) * (uint64_t)HPTS_USEC_IN_SEC;
14233         lentim /= fill_bw;
14234         *rate_wanted = fill_bw;
14235         if (non_paced || (lentim < slot)) {
14236                 rack_log_pacing_delay_calc(rack, len, slot, fill_bw,
14237                                            0, lentim, 12, __LINE__, NULL);
14238                 return ((int32_t)lentim);
14239         } else
14240                 return (slot);
14241 }
14242
14243 static int32_t
14244 rack_get_pacing_delay(struct tcp_rack *rack, struct tcpcb *tp, uint32_t len, struct rack_sendmap *rsm, uint32_t segsiz)
14245 {
14246         struct rack_sendmap *lrsm;
14247         int32_t slot = 0;
14248         int can_start_hw_pacing = 1;
14249         int err;
14250
14251         if (rack->rc_always_pace == 0) {
14252                 /*
14253                  * We use the most optimistic possible cwnd/srtt for
14254                  * sending calculations. This will make our
14255                  * calculation anticipate getting more through
14256                  * quicker then possible. But thats ok we don't want
14257                  * the peer to have a gap in data sending.
14258                  */
14259                 uint32_t srtt, cwnd, tr_perms = 0;
14260                 int32_t reduce = 0;
14261
14262         old_method:
14263                 /*
14264                  * We keep no precise pacing with the old method
14265                  * instead we use the pacer to mitigate bursts.
14266                  */
14267                 if (rack->r_ctl.rc_rack_min_rtt)
14268                         srtt = rack->r_ctl.rc_rack_min_rtt;
14269                 else
14270                         srtt = max(tp->t_srtt, 1);
14271                 if (rack->r_ctl.rc_rack_largest_cwnd)
14272                         cwnd = rack->r_ctl.rc_rack_largest_cwnd;
14273                 else
14274                         cwnd = rack->r_ctl.cwnd_to_use;
14275                 /* Inflate cwnd by 1000 so srtt of usecs is in ms */
14276                 tr_perms = (cwnd * 1000) / srtt;
14277                 if (tr_perms == 0) {
14278                         tr_perms = ctf_fixed_maxseg(tp);
14279                 }
14280                 /*
14281                  * Calculate how long this will take to drain, if
14282                  * the calculation comes out to zero, thats ok we
14283                  * will use send_a_lot to possibly spin around for
14284                  * more increasing tot_len_this_send to the point
14285                  * that its going to require a pace, or we hit the
14286                  * cwnd. Which in that case we are just waiting for
14287                  * a ACK.
14288                  */
14289                 slot = len / tr_perms;
14290                 /* Now do we reduce the time so we don't run dry? */
14291                 if (slot && rack_slot_reduction) {
14292                         reduce = (slot / rack_slot_reduction);
14293                         if (reduce < slot) {
14294                                 slot -= reduce;
14295                         } else
14296                                 slot = 0;
14297                 }
14298                 slot *= HPTS_USEC_IN_MSEC;
14299                 if (rsm == NULL) {
14300                         /*
14301                          * We always consider ourselves app limited with old style
14302                          * that are not retransmits. This could be the initial
14303                          * measurement, but thats ok its all setup and specially
14304                          * handled. If another send leaks out, then that too will
14305                          * be mark app-limited.
14306                          */
14307                         lrsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
14308                         if (lrsm && ((lrsm->r_flags & RACK_APP_LIMITED) == 0)) {
14309                                 rack->r_ctl.rc_first_appl = lrsm;
14310                                 lrsm->r_flags |= RACK_APP_LIMITED;
14311                                 rack->r_ctl.rc_app_limited_cnt++;
14312                         }
14313                 }
14314                 if (rack->rc_pace_to_cwnd) {
14315                         uint64_t rate_wanted = 0;
14316
14317                         slot = pace_to_fill_cwnd(rack, slot, len, segsiz, NULL, &rate_wanted, 1);
14318                         rack->rc_ack_can_sendout_data = 1;
14319                         rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, 0, 0, 14, __LINE__, NULL);
14320                 } else
14321                         rack_log_pacing_delay_calc(rack, len, slot, tr_perms, reduce, 0, 7, __LINE__, NULL);
14322         } else {
14323                 uint64_t bw_est, res, lentim, rate_wanted;
14324                 uint32_t orig_val, srtt, segs, oh;
14325                 int capped = 0;
14326                 int prev_fill;
14327
14328                 if ((rack->r_rr_config == 1) && rsm) {
14329                         return (rack->r_ctl.rc_min_to);
14330                 }
14331                 if (rack->use_fixed_rate) {
14332                         rate_wanted = bw_est = rack_get_fixed_pacing_bw(rack);
14333                 } else if ((rack->r_ctl.init_rate == 0) &&
14334 #ifdef NETFLIX_PEAKRATE
14335                            (rack->rc_tp->t_maxpeakrate == 0) &&
14336 #endif
14337                            (rack->r_ctl.gp_bw == 0)) {
14338                         /* no way to yet do an estimate */
14339                         bw_est = rate_wanted = 0;
14340                 } else {
14341                         bw_est = rack_get_bw(rack);
14342                         rate_wanted = rack_get_output_bw(rack, bw_est, rsm, &capped);
14343                 }
14344                 if ((bw_est == 0) || (rate_wanted == 0) ||
14345                     ((rack->gp_ready == 0) && (rack->use_fixed_rate == 0))) {
14346                         /*
14347                          * No way yet to make a b/w estimate or
14348                          * our raise is set incorrectly.
14349                          */
14350                         goto old_method;
14351                 }
14352                 /* We need to account for all the overheads */
14353                 segs = (len + segsiz - 1) / segsiz;
14354                 /*
14355                  * We need the diff between 1514 bytes (e-mtu with e-hdr)
14356                  * and how much data we put in each packet. Yes this
14357                  * means we may be off if we are larger than 1500 bytes
14358                  * or smaller. But this just makes us more conservative.
14359                  */
14360                 if (rack_hw_rate_min &&
14361                     (bw_est < rack_hw_rate_min))
14362                         can_start_hw_pacing = 0;
14363                 if (ETHERNET_SEGMENT_SIZE > segsiz)
14364                         oh = ETHERNET_SEGMENT_SIZE - segsiz;
14365                 else
14366                         oh = 0;
14367                 segs *= oh;
14368                 lentim = (uint64_t)(len + segs) * (uint64_t)HPTS_USEC_IN_SEC;
14369                 res = lentim / rate_wanted;
14370                 slot = (uint32_t)res;
14371                 orig_val = rack->r_ctl.rc_pace_max_segs;
14372                 if (rack->r_ctl.crte == NULL) {
14373                         /*
14374                          * Only do this if we are not hardware pacing
14375                          * since if we are doing hw-pacing below we will
14376                          * set make a call after setting up or changing
14377                          * the rate.
14378                          */
14379                         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
14380                 } else if (rack->rc_inp->inp_snd_tag == NULL) {
14381                         /*
14382                          * We lost our rate somehow, this can happen
14383                          * if the interface changed underneath us.
14384                          */
14385                         tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
14386                         rack->r_ctl.crte = NULL;
14387                         /* Lets re-allow attempting to setup pacing */
14388                         rack->rack_hdrw_pacing = 0;
14389                         rack->rack_attempt_hdwr_pace = 0;
14390                         rack_log_hdwr_pacing(rack,
14391                                              rate_wanted, bw_est, __LINE__,
14392                                              0, 6);
14393                 }
14394                 /* Did we change the TSO size, if so log it */
14395                 if (rack->r_ctl.rc_pace_max_segs != orig_val)
14396                         rack_log_pacing_delay_calc(rack, len, slot, orig_val, 0, 0, 15, __LINE__, NULL);
14397                 prev_fill = rack->r_via_fill_cw;
14398                 if ((rack->rc_pace_to_cwnd) &&
14399                     (capped == 0) &&
14400                     (rack->use_fixed_rate == 0) &&
14401                     (rack->in_probe_rtt == 0) &&
14402                     (IN_FASTRECOVERY(rack->rc_tp->t_flags) == 0)) {
14403                         /*
14404                          * We want to pace at our rate *or* faster to
14405                          * fill the cwnd to the max if its not full.
14406                          */
14407                         slot = pace_to_fill_cwnd(rack, slot, (len+segs), segsiz, &capped, &rate_wanted, 0);
14408                 }
14409                 if ((rack->rc_inp->inp_route.ro_nh != NULL) &&
14410                     (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
14411                         if ((rack->rack_hdw_pace_ena) &&
14412                             (can_start_hw_pacing > 0) &&
14413                             (rack->rack_hdrw_pacing == 0) &&
14414                             (rack->rack_attempt_hdwr_pace == 0)) {
14415                                 /*
14416                                  * Lets attempt to turn on hardware pacing
14417                                  * if we can.
14418                                  */
14419                                 rack->rack_attempt_hdwr_pace = 1;
14420                                 rack->r_ctl.crte = tcp_set_pacing_rate(rack->rc_tp,
14421                                                                        rack->rc_inp->inp_route.ro_nh->nh_ifp,
14422                                                                        rate_wanted,
14423                                                                        RS_PACING_GEQ,
14424                                                                        &err, &rack->r_ctl.crte_prev_rate);
14425                                 if (rack->r_ctl.crte) {
14426                                         rack->rack_hdrw_pacing = 1;
14427                                         rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted, segsiz,
14428                                                                                                  0, rack->r_ctl.crte,
14429                                                                                                  NULL);
14430                                         rack_log_hdwr_pacing(rack,
14431                                                              rate_wanted, rack->r_ctl.crte->rate, __LINE__,
14432                                                              err, 0);
14433                                         rack->r_ctl.last_hw_bw_req = rate_wanted;
14434                                 } else {
14435                                         counter_u64_add(rack_hw_pace_init_fail, 1);
14436                                 }
14437                         } else if (rack->rack_hdrw_pacing &&
14438                                    (rack->r_ctl.last_hw_bw_req != rate_wanted)) {
14439                                 /* Do we need to adjust our rate? */
14440                                 const struct tcp_hwrate_limit_table *nrte;
14441
14442                                 if (rack->r_up_only &&
14443                                     (rate_wanted < rack->r_ctl.crte->rate)) {
14444                                         /**
14445                                          * We have four possible states here
14446                                          * having to do with the previous time
14447                                          * and this time.
14448                                          *   previous  |  this-time
14449                                          * A)     0      |     0   -- fill_cw not in the picture
14450                                          * B)     1      |     0   -- we were doing a fill-cw but now are not
14451                                          * C)     1      |     1   -- all rates from fill_cw
14452                                          * D)     0      |     1   -- we were doing non-fill and now we are filling
14453                                          *
14454                                          * For case A, C and D we don't allow a drop. But for
14455                                          * case B where we now our on our steady rate we do
14456                                          * allow a drop.
14457                                          *
14458                                          */
14459                                         if (!((prev_fill == 1) && (rack->r_via_fill_cw == 0)))
14460                                                 goto done_w_hdwr;
14461                                 }
14462                                 if ((rate_wanted > rack->r_ctl.crte->rate) ||
14463                                     (rate_wanted <= rack->r_ctl.crte_prev_rate)) {
14464                                         if (rack_hw_rate_to_low &&
14465                                             (bw_est < rack_hw_rate_to_low)) {
14466                                                 /*
14467                                                  * The pacing rate is too low for hardware, but
14468                                                  * do allow hardware pacing to be restarted.
14469                                                  */
14470                                                 rack_log_hdwr_pacing(rack,
14471                                                              bw_est, rack->r_ctl.crte->rate, __LINE__,
14472                                                              0, 5);
14473                                                 tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
14474                                                 rack->r_ctl.crte = NULL;
14475                                                 rack->rack_attempt_hdwr_pace = 0;
14476                                                 rack->rack_hdrw_pacing = 0;
14477                                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
14478                                                 goto done_w_hdwr;
14479                                         }
14480                                         nrte = tcp_chg_pacing_rate(rack->r_ctl.crte,
14481                                                                    rack->rc_tp,
14482                                                                    rack->rc_inp->inp_route.ro_nh->nh_ifp,
14483                                                                    rate_wanted,
14484                                                                    RS_PACING_GEQ,
14485                                                                    &err, &rack->r_ctl.crte_prev_rate);
14486                                         if (nrte == NULL) {
14487                                                 /* Lost the rate */
14488                                                 rack->rack_hdrw_pacing = 0;
14489                                                 rack->r_ctl.crte = NULL;
14490                                                 rack_log_hdwr_pacing(rack,
14491                                                                      rate_wanted, 0, __LINE__,
14492                                                                      err, 1);
14493                                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
14494                                                 counter_u64_add(rack_hw_pace_lost, 1);
14495                                         } else if (nrte != rack->r_ctl.crte) {
14496                                                 rack->r_ctl.crte = nrte;
14497                                                 rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted,
14498                                                                                                          segsiz, 0,
14499                                                                                                          rack->r_ctl.crte,
14500                                                                                                          NULL);
14501                                                 rack_log_hdwr_pacing(rack,
14502                                                                      rate_wanted, rack->r_ctl.crte->rate, __LINE__,
14503                                                                      err, 2);
14504                                                 rack->r_ctl.last_hw_bw_req = rate_wanted;
14505                                         }
14506                                 } else {
14507                                         /* We just need to adjust the segment size */
14508                                         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
14509                                         rack_log_hdwr_pacing(rack,
14510                                                              rate_wanted, rack->r_ctl.crte->rate, __LINE__,
14511                                                              0, 4);
14512                                         rack->r_ctl.last_hw_bw_req = rate_wanted;
14513                                 }
14514                         }
14515                 }
14516                 if ((rack->r_ctl.crte != NULL) &&
14517                     (rack->r_ctl.crte->rate == rate_wanted)) {
14518                         /*
14519                          * We need to add a extra if the rates
14520                          * are exactly matched. The idea is
14521                          * we want the software to make sure the
14522                          * queue is empty before adding more, this
14523                          * gives us N MSS extra pace times where
14524                          * N is our sysctl
14525                          */
14526                         slot += (rack->r_ctl.crte->time_between * rack_hw_pace_extra_slots);
14527                 }
14528 done_w_hdwr:
14529                 if (rack_limit_time_with_srtt &&
14530                     (rack->use_fixed_rate == 0) &&
14531 #ifdef NETFLIX_PEAKRATE
14532                     (rack->rc_tp->t_maxpeakrate == 0) &&
14533 #endif
14534                     (rack->rack_hdrw_pacing == 0)) {
14535                         /*
14536                          * Sanity check, we do not allow the pacing delay
14537                          * to be longer than the SRTT of the path. If it is
14538                          * a slow path, then adding a packet should increase
14539                          * the RTT and compensate for this i.e. the srtt will
14540                          * be greater so the allowed pacing time will be greater.
14541                          *
14542                          * Note this restriction is not for where a peak rate
14543                          * is set, we are doing fixed pacing or hardware pacing.
14544                          */
14545                         if (rack->rc_tp->t_srtt)
14546                                 srtt = rack->rc_tp->t_srtt;
14547                         else
14548                                 srtt = RACK_INITIAL_RTO * HPTS_USEC_IN_MSEC;    /* its in ms convert */
14549                         if (srtt < slot) {
14550                                 rack_log_pacing_delay_calc(rack, srtt, slot, rate_wanted, bw_est, lentim, 99, __LINE__, NULL);
14551                                 slot = srtt;
14552                         }
14553                 }
14554                 rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, bw_est, lentim, 2, __LINE__, rsm);
14555         }
14556         if (rack->r_ctl.crte && (rack->r_ctl.crte->rs_num_enobufs > 0)) {
14557                 /*
14558                  * If this rate is seeing enobufs when it
14559                  * goes to send then either the nic is out
14560                  * of gas or we are mis-estimating the time
14561                  * somehow and not letting the queue empty
14562                  * completely. Lets add to the pacing time.
14563                  */
14564                 int hw_boost_delay;
14565
14566                 hw_boost_delay = rack->r_ctl.crte->time_between * rack_enobuf_hw_boost_mult;
14567                 if (hw_boost_delay > rack_enobuf_hw_max)
14568                         hw_boost_delay = rack_enobuf_hw_max;
14569                 else if (hw_boost_delay < rack_enobuf_hw_min)
14570                         hw_boost_delay = rack_enobuf_hw_min;
14571                 slot += hw_boost_delay;
14572         }
14573         if (slot)
14574                 counter_u64_add(rack_calc_nonzero, 1);
14575         else
14576                 counter_u64_add(rack_calc_zero, 1);
14577         return (slot);
14578 }
14579
14580 static void
14581 rack_start_gp_measurement(struct tcpcb *tp, struct tcp_rack *rack,
14582     tcp_seq startseq, uint32_t sb_offset)
14583 {
14584         struct rack_sendmap *my_rsm = NULL;
14585         struct rack_sendmap fe;
14586
14587         if (tp->t_state < TCPS_ESTABLISHED) {
14588                 /*
14589                  * We don't start any measurements if we are
14590                  * not at least established.
14591                  */
14592                 return;
14593         }
14594         tp->t_flags |= TF_GPUTINPROG;
14595         rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
14596         rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
14597         tp->gput_seq = startseq;
14598         rack->app_limited_needs_set = 0;
14599         if (rack->in_probe_rtt)
14600                 rack->measure_saw_probe_rtt = 1;
14601         else if ((rack->measure_saw_probe_rtt) &&
14602                  (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
14603                 rack->measure_saw_probe_rtt = 0;
14604         if (rack->rc_gp_filled)
14605                 tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
14606         else {
14607                 /* Special case initial measurement */
14608                 struct timeval tv;
14609
14610                 tp->gput_ts = tcp_get_usecs(&tv);
14611                 rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
14612         }
14613         /*
14614          * We take a guess out into the future,
14615          * if we have no measurement and no
14616          * initial rate, we measure the first
14617          * initial-windows worth of data to
14618          * speed up getting some GP measurement and
14619          * thus start pacing.
14620          */
14621         if ((rack->rc_gp_filled == 0) && (rack->r_ctl.init_rate == 0)) {
14622                 rack->app_limited_needs_set = 1;
14623                 tp->gput_ack = startseq + max(rc_init_window(rack),
14624                                               (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
14625                 rack_log_pacing_delay_calc(rack,
14626                                            tp->gput_seq,
14627                                            tp->gput_ack,
14628                                            0,
14629                                            tp->gput_ts,
14630                                            rack->r_ctl.rc_app_limited_cnt,
14631                                            9,
14632                                            __LINE__, NULL);
14633                 return;
14634         }
14635         if (sb_offset) {
14636                 /*
14637                  * We are out somewhere in the sb
14638                  * can we use the already outstanding data?
14639                  */
14640
14641                 if (rack->r_ctl.rc_app_limited_cnt == 0) {
14642                         /*
14643                          * Yes first one is good and in this case
14644                          * the tp->gput_ts is correctly set based on
14645                          * the last ack that arrived (no need to
14646                          * set things up when an ack comes in).
14647                          */
14648                         my_rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
14649                         if ((my_rsm == NULL) ||
14650                             (my_rsm->r_rtr_cnt != 1)) {
14651                                 /* retransmission? */
14652                                 goto use_latest;
14653                         }
14654                 } else {
14655                         if (rack->r_ctl.rc_first_appl == NULL) {
14656                                 /*
14657                                  * If rc_first_appl is NULL
14658                                  * then the cnt should be 0.
14659                                  * This is probably an error, maybe
14660                                  * a KASSERT would be approprate.
14661                                  */
14662                                 goto use_latest;
14663                         }
14664                         /*
14665                          * If we have a marker pointer to the last one that is
14666                          * app limited we can use that, but we need to set
14667                          * things up so that when it gets ack'ed we record
14668                          * the ack time (if its not already acked).
14669                          */
14670                         rack->app_limited_needs_set = 1;
14671                         /*
14672                          * We want to get to the rsm that is either
14673                          * next with space i.e. over 1 MSS or the one
14674                          * after that (after the app-limited).
14675                          */
14676                         my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
14677                                          rack->r_ctl.rc_first_appl);
14678                         if (my_rsm) {
14679                                 if ((my_rsm->r_end - my_rsm->r_start) <= ctf_fixed_maxseg(tp))
14680                                         /* Have to use the next one */
14681                                         my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
14682                                                          my_rsm);
14683                                 else {
14684                                         /* Use after the first MSS of it is acked */
14685                                         tp->gput_seq = my_rsm->r_start + ctf_fixed_maxseg(tp);
14686                                         goto start_set;
14687                                 }
14688                         }
14689                         if ((my_rsm == NULL) ||
14690                             (my_rsm->r_rtr_cnt != 1)) {
14691                                 /*
14692                                  * Either its a retransmit or
14693                                  * the last is the app-limited one.
14694                                  */
14695                                 goto use_latest;
14696                         }
14697                 }
14698                 tp->gput_seq = my_rsm->r_start;
14699 start_set:
14700                 if (my_rsm->r_flags & RACK_ACKED) {
14701                         /*
14702                          * This one has been acked use the arrival ack time
14703                          */
14704                         tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
14705                         rack->app_limited_needs_set = 0;
14706                 }
14707                 rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)];
14708                 tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
14709                 rack_log_pacing_delay_calc(rack,
14710                                            tp->gput_seq,
14711                                            tp->gput_ack,
14712                                            (uint64_t)my_rsm,
14713                                            tp->gput_ts,
14714                                            rack->r_ctl.rc_app_limited_cnt,
14715                                            9,
14716                                            __LINE__, NULL);
14717                 return;
14718         }
14719
14720 use_latest:
14721         /*
14722          * We don't know how long we may have been
14723          * idle or if this is the first-send. Lets
14724          * setup the flag so we will trim off
14725          * the first ack'd data so we get a true
14726          * measurement.
14727          */
14728         rack->app_limited_needs_set = 1;
14729         tp->gput_ack = startseq + rack_get_measure_window(tp, rack);
14730         /* Find this guy so we can pull the send time */
14731         fe.r_start = startseq;
14732         my_rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
14733         if (my_rsm) {
14734                 rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)];
14735                 if (my_rsm->r_flags & RACK_ACKED) {
14736                         /*
14737                          * Unlikely since its probably what was
14738                          * just transmitted (but I am paranoid).
14739                          */
14740                         tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
14741                         rack->app_limited_needs_set = 0;
14742                 }
14743                 if (SEQ_LT(my_rsm->r_start, tp->gput_seq)) {
14744                         /* This also is unlikely */
14745                         tp->gput_seq = my_rsm->r_start;
14746                 }
14747         } else {
14748                 /*
14749                  * TSNH unless we have some send-map limit,
14750                  * and even at that it should not be hitting
14751                  * that limit (we should have stopped sending).
14752                  */
14753                 struct timeval tv;
14754
14755                 microuptime(&tv);
14756                 rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
14757         }
14758         rack_log_pacing_delay_calc(rack,
14759                                    tp->gput_seq,
14760                                    tp->gput_ack,
14761                                    (uint64_t)my_rsm,
14762                                    tp->gput_ts,
14763                                    rack->r_ctl.rc_app_limited_cnt,
14764                                    9, __LINE__, NULL);
14765 }
14766
14767 static inline uint32_t
14768 rack_what_can_we_send(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t cwnd_to_use,
14769     uint32_t avail, int32_t sb_offset)
14770 {
14771         uint32_t len;
14772         uint32_t sendwin;
14773
14774         if (tp->snd_wnd > cwnd_to_use)
14775                 sendwin = cwnd_to_use;
14776         else
14777                 sendwin = tp->snd_wnd;
14778         if (ctf_outstanding(tp) >= tp->snd_wnd) {
14779                 /* We never want to go over our peers rcv-window */
14780                 len = 0;
14781         } else {
14782                 uint32_t flight;
14783
14784                 flight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
14785                 if (flight >= sendwin) {
14786                         /*
14787                          * We have in flight what we are allowed by cwnd (if
14788                          * it was rwnd blocking it would have hit above out
14789                          * >= tp->snd_wnd).
14790                          */
14791                         return (0);
14792                 }
14793                 len = sendwin - flight;
14794                 if ((len + ctf_outstanding(tp)) > tp->snd_wnd) {
14795                         /* We would send too much (beyond the rwnd) */
14796                         len = tp->snd_wnd - ctf_outstanding(tp);
14797                 }
14798                 if ((len + sb_offset) > avail) {
14799                         /*
14800                          * We don't have that much in the SB, how much is
14801                          * there?
14802                          */
14803                         len = avail - sb_offset;
14804                 }
14805         }
14806         return (len);
14807 }
14808
14809 static void
14810 rack_log_fsb(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t flags,
14811              unsigned ipoptlen, int32_t orig_len, int32_t len, int error,
14812              int rsm_is_null, int optlen, int line, uint16_t mode)
14813 {
14814         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
14815                 union tcp_log_stackspecific log;
14816                 struct timeval tv;
14817
14818                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
14819                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
14820                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
14821                 log.u_bbr.flex1 = error;
14822                 log.u_bbr.flex2 = flags;
14823                 log.u_bbr.flex3 = rsm_is_null;
14824                 log.u_bbr.flex4 = ipoptlen;
14825                 log.u_bbr.flex5 = tp->rcv_numsacks;
14826                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
14827                 log.u_bbr.flex7 = optlen;
14828                 log.u_bbr.flex8 = rack->r_fsb_inited;
14829                 log.u_bbr.applimited = rack->r_fast_output;
14830                 log.u_bbr.bw_inuse = rack_get_bw(rack);
14831                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
14832                 log.u_bbr.cwnd_gain = mode;
14833                 log.u_bbr.pkts_out = orig_len;
14834                 log.u_bbr.lt_epoch = len;
14835                 log.u_bbr.delivered = line;
14836                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14837                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14838                 tcp_log_event_(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_FSB, 0,
14839                                len, &log, false, NULL, NULL, 0, &tv);
14840         }
14841 }
14842
14843
14844 static struct mbuf *
14845 rack_fo_base_copym(struct mbuf *the_m, uint32_t the_off, int32_t *plen,
14846                    struct rack_fast_send_blk *fsb,
14847                    int32_t seglimit, int32_t segsize)
14848 {
14849 #ifdef KERN_TLS
14850         struct ktls_session *tls, *ntls;
14851         struct mbuf *start;
14852 #endif
14853         struct mbuf *m, *n, **np, *smb;
14854         struct mbuf *top;
14855         int32_t off, soff;
14856         int32_t len = *plen;
14857         int32_t fragsize;
14858         int32_t len_cp = 0;
14859         uint32_t mlen, frags;
14860
14861         soff = off = the_off;
14862         smb = m = the_m;
14863         np = &top;
14864         top = NULL;
14865 #ifdef KERN_TLS
14866         if (hw_tls && (m->m_flags & M_EXTPG))
14867                 tls = m->m_epg_tls;
14868         else
14869                 tls = NULL;
14870         start = m;
14871 #endif
14872         while (len > 0) {
14873                 if (m == NULL) {
14874                         *plen = len_cp;
14875                         break;
14876                 }
14877 #ifdef KERN_TLS
14878                 if (hw_tls) {
14879                         if (m->m_flags & M_EXTPG)
14880                                 ntls = m->m_epg_tls;
14881                         else
14882                                 ntls = NULL;
14883
14884                         /*
14885                          * Avoid mixing TLS records with handshake
14886                          * data or TLS records from different
14887                          * sessions.
14888                          */
14889                         if (tls != ntls) {
14890                                 MPASS(m != start);
14891                                 *plen = len_cp;
14892                                 break;
14893                         }
14894                 }
14895 #endif
14896                 mlen = min(len, m->m_len - off);
14897                 if (seglimit) {
14898                         /*
14899                          * For M_EXTPG mbufs, add 3 segments
14900                          * + 1 in case we are crossing page boundaries
14901                          * + 2 in case the TLS hdr/trailer are used
14902                          * It is cheaper to just add the segments
14903                          * than it is to take the cache miss to look
14904                          * at the mbuf ext_pgs state in detail.
14905                          */
14906                         if (m->m_flags & M_EXTPG) {
14907                                 fragsize = min(segsize, PAGE_SIZE);
14908                                 frags = 3;
14909                         } else {
14910                                 fragsize = segsize;
14911                                 frags = 0;
14912                         }
14913
14914                         /* Break if we really can't fit anymore. */
14915                         if ((frags + 1) >= seglimit) {
14916                                 *plen = len_cp;
14917                                 break;
14918                         }
14919
14920                         /*
14921                          * Reduce size if you can't copy the whole
14922                          * mbuf. If we can't copy the whole mbuf, also
14923                          * adjust len so the loop will end after this
14924                          * mbuf.
14925                          */
14926                         if ((frags + howmany(mlen, fragsize)) >= seglimit) {
14927                                 mlen = (seglimit - frags - 1) * fragsize;
14928                                 len = mlen;
14929                                 *plen = len_cp + len;
14930                         }
14931                         frags += howmany(mlen, fragsize);
14932                         if (frags == 0)
14933                                 frags++;
14934                         seglimit -= frags;
14935                         KASSERT(seglimit > 0,
14936                             ("%s: seglimit went too low", __func__));
14937                 }
14938                 n = m_get(M_NOWAIT, m->m_type);
14939                 *np = n;
14940                 if (n == NULL)
14941                         goto nospace;
14942                 n->m_len = mlen;
14943                 soff += mlen;
14944                 len_cp += n->m_len;
14945                 if (m->m_flags & (M_EXT|M_EXTPG)) {
14946                         n->m_data = m->m_data + off;
14947                         mb_dupcl(n, m);
14948                 } else {
14949                         bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
14950                             (u_int)n->m_len);
14951                 }
14952                 len -= n->m_len;
14953                 off = 0;
14954                 m = m->m_next;
14955                 np = &n->m_next;
14956                 if (len || (soff == smb->m_len)) {
14957                         /*
14958                          * We have more so we move forward  or
14959                          * we have consumed the entire mbuf and
14960                          * len has fell to 0.
14961                          */
14962                         soff = 0;
14963                         smb = m;
14964                 }
14965
14966         }
14967         if (fsb != NULL) {
14968                 fsb->m = smb;
14969                 fsb->off = soff;
14970                 if (smb) {
14971                         /*
14972                          * Save off the size of the mbuf. We do
14973                          * this so that we can recognize when it
14974                          * has been trimmed by sbcut() as acks
14975                          * come in.
14976                          */
14977                         fsb->o_m_len = smb->m_len;
14978                 } else {
14979                         /*
14980                          * This is the case where the next mbuf went to NULL. This
14981                          * means with this copy we have sent everything in the sb.
14982                          * In theory we could clear the fast_output flag, but lets
14983                          * not since its possible that we could get more added
14984                          * and acks that call the extend function which would let
14985                          * us send more.
14986                          */
14987                         fsb->o_m_len = 0;
14988                 }
14989         }
14990         return (top);
14991 nospace:
14992         if (top)
14993                 m_freem(top);
14994         return (NULL);
14995
14996 }
14997
14998 /*
14999  * This is a copy of m_copym(), taking the TSO segment size/limit
15000  * constraints into account, and advancing the sndptr as it goes.
15001  */
15002 static struct mbuf *
15003 rack_fo_m_copym(struct tcp_rack *rack, int32_t *plen,
15004                 int32_t seglimit, int32_t segsize, struct mbuf **s_mb, int *s_soff)
15005 {
15006         struct mbuf *m, *n;
15007         int32_t soff;
15008
15009         soff = rack->r_ctl.fsb.off;
15010         m = rack->r_ctl.fsb.m;
15011         if (rack->r_ctl.fsb.o_m_len != m->m_len) {
15012                 /*
15013                  * The mbuf had the front of it chopped off by an ack
15014                  * we need to adjust the soff/off by that difference.
15015                  */
15016                 uint32_t delta;
15017
15018                 delta = rack->r_ctl.fsb.o_m_len - m->m_len;
15019                 soff -= delta;
15020         }
15021         KASSERT(soff >= 0, ("%s, negative off %d", __FUNCTION__, soff));
15022         KASSERT(*plen >= 0, ("%s, negative len %d", __FUNCTION__, *plen));
15023         KASSERT(soff < m->m_len, ("%s rack:%p len:%u m:%p m->m_len:%u < off?",
15024                                  __FUNCTION__,
15025                                  rack, *plen, m, m->m_len));
15026         /* Save off the right location before we copy and advance */
15027         *s_soff = soff;
15028         *s_mb = rack->r_ctl.fsb.m;
15029         n = rack_fo_base_copym(m, soff, plen,
15030                                &rack->r_ctl.fsb,
15031                                seglimit, segsize);
15032         return (n);
15033 }
15034
15035 static int
15036 rack_fast_rsm_output(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm,
15037                      uint64_t ts_val, uint32_t cts, uint32_t ms_cts, struct timeval *tv, int len)
15038 {
15039         /*
15040          * Enter the fast retransmit path. We are given that a sched_pin is
15041          * in place (if accounting is compliled in) and the cycle count taken
15042          * at the entry is in the ts_val. The concept her is that the rsm
15043          * now holds the mbuf offsets and such so we can directly transmit
15044          * without a lot of overhead, the len field is already set for
15045          * us to prohibit us from sending too much (usually its 1MSS).
15046          */
15047         struct ip *ip = NULL;
15048         struct udphdr *udp = NULL;
15049         struct tcphdr *th = NULL;
15050         struct mbuf *m = NULL;
15051         struct inpcb *inp;
15052         uint8_t *cpto;
15053         struct tcp_log_buffer *lgb;
15054 #ifdef TCP_ACCOUNTING
15055         uint64_t crtsc;
15056         int cnt_thru = 1;
15057 #endif
15058         int doing_tlp = 0;
15059         struct tcpopt to;
15060         u_char opt[TCP_MAXOLEN];
15061         uint32_t hdrlen, optlen;
15062         int32_t slot, segsiz, max_val, tso = 0, error, flags, ulen = 0;
15063         uint32_t us_cts;
15064         uint32_t if_hw_tsomaxsegcount = 0, startseq;
15065         uint32_t if_hw_tsomaxsegsize;
15066 #ifdef INET6
15067         struct ip6_hdr *ip6 = NULL;
15068
15069         if (rack->r_is_v6) {
15070                 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
15071                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
15072         } else
15073 #endif                          /* INET6 */
15074         {
15075                 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
15076                 hdrlen = sizeof(struct tcpiphdr);
15077         }
15078         if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
15079                 goto failed;
15080         }
15081         if (rsm->r_flags & RACK_TLP)
15082                 doing_tlp = 1;
15083         startseq = rsm->r_start;
15084         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
15085         inp = rack->rc_inp;
15086         to.to_flags = 0;
15087         flags = tcp_outflags[tp->t_state];
15088         if (flags & (TH_SYN|TH_RST)) {
15089                 goto failed;
15090         }
15091         if (rsm->r_flags & RACK_HAS_FIN) {
15092                 /* We can't send a FIN here */
15093                 goto failed;
15094         }
15095         if (flags & TH_FIN) {
15096                 /* We never send a FIN */
15097                 flags &= ~TH_FIN;
15098         }
15099         if (tp->t_flags & TF_RCVD_TSTMP) {
15100                 to.to_tsval = ms_cts + tp->ts_offset;
15101                 to.to_tsecr = tp->ts_recent;
15102                 to.to_flags = TOF_TS;
15103         }
15104         optlen = tcp_addoptions(&to, opt);
15105         hdrlen += optlen;
15106         udp = rack->r_ctl.fsb.udp;
15107         if (udp)
15108                 hdrlen += sizeof(struct udphdr);
15109         if (rack->r_ctl.rc_pace_max_segs)
15110                 max_val = rack->r_ctl.rc_pace_max_segs;
15111         else if (rack->rc_user_set_max_segs)
15112                 max_val = rack->rc_user_set_max_segs * segsiz;
15113         else
15114                 max_val = len;
15115         if ((tp->t_flags & TF_TSO) &&
15116             V_tcp_do_tso &&
15117             (len > segsiz) &&
15118             (tp->t_port == 0))
15119                 tso = 1;
15120 #ifdef INET6
15121         if (MHLEN < hdrlen + max_linkhdr)
15122                 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
15123         else
15124 #endif
15125                 m = m_gethdr(M_NOWAIT, MT_DATA);
15126         if (m == NULL)
15127                 goto failed;
15128         m->m_data += max_linkhdr;
15129         m->m_len = hdrlen;
15130         th = rack->r_ctl.fsb.th;
15131         /* Establish the len to send */
15132         if (len > max_val)
15133                 len = max_val;
15134         if ((tso) && (len + optlen > tp->t_maxseg)) {
15135                 uint32_t if_hw_tsomax;
15136                 int32_t max_len;
15137
15138                 /* extract TSO information */
15139                 if_hw_tsomax = tp->t_tsomax;
15140                 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
15141                 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
15142                 /*
15143                  * Check if we should limit by maximum payload
15144                  * length:
15145                  */
15146                 if (if_hw_tsomax != 0) {
15147                         /* compute maximum TSO length */
15148                         max_len = (if_hw_tsomax - hdrlen -
15149                                    max_linkhdr);
15150                         if (max_len <= 0) {
15151                                 goto failed;
15152                         } else if (len > max_len) {
15153                                 len = max_len;
15154                         }
15155                 }
15156                 if (len <= segsiz) {
15157                         /*
15158                          * In case there are too many small fragments don't
15159                          * use TSO:
15160                          */
15161                         tso = 0;
15162                 }
15163         } else {
15164                 tso = 0;
15165         }
15166         if ((tso == 0) && (len > segsiz))
15167                 len = segsiz;
15168         us_cts = tcp_get_usecs(tv);
15169         if ((len == 0) ||
15170             (len <= MHLEN - hdrlen - max_linkhdr)) {
15171                 goto failed;
15172         }
15173         th->th_seq = htonl(rsm->r_start);
15174         th->th_ack = htonl(tp->rcv_nxt);
15175         if(rsm->r_flags & RACK_HAD_PUSH)
15176                 flags |= TH_PUSH;
15177         th->th_flags = flags;
15178         th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
15179         if (th->th_win == 0) {
15180                 tp->t_sndzerowin++;
15181                 tp->t_flags |= TF_RXWIN0SENT;
15182         } else
15183                 tp->t_flags &= ~TF_RXWIN0SENT;
15184         if (rsm->r_flags & RACK_TLP) {
15185                 /*
15186                  * TLP should not count in retran count, but
15187                  * in its own bin
15188                  */
15189                 counter_u64_add(rack_tlp_retran, 1);
15190                 counter_u64_add(rack_tlp_retran_bytes, len);
15191         } else {
15192                 tp->t_sndrexmitpack++;
15193                 KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
15194                 KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
15195         }
15196 #ifdef STATS
15197         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
15198                                  len);
15199 #endif
15200         if (rsm->m == NULL)
15201                 goto failed;
15202         if (rsm->orig_m_len != rsm->m->m_len) {
15203                 /* Fix up the orig_m_len and possibly the mbuf offset */
15204                 rack_adjust_orig_mlen(rsm);
15205         }
15206         m->m_next = rack_fo_base_copym(rsm->m, rsm->soff, &len, NULL, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize);
15207         if (len <= segsiz) {
15208                 /*
15209                  * Must have ran out of mbufs for the copy
15210                  * shorten it to no longer need tso. Lets
15211                  * not put on sendalot since we are low on
15212                  * mbufs.
15213                  */
15214                 tso = 0;
15215         }
15216         if ((m->m_next == NULL) || (len <= 0)){
15217                 goto failed;
15218         }
15219         if (udp) {
15220                 if (rack->r_is_v6)
15221                         ulen = hdrlen + len - sizeof(struct ip6_hdr);
15222                 else
15223                         ulen = hdrlen + len - sizeof(struct ip);
15224                 udp->uh_ulen = htons(ulen);
15225         }
15226         m->m_pkthdr.rcvif = (struct ifnet *)0;
15227         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
15228 #ifdef INET6
15229         if (rack->r_is_v6) {
15230                 if (tp->t_port) {
15231                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
15232                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15233                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
15234                         th->th_sum = htons(0);
15235                         UDPSTAT_INC(udps_opackets);
15236                 } else {
15237                         m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
15238                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15239                         th->th_sum = in6_cksum_pseudo(ip6,
15240                                                       sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
15241                                                       0);
15242                 }
15243         }
15244 #endif
15245 #if defined(INET6) && defined(INET)
15246         else
15247 #endif
15248 #ifdef INET
15249         {
15250                 if (tp->t_port) {
15251                         m->m_pkthdr.csum_flags = CSUM_UDP;
15252                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15253                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
15254                                                 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
15255                         th->th_sum = htons(0);
15256                         UDPSTAT_INC(udps_opackets);
15257                 } else {
15258                         m->m_pkthdr.csum_flags = CSUM_TCP;
15259                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15260                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
15261                                                ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
15262                                                                         IPPROTO_TCP + len + optlen));
15263                 }
15264                 /* IP version must be set here for ipv4/ipv6 checking later */
15265                 KASSERT(ip->ip_v == IPVERSION,
15266                         ("%s: IP version incorrect: %d", __func__, ip->ip_v));
15267         }
15268 #endif
15269         if (tso) {
15270                 KASSERT(len > tp->t_maxseg - optlen,
15271                         ("%s: len <= tso_segsz tp:%p", __func__, tp));
15272                 m->m_pkthdr.csum_flags |= CSUM_TSO;
15273                 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
15274         }
15275 #ifdef INET6
15276         if (rack->r_is_v6) {
15277                 ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
15278                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
15279                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
15280                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15281                 else
15282                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15283         }
15284 #endif
15285 #if defined(INET) && defined(INET6)
15286         else
15287 #endif
15288 #ifdef INET
15289         {
15290                 ip->ip_len = htons(m->m_pkthdr.len);
15291                 ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
15292                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
15293                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15294                         if (tp->t_port == 0 || len < V_tcp_minmss) {
15295                                 ip->ip_off |= htons(IP_DF);
15296                         }
15297                 } else {
15298                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15299                 }
15300         }
15301 #endif
15302         /* Time to copy in our header */
15303         cpto = mtod(m, uint8_t *);
15304         memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
15305         th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
15306         if (optlen) {
15307                 bcopy(opt, th + 1, optlen);
15308                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
15309         } else {
15310                 th->th_off = sizeof(struct tcphdr) >> 2;
15311         }
15312         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
15313                 union tcp_log_stackspecific log;
15314
15315                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15316                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
15317                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
15318                 if (rack->rack_no_prr)
15319                         log.u_bbr.flex1 = 0;
15320                 else
15321                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
15322                 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
15323                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
15324                 log.u_bbr.flex4 = max_val;
15325                 log.u_bbr.flex5 = 0;
15326                 /* Save off the early/late values */
15327                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
15328                 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
15329                 log.u_bbr.bw_inuse = rack_get_bw(rack);
15330                 log.u_bbr.flex8 = 1;
15331                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
15332                 log.u_bbr.flex7 = 55;
15333                 log.u_bbr.pkts_out = tp->t_maxseg;
15334                 log.u_bbr.timeStamp = cts;
15335                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15336                 log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
15337                 log.u_bbr.delivered = 0;
15338                 lgb = tcp_log_event_(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
15339                                      len, &log, false, NULL, NULL, 0, tv);
15340         } else
15341                 lgb = NULL;
15342 #ifdef INET6
15343         if (rack->r_is_v6) {
15344                 error = ip6_output(m, NULL,
15345                                    &inp->inp_route6,
15346                                    0, NULL, NULL, inp);
15347         }
15348 #endif
15349 #if defined(INET) && defined(INET6)
15350         else
15351 #endif
15352 #ifdef INET
15353         {
15354                 error = ip_output(m, NULL,
15355                                   &inp->inp_route,
15356                                   0, 0, inp);
15357         }
15358 #endif
15359         m = NULL;
15360         if (lgb) {
15361                 lgb->tlb_errno = error;
15362                 lgb = NULL;
15363         }
15364         if (error) {
15365                 goto failed;
15366         }
15367         rack_log_output(tp, &to, len, rsm->r_start, flags, error, rack_to_usec_ts(tv),
15368                         rsm, RACK_SENT_FP, rsm->m, rsm->soff);
15369         if (doing_tlp && (rack->fast_rsm_hack == 0)) {
15370                 rack->rc_tlp_in_progress = 1;
15371                 rack->r_ctl.rc_tlp_cnt_out++;
15372         }
15373         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
15374         rack->forced_ack = 0;   /* If we send something zap the FA flag */
15375         if (IN_FASTRECOVERY(tp->t_flags) && rsm)
15376                 rack->r_ctl.retran_during_recovery += len;
15377         {
15378                 int idx;
15379
15380                 idx = (len / segsiz) + 3;
15381                 if (idx >= TCP_MSS_ACCT_ATIMER)
15382                         counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
15383                 else
15384                         counter_u64_add(rack_out_size[idx], 1);
15385         }
15386         if (tp->t_rtttime == 0) {
15387                 tp->t_rtttime = ticks;
15388                 tp->t_rtseq = startseq;
15389                 KMOD_TCPSTAT_INC(tcps_segstimed);
15390         }
15391         counter_u64_add(rack_fto_rsm_send, 1);
15392         if (error && (error == ENOBUFS)) {
15393                 slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
15394                 if (rack->rc_enobuf < 0x7f)
15395                         rack->rc_enobuf++;
15396                 if (slot < (10 * HPTS_USEC_IN_MSEC))
15397                         slot = 10 * HPTS_USEC_IN_MSEC;
15398         } else
15399                 slot = rack_get_pacing_delay(rack, tp, len, NULL, segsiz);
15400         if ((slot == 0) ||
15401             (rack->rc_always_pace == 0) ||
15402             (rack->r_rr_config == 1)) {
15403                 /*
15404                  * We have no pacing set or we
15405                  * are using old-style rack or
15406                  * we are overriden to use the old 1ms pacing.
15407                  */
15408                 slot = rack->r_ctl.rc_min_to;
15409         }
15410         rack_start_hpts_timer(rack, tp, cts, slot, len, 0);
15411         if (rack->r_must_retran) {
15412                 rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
15413                 if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
15414                         /*
15415                          * We have retransmitted all we need.
15416                          */
15417                         rack->r_must_retran = 0;
15418                         rack->r_ctl.rc_out_at_rto = 0;
15419                 }
15420         }
15421 #ifdef TCP_ACCOUNTING
15422         crtsc = get_cyclecount();
15423         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15424                 tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
15425         }
15426         counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], cnt_thru);
15427         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15428                 tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
15429         }
15430         counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
15431         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15432                 tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((len + segsiz - 1) / segsiz);
15433         }
15434         counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((len + segsiz - 1) / segsiz));
15435         sched_unpin();
15436 #endif
15437         return (0);
15438 failed:
15439         if (m)
15440                 m_free(m);
15441         return (-1);
15442 }
15443
15444 static void
15445 rack_sndbuf_autoscale(struct tcp_rack *rack)
15446 {
15447         /*
15448          * Automatic sizing of send socket buffer.  Often the send buffer
15449          * size is not optimally adjusted to the actual network conditions
15450          * at hand (delay bandwidth product).  Setting the buffer size too
15451          * small limits throughput on links with high bandwidth and high
15452          * delay (eg. trans-continental/oceanic links).  Setting the
15453          * buffer size too big consumes too much real kernel memory,
15454          * especially with many connections on busy servers.
15455          *
15456          * The criteria to step up the send buffer one notch are:
15457          *  1. receive window of remote host is larger than send buffer
15458          *     (with a fudge factor of 5/4th);
15459          *  2. send buffer is filled to 7/8th with data (so we actually
15460          *     have data to make use of it);
15461          *  3. send buffer fill has not hit maximal automatic size;
15462          *  4. our send window (slow start and cogestion controlled) is
15463          *     larger than sent but unacknowledged data in send buffer.
15464          *
15465          * Note that the rack version moves things much faster since
15466          * we want to avoid hitting cache lines in the rack_fast_output()
15467          * path so this is called much less often and thus moves
15468          * the SB forward by a percentage.
15469          */
15470         struct socket *so;
15471         struct tcpcb *tp;
15472         uint32_t sendwin, scaleup;
15473
15474         tp = rack->rc_tp;
15475         so = rack->rc_inp->inp_socket;
15476         sendwin = min(rack->r_ctl.cwnd_to_use, tp->snd_wnd);
15477         if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
15478                 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
15479                     sbused(&so->so_snd) >=
15480                     (so->so_snd.sb_hiwat / 8 * 7) &&
15481                     sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
15482                     sendwin >= (sbused(&so->so_snd) -
15483                     (tp->snd_nxt - tp->snd_una))) {
15484                         if (rack_autosndbuf_inc)
15485                                 scaleup = (rack_autosndbuf_inc * so->so_snd.sb_hiwat) / 100;
15486                         else
15487                                 scaleup = V_tcp_autosndbuf_inc;
15488                         if (scaleup < V_tcp_autosndbuf_inc)
15489                                 scaleup = V_tcp_autosndbuf_inc;
15490                         scaleup += so->so_snd.sb_hiwat;
15491                         if (scaleup > V_tcp_autosndbuf_max)
15492                                 scaleup = V_tcp_autosndbuf_max;
15493                         if (!sbreserve_locked(&so->so_snd, scaleup, so, curthread))
15494                                 so->so_snd.sb_flags &= ~SB_AUTOSIZE;
15495                 }
15496         }
15497 }
15498
15499 static int
15500 rack_fast_output(struct tcpcb *tp, struct tcp_rack *rack, uint64_t ts_val,
15501                  uint32_t cts, uint32_t ms_cts, struct timeval *tv, long tot_len, int *send_err)
15502 {
15503         /*
15504          * Enter to do fast output. We are given that the sched_pin is
15505          * in place (if accounting is compiled in) and the cycle count taken
15506          * at entry is in place in ts_val. The idea here is that
15507          * we know how many more bytes needs to be sent (presumably either
15508          * during pacing or to fill the cwnd and that was greater than
15509          * the max-burst). We have how much to send and all the info we
15510          * need to just send.
15511          */
15512         struct ip *ip = NULL;
15513         struct udphdr *udp = NULL;
15514         struct tcphdr *th = NULL;
15515         struct mbuf *m, *s_mb;
15516         struct inpcb *inp;
15517         uint8_t *cpto;
15518         struct tcp_log_buffer *lgb;
15519 #ifdef TCP_ACCOUNTING
15520         uint64_t crtsc;
15521 #endif
15522         struct tcpopt to;
15523         u_char opt[TCP_MAXOLEN];
15524         uint32_t hdrlen, optlen;
15525         int cnt_thru = 1;
15526         int32_t slot, segsiz, len, max_val, tso = 0, sb_offset, error, flags, ulen = 0;
15527         uint32_t us_cts, s_soff;
15528         uint32_t if_hw_tsomaxsegcount = 0, startseq;
15529         uint32_t if_hw_tsomaxsegsize;
15530         uint16_t add_flag = RACK_SENT_FP;
15531 #ifdef INET6
15532         struct ip6_hdr *ip6 = NULL;
15533
15534         if (rack->r_is_v6) {
15535                 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
15536                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
15537         } else
15538 #endif                          /* INET6 */
15539         {
15540                 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
15541                 hdrlen = sizeof(struct tcpiphdr);
15542         }
15543         if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
15544                 m = NULL;
15545                 goto failed;
15546         }
15547         startseq = tp->snd_max;
15548         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
15549         inp = rack->rc_inp;
15550         len = rack->r_ctl.fsb.left_to_send;
15551         to.to_flags = 0;
15552         flags = rack->r_ctl.fsb.tcp_flags;
15553         if (tp->t_flags & TF_RCVD_TSTMP) {
15554                 to.to_tsval = ms_cts + tp->ts_offset;
15555                 to.to_tsecr = tp->ts_recent;
15556                 to.to_flags = TOF_TS;
15557         }
15558         optlen = tcp_addoptions(&to, opt);
15559         hdrlen += optlen;
15560         udp = rack->r_ctl.fsb.udp;
15561         if (udp)
15562                 hdrlen += sizeof(struct udphdr);
15563         if (rack->r_ctl.rc_pace_max_segs)
15564                 max_val = rack->r_ctl.rc_pace_max_segs;
15565         else if (rack->rc_user_set_max_segs)
15566                 max_val = rack->rc_user_set_max_segs * segsiz;
15567         else
15568                 max_val = len;
15569         if ((tp->t_flags & TF_TSO) &&
15570             V_tcp_do_tso &&
15571             (len > segsiz) &&
15572             (tp->t_port == 0))
15573                 tso = 1;
15574 again:
15575 #ifdef INET6
15576         if (MHLEN < hdrlen + max_linkhdr)
15577                 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
15578         else
15579 #endif
15580                 m = m_gethdr(M_NOWAIT, MT_DATA);
15581         if (m == NULL)
15582                 goto failed;
15583         m->m_data += max_linkhdr;
15584         m->m_len = hdrlen;
15585         th = rack->r_ctl.fsb.th;
15586         /* Establish the len to send */
15587         if (len > max_val)
15588                 len = max_val;
15589         if ((tso) && (len + optlen > tp->t_maxseg)) {
15590                 uint32_t if_hw_tsomax;
15591                 int32_t max_len;
15592
15593                 /* extract TSO information */
15594                 if_hw_tsomax = tp->t_tsomax;
15595                 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
15596                 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
15597                 /*
15598                  * Check if we should limit by maximum payload
15599                  * length:
15600                  */
15601                 if (if_hw_tsomax != 0) {
15602                         /* compute maximum TSO length */
15603                         max_len = (if_hw_tsomax - hdrlen -
15604                                    max_linkhdr);
15605                         if (max_len <= 0) {
15606                                 goto failed;
15607                         } else if (len > max_len) {
15608                                 len = max_len;
15609                         }
15610                 }
15611                 if (len <= segsiz) {
15612                         /*
15613                          * In case there are too many small fragments don't
15614                          * use TSO:
15615                          */
15616                         tso = 0;
15617                 }
15618         } else {
15619                 tso = 0;
15620         }
15621         if ((tso == 0) && (len > segsiz))
15622                 len = segsiz;
15623         us_cts = tcp_get_usecs(tv);
15624         if ((len == 0) ||
15625             (len <= MHLEN - hdrlen - max_linkhdr)) {
15626                 goto failed;
15627         }
15628         sb_offset = tp->snd_max - tp->snd_una;
15629         th->th_seq = htonl(tp->snd_max);
15630         th->th_ack = htonl(tp->rcv_nxt);
15631         th->th_flags = flags;
15632         th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
15633         if (th->th_win == 0) {
15634                 tp->t_sndzerowin++;
15635                 tp->t_flags |= TF_RXWIN0SENT;
15636         } else
15637                 tp->t_flags &= ~TF_RXWIN0SENT;
15638         tp->snd_up = tp->snd_una;       /* drag it along, its deprecated */
15639         KMOD_TCPSTAT_INC(tcps_sndpack);
15640         KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
15641 #ifdef STATS
15642         stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
15643                                  len);
15644 #endif
15645         if (rack->r_ctl.fsb.m == NULL)
15646                 goto failed;
15647
15648         /* s_mb and s_soff are saved for rack_log_output */
15649         m->m_next = rack_fo_m_copym(rack, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, &s_mb, &s_soff);
15650         if (len <= segsiz) {
15651                 /*
15652                  * Must have ran out of mbufs for the copy
15653                  * shorten it to no longer need tso. Lets
15654                  * not put on sendalot since we are low on
15655                  * mbufs.
15656                  */
15657                 tso = 0;
15658         }
15659         if (rack->r_ctl.fsb.rfo_apply_push &&
15660             (len == rack->r_ctl.fsb.left_to_send)) {
15661                 th->th_flags |= TH_PUSH;
15662                 add_flag |= RACK_HAD_PUSH;
15663         }
15664         if ((m->m_next == NULL) || (len <= 0)){
15665                 goto failed;
15666         }
15667         if (udp) {
15668                 if (rack->r_is_v6)
15669                         ulen = hdrlen + len - sizeof(struct ip6_hdr);
15670                 else
15671                         ulen = hdrlen + len - sizeof(struct ip);
15672                 udp->uh_ulen = htons(ulen);
15673         }
15674         m->m_pkthdr.rcvif = (struct ifnet *)0;
15675         if (tp->t_state == TCPS_ESTABLISHED &&
15676             (tp->t_flags2 & TF2_ECN_PERMIT)) {
15677                 /*
15678                  * If the peer has ECN, mark data packets with ECN capable
15679                  * transmission (ECT). Ignore pure ack packets,
15680                  * retransmissions.
15681                  */
15682                 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max)) {
15683 #ifdef INET6
15684                         if (rack->r_is_v6)
15685                                 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
15686                         else
15687 #endif
15688                                 ip->ip_tos |= IPTOS_ECN_ECT0;
15689                         KMOD_TCPSTAT_INC(tcps_ecn_ect0);
15690                         /*
15691                          * Reply with proper ECN notifications.
15692                          * Only set CWR on new data segments.
15693                          */
15694                         if (tp->t_flags2 & TF2_ECN_SND_CWR) {
15695                                 flags |= TH_CWR;
15696                                 tp->t_flags2 &= ~TF2_ECN_SND_CWR;
15697                         }
15698                 }
15699                 if (tp->t_flags2 & TF2_ECN_SND_ECE)
15700                         flags |= TH_ECE;
15701         }
15702         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
15703 #ifdef INET6
15704         if (rack->r_is_v6) {
15705                 if (tp->t_port) {
15706                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
15707                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15708                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
15709                         th->th_sum = htons(0);
15710                         UDPSTAT_INC(udps_opackets);
15711                 } else {
15712                         m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
15713                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15714                         th->th_sum = in6_cksum_pseudo(ip6,
15715                                                       sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
15716                                                       0);
15717                 }
15718         }
15719 #endif
15720 #if defined(INET6) && defined(INET)
15721         else
15722 #endif
15723 #ifdef INET
15724         {
15725                 if (tp->t_port) {
15726                         m->m_pkthdr.csum_flags = CSUM_UDP;
15727                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15728                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
15729                                                 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
15730                         th->th_sum = htons(0);
15731                         UDPSTAT_INC(udps_opackets);
15732                 } else {
15733                         m->m_pkthdr.csum_flags = CSUM_TCP;
15734                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15735                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
15736                                                ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
15737                                                                         IPPROTO_TCP + len + optlen));
15738                 }
15739                 /* IP version must be set here for ipv4/ipv6 checking later */
15740                 KASSERT(ip->ip_v == IPVERSION,
15741                         ("%s: IP version incorrect: %d", __func__, ip->ip_v));
15742         }
15743 #endif
15744         if (tso) {
15745                 KASSERT(len > tp->t_maxseg - optlen,
15746                         ("%s: len <= tso_segsz tp:%p", __func__, tp));
15747                 m->m_pkthdr.csum_flags |= CSUM_TSO;
15748                 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
15749         }
15750 #ifdef INET6
15751         if (rack->r_is_v6) {
15752                 ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
15753                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
15754                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
15755                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15756                 else
15757                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15758         }
15759 #endif
15760 #if defined(INET) && defined(INET6)
15761         else
15762 #endif
15763 #ifdef INET
15764         {
15765                 ip->ip_len = htons(m->m_pkthdr.len);
15766                 ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
15767                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
15768                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15769                         if (tp->t_port == 0 || len < V_tcp_minmss) {
15770                                 ip->ip_off |= htons(IP_DF);
15771                         }
15772                 } else {
15773                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15774                 }
15775         }
15776 #endif
15777         /* Time to copy in our header */
15778         cpto = mtod(m, uint8_t *);
15779         memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
15780         th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
15781         if (optlen) {
15782                 bcopy(opt, th + 1, optlen);
15783                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
15784         } else {
15785                 th->th_off = sizeof(struct tcphdr) >> 2;
15786         }
15787         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
15788                 union tcp_log_stackspecific log;
15789
15790                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15791                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
15792                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
15793                 if (rack->rack_no_prr)
15794                         log.u_bbr.flex1 = 0;
15795                 else
15796                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
15797                 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
15798                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
15799                 log.u_bbr.flex4 = max_val;
15800                 log.u_bbr.flex5 = 0;
15801                 /* Save off the early/late values */
15802                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
15803                 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
15804                 log.u_bbr.bw_inuse = rack_get_bw(rack);
15805                 log.u_bbr.flex8 = 0;
15806                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
15807                 log.u_bbr.flex7 = 44;
15808                 log.u_bbr.pkts_out = tp->t_maxseg;
15809                 log.u_bbr.timeStamp = cts;
15810                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15811                 log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
15812                 log.u_bbr.delivered = 0;
15813                 lgb = tcp_log_event_(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
15814                                      len, &log, false, NULL, NULL, 0, tv);
15815         } else
15816                 lgb = NULL;
15817 #ifdef INET6
15818         if (rack->r_is_v6) {
15819                 error = ip6_output(m, NULL,
15820                                    &inp->inp_route6,
15821                                    0, NULL, NULL, inp);
15822         }
15823 #endif
15824 #if defined(INET) && defined(INET6)
15825         else
15826 #endif
15827 #ifdef INET
15828         {
15829                 error = ip_output(m, NULL,
15830                                   &inp->inp_route,
15831                                   0, 0, inp);
15832         }
15833 #endif
15834         if (lgb) {
15835                 lgb->tlb_errno = error;
15836                 lgb = NULL;
15837         }
15838         if (error) {
15839                 *send_err = error;
15840                 m = NULL;
15841                 goto failed;
15842         }
15843         rack_log_output(tp, &to, len, tp->snd_max, flags, error, rack_to_usec_ts(tv),
15844                         NULL, add_flag, s_mb, s_soff);
15845         m = NULL;
15846         if (tp->snd_una == tp->snd_max) {
15847                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
15848                 rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
15849                 tp->t_acktime = ticks;
15850         }
15851         rack->forced_ack = 0;   /* If we send something zap the FA flag */
15852         tot_len += len;
15853         if ((tp->t_flags & TF_GPUTINPROG) == 0)
15854                 rack_start_gp_measurement(tp, rack, tp->snd_max, sb_offset);
15855         tp->snd_max += len;
15856         tp->snd_nxt = tp->snd_max;
15857         {
15858                 int idx;
15859
15860                 idx = (len / segsiz) + 3;
15861                 if (idx >= TCP_MSS_ACCT_ATIMER)
15862                         counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
15863                 else
15864                         counter_u64_add(rack_out_size[idx], 1);
15865         }
15866         if (len <= rack->r_ctl.fsb.left_to_send)
15867                 rack->r_ctl.fsb.left_to_send -= len;
15868         else
15869                 rack->r_ctl.fsb.left_to_send = 0;
15870         if (rack->r_ctl.fsb.left_to_send < segsiz) {
15871                 rack->r_fast_output = 0;
15872                 rack->r_ctl.fsb.left_to_send = 0;
15873                 /* At the end of fast_output scale up the sb */
15874                 SOCKBUF_LOCK(&rack->rc_inp->inp_socket->so_snd);
15875                 rack_sndbuf_autoscale(rack);
15876                 SOCKBUF_UNLOCK(&rack->rc_inp->inp_socket->so_snd);
15877         }
15878         if (tp->t_rtttime == 0) {
15879                 tp->t_rtttime = ticks;
15880                 tp->t_rtseq = startseq;
15881                 KMOD_TCPSTAT_INC(tcps_segstimed);
15882         }
15883         if ((rack->r_ctl.fsb.left_to_send >= segsiz) &&
15884             (max_val > len) &&
15885             (tso == 0)) {
15886                 max_val -= len;
15887                 len = segsiz;
15888                 th = rack->r_ctl.fsb.th;
15889                 cnt_thru++;
15890                 goto again;
15891         }
15892         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
15893         counter_u64_add(rack_fto_send, 1);
15894         slot = rack_get_pacing_delay(rack, tp, tot_len, NULL, segsiz);
15895         rack_start_hpts_timer(rack, tp, cts, slot, tot_len, 0);
15896 #ifdef TCP_ACCOUNTING
15897         crtsc = get_cyclecount();
15898         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15899                 tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
15900         }
15901         counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], cnt_thru);
15902         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15903                 tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
15904         }
15905         counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
15906         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15907                 tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len + segsiz - 1) / segsiz);
15908         }
15909         counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len + segsiz - 1) / segsiz));
15910         sched_unpin();
15911 #endif
15912         return (0);
15913 failed:
15914         if (m)
15915                 m_free(m);
15916         rack->r_fast_output = 0;
15917         return (-1);
15918 }
15919
15920 static int
15921 rack_output(struct tcpcb *tp)
15922 {
15923         struct socket *so;
15924         uint32_t recwin;
15925         uint32_t sb_offset, s_moff = 0;
15926         int32_t len, flags, error = 0;
15927         struct mbuf *m, *s_mb = NULL;
15928         struct mbuf *mb;
15929         uint32_t if_hw_tsomaxsegcount = 0;
15930         uint32_t if_hw_tsomaxsegsize;
15931         int32_t segsiz, minseg;
15932         long tot_len_this_send = 0;
15933 #ifdef INET
15934         struct ip *ip = NULL;
15935 #endif
15936 #ifdef TCPDEBUG
15937         struct ipovly *ipov = NULL;
15938 #endif
15939         struct udphdr *udp = NULL;
15940         struct tcp_rack *rack;
15941         struct tcphdr *th;
15942         uint8_t pass = 0;
15943         uint8_t mark = 0;
15944         uint8_t wanted_cookie = 0;
15945         u_char opt[TCP_MAXOLEN];
15946         unsigned ipoptlen, optlen, hdrlen, ulen=0;
15947         uint32_t rack_seq;
15948
15949 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
15950         unsigned ipsec_optlen = 0;
15951
15952 #endif
15953         int32_t idle, sendalot;
15954         int32_t sub_from_prr = 0;
15955         volatile int32_t sack_rxmit;
15956         struct rack_sendmap *rsm = NULL;
15957         int32_t tso, mtu;
15958         struct tcpopt to;
15959         int32_t slot = 0;
15960         int32_t sup_rack = 0;
15961         uint32_t cts, ms_cts, delayed, early;
15962         uint16_t add_flag = RACK_SENT_SP;
15963         uint8_t hpts_calling,  doing_tlp = 0;
15964         uint32_t cwnd_to_use, pace_max_seg;
15965         int32_t do_a_prefetch = 0;
15966         int32_t prefetch_rsm = 0;
15967         int32_t orig_len = 0;
15968         struct timeval tv;
15969         int32_t prefetch_so_done = 0;
15970         struct tcp_log_buffer *lgb;
15971         struct inpcb *inp;
15972         struct sockbuf *sb;
15973         uint64_t ts_val = 0;
15974 #ifdef TCP_ACCOUNTING
15975         uint64_t crtsc;
15976 #endif
15977 #ifdef INET6
15978         struct ip6_hdr *ip6 = NULL;
15979         int32_t isipv6;
15980 #endif
15981         uint8_t filled_all = 0;
15982         bool hw_tls = false;
15983
15984         /* setup and take the cache hits here */
15985         rack = (struct tcp_rack *)tp->t_fb_ptr;
15986 #ifdef TCP_ACCOUNTING
15987         sched_pin();
15988         ts_val = get_cyclecount();
15989 #endif
15990         hpts_calling = rack->rc_inp->inp_hpts_calls;
15991         NET_EPOCH_ASSERT();
15992         INP_WLOCK_ASSERT(rack->rc_inp);
15993 #ifdef TCP_OFFLOAD
15994         if (tp->t_flags & TF_TOE) {
15995 #ifdef TCP_ACCOUNTING
15996                 sched_unpin();
15997 #endif
15998                 return (tcp_offload_output(tp));
15999         }
16000 #endif
16001         /*
16002          * For TFO connections in SYN_RECEIVED, only allow the initial
16003          * SYN|ACK and those sent by the retransmit timer.
16004          */
16005         if (IS_FASTOPEN(tp->t_flags) &&
16006             (tp->t_state == TCPS_SYN_RECEIVED) &&
16007             SEQ_GT(tp->snd_max, tp->snd_una) &&    /* initial SYN|ACK sent */
16008             (rack->r_ctl.rc_resend == NULL)) {         /* not a retransmit */
16009 #ifdef TCP_ACCOUNTING
16010                 sched_unpin();
16011 #endif
16012                 return (0);
16013         }
16014 #ifdef INET6
16015         if (rack->r_state) {
16016                 /* Use the cache line loaded if possible */
16017                 isipv6 = rack->r_is_v6;
16018         } else {
16019                 isipv6 = (rack->rc_inp->inp_vflag & INP_IPV6) != 0;
16020         }
16021 #endif
16022         early = 0;
16023         cts = tcp_get_usecs(&tv);
16024         ms_cts = tcp_tv_to_mssectick(&tv);
16025         if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
16026             rack->rc_inp->inp_in_hpts) {
16027                 /*
16028                  * We are on the hpts for some timer but not hptsi output.
16029                  * Remove from the hpts unconditionally.
16030                  */
16031                 rack_timer_cancel(tp, rack, cts, __LINE__);
16032         }
16033         /* Are we pacing and late? */
16034         if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16035             TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to)) {
16036                 /* We are delayed */
16037                 delayed = cts - rack->r_ctl.rc_last_output_to;
16038         } else {
16039                 delayed = 0;
16040         }
16041         /* Do the timers, which may override the pacer */
16042         if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
16043                 if (rack_process_timers(tp, rack, cts, hpts_calling)) {
16044                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1);
16045 #ifdef TCP_ACCOUNTING
16046                         sched_unpin();
16047 #endif
16048                         return (0);
16049                 }
16050         }
16051         if (rack->rc_in_persist) {
16052                 if (rack->rc_inp->inp_in_hpts == 0) {
16053                         /* Timer is not running */
16054                         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
16055                 }
16056 #ifdef TCP_ACCOUNTING
16057                 sched_unpin();
16058 #endif
16059                 return (0);
16060         }
16061         if ((rack->r_timer_override) ||
16062             (rack->rc_ack_can_sendout_data) ||
16063             (delayed) ||
16064             (tp->t_state < TCPS_ESTABLISHED)) {
16065                 rack->rc_ack_can_sendout_data = 0;
16066                 if (rack->rc_inp->inp_in_hpts)
16067                         tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
16068         } else if (rack->rc_inp->inp_in_hpts) {
16069                 /*
16070                  * On the hpts you can't pass even if ACKNOW is on, we will
16071                  * when the hpts fires.
16072                  */
16073 #ifdef TCP_ACCOUNTING
16074                 crtsc = get_cyclecount();
16075                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16076                         tp->tcp_proc_time[SND_BLOCKED] += (crtsc - ts_val);
16077                 }
16078                 counter_u64_add(tcp_proc_time[SND_BLOCKED], (crtsc - ts_val));
16079                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16080                         tp->tcp_cnt_counters[SND_BLOCKED]++;
16081                 }
16082                 counter_u64_add(tcp_cnt_counters[SND_BLOCKED], 1);
16083                 sched_unpin();
16084 #endif
16085                 counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1);
16086                 return (0);
16087         }
16088         rack->rc_inp->inp_hpts_calls = 0;
16089         /* Finish out both pacing early and late accounting */
16090         if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16091             TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
16092                 early = rack->r_ctl.rc_last_output_to - cts;
16093         } else
16094                 early = 0;
16095         if (delayed) {
16096                 rack->r_ctl.rc_agg_delayed += delayed;
16097                 rack->r_late = 1;
16098         } else if (early) {
16099                 rack->r_ctl.rc_agg_early += early;
16100                 rack->r_early = 1;
16101         }
16102         /* Now that early/late accounting is done turn off the flag */
16103         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
16104         rack->r_wanted_output = 0;
16105         rack->r_timer_override = 0;
16106         if ((tp->t_state != rack->r_state) &&
16107             TCPS_HAVEESTABLISHED(tp->t_state)) {
16108                 rack_set_state(tp, rack);
16109         }
16110         if ((rack->r_fast_output) &&
16111             (tp->rcv_numsacks == 0)) {
16112                 int ret;
16113
16114                 error = 0;
16115                 ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
16116                 if (ret >= 0)
16117                         return(ret);
16118                 else if (error) {
16119                         inp = rack->rc_inp;
16120                         so = inp->inp_socket;
16121                         sb = &so->so_snd;
16122                         goto nomore;
16123                 }
16124         }
16125         inp = rack->rc_inp;
16126         /*
16127          * For TFO connections in SYN_SENT or SYN_RECEIVED,
16128          * only allow the initial SYN or SYN|ACK and those sent
16129          * by the retransmit timer.
16130          */
16131         if (IS_FASTOPEN(tp->t_flags) &&
16132             ((tp->t_state == TCPS_SYN_RECEIVED) ||
16133              (tp->t_state == TCPS_SYN_SENT)) &&
16134             SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
16135             (tp->t_rxtshift == 0)) {              /* not a retransmit */
16136                 cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
16137                 so = inp->inp_socket;
16138                 sb = &so->so_snd;
16139                 goto just_return_nolock;
16140         }
16141         /*
16142          * Determine length of data that should be transmitted, and flags
16143          * that will be used. If there is some data or critical controls
16144          * (SYN, RST) to send, then transmit; otherwise, investigate
16145          * further.
16146          */
16147         idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
16148         if (tp->t_idle_reduce) {
16149                 if (idle && ((ticks - tp->t_rcvtime) >= tp->t_rxtcur))
16150                         rack_cc_after_idle(rack, tp);
16151         }
16152         tp->t_flags &= ~TF_LASTIDLE;
16153         if (idle) {
16154                 if (tp->t_flags & TF_MORETOCOME) {
16155                         tp->t_flags |= TF_LASTIDLE;
16156                         idle = 0;
16157                 }
16158         }
16159         if ((tp->snd_una == tp->snd_max) &&
16160             rack->r_ctl.rc_went_idle_time &&
16161             TSTMP_GT(cts, rack->r_ctl.rc_went_idle_time)) {
16162                 idle = cts - rack->r_ctl.rc_went_idle_time;
16163                 if (idle > rack_min_probertt_hold) {
16164                         /* Count as a probe rtt */
16165                         if (rack->in_probe_rtt == 0) {
16166                                 rack->r_ctl.rc_lower_rtt_us_cts = cts;
16167                                 rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
16168                                 rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
16169                                 rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
16170                         } else {
16171                                 rack_exit_probertt(rack, cts);
16172                         }
16173                 }
16174                 idle = 0;
16175         }
16176         if (rack_use_fsb && (rack->r_fsb_inited == 0))
16177                 rack_init_fsb_block(tp, rack);
16178 again:
16179         /*
16180          * If we've recently taken a timeout, snd_max will be greater than
16181          * snd_nxt.  There may be SACK information that allows us to avoid
16182          * resending already delivered data.  Adjust snd_nxt accordingly.
16183          */
16184         sendalot = 0;
16185         cts = tcp_get_usecs(&tv);
16186         ms_cts = tcp_tv_to_mssectick(&tv);
16187         tso = 0;
16188         mtu = 0;
16189         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
16190         minseg = segsiz;
16191         if (rack->r_ctl.rc_pace_max_segs == 0)
16192                 pace_max_seg = rack->rc_user_set_max_segs * segsiz;
16193         else
16194                 pace_max_seg = rack->r_ctl.rc_pace_max_segs;
16195         sb_offset = tp->snd_max - tp->snd_una;
16196         cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
16197         flags = tcp_outflags[tp->t_state];
16198         while (rack->rc_free_cnt < rack_free_cache) {
16199                 rsm = rack_alloc(rack);
16200                 if (rsm == NULL) {
16201                         if (inp->inp_hpts_calls)
16202                                 /* Retry in a ms */
16203                                 slot = (1 * HPTS_USEC_IN_MSEC);
16204                         so = inp->inp_socket;
16205                         sb = &so->so_snd;
16206                         goto just_return_nolock;
16207                 }
16208                 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_tnext);
16209                 rack->rc_free_cnt++;
16210                 rsm = NULL;
16211         }
16212         if (inp->inp_hpts_calls)
16213                 inp->inp_hpts_calls = 0;
16214         sack_rxmit = 0;
16215         len = 0;
16216         rsm = NULL;
16217         if (flags & TH_RST) {
16218                 SOCKBUF_LOCK(&inp->inp_socket->so_snd);
16219                 so = inp->inp_socket;
16220                 sb = &so->so_snd;
16221                 goto send;
16222         }
16223         if (rack->r_ctl.rc_resend) {
16224                 /* Retransmit timer */
16225                 rsm = rack->r_ctl.rc_resend;
16226                 rack->r_ctl.rc_resend = NULL;
16227                 rsm->r_flags &= ~RACK_TLP;
16228                 len = rsm->r_end - rsm->r_start;
16229                 sack_rxmit = 1;
16230                 sendalot = 0;
16231                 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16232                         ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16233                          __func__, __LINE__,
16234                          rsm->r_start, tp->snd_una, tp, rack, rsm));
16235                 sb_offset = rsm->r_start - tp->snd_una;
16236                 if (len >= segsiz)
16237                         len = segsiz;
16238         } else if ((rsm = tcp_rack_output(tp, rack, cts)) != NULL) {
16239                 /* We have a retransmit that takes precedence */
16240                 rsm->r_flags &= ~RACK_TLP;
16241                 if ((!IN_FASTRECOVERY(tp->t_flags)) &&
16242                     ((tp->t_flags & TF_WASFRECOVERY) == 0)) {
16243                         /* Enter recovery if not induced by a time-out */
16244                         rack->r_ctl.rc_rsm_start = rsm->r_start;
16245                         rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
16246                         rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
16247                         rack_cong_signal(tp, CC_NDUPACK, tp->snd_una);
16248                 }
16249 #ifdef INVARIANTS
16250                 if (SEQ_LT(rsm->r_start, tp->snd_una)) {
16251                         panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n",
16252                               tp, rack, rsm, rsm->r_start, tp->snd_una);
16253                 }
16254 #endif
16255                 len = rsm->r_end - rsm->r_start;
16256                 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16257                         ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16258                          __func__, __LINE__,
16259                          rsm->r_start, tp->snd_una, tp, rack, rsm));
16260                 sb_offset = rsm->r_start - tp->snd_una;
16261                 sendalot = 0;
16262                 if (len >= segsiz)
16263                         len = segsiz;
16264                 if (len > 0) {
16265                         sack_rxmit = 1;
16266                         KMOD_TCPSTAT_INC(tcps_sack_rexmits);
16267                         KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes,
16268                             min(len, segsiz));
16269                         counter_u64_add(rack_rtm_prr_retran, 1);
16270                 }
16271         } else if (rack->r_ctl.rc_tlpsend) {
16272                 /* Tail loss probe */
16273                 long cwin;
16274                 long tlen;
16275
16276                 doing_tlp = 1;
16277                 /*
16278                  * Check if we can do a TLP with a RACK'd packet
16279                  * this can happen if we are not doing the rack
16280                  * cheat and we skipped to a TLP and it
16281                  * went off.
16282                  */
16283                 rsm = rack->r_ctl.rc_tlpsend;
16284                 rsm->r_flags |= RACK_TLP;
16285
16286                 rack->r_ctl.rc_tlpsend = NULL;
16287                 sack_rxmit = 1;
16288                 tlen = rsm->r_end - rsm->r_start;
16289                 if (tlen > segsiz)
16290                         tlen = segsiz;
16291                 tp->t_sndtlppack++;
16292                 tp->t_sndtlpbyte += tlen;
16293                 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16294                         ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16295                          __func__, __LINE__,
16296                          rsm->r_start, tp->snd_una, tp, rack, rsm));
16297                 sb_offset = rsm->r_start - tp->snd_una;
16298                 cwin = min(tp->snd_wnd, tlen);
16299                 len = cwin;
16300         }
16301         if (rack->r_must_retran &&
16302             (rsm == NULL)) {
16303                 /*
16304                  * Non-Sack and we had a RTO or MTU change, we
16305                  * need to retransmit until we reach
16306                  * the former snd_max (rack->r_ctl.rc_snd_max_at_rto).
16307                  */
16308                 if (SEQ_GT(tp->snd_max, tp->snd_una)) {
16309                         int sendwin, flight;
16310
16311                         sendwin = min(tp->snd_wnd, tp->snd_cwnd);
16312                         flight = ctf_flight_size(tp, rack->r_ctl.rc_out_at_rto);
16313                         if (flight >= sendwin) {
16314                                 so = inp->inp_socket;
16315                                 sb = &so->so_snd;
16316                                 goto just_return_nolock;
16317                         }
16318                         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
16319                         KASSERT(rsm != NULL, ("rsm is NULL rack:%p r_must_retran set", rack));
16320                         if (rsm == NULL) {
16321                                 /* TSNH */
16322                                 rack->r_must_retran = 0;
16323                                 rack->r_ctl.rc_out_at_rto = 0;
16324                                 rack->r_must_retran = 0;
16325                                 so = inp->inp_socket;
16326                                 sb = &so->so_snd;
16327                                 goto just_return_nolock;
16328                         }
16329                         sack_rxmit = 1;
16330                         len = rsm->r_end - rsm->r_start;
16331                         sendalot = 0;
16332                         sb_offset = rsm->r_start - tp->snd_una;
16333                         if (len >= segsiz)
16334                                 len = segsiz;
16335                 } else {
16336                         /* We must be done if there is nothing outstanding */
16337                         rack->r_must_retran = 0;
16338                         rack->r_ctl.rc_out_at_rto = 0;
16339                 }
16340         }
16341         /*
16342          * Enforce a connection sendmap count limit if set
16343          * as long as we are not retransmiting.
16344          */
16345         if ((rsm == NULL) &&
16346             (rack->do_detection == 0) &&
16347             (V_tcp_map_entries_limit > 0) &&
16348             (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
16349                 counter_u64_add(rack_to_alloc_limited, 1);
16350                 if (!rack->alloc_limit_reported) {
16351                         rack->alloc_limit_reported = 1;
16352                         counter_u64_add(rack_alloc_limited_conns, 1);
16353                 }
16354                 so = inp->inp_socket;
16355                 sb = &so->so_snd;
16356                 goto just_return_nolock;
16357         }
16358         if (rsm && (rsm->r_flags & RACK_HAS_FIN)) {
16359                 /* we are retransmitting the fin */
16360                 len--;
16361                 if (len) {
16362                         /*
16363                          * When retransmitting data do *not* include the
16364                          * FIN. This could happen from a TLP probe.
16365                          */
16366                         flags &= ~TH_FIN;
16367                 }
16368         }
16369 #ifdef INVARIANTS
16370         /* For debugging */
16371         rack->r_ctl.rc_rsm_at_retran = rsm;
16372 #endif
16373         if (rsm && rack->r_fsb_inited && rack_use_rsm_rfo &&
16374             ((rsm->r_flags & RACK_HAS_FIN) == 0)) {
16375                 int ret;
16376
16377                 ret = rack_fast_rsm_output(tp, rack, rsm, ts_val, cts, ms_cts, &tv, len);
16378                 if (ret == 0)
16379                         return (0);
16380         }
16381         so = inp->inp_socket;
16382         sb = &so->so_snd;
16383         if (do_a_prefetch == 0) {
16384                 kern_prefetch(sb, &do_a_prefetch);
16385                 do_a_prefetch = 1;
16386         }
16387 #ifdef NETFLIX_SHARED_CWND
16388         if ((tp->t_flags2 & TF2_TCP_SCWND_ALLOWED) &&
16389             rack->rack_enable_scwnd) {
16390                 /* We are doing cwnd sharing */
16391                 if (rack->gp_ready &&
16392                     (rack->rack_attempted_scwnd == 0) &&
16393                     (rack->r_ctl.rc_scw == NULL) &&
16394                     tp->t_lib) {
16395                         /* The pcbid is in, lets make an attempt */
16396                         counter_u64_add(rack_try_scwnd, 1);
16397                         rack->rack_attempted_scwnd = 1;
16398                         rack->r_ctl.rc_scw = tcp_shared_cwnd_alloc(tp,
16399                                                                    &rack->r_ctl.rc_scw_index,
16400                                                                    segsiz);
16401                 }
16402                 if (rack->r_ctl.rc_scw &&
16403                     (rack->rack_scwnd_is_idle == 1) &&
16404                     sbavail(&so->so_snd)) {
16405                         /* we are no longer out of data */
16406                         tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
16407                         rack->rack_scwnd_is_idle = 0;
16408                 }
16409                 if (rack->r_ctl.rc_scw) {
16410                         /* First lets update and get the cwnd */
16411                         rack->r_ctl.cwnd_to_use = cwnd_to_use = tcp_shared_cwnd_update(rack->r_ctl.rc_scw,
16412                                                                     rack->r_ctl.rc_scw_index,
16413                                                                     tp->snd_cwnd, tp->snd_wnd, segsiz);
16414                 }
16415         }
16416 #endif
16417         /*
16418          * Get standard flags, and add SYN or FIN if requested by 'hidden'
16419          * state flags.
16420          */
16421         if (tp->t_flags & TF_NEEDFIN)
16422                 flags |= TH_FIN;
16423         if (tp->t_flags & TF_NEEDSYN)
16424                 flags |= TH_SYN;
16425         if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
16426                 void *end_rsm;
16427                 end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
16428                 if (end_rsm)
16429                         kern_prefetch(end_rsm, &prefetch_rsm);
16430                 prefetch_rsm = 1;
16431         }
16432         SOCKBUF_LOCK(sb);
16433         /*
16434          * If snd_nxt == snd_max and we have transmitted a FIN, the
16435          * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a
16436          * negative length.  This can also occur when TCP opens up its
16437          * congestion window while receiving additional duplicate acks after
16438          * fast-retransmit because TCP will reset snd_nxt to snd_max after
16439          * the fast-retransmit.
16440          *
16441          * In the normal retransmit-FIN-only case, however, snd_nxt will be
16442          * set to snd_una, the sb_offset will be 0, and the length may wind
16443          * up 0.
16444          *
16445          * If sack_rxmit is true we are retransmitting from the scoreboard
16446          * in which case len is already set.
16447          */
16448         if ((sack_rxmit == 0) &&
16449             (TCPS_HAVEESTABLISHED(tp->t_state) || IS_FASTOPEN(tp->t_flags))) {
16450                 uint32_t avail;
16451
16452                 avail = sbavail(sb);
16453                 if (SEQ_GT(tp->snd_nxt, tp->snd_una) && avail)
16454                         sb_offset = tp->snd_nxt - tp->snd_una;
16455                 else
16456                         sb_offset = 0;
16457                 if ((IN_FASTRECOVERY(tp->t_flags) == 0) || rack->rack_no_prr) {
16458                         if (rack->r_ctl.rc_tlp_new_data) {
16459                                 /* TLP is forcing out new data */
16460                                 if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) {
16461                                         rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset);
16462                                 }
16463                                 if ((rack->r_ctl.rc_tlp_new_data + sb_offset) > tp->snd_wnd) {
16464                                         if (tp->snd_wnd > sb_offset)
16465                                                 len = tp->snd_wnd - sb_offset;
16466                                         else
16467                                                 len = 0;
16468                                 } else {
16469                                         len = rack->r_ctl.rc_tlp_new_data;
16470                                 }
16471                                 rack->r_ctl.rc_tlp_new_data = 0;
16472                                 doing_tlp = 1;
16473                         }  else {
16474                                 len = rack_what_can_we_send(tp, rack, cwnd_to_use, avail, sb_offset);
16475                         }
16476                         if ((rack->r_ctl.crte == NULL) && IN_FASTRECOVERY(tp->t_flags) && (len > segsiz)) {
16477                                 /*
16478                                  * For prr=off, we need to send only 1 MSS
16479                                  * at a time. We do this because another sack could
16480                                  * be arriving that causes us to send retransmits and
16481                                  * we don't want to be on a long pace due to a larger send
16482                                  * that keeps us from sending out the retransmit.
16483                                  */
16484                                 len = segsiz;
16485                         }
16486                 } else {
16487                         uint32_t outstanding;
16488                         /*
16489                          * We are inside of a Fast recovery episode, this
16490                          * is caused by a SACK or 3 dup acks. At this point
16491                          * we have sent all the retransmissions and we rely
16492                          * on PRR to dictate what we will send in the form of
16493                          * new data.
16494                          */
16495
16496                         outstanding = tp->snd_max - tp->snd_una;
16497                         if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) {
16498                                 if (tp->snd_wnd > outstanding) {
16499                                         len = tp->snd_wnd - outstanding;
16500                                         /* Check to see if we have the data */
16501                                         if ((sb_offset + len) > avail) {
16502                                                 /* It does not all fit */
16503                                                 if (avail > sb_offset)
16504                                                         len = avail - sb_offset;
16505                                                 else
16506                                                         len = 0;
16507                                         }
16508                                 } else {
16509                                         len = 0;
16510                                 }
16511                         } else if (avail > sb_offset) {
16512                                 len = avail - sb_offset;
16513                         } else {
16514                                 len = 0;
16515                         }
16516                         if (len > 0) {
16517                                 if (len > rack->r_ctl.rc_prr_sndcnt) {
16518                                         len = rack->r_ctl.rc_prr_sndcnt;
16519                                 }
16520                                 if (len > 0) {
16521                                         sub_from_prr = 1;
16522                                         counter_u64_add(rack_rtm_prr_newdata, 1);
16523                                 }
16524                         }
16525                         if (len > segsiz) {
16526                                 /*
16527                                  * We should never send more than a MSS when
16528                                  * retransmitting or sending new data in prr
16529                                  * mode unless the override flag is on. Most
16530                                  * likely the PRR algorithm is not going to
16531                                  * let us send a lot as well :-)
16532                                  */
16533                                 if (rack->r_ctl.rc_prr_sendalot == 0) {
16534                                         len = segsiz;
16535                                 }
16536                         } else if (len < segsiz) {
16537                                 /*
16538                                  * Do we send any? The idea here is if the
16539                                  * send empty's the socket buffer we want to
16540                                  * do it. However if not then lets just wait
16541                                  * for our prr_sndcnt to get bigger.
16542                                  */
16543                                 long leftinsb;
16544
16545                                 leftinsb = sbavail(sb) - sb_offset;
16546                                 if (leftinsb > len) {
16547                                         /* This send does not empty the sb */
16548                                         len = 0;
16549                                 }
16550                         }
16551                 }
16552         } else if (!TCPS_HAVEESTABLISHED(tp->t_state)) {
16553                 /*
16554                  * If you have not established
16555                  * and are not doing FAST OPEN
16556                  * no data please.
16557                  */
16558                 if ((sack_rxmit == 0) &&
16559                     (!IS_FASTOPEN(tp->t_flags))){
16560                         len = 0;
16561                         sb_offset = 0;
16562                 }
16563         }
16564         if (prefetch_so_done == 0) {
16565                 kern_prefetch(so, &prefetch_so_done);
16566                 prefetch_so_done = 1;
16567         }
16568         /*
16569          * Lop off SYN bit if it has already been sent.  However, if this is
16570          * SYN-SENT state and if segment contains data and if we don't know
16571          * that foreign host supports TAO, suppress sending segment.
16572          */
16573         if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una) &&
16574             ((sack_rxmit == 0) && (tp->t_rxtshift == 0))) {
16575                 /*
16576                  * When sending additional segments following a TFO SYN|ACK,
16577                  * do not include the SYN bit.
16578                  */
16579                 if (IS_FASTOPEN(tp->t_flags) &&
16580                     (tp->t_state == TCPS_SYN_RECEIVED))
16581                         flags &= ~TH_SYN;
16582         }
16583         /*
16584          * Be careful not to send data and/or FIN on SYN segments. This
16585          * measure is needed to prevent interoperability problems with not
16586          * fully conformant TCP implementations.
16587          */
16588         if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
16589                 len = 0;
16590                 flags &= ~TH_FIN;
16591         }
16592         /*
16593          * On TFO sockets, ensure no data is sent in the following cases:
16594          *
16595          *  - When retransmitting SYN|ACK on a passively-created socket
16596          *
16597          *  - When retransmitting SYN on an actively created socket
16598          *
16599          *  - When sending a zero-length cookie (cookie request) on an
16600          *    actively created socket
16601          *
16602          *  - When the socket is in the CLOSED state (RST is being sent)
16603          */
16604         if (IS_FASTOPEN(tp->t_flags) &&
16605             (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
16606              ((tp->t_state == TCPS_SYN_SENT) &&
16607               (tp->t_tfo_client_cookie_len == 0)) ||
16608              (flags & TH_RST))) {
16609                 sack_rxmit = 0;
16610                 len = 0;
16611         }
16612         /* Without fast-open there should never be data sent on a SYN */
16613         if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags))) {
16614                 tp->snd_nxt = tp->iss;
16615                 len = 0;
16616         }
16617         if ((len > segsiz) && (tcp_dsack_block_exists(tp))) {
16618                 /* We only send 1 MSS if we have a DSACK block */
16619                 add_flag |= RACK_SENT_W_DSACK;
16620                 len = segsiz;
16621         }
16622         orig_len = len;
16623         if (len <= 0) {
16624                 /*
16625                  * If FIN has been sent but not acked, but we haven't been
16626                  * called to retransmit, len will be < 0.  Otherwise, window
16627                  * shrank after we sent into it.  If window shrank to 0,
16628                  * cancel pending retransmit, pull snd_nxt back to (closed)
16629                  * window, and set the persist timer if it isn't already
16630                  * going.  If the window didn't close completely, just wait
16631                  * for an ACK.
16632                  *
16633                  * We also do a general check here to ensure that we will
16634                  * set the persist timer when we have data to send, but a
16635                  * 0-byte window. This makes sure the persist timer is set
16636                  * even if the packet hits one of the "goto send" lines
16637                  * below.
16638                  */
16639                 len = 0;
16640                 if ((tp->snd_wnd == 0) &&
16641                     (TCPS_HAVEESTABLISHED(tp->t_state)) &&
16642                     (tp->snd_una == tp->snd_max) &&
16643                     (sb_offset < (int)sbavail(sb))) {
16644                         rack_enter_persist(tp, rack, cts);
16645                 }
16646         } else if ((rsm == NULL) &&
16647                    (doing_tlp == 0) &&
16648                    (len < pace_max_seg)) {
16649                 /*
16650                  * We are not sending a maximum sized segment for
16651                  * some reason. Should we not send anything (think
16652                  * sws or persists)?
16653                  */
16654                 if ((tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
16655                     (TCPS_HAVEESTABLISHED(tp->t_state)) &&
16656                     (len < minseg) &&
16657                     (len < (int)(sbavail(sb) - sb_offset))) {
16658                         /*
16659                          * Here the rwnd is less than
16660                          * the minimum pacing size, this is not a retransmit,
16661                          * we are established and
16662                          * the send is not the last in the socket buffer
16663                          * we send nothing, and we may enter persists
16664                          * if nothing is outstanding.
16665                          */
16666                         len = 0;
16667                         if (tp->snd_max == tp->snd_una) {
16668                                 /*
16669                                  * Nothing out we can
16670                                  * go into persists.
16671                                  */
16672                                 rack_enter_persist(tp, rack, cts);
16673                         }
16674                      } else if ((cwnd_to_use >= max(minseg, (segsiz * 4))) &&
16675                            (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
16676                            (len < (int)(sbavail(sb) - sb_offset)) &&
16677                            (len < minseg)) {
16678                         /*
16679                          * Here we are not retransmitting, and
16680                          * the cwnd is not so small that we could
16681                          * not send at least a min size (rxt timer
16682                          * not having gone off), We have 2 segments or
16683                          * more already in flight, its not the tail end
16684                          * of the socket buffer  and the cwnd is blocking
16685                          * us from sending out a minimum pacing segment size.
16686                          * Lets not send anything.
16687                          */
16688                         len = 0;
16689                 } else if (((tp->snd_wnd - ctf_outstanding(tp)) <
16690                             min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
16691                            (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
16692                            (len < (int)(sbavail(sb) - sb_offset)) &&
16693                            (TCPS_HAVEESTABLISHED(tp->t_state))) {
16694                         /*
16695                          * Here we have a send window but we have
16696                          * filled it up and we can't send another pacing segment.
16697                          * We also have in flight more than 2 segments
16698                          * and we are not completing the sb i.e. we allow
16699                          * the last bytes of the sb to go out even if
16700                          * its not a full pacing segment.
16701                          */
16702                         len = 0;
16703                 } else if ((rack->r_ctl.crte != NULL) &&
16704                            (tp->snd_wnd >= (pace_max_seg * max(1, rack_hw_rwnd_factor))) &&
16705                            (cwnd_to_use >= (pace_max_seg + (4 * segsiz))) &&
16706                            (ctf_flight_size(tp, rack->r_ctl.rc_sacked) >= (2 * segsiz)) &&
16707                            (len < (int)(sbavail(sb) - sb_offset))) {
16708                         /*
16709                          * Here we are doing hardware pacing, this is not a TLP,
16710                          * we are not sending a pace max segment size, there is rwnd
16711                          * room to send at least N pace_max_seg, the cwnd is greater
16712                          * than or equal to a full pacing segments plus 4 mss and we have 2 or
16713                          * more segments in flight and its not the tail of the socket buffer.
16714                          *
16715                          * We don't want to send instead we need to get more ack's in to
16716                          * allow us to send a full pacing segment. Normally, if we are pacing
16717                          * about the right speed, we should have finished our pacing
16718                          * send as most of the acks have come back if we are at the
16719                          * right rate. This is a bit fuzzy since return path delay
16720                          * can delay the acks, which is why we want to make sure we
16721                          * have cwnd space to have a bit more than a max pace segments in flight.
16722                          *
16723                          * If we have not gotten our acks back we are pacing at too high a
16724                          * rate delaying will not hurt and will bring our GP estimate down by
16725                          * injecting the delay. If we don't do this we will send
16726                          * 2 MSS out in response to the acks being clocked in which
16727                          * defeats the point of hw-pacing (i.e. to help us get
16728                          * larger TSO's out).
16729                          */
16730                         len = 0;
16731
16732                 }
16733
16734         }
16735         /* len will be >= 0 after this point. */
16736         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
16737         rack_sndbuf_autoscale(rack);
16738         /*
16739          * Decide if we can use TCP Segmentation Offloading (if supported by
16740          * hardware).
16741          *
16742          * TSO may only be used if we are in a pure bulk sending state.  The
16743          * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
16744          * options prevent using TSO.  With TSO the TCP header is the same
16745          * (except for the sequence number) for all generated packets.  This
16746          * makes it impossible to transmit any options which vary per
16747          * generated segment or packet.
16748          *
16749          * IPv4 handling has a clear separation of ip options and ip header
16750          * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
16751          * the right thing below to provide length of just ip options and thus
16752          * checking for ipoptlen is enough to decide if ip options are present.
16753          */
16754         ipoptlen = 0;
16755 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
16756         /*
16757          * Pre-calculate here as we save another lookup into the darknesses
16758          * of IPsec that way and can actually decide if TSO is ok.
16759          */
16760 #ifdef INET6
16761         if (isipv6 && IPSEC_ENABLED(ipv6))
16762                 ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb);
16763 #ifdef INET
16764         else
16765 #endif
16766 #endif                          /* INET6 */
16767 #ifdef INET
16768                 if (IPSEC_ENABLED(ipv4))
16769                         ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb);
16770 #endif                          /* INET */
16771 #endif
16772
16773 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
16774         ipoptlen += ipsec_optlen;
16775 #endif
16776         if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > segsiz &&
16777             (tp->t_port == 0) &&
16778             ((tp->t_flags & TF_SIGNATURE) == 0) &&
16779             tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
16780             ipoptlen == 0)
16781                 tso = 1;
16782         {
16783                 uint32_t outstanding;
16784
16785                 outstanding = tp->snd_max - tp->snd_una;
16786                 if (tp->t_flags & TF_SENTFIN) {
16787                         /*
16788                          * If we sent a fin, snd_max is 1 higher than
16789                          * snd_una
16790                          */
16791                         outstanding--;
16792                 }
16793                 if (sack_rxmit) {
16794                         if ((rsm->r_flags & RACK_HAS_FIN) == 0)
16795                                 flags &= ~TH_FIN;
16796                 } else {
16797                         if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
16798                                    sbused(sb)))
16799                                 flags &= ~TH_FIN;
16800                 }
16801         }
16802         recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
16803             (long)TCP_MAXWIN << tp->rcv_scale);
16804
16805         /*
16806          * Sender silly window avoidance.   We transmit under the following
16807          * conditions when len is non-zero:
16808          *
16809          * - We have a full segment (or more with TSO) - This is the last
16810          * buffer in a write()/send() and we are either idle or running
16811          * NODELAY - we've timed out (e.g. persist timer) - we have more
16812          * then 1/2 the maximum send window's worth of data (receiver may be
16813          * limited the window size) - we need to retransmit
16814          */
16815         if (len) {
16816                 if (len >= segsiz) {
16817                         goto send;
16818                 }
16819                 /*
16820                  * NOTE! on localhost connections an 'ack' from the remote
16821                  * end may occur synchronously with the output and cause us
16822                  * to flush a buffer queued with moretocome.  XXX
16823                  *
16824                  */
16825                 if (!(tp->t_flags & TF_MORETOCOME) &&   /* normal case */
16826                     (idle || (tp->t_flags & TF_NODELAY)) &&
16827                     ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
16828                     (tp->t_flags & TF_NOPUSH) == 0) {
16829                         pass = 2;
16830                         goto send;
16831                 }
16832                 if ((tp->snd_una == tp->snd_max) && len) {      /* Nothing outstanding */
16833                         pass = 22;
16834                         goto send;
16835                 }
16836                 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) {
16837                         pass = 4;
16838                         goto send;
16839                 }
16840                 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { /* retransmit case */
16841                         pass = 5;
16842                         goto send;
16843                 }
16844                 if (sack_rxmit) {
16845                         pass = 6;
16846                         goto send;
16847                 }
16848                 if (((tp->snd_wnd - ctf_outstanding(tp)) < segsiz) &&
16849                     (ctf_outstanding(tp) < (segsiz * 2))) {
16850                         /*
16851                          * We have less than two MSS outstanding (delayed ack)
16852                          * and our rwnd will not let us send a full sized
16853                          * MSS. Lets go ahead and let this small segment
16854                          * out because we want to try to have at least two
16855                          * packets inflight to not be caught by delayed ack.
16856                          */
16857                         pass = 12;
16858                         goto send;
16859                 }
16860         }
16861         /*
16862          * Sending of standalone window updates.
16863          *
16864          * Window updates are important when we close our window due to a
16865          * full socket buffer and are opening it again after the application
16866          * reads data from it.  Once the window has opened again and the
16867          * remote end starts to send again the ACK clock takes over and
16868          * provides the most current window information.
16869          *
16870          * We must avoid the silly window syndrome whereas every read from
16871          * the receive buffer, no matter how small, causes a window update
16872          * to be sent.  We also should avoid sending a flurry of window
16873          * updates when the socket buffer had queued a lot of data and the
16874          * application is doing small reads.
16875          *
16876          * Prevent a flurry of pointless window updates by only sending an
16877          * update when we can increase the advertized window by more than
16878          * 1/4th of the socket buffer capacity.  When the buffer is getting
16879          * full or is very small be more aggressive and send an update
16880          * whenever we can increase by two mss sized segments. In all other
16881          * situations the ACK's to new incoming data will carry further
16882          * window increases.
16883          *
16884          * Don't send an independent window update if a delayed ACK is
16885          * pending (it will get piggy-backed on it) or the remote side
16886          * already has done a half-close and won't send more data.  Skip
16887          * this if the connection is in T/TCP half-open state.
16888          */
16889         if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
16890             !(tp->t_flags & TF_DELACK) &&
16891             !TCPS_HAVERCVDFIN(tp->t_state)) {
16892                 /*
16893                  * "adv" is the amount we could increase the window, taking
16894                  * into account that we are limited by TCP_MAXWIN <<
16895                  * tp->rcv_scale.
16896                  */
16897                 int32_t adv;
16898                 int oldwin;
16899
16900                 adv = recwin;
16901                 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
16902                         oldwin = (tp->rcv_adv - tp->rcv_nxt);
16903                         if (adv > oldwin)
16904                             adv -= oldwin;
16905                         else {
16906                                 /* We can't increase the window */
16907                                 adv = 0;
16908                         }
16909                 } else
16910                         oldwin = 0;
16911
16912                 /*
16913                  * If the new window size ends up being the same as or less
16914                  * than the old size when it is scaled, then don't force
16915                  * a window update.
16916                  */
16917                 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
16918                         goto dontupdate;
16919
16920                 if (adv >= (int32_t)(2 * segsiz) &&
16921                     (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
16922                      recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) ||
16923                      so->so_rcv.sb_hiwat <= 8 * segsiz)) {
16924                         pass = 7;
16925                         goto send;
16926                 }
16927                 if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) {
16928                         pass = 23;
16929                         goto send;
16930                 }
16931         }
16932 dontupdate:
16933
16934         /*
16935          * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
16936          * is also a catch-all for the retransmit timer timeout case.
16937          */
16938         if (tp->t_flags & TF_ACKNOW) {
16939                 pass = 8;
16940                 goto send;
16941         }
16942         if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) {
16943                 pass = 9;
16944                 goto send;
16945         }
16946         /*
16947          * If our state indicates that FIN should be sent and we have not
16948          * yet done so, then we need to send.
16949          */
16950         if ((flags & TH_FIN) &&
16951             (tp->snd_nxt == tp->snd_una)) {
16952                 pass = 11;
16953                 goto send;
16954         }
16955         /*
16956          * No reason to send a segment, just return.
16957          */
16958 just_return:
16959         SOCKBUF_UNLOCK(sb);
16960 just_return_nolock:
16961         {
16962                 int app_limited = CTF_JR_SENT_DATA;
16963
16964                 if (tot_len_this_send > 0) {
16965                         /* Make sure snd_nxt is up to max */
16966                         rack->r_ctl.fsb.recwin = recwin;
16967                         slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, NULL, segsiz);
16968                         if ((error == 0) &&
16969                             rack_use_rfo &&
16970                             ((flags & (TH_SYN|TH_FIN)) == 0) &&
16971                             (ipoptlen == 0) &&
16972                             (tp->snd_nxt == tp->snd_max) &&
16973                             (tp->rcv_numsacks == 0) &&
16974                             rack->r_fsb_inited &&
16975                             TCPS_HAVEESTABLISHED(tp->t_state) &&
16976                             (rack->r_must_retran == 0) &&
16977                             ((tp->t_flags & TF_NEEDFIN) == 0) &&
16978                             (len > 0) && (orig_len > 0) &&
16979                             (orig_len > len) &&
16980                             ((orig_len - len) >= segsiz) &&
16981                             ((optlen == 0) ||
16982                              ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
16983                                 /* We can send at least one more MSS using our fsb */
16984
16985                                 rack->r_fast_output = 1;
16986                                 rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
16987                                 rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
16988                                 rack->r_ctl.fsb.tcp_flags = flags;
16989                                 rack->r_ctl.fsb.left_to_send = orig_len - len;
16990                                 KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
16991                                         ("rack:%p left_to_send:%u sbavail:%u out:%u",
16992                                         rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
16993                                          (tp->snd_max - tp->snd_una)));
16994                                 if (rack->r_ctl.fsb.left_to_send < segsiz)
16995                                         rack->r_fast_output = 0;
16996                                 else {
16997                                         if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
16998                                                 rack->r_ctl.fsb.rfo_apply_push = 1;
16999                                         else
17000                                                 rack->r_ctl.fsb.rfo_apply_push = 0;
17001                                 }
17002                         } else
17003                                 rack->r_fast_output = 0;
17004
17005
17006                         rack_log_fsb(rack, tp, so, flags,
17007                                      ipoptlen, orig_len, len, 0,
17008                                      1, optlen, __LINE__, 1);
17009                         if (SEQ_GT(tp->snd_max, tp->snd_nxt))
17010                                 tp->snd_nxt = tp->snd_max;
17011                 } else {
17012                         int end_window = 0;
17013                         uint32_t seq = tp->gput_ack;
17014
17015                         rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
17016                         if (rsm) {
17017                                 /*
17018                                  * Mark the last sent that we just-returned (hinting
17019                                  * that delayed ack may play a role in any rtt measurement).
17020                                  */
17021                                 rsm->r_just_ret = 1;
17022                         }
17023                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1);
17024                         rack->r_ctl.rc_agg_delayed = 0;
17025                         rack->r_early = 0;
17026                         rack->r_late = 0;
17027                         rack->r_ctl.rc_agg_early = 0;
17028                         if ((ctf_outstanding(tp) +
17029                              min(max(segsiz, (rack->r_ctl.rc_high_rwnd/2)),
17030                                  minseg)) >= tp->snd_wnd) {
17031                                 /* We are limited by the rwnd */
17032                                 app_limited = CTF_JR_RWND_LIMITED;
17033                                 if (IN_FASTRECOVERY(tp->t_flags))
17034                                     rack->r_ctl.rc_prr_sndcnt = 0;
17035                         } else if (ctf_outstanding(tp) >= sbavail(sb)) {
17036                                 /* We are limited by whats available -- app limited */
17037                                 app_limited = CTF_JR_APP_LIMITED;
17038                                 if (IN_FASTRECOVERY(tp->t_flags))
17039                                     rack->r_ctl.rc_prr_sndcnt = 0;
17040                         } else if ((idle == 0) &&
17041                                    ((tp->t_flags & TF_NODELAY) == 0) &&
17042                                    ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
17043                                    (len < segsiz)) {
17044                                 /*
17045                                  * No delay is not on and the
17046                                  * user is sending less than 1MSS. This
17047                                  * brings out SWS avoidance so we
17048                                  * don't send. Another app-limited case.
17049                                  */
17050                                 app_limited = CTF_JR_APP_LIMITED;
17051                         } else if (tp->t_flags & TF_NOPUSH) {
17052                                 /*
17053                                  * The user has requested no push of
17054                                  * the last segment and we are
17055                                  * at the last segment. Another app
17056                                  * limited case.
17057                                  */
17058                                 app_limited = CTF_JR_APP_LIMITED;
17059                         } else if ((ctf_outstanding(tp) + minseg) > cwnd_to_use) {
17060                                 /* Its the cwnd */
17061                                 app_limited = CTF_JR_CWND_LIMITED;
17062                         } else if (IN_FASTRECOVERY(tp->t_flags) &&
17063                                    (rack->rack_no_prr == 0) &&
17064                                    (rack->r_ctl.rc_prr_sndcnt < segsiz)) {
17065                                 app_limited = CTF_JR_PRR;
17066                         } else {
17067                                 /* Now why here are we not sending? */
17068 #ifdef NOW
17069 #ifdef INVARIANTS
17070                                 panic("rack:%p hit JR_ASSESSING case cwnd_to_use:%u?", rack, cwnd_to_use);
17071 #endif
17072 #endif
17073                                 app_limited = CTF_JR_ASSESSING;
17074                         }
17075                         /*
17076                          * App limited in some fashion, for our pacing GP
17077                          * measurements we don't want any gap (even cwnd).
17078                          * Close  down the measurement window.
17079                          */
17080                         if (rack_cwnd_block_ends_measure &&
17081                             ((app_limited == CTF_JR_CWND_LIMITED) ||
17082                              (app_limited == CTF_JR_PRR))) {
17083                                 /*
17084                                  * The reason we are not sending is
17085                                  * the cwnd (or prr). We have been configured
17086                                  * to end the measurement window in
17087                                  * this case.
17088                                  */
17089                                 end_window = 1;
17090                         } else if (rack_rwnd_block_ends_measure &&
17091                                    (app_limited == CTF_JR_RWND_LIMITED)) {
17092                                 /*
17093                                  * We are rwnd limited and have been
17094                                  * configured to end the measurement
17095                                  * window in this case.
17096                                  */
17097                                 end_window = 1;
17098                         } else if (app_limited == CTF_JR_APP_LIMITED) {
17099                                 /*
17100                                  * A true application limited period, we have
17101                                  * ran out of data.
17102                                  */
17103                                 end_window = 1;
17104                         } else if (app_limited == CTF_JR_ASSESSING) {
17105                                 /*
17106                                  * In the assessing case we hit the end of
17107                                  * the if/else and had no known reason
17108                                  * This will panic us under invariants..
17109                                  *
17110                                  * If we get this out in logs we need to
17111                                  * investagate which reason we missed.
17112                                  */
17113                                 end_window = 1;
17114                         }
17115                         if (end_window) {
17116                                 uint8_t log = 0;
17117
17118                                 if ((tp->t_flags & TF_GPUTINPROG) &&
17119                                     SEQ_GT(tp->gput_ack, tp->snd_max)) {
17120                                         /* Mark the last packet has app limited */
17121                                         tp->gput_ack = tp->snd_max;
17122                                         log = 1;
17123                                 }
17124                                 rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
17125                                 if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
17126                                         if (rack->r_ctl.rc_app_limited_cnt == 0)
17127                                                 rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
17128                                         else {
17129                                                 /*
17130                                                  * Go out to the end app limited and mark
17131                                                  * this new one as next and move the end_appl up
17132                                                  * to this guy.
17133                                                  */
17134                                                 if (rack->r_ctl.rc_end_appl)
17135                                                         rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
17136                                                 rack->r_ctl.rc_end_appl = rsm;
17137                                         }
17138                                         rsm->r_flags |= RACK_APP_LIMITED;
17139                                         rack->r_ctl.rc_app_limited_cnt++;
17140                                 }
17141                                 if (log)
17142                                         rack_log_pacing_delay_calc(rack,
17143                                                                    rack->r_ctl.rc_app_limited_cnt, seq,
17144                                                                    tp->gput_ack, 0, 0, 4, __LINE__, NULL);
17145                         }
17146                 }
17147                 if (slot) {
17148                         /* set the rack tcb into the slot N */
17149                         counter_u64_add(rack_paced_segments, 1);
17150                 } else if (tot_len_this_send) {
17151                         counter_u64_add(rack_unpaced_segments, 1);
17152                 }
17153                 /* Check if we need to go into persists or not */
17154                 if ((tp->snd_max == tp->snd_una) &&
17155                     TCPS_HAVEESTABLISHED(tp->t_state) &&
17156                     sbavail(sb) &&
17157                     (sbavail(sb) > tp->snd_wnd) &&
17158                     (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg))) {
17159                         /* Yes lets make sure to move to persist before timer-start */
17160                         rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
17161                 }
17162                 rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, sup_rack);
17163                 rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling, app_limited, cwnd_to_use);
17164         }
17165 #ifdef NETFLIX_SHARED_CWND
17166         if ((sbavail(sb) == 0) &&
17167             rack->r_ctl.rc_scw) {
17168                 tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
17169                 rack->rack_scwnd_is_idle = 1;
17170         }
17171 #endif
17172 #ifdef TCP_ACCOUNTING
17173         if (tot_len_this_send > 0) {
17174                 crtsc = get_cyclecount();
17175                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17176                         tp->tcp_cnt_counters[SND_OUT_DATA]++;
17177                 }
17178                 counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], 1);
17179                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17180                         tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
17181                 }
17182                 counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
17183                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17184                         tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) / segsiz);
17185                 }
17186                 counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len_this_send + segsiz - 1) / segsiz));
17187         } else {
17188                 crtsc = get_cyclecount();
17189                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17190                         tp->tcp_cnt_counters[SND_LIMITED]++;
17191                 }
17192                 counter_u64_add(tcp_cnt_counters[SND_LIMITED], 1);
17193                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17194                         tp->tcp_proc_time[SND_LIMITED] += (crtsc - ts_val);
17195                 }
17196                 counter_u64_add(tcp_proc_time[SND_LIMITED], (crtsc - ts_val));
17197         }
17198         sched_unpin();
17199 #endif
17200         return (0);
17201
17202 send:
17203         if (rsm || sack_rxmit)
17204                 counter_u64_add(rack_nfto_resend, 1);
17205         else
17206                 counter_u64_add(rack_non_fto_send, 1);
17207         if ((flags & TH_FIN) &&
17208             sbavail(sb)) {
17209                 /*
17210                  * We do not transmit a FIN
17211                  * with data outstanding. We
17212                  * need to make it so all data
17213                  * is acked first.
17214                  */
17215                 flags &= ~TH_FIN;
17216         }
17217         /* Enforce stack imposed max seg size if we have one */
17218         if (rack->r_ctl.rc_pace_max_segs &&
17219             (len > rack->r_ctl.rc_pace_max_segs)) {
17220                 mark = 1;
17221                 len = rack->r_ctl.rc_pace_max_segs;
17222         }
17223         SOCKBUF_LOCK_ASSERT(sb);
17224         if (len > 0) {
17225                 if (len >= segsiz)
17226                         tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
17227                 else
17228                         tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
17229         }
17230         /*
17231          * Before ESTABLISHED, force sending of initial options unless TCP
17232          * set not to do any options. NOTE: we assume that the IP/TCP header
17233          * plus TCP options always fit in a single mbuf, leaving room for a
17234          * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr)
17235          * + optlen <= MCLBYTES
17236          */
17237         optlen = 0;
17238 #ifdef INET6
17239         if (isipv6)
17240                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
17241         else
17242 #endif
17243                 hdrlen = sizeof(struct tcpiphdr);
17244
17245         /*
17246          * Compute options for segment. We only have to care about SYN and
17247          * established connection segments.  Options for SYN-ACK segments
17248          * are handled in TCP syncache.
17249          */
17250         to.to_flags = 0;
17251         if ((tp->t_flags & TF_NOOPT) == 0) {
17252                 /* Maximum segment size. */
17253                 if (flags & TH_SYN) {
17254                         tp->snd_nxt = tp->iss;
17255                         to.to_mss = tcp_mssopt(&inp->inp_inc);
17256                         if (tp->t_port)
17257                                 to.to_mss -= V_tcp_udp_tunneling_overhead;
17258                         to.to_flags |= TOF_MSS;
17259
17260                         /*
17261                          * On SYN or SYN|ACK transmits on TFO connections,
17262                          * only include the TFO option if it is not a
17263                          * retransmit, as the presence of the TFO option may
17264                          * have caused the original SYN or SYN|ACK to have
17265                          * been dropped by a middlebox.
17266                          */
17267                         if (IS_FASTOPEN(tp->t_flags) &&
17268                             (tp->t_rxtshift == 0)) {
17269                                 if (tp->t_state == TCPS_SYN_RECEIVED) {
17270                                         to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
17271                                         to.to_tfo_cookie =
17272                                                 (u_int8_t *)&tp->t_tfo_cookie.server;
17273                                         to.to_flags |= TOF_FASTOPEN;
17274                                         wanted_cookie = 1;
17275                                 } else if (tp->t_state == TCPS_SYN_SENT) {
17276                                         to.to_tfo_len =
17277                                                 tp->t_tfo_client_cookie_len;
17278                                         to.to_tfo_cookie =
17279                                                 tp->t_tfo_cookie.client;
17280                                         to.to_flags |= TOF_FASTOPEN;
17281                                         wanted_cookie = 1;
17282                                         /*
17283                                          * If we wind up having more data to
17284                                          * send with the SYN than can fit in
17285                                          * one segment, don't send any more
17286                                          * until the SYN|ACK comes back from
17287                                          * the other end.
17288                                          */
17289                                         sendalot = 0;
17290                                 }
17291                         }
17292                 }
17293                 /* Window scaling. */
17294                 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
17295                         to.to_wscale = tp->request_r_scale;
17296                         to.to_flags |= TOF_SCALE;
17297                 }
17298                 /* Timestamps. */
17299                 if ((tp->t_flags & TF_RCVD_TSTMP) ||
17300                     ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
17301                         to.to_tsval = ms_cts + tp->ts_offset;
17302                         to.to_tsecr = tp->ts_recent;
17303                         to.to_flags |= TOF_TS;
17304                 }
17305                 /* Set receive buffer autosizing timestamp. */
17306                 if (tp->rfbuf_ts == 0 &&
17307                     (so->so_rcv.sb_flags & SB_AUTOSIZE))
17308                         tp->rfbuf_ts = tcp_ts_getticks();
17309                 /* Selective ACK's. */
17310                 if (tp->t_flags & TF_SACK_PERMIT) {
17311                         if (flags & TH_SYN)
17312                                 to.to_flags |= TOF_SACKPERM;
17313                         else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
17314                                  tp->rcv_numsacks > 0) {
17315                                 to.to_flags |= TOF_SACK;
17316                                 to.to_nsacks = tp->rcv_numsacks;
17317                                 to.to_sacks = (u_char *)tp->sackblks;
17318                         }
17319                 }
17320 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
17321                 /* TCP-MD5 (RFC2385). */
17322                 if (tp->t_flags & TF_SIGNATURE)
17323                         to.to_flags |= TOF_SIGNATURE;
17324 #endif                          /* TCP_SIGNATURE */
17325
17326                 /* Processing the options. */
17327                 hdrlen += optlen = tcp_addoptions(&to, opt);
17328                 /*
17329                  * If we wanted a TFO option to be added, but it was unable
17330                  * to fit, ensure no data is sent.
17331                  */
17332                 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
17333                     !(to.to_flags & TOF_FASTOPEN))
17334                         len = 0;
17335         }
17336         if (tp->t_port) {
17337                 if (V_tcp_udp_tunneling_port == 0) {
17338                         /* The port was removed?? */
17339                         SOCKBUF_UNLOCK(&so->so_snd);
17340 #ifdef TCP_ACCOUNTING
17341                         crtsc = get_cyclecount();
17342                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17343                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
17344                         }
17345                         counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
17346                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17347                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
17348                         }
17349                         counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
17350                         sched_unpin();
17351 #endif
17352                         return (EHOSTUNREACH);
17353                 }
17354                 hdrlen += sizeof(struct udphdr);
17355         }
17356 #ifdef INET6
17357         if (isipv6)
17358                 ipoptlen = ip6_optlen(tp->t_inpcb);
17359         else
17360 #endif
17361                 if (tp->t_inpcb->inp_options)
17362                         ipoptlen = tp->t_inpcb->inp_options->m_len -
17363                                 offsetof(struct ipoption, ipopt_list);
17364                 else
17365                         ipoptlen = 0;
17366 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
17367         ipoptlen += ipsec_optlen;
17368 #endif
17369
17370         /*
17371          * Adjust data length if insertion of options will bump the packet
17372          * length beyond the t_maxseg length. Clear the FIN bit because we
17373          * cut off the tail of the segment.
17374          */
17375         if (len + optlen + ipoptlen > tp->t_maxseg) {
17376                 if (tso) {
17377                         uint32_t if_hw_tsomax;
17378                         uint32_t moff;
17379                         int32_t max_len;
17380
17381                         /* extract TSO information */
17382                         if_hw_tsomax = tp->t_tsomax;
17383                         if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
17384                         if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
17385                         KASSERT(ipoptlen == 0,
17386                                 ("%s: TSO can't do IP options", __func__));
17387
17388                         /*
17389                          * Check if we should limit by maximum payload
17390                          * length:
17391                          */
17392                         if (if_hw_tsomax != 0) {
17393                                 /* compute maximum TSO length */
17394                                 max_len = (if_hw_tsomax - hdrlen -
17395                                            max_linkhdr);
17396                                 if (max_len <= 0) {
17397                                         len = 0;
17398                                 } else if (len > max_len) {
17399                                         sendalot = 1;
17400                                         len = max_len;
17401                                         mark = 2;
17402                                 }
17403                         }
17404                         /*
17405                          * Prevent the last segment from being fractional
17406                          * unless the send sockbuf can be emptied:
17407                          */
17408                         max_len = (tp->t_maxseg - optlen);
17409                         if ((sb_offset + len) < sbavail(sb)) {
17410                                 moff = len % (u_int)max_len;
17411                                 if (moff != 0) {
17412                                         mark = 3;
17413                                         len -= moff;
17414                                 }
17415                         }
17416                         /*
17417                          * In case there are too many small fragments don't
17418                          * use TSO:
17419                          */
17420                         if (len <= segsiz) {
17421                                 mark = 4;
17422                                 tso = 0;
17423                         }
17424                         /*
17425                          * Send the FIN in a separate segment after the bulk
17426                          * sending is done. We don't trust the TSO
17427                          * implementations to clear the FIN flag on all but
17428                          * the last segment.
17429                          */
17430                         if (tp->t_flags & TF_NEEDFIN) {
17431                                 sendalot = 4;
17432                         }
17433                 } else {
17434                         mark = 5;
17435                         if (optlen + ipoptlen >= tp->t_maxseg) {
17436                                 /*
17437                                  * Since we don't have enough space to put
17438                                  * the IP header chain and the TCP header in
17439                                  * one packet as required by RFC 7112, don't
17440                                  * send it. Also ensure that at least one
17441                                  * byte of the payload can be put into the
17442                                  * TCP segment.
17443                                  */
17444                                 SOCKBUF_UNLOCK(&so->so_snd);
17445                                 error = EMSGSIZE;
17446                                 sack_rxmit = 0;
17447                                 goto out;
17448                         }
17449                         len = tp->t_maxseg - optlen - ipoptlen;
17450                         sendalot = 5;
17451                 }
17452         } else {
17453                 tso = 0;
17454                 mark = 6;
17455         }
17456         KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
17457                 ("%s: len > IP_MAXPACKET", __func__));
17458 #ifdef DIAGNOSTIC
17459 #ifdef INET6
17460         if (max_linkhdr + hdrlen > MCLBYTES)
17461 #else
17462                 if (max_linkhdr + hdrlen > MHLEN)
17463 #endif
17464                         panic("tcphdr too big");
17465 #endif
17466
17467         /*
17468          * This KASSERT is here to catch edge cases at a well defined place.
17469          * Before, those had triggered (random) panic conditions further
17470          * down.
17471          */
17472         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
17473         if ((len == 0) &&
17474             (flags & TH_FIN) &&
17475             (sbused(sb))) {
17476                 /*
17477                  * We have outstanding data, don't send a fin by itself!.
17478                  */
17479                 goto just_return;
17480         }
17481         /*
17482          * Grab a header mbuf, attaching a copy of data to be transmitted,
17483          * and initialize the header from the template for sends on this
17484          * connection.
17485          */
17486         hw_tls = (sb->sb_flags & SB_TLS_IFNET) != 0;
17487         if (len) {
17488                 uint32_t max_val;
17489                 uint32_t moff;
17490
17491                 if (rack->r_ctl.rc_pace_max_segs)
17492                         max_val = rack->r_ctl.rc_pace_max_segs;
17493                 else if (rack->rc_user_set_max_segs)
17494                         max_val = rack->rc_user_set_max_segs * segsiz;
17495                 else
17496                         max_val = len;
17497                 /*
17498                  * We allow a limit on sending with hptsi.
17499                  */
17500                 if (len > max_val) {
17501                         mark = 7;
17502                         len = max_val;
17503                 }
17504 #ifdef INET6
17505                 if (MHLEN < hdrlen + max_linkhdr)
17506                         m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
17507                 else
17508 #endif
17509                         m = m_gethdr(M_NOWAIT, MT_DATA);
17510
17511                 if (m == NULL) {
17512                         SOCKBUF_UNLOCK(sb);
17513                         error = ENOBUFS;
17514                         sack_rxmit = 0;
17515                         goto out;
17516                 }
17517                 m->m_data += max_linkhdr;
17518                 m->m_len = hdrlen;
17519
17520                 /*
17521                  * Start the m_copy functions from the closest mbuf to the
17522                  * sb_offset in the socket buffer chain.
17523                  */
17524                 mb = sbsndptr_noadv(sb, sb_offset, &moff);
17525                 s_mb = mb;
17526                 s_moff = moff;
17527                 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
17528                         m_copydata(mb, moff, (int)len,
17529                                    mtod(m, caddr_t)+hdrlen);
17530                         if (SEQ_LT(tp->snd_nxt, tp->snd_max))
17531                                 sbsndptr_adv(sb, mb, len);
17532                         m->m_len += len;
17533                 } else {
17534                         struct sockbuf *msb;
17535
17536                         if (SEQ_LT(tp->snd_nxt, tp->snd_max))
17537                                 msb = NULL;
17538                         else
17539                                 msb = sb;
17540                         m->m_next = tcp_m_copym(
17541                                 mb, moff, &len,
17542                                 if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb,
17543                                 ((rsm == NULL) ? hw_tls : 0)
17544 #ifdef NETFLIX_COPY_ARGS
17545                                 , &filled_all
17546 #endif
17547                                 );
17548                         if (len <= (tp->t_maxseg - optlen)) {
17549                                 /*
17550                                  * Must have ran out of mbufs for the copy
17551                                  * shorten it to no longer need tso. Lets
17552                                  * not put on sendalot since we are low on
17553                                  * mbufs.
17554                                  */
17555                                 tso = 0;
17556                         }
17557                         if (m->m_next == NULL) {
17558                                 SOCKBUF_UNLOCK(sb);
17559                                 (void)m_free(m);
17560                                 error = ENOBUFS;
17561                                 sack_rxmit = 0;
17562                                 goto out;
17563                         }
17564                 }
17565                 if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
17566                         if (rsm && (rsm->r_flags & RACK_TLP)) {
17567                                 /*
17568                                  * TLP should not count in retran count, but
17569                                  * in its own bin
17570                                  */
17571                                 counter_u64_add(rack_tlp_retran, 1);
17572                                 counter_u64_add(rack_tlp_retran_bytes, len);
17573                         } else {
17574                                 tp->t_sndrexmitpack++;
17575                                 KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
17576                                 KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
17577                         }
17578 #ifdef STATS
17579                         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
17580                                                  len);
17581 #endif
17582                 } else {
17583                         KMOD_TCPSTAT_INC(tcps_sndpack);
17584                         KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
17585 #ifdef STATS
17586                         stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
17587                                                  len);
17588 #endif
17589                 }
17590                 /*
17591                  * If we're sending everything we've got, set PUSH. (This
17592                  * will keep happy those implementations which only give
17593                  * data to the user when a buffer fills or a PUSH comes in.)
17594                  */
17595                 if (sb_offset + len == sbused(sb) &&
17596                     sbused(sb) &&
17597                     !(flags & TH_SYN)) {
17598                         flags |= TH_PUSH;
17599                         add_flag |= RACK_HAD_PUSH;
17600                 }
17601
17602                 SOCKBUF_UNLOCK(sb);
17603         } else {
17604                 SOCKBUF_UNLOCK(sb);
17605                 if (tp->t_flags & TF_ACKNOW)
17606                         KMOD_TCPSTAT_INC(tcps_sndacks);
17607                 else if (flags & (TH_SYN | TH_FIN | TH_RST))
17608                         KMOD_TCPSTAT_INC(tcps_sndctrl);
17609                 else
17610                         KMOD_TCPSTAT_INC(tcps_sndwinup);
17611
17612                 m = m_gethdr(M_NOWAIT, MT_DATA);
17613                 if (m == NULL) {
17614                         error = ENOBUFS;
17615                         sack_rxmit = 0;
17616                         goto out;
17617                 }
17618 #ifdef INET6
17619                 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
17620                     MHLEN >= hdrlen) {
17621                         M_ALIGN(m, hdrlen);
17622                 } else
17623 #endif
17624                         m->m_data += max_linkhdr;
17625                 m->m_len = hdrlen;
17626         }
17627         SOCKBUF_UNLOCK_ASSERT(sb);
17628         m->m_pkthdr.rcvif = (struct ifnet *)0;
17629 #ifdef MAC
17630         mac_inpcb_create_mbuf(inp, m);
17631 #endif
17632         if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) &&  rack->r_fsb_inited) {
17633 #ifdef INET6
17634                 if (isipv6)
17635                         ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
17636                 else
17637 #endif                          /* INET6 */
17638                         ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
17639                 th = rack->r_ctl.fsb.th;
17640                 udp = rack->r_ctl.fsb.udp;
17641                 if (udp) {
17642                         if (isipv6)
17643                                 ulen = hdrlen + len - sizeof(struct ip6_hdr);
17644                         else
17645                                 ulen = hdrlen + len - sizeof(struct ip);
17646                         udp->uh_ulen = htons(ulen);
17647                 }
17648         } else {
17649 #ifdef INET6
17650                 if (isipv6) {
17651                         ip6 = mtod(m, struct ip6_hdr *);
17652                         if (tp->t_port) {
17653                                 udp = (struct udphdr *)((caddr_t)ip6 + ipoptlen + sizeof(struct ip6_hdr));
17654                                 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
17655                                 udp->uh_dport = tp->t_port;
17656                                 ulen = hdrlen + len - sizeof(struct ip6_hdr);
17657                                 udp->uh_ulen = htons(ulen);
17658                                 th = (struct tcphdr *)(udp + 1);
17659                         } else
17660                                 th = (struct tcphdr *)(ip6 + 1);
17661                         tcpip_fillheaders(inp, tp->t_port, ip6, th);
17662                 } else
17663 #endif                          /* INET6 */
17664                 {
17665                         ip = mtod(m, struct ip *);
17666 #ifdef TCPDEBUG
17667                         ipov = (struct ipovly *)ip;
17668 #endif
17669                         if (tp->t_port) {
17670                                 udp = (struct udphdr *)((caddr_t)ip + ipoptlen + sizeof(struct ip));
17671                                 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
17672                                 udp->uh_dport = tp->t_port;
17673                                 ulen = hdrlen + len - sizeof(struct ip);
17674                                 udp->uh_ulen = htons(ulen);
17675                                 th = (struct tcphdr *)(udp + 1);
17676                         } else
17677                                 th = (struct tcphdr *)(ip + 1);
17678                         tcpip_fillheaders(inp, tp->t_port, ip, th);
17679                 }
17680         }
17681         /*
17682          * Fill in fields, remembering maximum advertised window for use in
17683          * delaying messages about window sizes. If resending a FIN, be sure
17684          * not to use a new sequence number.
17685          */
17686         if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
17687             tp->snd_nxt == tp->snd_max)
17688                 tp->snd_nxt--;
17689         /*
17690          * If we are starting a connection, send ECN setup SYN packet. If we
17691          * are on a retransmit, we may resend those bits a number of times
17692          * as per RFC 3168.
17693          */
17694         if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) {
17695                 if (tp->t_rxtshift >= 1) {
17696                         if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
17697                                 flags |= TH_ECE | TH_CWR;
17698                 } else
17699                         flags |= TH_ECE | TH_CWR;
17700         }
17701         /* Handle parallel SYN for ECN */
17702         if ((tp->t_state == TCPS_SYN_RECEIVED) &&
17703             (tp->t_flags2 & TF2_ECN_SND_ECE)) {
17704                 flags |= TH_ECE;
17705                 tp->t_flags2 &= ~TF2_ECN_SND_ECE;
17706         }
17707         if (TCPS_HAVEESTABLISHED(tp->t_state) &&
17708             (tp->t_flags2 & TF2_ECN_PERMIT)) {
17709                 /*
17710                  * If the peer has ECN, mark data packets with ECN capable
17711                  * transmission (ECT). Ignore pure ack packets,
17712                  * retransmissions.
17713                  */
17714                 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
17715                     (sack_rxmit == 0)) {
17716 #ifdef INET6
17717                         if (isipv6)
17718                                 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
17719                         else
17720 #endif
17721                                 ip->ip_tos |= IPTOS_ECN_ECT0;
17722                         KMOD_TCPSTAT_INC(tcps_ecn_ect0);
17723                         /*
17724                          * Reply with proper ECN notifications.
17725                          * Only set CWR on new data segments.
17726                          */
17727                         if (tp->t_flags2 & TF2_ECN_SND_CWR) {
17728                                 flags |= TH_CWR;
17729                                 tp->t_flags2 &= ~TF2_ECN_SND_CWR;
17730                         }
17731                 }
17732                 if (tp->t_flags2 & TF2_ECN_SND_ECE)
17733                         flags |= TH_ECE;
17734         }
17735         /*
17736          * If we are doing retransmissions, then snd_nxt will not reflect
17737          * the first unsent octet.  For ACK only packets, we do not want the
17738          * sequence number of the retransmitted packet, we want the sequence
17739          * number of the next unsent octet.  So, if there is no data (and no
17740          * SYN or FIN), use snd_max instead of snd_nxt when filling in
17741          * ti_seq.  But if we are in persist state, snd_max might reflect
17742          * one byte beyond the right edge of the window, so use snd_nxt in
17743          * that case, since we know we aren't doing a retransmission.
17744          * (retransmit and persist are mutually exclusive...)
17745          */
17746         if (sack_rxmit == 0) {
17747                 if (len || (flags & (TH_SYN | TH_FIN))) {
17748                         th->th_seq = htonl(tp->snd_nxt);
17749                         rack_seq = tp->snd_nxt;
17750                 } else {
17751                         th->th_seq = htonl(tp->snd_max);
17752                         rack_seq = tp->snd_max;
17753                 }
17754         } else {
17755                 th->th_seq = htonl(rsm->r_start);
17756                 rack_seq = rsm->r_start;
17757         }
17758         th->th_ack = htonl(tp->rcv_nxt);
17759         th->th_flags = flags;
17760         /*
17761          * Calculate receive window.  Don't shrink window, but avoid silly
17762          * window syndrome.
17763          * If a RST segment is sent, advertise a window of zero.
17764          */
17765         if (flags & TH_RST) {
17766                 recwin = 0;
17767         } else {
17768                 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
17769                     recwin < (long)segsiz) {
17770                         recwin = 0;
17771                 }
17772                 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
17773                     recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
17774                         recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
17775         }
17776
17777         /*
17778          * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
17779          * <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK> case is
17780          * handled in syncache.
17781          */
17782         if (flags & TH_SYN)
17783                 th->th_win = htons((u_short)
17784                                    (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
17785         else {
17786                 /* Avoid shrinking window with window scaling. */
17787                 recwin = roundup2(recwin, 1 << tp->rcv_scale);
17788                 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
17789         }
17790         /*
17791          * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
17792          * window.  This may cause the remote transmitter to stall.  This
17793          * flag tells soreceive() to disable delayed acknowledgements when
17794          * draining the buffer.  This can occur if the receiver is
17795          * attempting to read more data than can be buffered prior to
17796          * transmitting on the connection.
17797          */
17798         if (th->th_win == 0) {
17799                 tp->t_sndzerowin++;
17800                 tp->t_flags |= TF_RXWIN0SENT;
17801         } else
17802                 tp->t_flags &= ~TF_RXWIN0SENT;
17803         tp->snd_up = tp->snd_una;       /* drag it along, its deprecated */
17804         /* Now are we using fsb?, if so copy the template data to the mbuf */
17805         if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) && rack->r_fsb_inited) {
17806                 uint8_t *cpto;
17807
17808                 cpto = mtod(m, uint8_t *);
17809                 memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
17810                 /*
17811                  * We have just copied in:
17812                  * IP/IP6
17813                  * <optional udphdr>
17814                  * tcphdr (no options)
17815                  *
17816                  * We need to grab the correct pointers into the mbuf
17817                  * for both the tcp header, and possibly the udp header (if tunneling).
17818                  * We do this by using the offset in the copy buffer and adding it
17819                  * to the mbuf base pointer (cpto).
17820                  */
17821 #ifdef INET6
17822                 if (isipv6)
17823                         ip6 = mtod(m, struct ip6_hdr *);
17824                 else
17825 #endif                          /* INET6 */
17826                         ip = mtod(m, struct ip *);
17827                 th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
17828                 /* If we have a udp header lets set it into the mbuf as well */
17829                 if (udp)
17830                         udp = (struct udphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.udp - rack->r_ctl.fsb.tcp_ip_hdr));
17831         }
17832 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
17833         if (to.to_flags & TOF_SIGNATURE) {
17834                 /*
17835                  * Calculate MD5 signature and put it into the place
17836                  * determined before.
17837                  * NOTE: since TCP options buffer doesn't point into
17838                  * mbuf's data, calculate offset and use it.
17839                  */
17840                 if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
17841                                                        (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
17842                         /*
17843                          * Do not send segment if the calculation of MD5
17844                          * digest has failed.
17845                          */
17846                         goto out;
17847                 }
17848         }
17849 #endif
17850         if (optlen) {
17851                 bcopy(opt, th + 1, optlen);
17852                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
17853         }
17854         /*
17855          * Put TCP length in extended header, and then checksum extended
17856          * header and data.
17857          */
17858         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
17859 #ifdef INET6
17860         if (isipv6) {
17861                 /*
17862                  * ip6_plen is not need to be filled now, and will be filled
17863                  * in ip6_output.
17864                  */
17865                 if (tp->t_port) {
17866                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
17867                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
17868                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
17869                         th->th_sum = htons(0);
17870                         UDPSTAT_INC(udps_opackets);
17871                 } else {
17872                         m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
17873                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
17874                         th->th_sum = in6_cksum_pseudo(ip6,
17875                                                       sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
17876                                                       0);
17877                 }
17878         }
17879 #endif
17880 #if defined(INET6) && defined(INET)
17881         else
17882 #endif
17883 #ifdef INET
17884         {
17885                 if (tp->t_port) {
17886                         m->m_pkthdr.csum_flags = CSUM_UDP;
17887                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
17888                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
17889                                                 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
17890                         th->th_sum = htons(0);
17891                         UDPSTAT_INC(udps_opackets);
17892                 } else {
17893                         m->m_pkthdr.csum_flags = CSUM_TCP;
17894                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
17895                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
17896                                                ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
17897                                                                         IPPROTO_TCP + len + optlen));
17898                 }
17899                 /* IP version must be set here for ipv4/ipv6 checking later */
17900                 KASSERT(ip->ip_v == IPVERSION,
17901                         ("%s: IP version incorrect: %d", __func__, ip->ip_v));
17902         }
17903 #endif
17904         /*
17905          * Enable TSO and specify the size of the segments. The TCP pseudo
17906          * header checksum is always provided. XXX: Fixme: This is currently
17907          * not the case for IPv6.
17908          */
17909         if (tso) {
17910                 KASSERT(len > tp->t_maxseg - optlen,
17911                         ("%s: len <= tso_segsz", __func__));
17912                 m->m_pkthdr.csum_flags |= CSUM_TSO;
17913                 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
17914         }
17915         KASSERT(len + hdrlen == m_length(m, NULL),
17916                 ("%s: mbuf chain different than expected: %d + %u != %u",
17917                  __func__, len, hdrlen, m_length(m, NULL)));
17918
17919 #ifdef TCP_HHOOK
17920         /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
17921         hhook_run_tcp_est_out(tp, th, &to, len, tso);
17922 #endif
17923         /* We're getting ready to send; log now. */
17924         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
17925                 union tcp_log_stackspecific log;
17926
17927                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
17928                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
17929                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
17930                 if (rack->rack_no_prr)
17931                         log.u_bbr.flex1 = 0;
17932                 else
17933                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
17934                 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
17935                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
17936                 log.u_bbr.flex4 = orig_len;
17937                 if (filled_all)
17938                         log.u_bbr.flex5 = 0x80000000;
17939                 else
17940                         log.u_bbr.flex5 = 0;
17941                 /* Save off the early/late values */
17942                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
17943                 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
17944                 log.u_bbr.bw_inuse = rack_get_bw(rack);
17945                 if (rsm || sack_rxmit) {
17946                         if (doing_tlp)
17947                                 log.u_bbr.flex8 = 2;
17948                         else
17949                                 log.u_bbr.flex8 = 1;
17950                 } else {
17951                         log.u_bbr.flex8 = 0;
17952                 }
17953                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
17954                 log.u_bbr.flex7 = mark;
17955                 log.u_bbr.flex7 <<= 8;
17956                 log.u_bbr.flex7 |= pass;
17957                 log.u_bbr.pkts_out = tp->t_maxseg;
17958                 log.u_bbr.timeStamp = cts;
17959                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
17960                 log.u_bbr.lt_epoch = cwnd_to_use;
17961                 log.u_bbr.delivered = sendalot;
17962                 lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
17963                                      len, &log, false, NULL, NULL, 0, &tv);
17964         } else
17965                 lgb = NULL;
17966
17967         /*
17968          * Fill in IP length and desired time to live and send to IP level.
17969          * There should be a better way to handle ttl and tos; we could keep
17970          * them in the template, but need a way to checksum without them.
17971          */
17972         /*
17973          * m->m_pkthdr.len should have been set before cksum calcuration,
17974          * because in6_cksum() need it.
17975          */
17976 #ifdef INET6
17977         if (isipv6) {
17978                 /*
17979                  * we separately set hoplimit for every segment, since the
17980                  * user might want to change the value via setsockopt. Also,
17981                  * desired default hop limit might be changed via Neighbor
17982                  * Discovery.
17983                  */
17984                 rack->r_ctl.fsb.hoplimit = ip6->ip6_hlim = in6_selecthlim(inp, NULL);
17985
17986                 /*
17987                  * Set the packet size here for the benefit of DTrace
17988                  * probes. ip6_output() will set it properly; it's supposed
17989                  * to include the option header lengths as well.
17990                  */
17991                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
17992
17993                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
17994                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
17995                 else
17996                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
17997
17998                 if (tp->t_state == TCPS_SYN_SENT)
17999                         TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
18000
18001                 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
18002                 /* TODO: IPv6 IP6TOS_ECT bit on */
18003                 error = ip6_output(m,
18004 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18005                                    inp->in6p_outputopts,
18006 #else
18007                                    NULL,
18008 #endif
18009                                    &inp->inp_route6,
18010                                    ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0),
18011                                    NULL, NULL, inp);
18012
18013                 if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
18014                         mtu = inp->inp_route6.ro_nh->nh_mtu;
18015         }
18016 #endif                          /* INET6 */
18017 #if defined(INET) && defined(INET6)
18018         else
18019 #endif
18020 #ifdef INET
18021         {
18022                 ip->ip_len = htons(m->m_pkthdr.len);
18023 #ifdef INET6
18024                 if (inp->inp_vflag & INP_IPV6PROTO)
18025                         ip->ip_ttl = in6_selecthlim(inp, NULL);
18026 #endif                          /* INET6 */
18027                 rack->r_ctl.fsb.hoplimit = ip->ip_ttl;
18028                 /*
18029                  * If we do path MTU discovery, then we set DF on every
18030                  * packet. This might not be the best thing to do according
18031                  * to RFC3390 Section 2. However the tcp hostcache migitates
18032                  * the problem so it affects only the first tcp connection
18033                  * with a host.
18034                  *
18035                  * NB: Don't set DF on small MTU/MSS to have a safe
18036                  * fallback.
18037                  */
18038                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
18039                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18040                         if (tp->t_port == 0 || len < V_tcp_minmss) {
18041                                 ip->ip_off |= htons(IP_DF);
18042                         }
18043                 } else {
18044                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18045                 }
18046
18047                 if (tp->t_state == TCPS_SYN_SENT)
18048                         TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
18049
18050                 TCP_PROBE5(send, NULL, tp, ip, tp, th);
18051
18052                 error = ip_output(m,
18053 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18054                                   inp->inp_options,
18055 #else
18056                                   NULL,
18057 #endif
18058                                   &inp->inp_route,
18059                                   ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0,
18060                                   inp);
18061                 if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
18062                         mtu = inp->inp_route.ro_nh->nh_mtu;
18063         }
18064 #endif                          /* INET */
18065
18066 out:
18067         if (lgb) {
18068                 lgb->tlb_errno = error;
18069                 lgb = NULL;
18070         }
18071         /*
18072          * In transmit state, time the transmission and arrange for the
18073          * retransmit.  In persist state, just set snd_max.
18074          */
18075         if (error == 0) {
18076                 rack->forced_ack = 0;   /* If we send something zap the FA flag */
18077                 if (rsm && (doing_tlp == 0)) {
18078                         /* Set we retransmitted */
18079                         rack->rc_gp_saw_rec = 1;
18080                 } else {
18081                         if (cwnd_to_use > tp->snd_ssthresh) {
18082                                 /* Set we sent in CA */
18083                                 rack->rc_gp_saw_ca = 1;
18084                         } else {
18085                                 /* Set we sent in SS */
18086                                 rack->rc_gp_saw_ss = 1;
18087                         }
18088                 }
18089                 if (TCPS_HAVEESTABLISHED(tp->t_state) &&
18090                     (tp->t_flags & TF_SACK_PERMIT) &&
18091                     tp->rcv_numsacks > 0)
18092                         tcp_clean_dsack_blocks(tp);
18093                 tot_len_this_send += len;
18094                 if (len == 0)
18095                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1);
18096                 else if (len == 1) {
18097                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1);
18098                 } else if (len > 1) {
18099                         int idx;
18100
18101                         idx = (len / segsiz) + 3;
18102                         if (idx >= TCP_MSS_ACCT_ATIMER)
18103                                 counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
18104                         else
18105                                 counter_u64_add(rack_out_size[idx], 1);
18106                 }
18107         }
18108         if ((rack->rack_no_prr == 0) &&
18109             sub_from_prr &&
18110             (error == 0)) {
18111                 if (rack->r_ctl.rc_prr_sndcnt >= len)
18112                         rack->r_ctl.rc_prr_sndcnt -= len;
18113                 else
18114                         rack->r_ctl.rc_prr_sndcnt = 0;
18115         }
18116         sub_from_prr = 0;
18117         if (doing_tlp && (rsm == NULL)) {
18118                 /* New send doing a TLP */
18119                 add_flag |= RACK_TLP;
18120                 tp->t_sndtlppack++;
18121                 tp->t_sndtlpbyte += len;
18122         }
18123         rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error,
18124                         rack_to_usec_ts(&tv),
18125                         rsm, add_flag, s_mb, s_moff);
18126
18127
18128         if ((error == 0) &&
18129             (len > 0) &&
18130             (tp->snd_una == tp->snd_max))
18131                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
18132         {
18133                 tcp_seq startseq = tp->snd_nxt;
18134
18135                 /* Track our lost count */
18136                 if (rsm && (doing_tlp == 0))
18137                         rack->r_ctl.rc_loss_count += rsm->r_end - rsm->r_start;
18138                 /*
18139                  * Advance snd_nxt over sequence space of this segment.
18140                  */
18141                 if (error)
18142                         /* We don't log or do anything with errors */
18143                         goto nomore;
18144                 if (doing_tlp == 0) {
18145                         if (rsm == NULL) {
18146                                 /*
18147                                  * Not a retransmission of some
18148                                  * sort, new data is going out so
18149                                  * clear our TLP count and flag.
18150                                  */
18151                                 rack->rc_tlp_in_progress = 0;
18152                                 rack->r_ctl.rc_tlp_cnt_out = 0;
18153                         }
18154                 } else {
18155                         /*
18156                          * We have just sent a TLP, mark that it is true
18157                          * and make sure our in progress is set so we
18158                          * continue to check the count.
18159                          */
18160                         rack->rc_tlp_in_progress = 1;
18161                         rack->r_ctl.rc_tlp_cnt_out++;
18162                 }
18163                 if (flags & (TH_SYN | TH_FIN)) {
18164                         if (flags & TH_SYN)
18165                                 tp->snd_nxt++;
18166                         if (flags & TH_FIN) {
18167                                 tp->snd_nxt++;
18168                                 tp->t_flags |= TF_SENTFIN;
18169                         }
18170                 }
18171                 /* In the ENOBUFS case we do *not* update snd_max */
18172                 if (sack_rxmit)
18173                         goto nomore;
18174
18175                 tp->snd_nxt += len;
18176                 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
18177                         if (tp->snd_una == tp->snd_max) {
18178                                 /*
18179                                  * Update the time we just added data since
18180                                  * none was outstanding.
18181                                  */
18182                                 rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
18183                                 tp->t_acktime = ticks;
18184                         }
18185                         tp->snd_max = tp->snd_nxt;
18186                         /*
18187                          * Time this transmission if not a retransmission and
18188                          * not currently timing anything.
18189                          * This is only relevant in case of switching back to
18190                          * the base stack.
18191                          */
18192                         if (tp->t_rtttime == 0) {
18193                                 tp->t_rtttime = ticks;
18194                                 tp->t_rtseq = startseq;
18195                                 KMOD_TCPSTAT_INC(tcps_segstimed);
18196                         }
18197                         if (len &&
18198                             ((tp->t_flags & TF_GPUTINPROG) == 0))
18199                                 rack_start_gp_measurement(tp, rack, startseq, sb_offset);
18200                 }
18201                 /*
18202                  * If we are doing FO we need to update the mbuf position and subtract
18203                  * this happens when the peer sends us duplicate information and
18204                  * we thus want to send a DSACK.
18205                  *
18206                  * XXXRRS: This brings to mind a ?, when we send a DSACK block is TSO
18207                  * turned off? If not then we are going to echo multiple DSACK blocks
18208                  * out (with the TSO), which we should not be doing.
18209                  */
18210                 if (rack->r_fast_output && len) {
18211                         if (rack->r_ctl.fsb.left_to_send > len)
18212                                 rack->r_ctl.fsb.left_to_send -= len;
18213                         else
18214                                 rack->r_ctl.fsb.left_to_send = 0;
18215                         if (rack->r_ctl.fsb.left_to_send < segsiz)
18216                                 rack->r_fast_output = 0;
18217                         if (rack->r_fast_output) {
18218                                 rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
18219                                 rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
18220                         }
18221                 }
18222         }
18223 nomore:
18224         if (error) {
18225                 rack->r_ctl.rc_agg_delayed = 0;
18226                 rack->r_early = 0;
18227                 rack->r_late = 0;
18228                 rack->r_ctl.rc_agg_early = 0;
18229                 SOCKBUF_UNLOCK_ASSERT(sb);      /* Check gotos. */
18230                 /*
18231                  * Failures do not advance the seq counter above. For the
18232                  * case of ENOBUFS we will fall out and retry in 1ms with
18233                  * the hpts. Everything else will just have to retransmit
18234                  * with the timer.
18235                  *
18236                  * In any case, we do not want to loop around for another
18237                  * send without a good reason.
18238                  */
18239                 sendalot = 0;
18240                 switch (error) {
18241                 case EPERM:
18242                         tp->t_softerror = error;
18243 #ifdef TCP_ACCOUNTING
18244                         crtsc = get_cyclecount();
18245                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18246                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
18247                         }
18248                         counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
18249                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18250                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
18251                         }
18252                         counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
18253                         sched_unpin();
18254 #endif
18255                         return (error);
18256                 case ENOBUFS:
18257                         /*
18258                          * Pace us right away to retry in a some
18259                          * time
18260                          */
18261                         slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
18262                         if (rack->rc_enobuf < 0x7f)
18263                                 rack->rc_enobuf++;
18264                         if (slot < (10 * HPTS_USEC_IN_MSEC))
18265                                 slot = 10 * HPTS_USEC_IN_MSEC;
18266                         if (rack->r_ctl.crte != NULL) {
18267                                 counter_u64_add(rack_saw_enobuf_hw, 1);
18268                                 tcp_rl_log_enobuf(rack->r_ctl.crte);
18269                         }
18270                         counter_u64_add(rack_saw_enobuf, 1);
18271                         goto enobufs;
18272                 case EMSGSIZE:
18273                         /*
18274                          * For some reason the interface we used initially
18275                          * to send segments changed to another or lowered
18276                          * its MTU. If TSO was active we either got an
18277                          * interface without TSO capabilits or TSO was
18278                          * turned off. If we obtained mtu from ip_output()
18279                          * then update it and try again.
18280                          */
18281                         if (tso)
18282                                 tp->t_flags &= ~TF_TSO;
18283                         if (mtu != 0) {
18284                                 tcp_mss_update(tp, -1, mtu, NULL, NULL);
18285                                 goto again;
18286                         }
18287                         slot = 10 * HPTS_USEC_IN_MSEC;
18288                         rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
18289 #ifdef TCP_ACCOUNTING
18290                         crtsc = get_cyclecount();
18291                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18292                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
18293                         }
18294                         counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
18295                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18296                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
18297                         }
18298                         counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
18299                         sched_unpin();
18300 #endif
18301                         return (error);
18302                 case ENETUNREACH:
18303                         counter_u64_add(rack_saw_enetunreach, 1);
18304                 case EHOSTDOWN:
18305                 case EHOSTUNREACH:
18306                 case ENETDOWN:
18307                         if (TCPS_HAVERCVDSYN(tp->t_state)) {
18308                                 tp->t_softerror = error;
18309                         }
18310                         /* FALLTHROUGH */
18311                 default:
18312                         slot = 10 * HPTS_USEC_IN_MSEC;
18313                         rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
18314 #ifdef TCP_ACCOUNTING
18315                         crtsc = get_cyclecount();
18316                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18317                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
18318                         }
18319                         counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
18320                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18321                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
18322                         }
18323                         counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
18324                         sched_unpin();
18325 #endif
18326                         return (error);
18327                 }
18328         } else {
18329                 rack->rc_enobuf = 0;
18330                 if (IN_FASTRECOVERY(tp->t_flags) && rsm)
18331                         rack->r_ctl.retran_during_recovery += len;
18332         }
18333         KMOD_TCPSTAT_INC(tcps_sndtotal);
18334
18335         /*
18336          * Data sent (as far as we can tell). If this advertises a larger
18337          * window than any other segment, then remember the size of the
18338          * advertised window. Any pending ACK has now been sent.
18339          */
18340         if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
18341                 tp->rcv_adv = tp->rcv_nxt + recwin;
18342
18343         tp->last_ack_sent = tp->rcv_nxt;
18344         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
18345 enobufs:
18346         if (sendalot) {
18347                 /* Do we need to turn off sendalot? */
18348                 if (rack->r_ctl.rc_pace_max_segs &&
18349                     (tot_len_this_send >= rack->r_ctl.rc_pace_max_segs)) {
18350                         /* We hit our max. */
18351                         sendalot = 0;
18352                 } else if ((rack->rc_user_set_max_segs) &&
18353                            (tot_len_this_send >= (rack->rc_user_set_max_segs * segsiz))) {
18354                         /* We hit the user defined max */
18355                         sendalot = 0;
18356                 }
18357         }
18358         if ((error == 0) && (flags & TH_FIN))
18359                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN);
18360         if (flags & TH_RST) {
18361                 /*
18362                  * We don't send again after sending a RST.
18363                  */
18364                 slot = 0;
18365                 sendalot = 0;
18366                 if (error == 0)
18367                         tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
18368         } else if ((slot == 0) && (sendalot == 0) && tot_len_this_send) {
18369                 /*
18370                  * Get our pacing rate, if an error
18371                  * occurred in sending (ENOBUF) we would
18372                  * hit the else if with slot preset. Other
18373                  * errors return.
18374                  */
18375                 slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, rsm, segsiz);
18376         }
18377         if (rsm &&
18378             (rsm->r_flags & RACK_HAS_SYN) == 0 &&
18379             rack->use_rack_rr) {
18380                 /* Its a retransmit and we use the rack cheat? */
18381                 if ((slot == 0) ||
18382                     (rack->rc_always_pace == 0) ||
18383                     (rack->r_rr_config == 1)) {
18384                         /*
18385                          * We have no pacing set or we
18386                          * are using old-style rack or
18387                          * we are overriden to use the old 1ms pacing.
18388                          */
18389                         slot = rack->r_ctl.rc_min_to;
18390                 }
18391         }
18392         /* We have sent clear the flag */
18393         rack->r_ent_rec_ns = 0;
18394         if (rack->r_must_retran) {
18395                 if (rsm) {
18396                         rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
18397                         if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
18398                                 /*
18399                                  * We have retransmitted all.
18400                                  */
18401                                 rack->r_must_retran = 0;
18402                                 rack->r_ctl.rc_out_at_rto = 0;
18403                         }
18404                 } else if (SEQ_GEQ(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
18405                         /*
18406                          * Sending new data will also kill
18407                          * the loop.
18408                          */
18409                         rack->r_must_retran = 0;
18410                         rack->r_ctl.rc_out_at_rto = 0;
18411                 }
18412         }
18413         rack->r_ctl.fsb.recwin = recwin;
18414         if ((tp->t_flags & (TF_WASCRECOVERY|TF_WASFRECOVERY)) &&
18415             SEQ_GT(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
18416                 /*
18417                  * We hit an RTO and now have past snd_max at the RTO
18418                  * clear all the WAS flags.
18419                  */
18420                 tp->t_flags &= ~(TF_WASCRECOVERY|TF_WASFRECOVERY);
18421         }
18422         if (slot) {
18423                 /* set the rack tcb into the slot N */
18424                 counter_u64_add(rack_paced_segments, 1);
18425                 if ((error == 0) &&
18426                     rack_use_rfo &&
18427                     ((flags & (TH_SYN|TH_FIN)) == 0) &&
18428                     (rsm == NULL) &&
18429                     (tp->snd_nxt == tp->snd_max) &&
18430                     (ipoptlen == 0) &&
18431                     (tp->rcv_numsacks == 0) &&
18432                     rack->r_fsb_inited &&
18433                     TCPS_HAVEESTABLISHED(tp->t_state) &&
18434                     (rack->r_must_retran == 0) &&
18435                     ((tp->t_flags & TF_NEEDFIN) == 0) &&
18436                     (len > 0) && (orig_len > 0) &&
18437                     (orig_len > len) &&
18438                     ((orig_len - len) >= segsiz) &&
18439                     ((optlen == 0) ||
18440                      ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
18441                         /* We can send at least one more MSS using our fsb */
18442
18443                         rack->r_fast_output = 1;
18444                         rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
18445                         rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
18446                         rack->r_ctl.fsb.tcp_flags = flags;
18447                         rack->r_ctl.fsb.left_to_send = orig_len - len;
18448                         KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
18449                                 ("rack:%p left_to_send:%u sbavail:%u out:%u",
18450                                  rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
18451                                  (tp->snd_max - tp->snd_una)));
18452                         if (rack->r_ctl.fsb.left_to_send < segsiz)
18453                                 rack->r_fast_output = 0;
18454                         else {
18455                                 if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
18456                                         rack->r_ctl.fsb.rfo_apply_push = 1;
18457                                 else
18458                                         rack->r_ctl.fsb.rfo_apply_push = 0;
18459                         }
18460                 } else
18461                         rack->r_fast_output = 0;
18462                 rack_log_fsb(rack, tp, so, flags,
18463                              ipoptlen, orig_len, len, error,
18464                              (rsm == NULL), optlen, __LINE__, 2);
18465         } else if (sendalot) {
18466                 int ret;
18467
18468                 if (len)
18469                         counter_u64_add(rack_unpaced_segments, 1);
18470                 sack_rxmit = 0;
18471                 if ((error == 0) &&
18472                     rack_use_rfo &&
18473                     ((flags & (TH_SYN|TH_FIN)) == 0) &&
18474                     (rsm == NULL) &&
18475                     (ipoptlen == 0) &&
18476                     (tp->rcv_numsacks == 0) &&
18477                     (tp->snd_nxt == tp->snd_max) &&
18478                     (rack->r_must_retran == 0) &&
18479                     rack->r_fsb_inited &&
18480                     TCPS_HAVEESTABLISHED(tp->t_state) &&
18481                     ((tp->t_flags & TF_NEEDFIN) == 0) &&
18482                     (len > 0) && (orig_len > 0) &&
18483                     (orig_len > len) &&
18484                     ((orig_len - len) >= segsiz) &&
18485                     ((optlen == 0) ||
18486                      ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
18487                         /* we can use fast_output for more */
18488
18489                         rack->r_fast_output = 1;
18490                         rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
18491                         rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
18492                         rack->r_ctl.fsb.tcp_flags = flags;
18493                         rack->r_ctl.fsb.left_to_send = orig_len - len;
18494                         KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
18495                                 ("rack:%p left_to_send:%u sbavail:%u out:%u",
18496                                  rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
18497                                  (tp->snd_max - tp->snd_una)));
18498                         if (rack->r_ctl.fsb.left_to_send < segsiz) {
18499                                 rack->r_fast_output = 0;
18500                         }
18501                         if (rack->r_fast_output) {
18502                                 if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
18503                                         rack->r_ctl.fsb.rfo_apply_push = 1;
18504                                 else
18505                                         rack->r_ctl.fsb.rfo_apply_push = 0;
18506                                 rack_log_fsb(rack, tp, so, flags,
18507                                              ipoptlen, orig_len, len, error,
18508                                              (rsm == NULL), optlen, __LINE__, 3);
18509                                 error = 0;
18510                                 ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
18511                                 if (ret >= 0)
18512                                         return (ret);
18513                                 else if (error)
18514                                         goto nomore;
18515
18516                         }
18517                 }
18518                 goto again;
18519         } else if (len) {
18520                 counter_u64_add(rack_unpaced_segments, 1);
18521         }
18522         /* Assure when we leave that snd_nxt will point to top */
18523         if (SEQ_GT(tp->snd_max, tp->snd_nxt))
18524                 tp->snd_nxt = tp->snd_max;
18525         rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, 0);
18526 #ifdef TCP_ACCOUNTING
18527         crtsc = get_cyclecount() - ts_val;
18528         if (tot_len_this_send) {
18529                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18530                         tp->tcp_cnt_counters[SND_OUT_DATA]++;
18531                 }
18532                 counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], 1);
18533                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18534                         tp->tcp_proc_time[SND_OUT_DATA] += crtsc;
18535                 }
18536                 counter_u64_add(tcp_proc_time[SND_OUT_DATA], crtsc);
18537                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18538                         tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) /segsiz);
18539                 }
18540                 counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len_this_send + segsiz - 1) /segsiz));
18541         } else {
18542                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18543                         tp->tcp_cnt_counters[SND_OUT_ACK]++;
18544                 }
18545                 counter_u64_add(tcp_cnt_counters[SND_OUT_ACK], 1);
18546                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18547                         tp->tcp_proc_time[SND_OUT_ACK] += crtsc;
18548                 }
18549                 counter_u64_add(tcp_proc_time[SND_OUT_ACK], crtsc);
18550         }
18551         sched_unpin();
18552 #endif
18553         if (error == ENOBUFS)
18554                 error = 0;
18555         return (error);
18556 }
18557
18558 static void
18559 rack_update_seg(struct tcp_rack *rack)
18560 {
18561         uint32_t orig_val;
18562
18563         orig_val = rack->r_ctl.rc_pace_max_segs;
18564         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
18565         if (orig_val != rack->r_ctl.rc_pace_max_segs)
18566                 rack_log_pacing_delay_calc(rack, 0, 0, orig_val, 0, 0, 15, __LINE__, NULL);
18567 }
18568
18569 static void
18570 rack_mtu_change(struct tcpcb *tp)
18571 {
18572         /*
18573          * The MSS may have changed
18574          */
18575         struct tcp_rack *rack;
18576
18577         rack = (struct tcp_rack *)tp->t_fb_ptr;
18578         if (rack->r_ctl.rc_pace_min_segs != ctf_fixed_maxseg(tp)) {
18579                 /*
18580                  * The MTU has changed we need to resend everything
18581                  * since all we have sent is lost. We first fix
18582                  * up the mtu though.
18583                  */
18584                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
18585                 /* We treat this like a full retransmit timeout without the cwnd adjustment */
18586                 rack_remxt_tmr(tp);
18587                 rack->r_fast_output = 0;
18588                 rack->r_ctl.rc_out_at_rto = ctf_flight_size(tp,
18589                                                 rack->r_ctl.rc_sacked);
18590                 rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
18591                 rack->r_must_retran = 1;
18592
18593         }
18594         sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
18595         /* We don't use snd_nxt to retransmit */
18596         tp->snd_nxt = tp->snd_max;
18597 }
18598
18599 static int
18600 rack_set_profile(struct tcp_rack *rack, int prof)
18601 {
18602         int err = EINVAL;
18603         if (prof == 1) {
18604                 /* pace_always=1 */
18605                 if (rack->rc_always_pace == 0) {
18606                         if (tcp_can_enable_pacing() == 0)
18607                                 return (EBUSY);
18608                 }
18609                 rack->rc_always_pace = 1;
18610                 if (rack->use_fixed_rate || rack->gp_ready)
18611                         rack_set_cc_pacing(rack);
18612                 rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
18613                 rack->rack_attempt_hdwr_pace = 0;
18614                 /* cmpack=1 */
18615                 if (rack_use_cmp_acks)
18616                         rack->r_use_cmp_ack = 1;
18617                 if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
18618                     rack->r_use_cmp_ack)
18619                         rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
18620                 /* scwnd=1 */
18621                 rack->rack_enable_scwnd = 1;
18622                 /* dynamic=100 */
18623                 rack->rc_gp_dyn_mul = 1;
18624                 /* gp_inc_ca */
18625                 rack->r_ctl.rack_per_of_gp_ca = 100;
18626                 /* rrr_conf=3 */
18627                 rack->r_rr_config = 3;
18628                 /* npush=2 */
18629                 rack->r_ctl.rc_no_push_at_mrtt = 2;
18630                 /* fillcw=1 */
18631                 rack->rc_pace_to_cwnd = 1;
18632                 rack->rc_pace_fill_if_rttin_range = 0;
18633                 rack->rtt_limit_mul = 0;
18634                 /* noprr=1 */
18635                 rack->rack_no_prr = 1;
18636                 /* lscwnd=1 */
18637                 rack->r_limit_scw = 1;
18638                 /* gp_inc_rec */
18639                 rack->r_ctl.rack_per_of_gp_rec = 90;
18640                 err = 0;
18641
18642         } else if (prof == 3) {
18643                 /* Same as profile one execept fill_cw becomes 2 (less aggressive set) */
18644                 /* pace_always=1 */
18645                 if (rack->rc_always_pace == 0) {
18646                         if (tcp_can_enable_pacing() == 0)
18647                                 return (EBUSY);
18648                 }
18649                 rack->rc_always_pace = 1;
18650                 if (rack->use_fixed_rate || rack->gp_ready)
18651                         rack_set_cc_pacing(rack);
18652                 rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
18653                 rack->rack_attempt_hdwr_pace = 0;
18654                 /* cmpack=1 */
18655                 if (rack_use_cmp_acks)
18656                         rack->r_use_cmp_ack = 1;
18657                 if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
18658                     rack->r_use_cmp_ack)
18659                         rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
18660                 /* scwnd=1 */
18661                 rack->rack_enable_scwnd = 1;
18662                 /* dynamic=100 */
18663                 rack->rc_gp_dyn_mul = 1;
18664                 /* gp_inc_ca */
18665                 rack->r_ctl.rack_per_of_gp_ca = 100;
18666                 /* rrr_conf=3 */
18667                 rack->r_rr_config = 3;
18668                 /* npush=2 */
18669                 rack->r_ctl.rc_no_push_at_mrtt = 2;
18670                 /* fillcw=2 */
18671                 rack->rc_pace_to_cwnd = 1;
18672                 rack->r_fill_less_agg = 1;
18673                 rack->rc_pace_fill_if_rttin_range = 0;
18674                 rack->rtt_limit_mul = 0;
18675                 /* noprr=1 */
18676                 rack->rack_no_prr = 1;
18677                 /* lscwnd=1 */
18678                 rack->r_limit_scw = 1;
18679                 /* gp_inc_rec */
18680                 rack->r_ctl.rack_per_of_gp_rec = 90;
18681                 err = 0;
18682
18683
18684         } else if (prof == 2) {
18685                 /* cmpack=1 */
18686                 if (rack->rc_always_pace == 0) {
18687                         if (tcp_can_enable_pacing() == 0)
18688                                 return (EBUSY);
18689                 }
18690                 rack->rc_always_pace = 1;
18691                 if (rack->use_fixed_rate || rack->gp_ready)
18692                         rack_set_cc_pacing(rack);
18693                 rack->r_use_cmp_ack = 1;
18694                 if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
18695                         rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
18696                 /* pace_always=1 */
18697                 rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
18698                 /* scwnd=1 */
18699                 rack->rack_enable_scwnd = 1;
18700                 /* dynamic=100 */
18701                 rack->rc_gp_dyn_mul = 1;
18702                 rack->r_ctl.rack_per_of_gp_ca = 100;
18703                 /* rrr_conf=3 */
18704                 rack->r_rr_config = 3;
18705                 /* npush=2 */
18706                 rack->r_ctl.rc_no_push_at_mrtt = 2;
18707                 /* fillcw=1 */
18708                 rack->rc_pace_to_cwnd = 1;
18709                 rack->rc_pace_fill_if_rttin_range = 0;
18710                 rack->rtt_limit_mul = 0;
18711                 /* noprr=1 */
18712                 rack->rack_no_prr = 1;
18713                 /* lscwnd=0 */
18714                 rack->r_limit_scw = 0;
18715                 err = 0;
18716         } else if (prof == 0) {
18717                 /* This changes things back to the default settings */
18718                 err = 0;
18719                 if (rack->rc_always_pace) {
18720                         tcp_decrement_paced_conn();
18721                         rack_undo_cc_pacing(rack);
18722                         rack->rc_always_pace = 0;
18723                 }
18724                 if (rack_pace_every_seg && tcp_can_enable_pacing()) {
18725                         rack->rc_always_pace = 1;
18726                         if (rack->use_fixed_rate || rack->gp_ready)
18727                                 rack_set_cc_pacing(rack);
18728                 } else
18729                         rack->rc_always_pace = 0;
18730                 if (rack_use_cmp_acks)
18731                         rack->r_use_cmp_ack = 1;
18732                 else
18733                         rack->r_use_cmp_ack = 0;
18734                 if (rack_disable_prr)
18735                         rack->rack_no_prr = 1;
18736                 else
18737                         rack->rack_no_prr = 0;
18738                 if (rack_gp_no_rec_chg)
18739                         rack->rc_gp_no_rec_chg = 1;
18740                 else
18741                         rack->rc_gp_no_rec_chg = 0;
18742                 if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack) {
18743                         rack->r_mbuf_queue = 1;
18744                         if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
18745                                 rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
18746                         rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
18747                 } else {
18748                         rack->r_mbuf_queue = 0;
18749                         rack->rc_inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
18750                 }
18751                 if (rack_enable_shared_cwnd)
18752                         rack->rack_enable_scwnd = 1;
18753                 else
18754                         rack->rack_enable_scwnd = 0;
18755                 if (rack_do_dyn_mul) {
18756                         /* When dynamic adjustment is on CA needs to start at 100% */
18757                         rack->rc_gp_dyn_mul = 1;
18758                         if (rack_do_dyn_mul >= 100)
18759                                 rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
18760                 } else {
18761                         rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
18762                         rack->rc_gp_dyn_mul = 0;
18763                 }
18764                 rack->r_rr_config = 0;
18765                 rack->r_ctl.rc_no_push_at_mrtt = 0;
18766                 rack->rc_pace_to_cwnd = 0;
18767                 rack->rc_pace_fill_if_rttin_range = 0;
18768                 rack->rtt_limit_mul = 0;
18769
18770                 if (rack_enable_hw_pacing)
18771                         rack->rack_hdw_pace_ena = 1;
18772                 else
18773                         rack->rack_hdw_pace_ena = 0;
18774                 if (rack_disable_prr)
18775                         rack->rack_no_prr = 1;
18776                 else
18777                         rack->rack_no_prr = 0;
18778                 if (rack_limits_scwnd)
18779                         rack->r_limit_scw  = 1;
18780                 else
18781                         rack->r_limit_scw  = 0;
18782                 err = 0;
18783         }
18784         return (err);
18785 }
18786
18787 static int
18788 rack_add_deferred_option(struct tcp_rack *rack, int sopt_name, uint64_t loptval)
18789 {
18790         struct deferred_opt_list *dol;
18791
18792         dol = malloc(sizeof(struct deferred_opt_list),
18793                      M_TCPFSB, M_NOWAIT|M_ZERO);
18794         if (dol == NULL) {
18795                 /*
18796                  * No space yikes -- fail out..
18797                  */
18798                 return (0);
18799         }
18800         dol->optname = sopt_name;
18801         dol->optval = loptval;
18802         TAILQ_INSERT_TAIL(&rack->r_ctl.opt_list, dol, next);
18803         return (1);
18804 }
18805
18806 static int
18807 rack_process_option(struct tcpcb *tp, struct tcp_rack *rack, int sopt_name,
18808                     uint32_t optval, uint64_t loptval)
18809 {
18810         struct epoch_tracker et;
18811         struct sockopt sopt;
18812         struct cc_newreno_opts opt;
18813         uint64_t val;
18814         int error = 0;
18815         uint16_t ca, ss;
18816
18817         switch (sopt_name) {
18818
18819         case TCP_RACK_PACING_BETA:
18820                 RACK_OPTS_INC(tcp_rack_beta);
18821                 if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
18822                         /* This only works for newreno. */
18823                         error = EINVAL;
18824                         break;
18825                 }
18826                 if (rack->rc_pacing_cc_set) {
18827                         /*
18828                          * Set them into the real CC module
18829                          * whats in the rack pcb is the old values
18830                          * to be used on restoral/
18831                          */
18832                         sopt.sopt_dir = SOPT_SET;
18833                         opt.name = CC_NEWRENO_BETA;
18834                         opt.val = optval;
18835                         if (CC_ALGO(tp)->ctl_output != NULL)
18836                                 error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
18837                         else {
18838                                 error = ENOENT;
18839                                 break;
18840                         }
18841                 } else {
18842                         /*
18843                          * Not pacing yet so set it into our local
18844                          * rack pcb storage.
18845                          */
18846                         rack->r_ctl.rc_saved_beta.beta = optval;
18847                 }
18848                 break;
18849         case TCP_RACK_PACING_BETA_ECN:
18850                 RACK_OPTS_INC(tcp_rack_beta_ecn);
18851                 if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
18852                         /* This only works for newreno. */
18853                         error = EINVAL;
18854                         break;
18855                 }
18856                 if (rack->rc_pacing_cc_set) {
18857                         /*
18858                          * Set them into the real CC module
18859                          * whats in the rack pcb is the old values
18860                          * to be used on restoral/
18861                          */
18862                         sopt.sopt_dir = SOPT_SET;
18863                         opt.name = CC_NEWRENO_BETA_ECN;
18864                         opt.val = optval;
18865                         if (CC_ALGO(tp)->ctl_output != NULL)
18866                                 error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
18867                         else
18868                                 error = ENOENT;
18869                 } else {
18870                         /*
18871                          * Not pacing yet so set it into our local
18872                          * rack pcb storage.
18873                          */
18874                         rack->r_ctl.rc_saved_beta.beta_ecn = optval;
18875                         rack->r_ctl.rc_saved_beta.newreno_flags = CC_NEWRENO_BETA_ECN;
18876                 }
18877                 break;
18878         case TCP_DEFER_OPTIONS:
18879                 RACK_OPTS_INC(tcp_defer_opt);
18880                 if (optval) {
18881                         if (rack->gp_ready) {
18882                                 /* Too late */
18883                                 error = EINVAL;
18884                                 break;
18885                         }
18886                         rack->defer_options = 1;
18887                 } else
18888                         rack->defer_options = 0;
18889                 break;
18890         case TCP_RACK_MEASURE_CNT:
18891                 RACK_OPTS_INC(tcp_rack_measure_cnt);
18892                 if (optval && (optval <= 0xff)) {
18893                         rack->r_ctl.req_measurements = optval;
18894                 } else
18895                         error = EINVAL;
18896                 break;
18897         case TCP_REC_ABC_VAL:
18898                 RACK_OPTS_INC(tcp_rec_abc_val);
18899                 if (optval > 0)
18900                         rack->r_use_labc_for_rec = 1;
18901                 else
18902                         rack->r_use_labc_for_rec = 0;
18903                 break;
18904         case TCP_RACK_ABC_VAL:
18905                 RACK_OPTS_INC(tcp_rack_abc_val);
18906                 if ((optval > 0) && (optval < 255))
18907                         rack->rc_labc = optval;
18908                 else
18909                         error = EINVAL;
18910                 break;
18911         case TCP_HDWR_UP_ONLY:
18912                 RACK_OPTS_INC(tcp_pacing_up_only);
18913                 if (optval)
18914                         rack->r_up_only = 1;
18915                 else
18916                         rack->r_up_only = 0;
18917                 break;
18918         case TCP_PACING_RATE_CAP:
18919                 RACK_OPTS_INC(tcp_pacing_rate_cap);
18920                 rack->r_ctl.bw_rate_cap = loptval;
18921                 break;
18922         case TCP_RACK_PROFILE:
18923                 RACK_OPTS_INC(tcp_profile);
18924                 error = rack_set_profile(rack, optval);
18925                 break;
18926         case TCP_USE_CMP_ACKS:
18927                 RACK_OPTS_INC(tcp_use_cmp_acks);
18928                 if ((optval == 0) && (rack->rc_inp->inp_flags2 & INP_MBUF_ACKCMP)) {
18929                         /* You can't turn it off once its on! */
18930                         error = EINVAL;
18931                 } else if ((optval == 1) && (rack->r_use_cmp_ack == 0)) {
18932                         rack->r_use_cmp_ack = 1;
18933                         rack->r_mbuf_queue = 1;
18934                         tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
18935                 }
18936                 if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
18937                         rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
18938                 break;
18939         case TCP_SHARED_CWND_TIME_LIMIT:
18940                 RACK_OPTS_INC(tcp_lscwnd);
18941                 if (optval)
18942                         rack->r_limit_scw = 1;
18943                 else
18944                         rack->r_limit_scw = 0;
18945                 break;
18946         case TCP_RACK_PACE_TO_FILL:
18947                 RACK_OPTS_INC(tcp_fillcw);
18948                 if (optval == 0)
18949                         rack->rc_pace_to_cwnd = 0;
18950                 else {
18951                         rack->rc_pace_to_cwnd = 1;
18952                         if (optval > 1)
18953                                 rack->r_fill_less_agg = 1;
18954                 }
18955                 if ((optval >= rack_gp_rtt_maxmul) &&
18956                     rack_gp_rtt_maxmul &&
18957                     (optval < 0xf)) {
18958                         rack->rc_pace_fill_if_rttin_range = 1;
18959                         rack->rtt_limit_mul = optval;
18960                 } else {
18961                         rack->rc_pace_fill_if_rttin_range = 0;
18962                         rack->rtt_limit_mul = 0;
18963                 }
18964                 break;
18965         case TCP_RACK_NO_PUSH_AT_MAX:
18966                 RACK_OPTS_INC(tcp_npush);
18967                 if (optval == 0)
18968                         rack->r_ctl.rc_no_push_at_mrtt = 0;
18969                 else if (optval < 0xff)
18970                         rack->r_ctl.rc_no_push_at_mrtt = optval;
18971                 else
18972                         error = EINVAL;
18973                 break;
18974         case TCP_SHARED_CWND_ENABLE:
18975                 RACK_OPTS_INC(tcp_rack_scwnd);
18976                 if (optval == 0)
18977                         rack->rack_enable_scwnd = 0;
18978                 else
18979                         rack->rack_enable_scwnd = 1;
18980                 break;
18981         case TCP_RACK_MBUF_QUEUE:
18982                 /* Now do we use the LRO mbuf-queue feature */
18983                 RACK_OPTS_INC(tcp_rack_mbufq);
18984                 if (optval || rack->r_use_cmp_ack)
18985                         rack->r_mbuf_queue = 1;
18986                 else
18987                         rack->r_mbuf_queue = 0;
18988                 if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
18989                         tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
18990                 else
18991                         tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
18992                 break;
18993         case TCP_RACK_NONRXT_CFG_RATE:
18994                 RACK_OPTS_INC(tcp_rack_cfg_rate);
18995                 if (optval == 0)
18996                         rack->rack_rec_nonrxt_use_cr = 0;
18997                 else
18998                         rack->rack_rec_nonrxt_use_cr = 1;
18999                 break;
19000         case TCP_NO_PRR:
19001                 RACK_OPTS_INC(tcp_rack_noprr);
19002                 if (optval == 0)
19003                         rack->rack_no_prr = 0;
19004                 else if (optval == 1)
19005                         rack->rack_no_prr = 1;
19006                 else if (optval == 2)
19007                         rack->no_prr_addback = 1;
19008                 else
19009                         error = EINVAL;
19010                 break;
19011         case TCP_TIMELY_DYN_ADJ:
19012                 RACK_OPTS_INC(tcp_timely_dyn);
19013                 if (optval == 0)
19014                         rack->rc_gp_dyn_mul = 0;
19015                 else {
19016                         rack->rc_gp_dyn_mul = 1;
19017                         if (optval >= 100) {
19018                                 /*
19019                                  * If the user sets something 100 or more
19020                                  * its the gp_ca value.
19021                                  */
19022                                 rack->r_ctl.rack_per_of_gp_ca  = optval;
19023                         }
19024                 }
19025                 break;
19026         case TCP_RACK_DO_DETECTION:
19027                 RACK_OPTS_INC(tcp_rack_do_detection);
19028                 if (optval == 0)
19029                         rack->do_detection = 0;
19030                 else
19031                         rack->do_detection = 1;
19032                 break;
19033         case TCP_RACK_TLP_USE:
19034                 if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) {
19035                         error = EINVAL;
19036                         break;
19037                 }
19038                 RACK_OPTS_INC(tcp_tlp_use);
19039                 rack->rack_tlp_threshold_use = optval;
19040                 break;
19041         case TCP_RACK_TLP_REDUCE:
19042                 /* RACK TLP cwnd reduction (bool) */
19043                 RACK_OPTS_INC(tcp_rack_tlp_reduce);
19044                 rack->r_ctl.rc_tlp_cwnd_reduce = optval;
19045                 break;
19046         /*  Pacing related ones */
19047         case TCP_RACK_PACE_ALWAYS:
19048                 /*
19049                  * zero is old rack method, 1 is new
19050                  * method using a pacing rate.
19051                  */
19052                 RACK_OPTS_INC(tcp_rack_pace_always);
19053                 if (optval > 0) {
19054                         if (rack->rc_always_pace) {
19055                                 error = EALREADY;
19056                                 break;
19057                         } else if (tcp_can_enable_pacing()) {
19058                                 rack->rc_always_pace = 1;
19059                                 if (rack->use_fixed_rate || rack->gp_ready)
19060                                         rack_set_cc_pacing(rack);
19061                         }
19062                         else {
19063                                 error = ENOSPC;
19064                                 break;
19065                         }
19066                 } else {
19067                         if (rack->rc_always_pace) {
19068                                 tcp_decrement_paced_conn();
19069                                 rack->rc_always_pace = 0;
19070                                 rack_undo_cc_pacing(rack);
19071                         }
19072                 }
19073                 if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
19074                         tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19075                 else
19076                         tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19077                 /* A rate may be set irate or other, if so set seg size */
19078                 rack_update_seg(rack);
19079                 break;
19080         case TCP_BBR_RACK_INIT_RATE:
19081                 RACK_OPTS_INC(tcp_initial_rate);
19082                 val = optval;
19083                 /* Change from kbits per second to bytes per second */
19084                 val *= 1000;
19085                 val /= 8;
19086                 rack->r_ctl.init_rate = val;
19087                 if (rack->rc_init_win != rack_default_init_window) {
19088                         uint32_t win, snt;
19089
19090                         /*
19091                          * Options don't always get applied
19092                          * in the order you think. So in order
19093                          * to assure we update a cwnd we need
19094                          * to check and see if we are still
19095                          * where we should raise the cwnd.
19096                          */
19097                         win = rc_init_window(rack);
19098                         if (SEQ_GT(tp->snd_max, tp->iss))
19099                                 snt = tp->snd_max - tp->iss;
19100                         else
19101                                 snt = 0;
19102                         if ((snt < win) &&
19103                             (tp->snd_cwnd < win))
19104                                 tp->snd_cwnd = win;
19105                 }
19106                 if (rack->rc_always_pace)
19107                         rack_update_seg(rack);
19108                 break;
19109         case TCP_BBR_IWINTSO:
19110                 RACK_OPTS_INC(tcp_initial_win);
19111                 if (optval && (optval <= 0xff)) {
19112                         uint32_t win, snt;
19113
19114                         rack->rc_init_win = optval;
19115                         win = rc_init_window(rack);
19116                         if (SEQ_GT(tp->snd_max, tp->iss))
19117                                 snt = tp->snd_max - tp->iss;
19118                         else
19119                                 snt = 0;
19120                         if ((snt < win) &&
19121                             (tp->t_srtt |
19122 #ifdef NETFLIX_PEAKRATE
19123                              tp->t_maxpeakrate |
19124 #endif
19125                              rack->r_ctl.init_rate)) {
19126                                 /*
19127                                  * We are not past the initial window
19128                                  * and we have some bases for pacing,
19129                                  * so we need to possibly adjust up
19130                                  * the cwnd. Note even if we don't set
19131                                  * the cwnd, its still ok to raise the rc_init_win
19132                                  * which can be used coming out of idle when we
19133                                  * would have a rate.
19134                                  */
19135                                 if (tp->snd_cwnd < win)
19136                                         tp->snd_cwnd = win;
19137                         }
19138                         if (rack->rc_always_pace)
19139                                 rack_update_seg(rack);
19140                 } else
19141                         error = EINVAL;
19142                 break;
19143         case TCP_RACK_FORCE_MSEG:
19144                 RACK_OPTS_INC(tcp_rack_force_max_seg);
19145                 if (optval)
19146                         rack->rc_force_max_seg = 1;
19147                 else
19148                         rack->rc_force_max_seg = 0;
19149                 break;
19150         case TCP_RACK_PACE_MAX_SEG:
19151                 /* Max segments size in a pace in bytes */
19152                 RACK_OPTS_INC(tcp_rack_max_seg);
19153                 rack->rc_user_set_max_segs = optval;
19154                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
19155                 break;
19156         case TCP_RACK_PACE_RATE_REC:
19157                 /* Set the fixed pacing rate in Bytes per second ca */
19158                 RACK_OPTS_INC(tcp_rack_pace_rate_rec);
19159                 rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19160                 if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
19161                         rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19162                 if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
19163                         rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19164                 rack->use_fixed_rate = 1;
19165                 if (rack->rc_always_pace)
19166                         rack_set_cc_pacing(rack);
19167                 rack_log_pacing_delay_calc(rack,
19168                                            rack->r_ctl.rc_fixed_pacing_rate_ss,
19169                                            rack->r_ctl.rc_fixed_pacing_rate_ca,
19170                                            rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19171                                            __LINE__, NULL);
19172                 break;
19173
19174         case TCP_RACK_PACE_RATE_SS:
19175                 /* Set the fixed pacing rate in Bytes per second ca */
19176                 RACK_OPTS_INC(tcp_rack_pace_rate_ss);
19177                 rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19178                 if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
19179                         rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19180                 if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
19181                         rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19182                 rack->use_fixed_rate = 1;
19183                 if (rack->rc_always_pace)
19184                         rack_set_cc_pacing(rack);
19185                 rack_log_pacing_delay_calc(rack,
19186                                            rack->r_ctl.rc_fixed_pacing_rate_ss,
19187                                            rack->r_ctl.rc_fixed_pacing_rate_ca,
19188                                            rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19189                                            __LINE__, NULL);
19190                 break;
19191
19192         case TCP_RACK_PACE_RATE_CA:
19193                 /* Set the fixed pacing rate in Bytes per second ca */
19194                 RACK_OPTS_INC(tcp_rack_pace_rate_ca);
19195                 rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19196                 if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
19197                         rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19198                 if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
19199                         rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19200                 rack->use_fixed_rate = 1;
19201                 if (rack->rc_always_pace)
19202                         rack_set_cc_pacing(rack);
19203                 rack_log_pacing_delay_calc(rack,
19204                                            rack->r_ctl.rc_fixed_pacing_rate_ss,
19205                                            rack->r_ctl.rc_fixed_pacing_rate_ca,
19206                                            rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19207                                            __LINE__, NULL);
19208                 break;
19209         case TCP_RACK_GP_INCREASE_REC:
19210                 RACK_OPTS_INC(tcp_gp_inc_rec);
19211                 rack->r_ctl.rack_per_of_gp_rec = optval;
19212                 rack_log_pacing_delay_calc(rack,
19213                                            rack->r_ctl.rack_per_of_gp_ss,
19214                                            rack->r_ctl.rack_per_of_gp_ca,
19215                                            rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
19216                                            __LINE__, NULL);
19217                 break;
19218         case TCP_RACK_GP_INCREASE_CA:
19219                 RACK_OPTS_INC(tcp_gp_inc_ca);
19220                 ca = optval;
19221                 if (ca < 100) {
19222                         /*
19223                          * We don't allow any reduction
19224                          * over the GP b/w.
19225                          */
19226                         error = EINVAL;
19227                         break;
19228                 }
19229                 rack->r_ctl.rack_per_of_gp_ca = ca;
19230                 rack_log_pacing_delay_calc(rack,
19231                                            rack->r_ctl.rack_per_of_gp_ss,
19232                                            rack->r_ctl.rack_per_of_gp_ca,
19233                                            rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
19234                                            __LINE__, NULL);
19235                 break;
19236         case TCP_RACK_GP_INCREASE_SS:
19237                 RACK_OPTS_INC(tcp_gp_inc_ss);
19238                 ss = optval;
19239                 if (ss < 100) {
19240                         /*
19241                          * We don't allow any reduction
19242                          * over the GP b/w.
19243                          */
19244                         error = EINVAL;
19245                         break;
19246                 }
19247                 rack->r_ctl.rack_per_of_gp_ss = ss;
19248                 rack_log_pacing_delay_calc(rack,
19249                                            rack->r_ctl.rack_per_of_gp_ss,
19250                                            rack->r_ctl.rack_per_of_gp_ca,
19251                                            rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
19252                                            __LINE__, NULL);
19253                 break;
19254         case TCP_RACK_RR_CONF:
19255                 RACK_OPTS_INC(tcp_rack_rrr_no_conf_rate);
19256                 if (optval && optval <= 3)
19257                         rack->r_rr_config = optval;
19258                 else
19259                         rack->r_rr_config = 0;
19260                 break;
19261         case TCP_HDWR_RATE_CAP:
19262                 RACK_OPTS_INC(tcp_hdwr_rate_cap);
19263                 if (optval) {
19264                         if (rack->r_rack_hw_rate_caps == 0)
19265                                 rack->r_rack_hw_rate_caps = 1;
19266                         else
19267                                 error = EALREADY;
19268                 } else {
19269                         rack->r_rack_hw_rate_caps = 0;
19270                 }
19271                 break;
19272         case TCP_BBR_HDWR_PACE:
19273                 RACK_OPTS_INC(tcp_hdwr_pacing);
19274                 if (optval){
19275                         if (rack->rack_hdrw_pacing == 0) {
19276                                 rack->rack_hdw_pace_ena = 1;
19277                                 rack->rack_attempt_hdwr_pace = 0;
19278                         } else
19279                                 error = EALREADY;
19280                 } else {
19281                         rack->rack_hdw_pace_ena = 0;
19282 #ifdef RATELIMIT
19283                         if (rack->r_ctl.crte != NULL) {
19284                                 rack->rack_hdrw_pacing = 0;
19285                                 rack->rack_attempt_hdwr_pace = 0;
19286                                 tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
19287                                 rack->r_ctl.crte = NULL;
19288                         }
19289 #endif
19290                 }
19291                 break;
19292         /*  End Pacing related ones */
19293         case TCP_RACK_PRR_SENDALOT:
19294                 /* Allow PRR to send more than one seg */
19295                 RACK_OPTS_INC(tcp_rack_prr_sendalot);
19296                 rack->r_ctl.rc_prr_sendalot = optval;
19297                 break;
19298         case TCP_RACK_MIN_TO:
19299                 /* Minimum time between rack t-o's in ms */
19300                 RACK_OPTS_INC(tcp_rack_min_to);
19301                 rack->r_ctl.rc_min_to = optval;
19302                 break;
19303         case TCP_RACK_EARLY_SEG:
19304                 /* If early recovery max segments */
19305                 RACK_OPTS_INC(tcp_rack_early_seg);
19306                 rack->r_ctl.rc_early_recovery_segs = optval;
19307                 break;
19308         case TCP_RACK_REORD_THRESH:
19309                 /* RACK reorder threshold (shift amount) */
19310                 RACK_OPTS_INC(tcp_rack_reord_thresh);
19311                 if ((optval > 0) && (optval < 31))
19312                         rack->r_ctl.rc_reorder_shift = optval;
19313                 else
19314                         error = EINVAL;
19315                 break;
19316         case TCP_RACK_REORD_FADE:
19317                 /* Does reordering fade after ms time */
19318                 RACK_OPTS_INC(tcp_rack_reord_fade);
19319                 rack->r_ctl.rc_reorder_fade = optval;
19320                 break;
19321         case TCP_RACK_TLP_THRESH:
19322                 /* RACK TLP theshold i.e. srtt+(srtt/N) */
19323                 RACK_OPTS_INC(tcp_rack_tlp_thresh);
19324                 if (optval)
19325                         rack->r_ctl.rc_tlp_threshold = optval;
19326                 else
19327                         error = EINVAL;
19328                 break;
19329         case TCP_BBR_USE_RACK_RR:
19330                 RACK_OPTS_INC(tcp_rack_rr);
19331                 if (optval)
19332                         rack->use_rack_rr = 1;
19333                 else
19334                         rack->use_rack_rr = 0;
19335                 break;
19336         case TCP_FAST_RSM_HACK:
19337                 RACK_OPTS_INC(tcp_rack_fastrsm_hack);
19338                 if (optval)
19339                         rack->fast_rsm_hack = 1;
19340                 else
19341                         rack->fast_rsm_hack = 0;
19342                 break;
19343         case TCP_RACK_PKT_DELAY:
19344                 /* RACK added ms i.e. rack-rtt + reord + N */
19345                 RACK_OPTS_INC(tcp_rack_pkt_delay);
19346                 rack->r_ctl.rc_pkt_delay = optval;
19347                 break;
19348         case TCP_DELACK:
19349                 RACK_OPTS_INC(tcp_rack_delayed_ack);
19350                 if (optval == 0)
19351                         tp->t_delayed_ack = 0;
19352                 else
19353                         tp->t_delayed_ack = 1;
19354                 if (tp->t_flags & TF_DELACK) {
19355                         tp->t_flags &= ~TF_DELACK;
19356                         tp->t_flags |= TF_ACKNOW;
19357                         NET_EPOCH_ENTER(et);
19358                         rack_output(tp);
19359                         NET_EPOCH_EXIT(et);
19360                 }
19361                 break;
19362
19363         case TCP_BBR_RACK_RTT_USE:
19364                 RACK_OPTS_INC(tcp_rack_rtt_use);
19365                 if ((optval != USE_RTT_HIGH) &&
19366                     (optval != USE_RTT_LOW) &&
19367                     (optval != USE_RTT_AVG))
19368                         error = EINVAL;
19369                 else
19370                         rack->r_ctl.rc_rate_sample_method = optval;
19371                 break;
19372         case TCP_DATA_AFTER_CLOSE:
19373                 RACK_OPTS_INC(tcp_data_after_close);
19374                 if (optval)
19375                         rack->rc_allow_data_af_clo = 1;
19376                 else
19377                         rack->rc_allow_data_af_clo = 0;
19378                 break;
19379         default:
19380                 break;
19381         }
19382 #ifdef NETFLIX_STATS
19383         tcp_log_socket_option(tp, sopt_name, optval, error);
19384 #endif
19385         return (error);
19386 }
19387
19388
19389 static void
19390 rack_apply_deferred_options(struct tcp_rack *rack)
19391 {
19392         struct deferred_opt_list *dol, *sdol;
19393         uint32_t s_optval;
19394
19395         TAILQ_FOREACH_SAFE(dol, &rack->r_ctl.opt_list, next, sdol) {
19396                 TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
19397                 /* Disadvantage of deferal is you loose the error return */
19398                 s_optval = (uint32_t)dol->optval;
19399                 (void)rack_process_option(rack->rc_tp, rack, dol->optname, s_optval, dol->optval);
19400                 free(dol, M_TCPDO);
19401         }
19402 }
19403
19404 /*
19405  * rack_ctloutput() must drop the inpcb lock before performing copyin on
19406  * socket option arguments.  When it re-acquires the lock after the copy, it
19407  * has to revalidate that the connection is still valid for the socket
19408  * option.
19409  */
19410 static int
19411 rack_set_sockopt(struct socket *so, struct sockopt *sopt,
19412     struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack)
19413 {
19414         uint64_t loptval;
19415         int32_t error = 0, optval;
19416
19417         switch (sopt->sopt_name) {
19418         case TCP_RACK_TLP_REDUCE:               /*  URL:tlp_reduce */
19419         /*  Pacing related ones */
19420         case TCP_RACK_PACE_ALWAYS:              /*  URL:pace_always */
19421         case TCP_BBR_RACK_INIT_RATE:            /*  URL:irate */
19422         case TCP_BBR_IWINTSO:                   /*  URL:tso_iwin */
19423         case TCP_RACK_PACE_MAX_SEG:             /*  URL:pace_max_seg */
19424         case TCP_RACK_FORCE_MSEG:               /*  URL:force_max_seg */
19425         case TCP_RACK_PACE_RATE_CA:             /*  URL:pr_ca */
19426         case TCP_RACK_PACE_RATE_SS:             /*  URL:pr_ss*/
19427         case TCP_RACK_PACE_RATE_REC:            /*  URL:pr_rec */
19428         case TCP_RACK_GP_INCREASE_CA:           /*  URL:gp_inc_ca */
19429         case TCP_RACK_GP_INCREASE_SS:           /*  URL:gp_inc_ss */
19430         case TCP_RACK_GP_INCREASE_REC:          /*  URL:gp_inc_rec */
19431         case TCP_RACK_RR_CONF:                  /*  URL:rrr_conf */
19432         case TCP_BBR_HDWR_PACE:                 /*  URL:hdwrpace */
19433         case TCP_HDWR_RATE_CAP:                 /*  URL: hdwrcap boolean */
19434         case TCP_PACING_RATE_CAP:               /*  URL:cap-- used by side-channel */
19435         case TCP_HDWR_UP_ONLY:                  /*  URL:uponly -- hardware pacing  boolean */
19436        /* End pacing related */
19437         case TCP_FAST_RSM_HACK:                 /*  URL:frsm_hack */
19438         case TCP_DELACK:                        /*  URL:delack (in base TCP i.e. tcp_hints along with cc etc ) */
19439         case TCP_RACK_PRR_SENDALOT:             /*  URL:prr_sendalot */
19440         case TCP_RACK_MIN_TO:                   /*  URL:min_to */
19441         case TCP_RACK_EARLY_SEG:                /*  URL:early_seg */
19442         case TCP_RACK_REORD_THRESH:             /*  URL:reord_thresh */
19443         case TCP_RACK_REORD_FADE:               /*  URL:reord_fade */
19444         case TCP_RACK_TLP_THRESH:               /*  URL:tlp_thresh */
19445         case TCP_RACK_PKT_DELAY:                /*  URL:pkt_delay */
19446         case TCP_RACK_TLP_USE:                  /*  URL:tlp_use */
19447         case TCP_BBR_RACK_RTT_USE:              /*  URL:rttuse */
19448         case TCP_BBR_USE_RACK_RR:               /*  URL:rackrr */
19449         case TCP_RACK_DO_DETECTION:             /*  URL:detect */
19450         case TCP_NO_PRR:                        /*  URL:noprr */
19451         case TCP_TIMELY_DYN_ADJ:                /*  URL:dynamic */
19452         case TCP_DATA_AFTER_CLOSE:              /*  no URL */
19453         case TCP_RACK_NONRXT_CFG_RATE:          /*  URL:nonrxtcr */
19454         case TCP_SHARED_CWND_ENABLE:            /*  URL:scwnd */
19455         case TCP_RACK_MBUF_QUEUE:               /*  URL:mqueue */
19456         case TCP_RACK_NO_PUSH_AT_MAX:           /*  URL:npush */
19457         case TCP_RACK_PACE_TO_FILL:             /*  URL:fillcw */
19458         case TCP_SHARED_CWND_TIME_LIMIT:        /*  URL:lscwnd */
19459         case TCP_RACK_PROFILE:                  /*  URL:profile */
19460         case TCP_USE_CMP_ACKS:                  /*  URL:cmpack */
19461         case TCP_RACK_ABC_VAL:                  /*  URL:labc */
19462         case TCP_REC_ABC_VAL:                   /*  URL:reclabc */
19463         case TCP_RACK_MEASURE_CNT:              /*  URL:measurecnt */
19464         case TCP_DEFER_OPTIONS:                 /*  URL:defer */
19465         case TCP_RACK_PACING_BETA:              /*  URL:pacing_beta */
19466         case TCP_RACK_PACING_BETA_ECN:          /*  URL:pacing_beta_ecn */
19467                 break;
19468         default:
19469                 /* Filter off all unknown options to the base stack */
19470                 return (tcp_default_ctloutput(so, sopt, inp, tp));
19471                 break;
19472         }
19473         INP_WUNLOCK(inp);
19474         if (sopt->sopt_name == TCP_PACING_RATE_CAP) {
19475                 error = sooptcopyin(sopt, &loptval, sizeof(loptval), sizeof(loptval));
19476                 /*
19477                  * We truncate it down to 32 bits for the socket-option trace this
19478                  * means rates > 34Gbps won't show right, but thats probably ok.
19479                  */
19480                 optval = (uint32_t)loptval;
19481         } else {
19482                 error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
19483                 /* Save it in 64 bit form too */
19484                 loptval = optval;
19485         }
19486         if (error)
19487                 return (error);
19488         INP_WLOCK(inp);
19489         if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
19490                 INP_WUNLOCK(inp);
19491                 return (ECONNRESET);
19492         }
19493         if (rack->defer_options && (rack->gp_ready == 0) &&
19494             (sopt->sopt_name != TCP_DEFER_OPTIONS) &&
19495             (sopt->sopt_name != TCP_RACK_PACING_BETA) &&
19496             (sopt->sopt_name != TCP_RACK_PACING_BETA_ECN) &&
19497             (sopt->sopt_name != TCP_RACK_MEASURE_CNT)) {
19498                 /* Options are beind deferred */
19499                 if (rack_add_deferred_option(rack, sopt->sopt_name, loptval)) {
19500                         INP_WUNLOCK(inp);
19501                         return (0);
19502                 } else {
19503                         /* No memory to defer, fail */
19504                         INP_WUNLOCK(inp);
19505                         return (ENOMEM);
19506                 }
19507         }
19508         error = rack_process_option(tp, rack, sopt->sopt_name, optval, loptval);
19509         INP_WUNLOCK(inp);
19510         return (error);
19511 }
19512
19513 static void
19514 rack_fill_info(struct tcpcb *tp, struct tcp_info *ti)
19515 {
19516
19517         INP_WLOCK_ASSERT(tp->t_inpcb);
19518         bzero(ti, sizeof(*ti));
19519
19520         ti->tcpi_state = tp->t_state;
19521         if ((tp->t_flags & TF_REQ_TSTMP) && (tp->t_flags & TF_RCVD_TSTMP))
19522                 ti->tcpi_options |= TCPI_OPT_TIMESTAMPS;
19523         if (tp->t_flags & TF_SACK_PERMIT)
19524                 ti->tcpi_options |= TCPI_OPT_SACK;
19525         if ((tp->t_flags & TF_REQ_SCALE) && (tp->t_flags & TF_RCVD_SCALE)) {
19526                 ti->tcpi_options |= TCPI_OPT_WSCALE;
19527                 ti->tcpi_snd_wscale = tp->snd_scale;
19528                 ti->tcpi_rcv_wscale = tp->rcv_scale;
19529         }
19530         if (tp->t_flags2 & TF2_ECN_PERMIT)
19531                 ti->tcpi_options |= TCPI_OPT_ECN;
19532         if (tp->t_flags & TF_FASTOPEN)
19533                 ti->tcpi_options |= TCPI_OPT_TFO;
19534         /* still kept in ticks is t_rcvtime */
19535         ti->tcpi_last_data_recv = ((uint32_t)ticks - tp->t_rcvtime) * tick;
19536         /* Since we hold everything in precise useconds this is easy */
19537         ti->tcpi_rtt = tp->t_srtt;
19538         ti->tcpi_rttvar = tp->t_rttvar;
19539         ti->tcpi_rto = tp->t_rxtcur;
19540         ti->tcpi_snd_ssthresh = tp->snd_ssthresh;
19541         ti->tcpi_snd_cwnd = tp->snd_cwnd;
19542         /*
19543          * FreeBSD-specific extension fields for tcp_info.
19544          */
19545         ti->tcpi_rcv_space = tp->rcv_wnd;
19546         ti->tcpi_rcv_nxt = tp->rcv_nxt;
19547         ti->tcpi_snd_wnd = tp->snd_wnd;
19548         ti->tcpi_snd_bwnd = 0;          /* Unused, kept for compat. */
19549         ti->tcpi_snd_nxt = tp->snd_nxt;
19550         ti->tcpi_snd_mss = tp->t_maxseg;
19551         ti->tcpi_rcv_mss = tp->t_maxseg;
19552         ti->tcpi_snd_rexmitpack = tp->t_sndrexmitpack;
19553         ti->tcpi_rcv_ooopack = tp->t_rcvoopack;
19554         ti->tcpi_snd_zerowin = tp->t_sndzerowin;
19555 #ifdef NETFLIX_STATS
19556         ti->tcpi_total_tlp = tp->t_sndtlppack;
19557         ti->tcpi_total_tlp_bytes = tp->t_sndtlpbyte;
19558         memcpy(&ti->tcpi_rxsyninfo, &tp->t_rxsyninfo, sizeof(struct tcpsyninfo));
19559 #endif
19560 #ifdef TCP_OFFLOAD
19561         if (tp->t_flags & TF_TOE) {
19562                 ti->tcpi_options |= TCPI_OPT_TOE;
19563                 tcp_offload_tcp_info(tp, ti);
19564         }
19565 #endif
19566 }
19567
19568 static int
19569 rack_get_sockopt(struct socket *so, struct sockopt *sopt,
19570     struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack)
19571 {
19572         int32_t error, optval;
19573         uint64_t val, loptval;
19574         struct  tcp_info ti;
19575         /*
19576          * Because all our options are either boolean or an int, we can just
19577          * pull everything into optval and then unlock and copy. If we ever
19578          * add a option that is not a int, then this will have quite an
19579          * impact to this routine.
19580          */
19581         error = 0;
19582         switch (sopt->sopt_name) {
19583         case TCP_INFO:
19584                 /* First get the info filled */
19585                 rack_fill_info(tp, &ti);
19586                 /* Fix up the rtt related fields if needed */
19587                 INP_WUNLOCK(inp);
19588                 error = sooptcopyout(sopt, &ti, sizeof ti);
19589                 return (error);
19590         /*
19591          * Beta is the congestion control value for NewReno that influences how
19592          * much of a backoff happens when loss is detected. It is normally set
19593          * to 50 for 50% i.e. the cwnd is reduced to 50% of its previous value
19594          * when you exit recovery.
19595          */
19596         case TCP_RACK_PACING_BETA:
19597                 if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0)
19598                         error = EINVAL;
19599                 else if (rack->rc_pacing_cc_set == 0)
19600                         optval = rack->r_ctl.rc_saved_beta.beta;
19601                 else {
19602                         /*
19603                          * Reach out into the CC data and report back what
19604                          * I have previously set. Yeah it looks hackish but
19605                          * we don't want to report the saved values.
19606                          */
19607                         if (tp->ccv->cc_data)
19608                                 optval = ((struct newreno *)tp->ccv->cc_data)->beta;
19609                         else
19610                                 error = EINVAL;
19611                 }
19612                 break;
19613                 /*
19614                  * Beta_ecn is the congestion control value for NewReno that influences how
19615                  * much of a backoff happens when a ECN mark is detected. It is normally set
19616                  * to 80 for 80% i.e. the cwnd is reduced by 20% of its previous value when
19617                  * you exit recovery. Note that classic ECN has a beta of 50, it is only
19618                  * ABE Ecn that uses this "less" value, but we do too with pacing :)
19619                  */
19620
19621         case TCP_RACK_PACING_BETA_ECN:
19622                 if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0)
19623                         error = EINVAL;
19624                 else if (rack->rc_pacing_cc_set == 0)
19625                         optval = rack->r_ctl.rc_saved_beta.beta_ecn;
19626                 else {
19627                         /*
19628                          * Reach out into the CC data and report back what
19629                          * I have previously set. Yeah it looks hackish but
19630                          * we don't want to report the saved values.
19631                          */
19632                         if (tp->ccv->cc_data)
19633                                 optval = ((struct newreno *)tp->ccv->cc_data)->beta_ecn;
19634                         else
19635                                 error = EINVAL;
19636                 }
19637                 break;
19638         case TCP_FAST_RSM_HACK:
19639                 optval = rack->fast_rsm_hack;
19640                 break;
19641         case TCP_DEFER_OPTIONS:
19642                 optval = rack->defer_options;
19643                 break;
19644         case TCP_RACK_MEASURE_CNT:
19645                 optval = rack->r_ctl.req_measurements;
19646                 break;
19647         case TCP_REC_ABC_VAL:
19648                 optval = rack->r_use_labc_for_rec;
19649                 break;
19650         case TCP_RACK_ABC_VAL:
19651                 optval = rack->rc_labc;
19652                 break;
19653         case TCP_HDWR_UP_ONLY:
19654                 optval= rack->r_up_only;
19655                 break;
19656         case TCP_PACING_RATE_CAP:
19657                 loptval = rack->r_ctl.bw_rate_cap;
19658                 break;
19659         case TCP_RACK_PROFILE:
19660                 /* You cannot retrieve a profile, its write only */
19661                 error = EINVAL;
19662                 break;
19663         case TCP_USE_CMP_ACKS:
19664                 optval = rack->r_use_cmp_ack;
19665                 break;
19666         case TCP_RACK_PACE_TO_FILL:
19667                 optval = rack->rc_pace_to_cwnd;
19668                 if (optval && rack->r_fill_less_agg)
19669                         optval++;
19670                 break;
19671         case TCP_RACK_NO_PUSH_AT_MAX:
19672                 optval = rack->r_ctl.rc_no_push_at_mrtt;
19673                 break;
19674         case TCP_SHARED_CWND_ENABLE:
19675                 optval = rack->rack_enable_scwnd;
19676                 break;
19677         case TCP_RACK_NONRXT_CFG_RATE:
19678                 optval = rack->rack_rec_nonrxt_use_cr;
19679                 break;
19680         case TCP_NO_PRR:
19681                 if (rack->rack_no_prr  == 1)
19682                         optval = 1;
19683                 else if (rack->no_prr_addback == 1)
19684                         optval = 2;
19685                 else
19686                         optval = 0;
19687                 break;
19688         case TCP_RACK_DO_DETECTION:
19689                 optval = rack->do_detection;
19690                 break;
19691         case TCP_RACK_MBUF_QUEUE:
19692                 /* Now do we use the LRO mbuf-queue feature */
19693                 optval = rack->r_mbuf_queue;
19694                 break;
19695         case TCP_TIMELY_DYN_ADJ:
19696                 optval = rack->rc_gp_dyn_mul;
19697                 break;
19698         case TCP_BBR_IWINTSO:
19699                 optval = rack->rc_init_win;
19700                 break;
19701         case TCP_RACK_TLP_REDUCE:
19702                 /* RACK TLP cwnd reduction (bool) */
19703                 optval = rack->r_ctl.rc_tlp_cwnd_reduce;
19704                 break;
19705         case TCP_BBR_RACK_INIT_RATE:
19706                 val = rack->r_ctl.init_rate;
19707                 /* convert to kbits per sec */
19708                 val *= 8;
19709                 val /= 1000;
19710                 optval = (uint32_t)val;
19711                 break;
19712         case TCP_RACK_FORCE_MSEG:
19713                 optval = rack->rc_force_max_seg;
19714                 break;
19715         case TCP_RACK_PACE_MAX_SEG:
19716                 /* Max segments in a pace */
19717                 optval = rack->rc_user_set_max_segs;
19718                 break;
19719         case TCP_RACK_PACE_ALWAYS:
19720                 /* Use the always pace method */
19721                 optval = rack->rc_always_pace;
19722                 break;
19723         case TCP_RACK_PRR_SENDALOT:
19724                 /* Allow PRR to send more than one seg */
19725                 optval = rack->r_ctl.rc_prr_sendalot;
19726                 break;
19727         case TCP_RACK_MIN_TO:
19728                 /* Minimum time between rack t-o's in ms */
19729                 optval = rack->r_ctl.rc_min_to;
19730                 break;
19731         case TCP_RACK_EARLY_SEG:
19732                 /* If early recovery max segments */
19733                 optval = rack->r_ctl.rc_early_recovery_segs;
19734                 break;
19735         case TCP_RACK_REORD_THRESH:
19736                 /* RACK reorder threshold (shift amount) */
19737                 optval = rack->r_ctl.rc_reorder_shift;
19738                 break;
19739         case TCP_RACK_REORD_FADE:
19740                 /* Does reordering fade after ms time */
19741                 optval = rack->r_ctl.rc_reorder_fade;
19742                 break;
19743         case TCP_BBR_USE_RACK_RR:
19744                 /* Do we use the rack cheat for rxt */
19745                 optval = rack->use_rack_rr;
19746                 break;
19747         case TCP_RACK_RR_CONF:
19748                 optval = rack->r_rr_config;
19749                 break;
19750         case TCP_HDWR_RATE_CAP:
19751                 optval = rack->r_rack_hw_rate_caps;
19752                 break;
19753         case TCP_BBR_HDWR_PACE:
19754                 optval = rack->rack_hdw_pace_ena;
19755                 break;
19756         case TCP_RACK_TLP_THRESH:
19757                 /* RACK TLP theshold i.e. srtt+(srtt/N) */
19758                 optval = rack->r_ctl.rc_tlp_threshold;
19759                 break;
19760         case TCP_RACK_PKT_DELAY:
19761                 /* RACK added ms i.e. rack-rtt + reord + N */
19762                 optval = rack->r_ctl.rc_pkt_delay;
19763                 break;
19764         case TCP_RACK_TLP_USE:
19765                 optval = rack->rack_tlp_threshold_use;
19766                 break;
19767         case TCP_RACK_PACE_RATE_CA:
19768                 optval = rack->r_ctl.rc_fixed_pacing_rate_ca;
19769                 break;
19770         case TCP_RACK_PACE_RATE_SS:
19771                 optval = rack->r_ctl.rc_fixed_pacing_rate_ss;
19772                 break;
19773         case TCP_RACK_PACE_RATE_REC:
19774                 optval = rack->r_ctl.rc_fixed_pacing_rate_rec;
19775                 break;
19776         case TCP_RACK_GP_INCREASE_SS:
19777                 optval = rack->r_ctl.rack_per_of_gp_ca;
19778                 break;
19779         case TCP_RACK_GP_INCREASE_CA:
19780                 optval = rack->r_ctl.rack_per_of_gp_ss;
19781                 break;
19782         case TCP_BBR_RACK_RTT_USE:
19783                 optval = rack->r_ctl.rc_rate_sample_method;
19784                 break;
19785         case TCP_DELACK:
19786                 optval = tp->t_delayed_ack;
19787                 break;
19788         case TCP_DATA_AFTER_CLOSE:
19789                 optval = rack->rc_allow_data_af_clo;
19790                 break;
19791         case TCP_SHARED_CWND_TIME_LIMIT:
19792                 optval = rack->r_limit_scw;
19793                 break;
19794         default:
19795                 return (tcp_default_ctloutput(so, sopt, inp, tp));
19796                 break;
19797         }
19798         INP_WUNLOCK(inp);
19799         if (error == 0) {
19800                 if (TCP_PACING_RATE_CAP)
19801                         error = sooptcopyout(sopt, &loptval, sizeof loptval);
19802                 else
19803                         error = sooptcopyout(sopt, &optval, sizeof optval);
19804         }
19805         return (error);
19806 }
19807
19808 static int
19809 rack_ctloutput(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp)
19810 {
19811         int32_t error = EINVAL;
19812         struct tcp_rack *rack;
19813
19814         rack = (struct tcp_rack *)tp->t_fb_ptr;
19815         if (rack == NULL) {
19816                 /* Huh? */
19817                 goto out;
19818         }
19819         if (sopt->sopt_dir == SOPT_SET) {
19820                 return (rack_set_sockopt(so, sopt, inp, tp, rack));
19821         } else if (sopt->sopt_dir == SOPT_GET) {
19822                 return (rack_get_sockopt(so, sopt, inp, tp, rack));
19823         }
19824 out:
19825         INP_WUNLOCK(inp);
19826         return (error);
19827 }
19828
19829 static int
19830 rack_pru_options(struct tcpcb *tp, int flags)
19831 {
19832         if (flags & PRUS_OOB)
19833                 return (EOPNOTSUPP);
19834         return (0);
19835 }
19836
19837 static struct tcp_function_block __tcp_rack = {
19838         .tfb_tcp_block_name = __XSTRING(STACKNAME),
19839         .tfb_tcp_output = rack_output,
19840         .tfb_do_queued_segments = ctf_do_queued_segments,
19841         .tfb_do_segment_nounlock = rack_do_segment_nounlock,
19842         .tfb_tcp_do_segment = rack_do_segment,
19843         .tfb_tcp_ctloutput = rack_ctloutput,
19844         .tfb_tcp_fb_init = rack_init,
19845         .tfb_tcp_fb_fini = rack_fini,
19846         .tfb_tcp_timer_stop_all = rack_stopall,
19847         .tfb_tcp_timer_activate = rack_timer_activate,
19848         .tfb_tcp_timer_active = rack_timer_active,
19849         .tfb_tcp_timer_stop = rack_timer_stop,
19850         .tfb_tcp_rexmit_tmr = rack_remxt_tmr,
19851         .tfb_tcp_handoff_ok = rack_handoff_ok,
19852         .tfb_tcp_mtu_chg = rack_mtu_change,
19853         .tfb_pru_options = rack_pru_options,
19854
19855 };
19856
19857 static const char *rack_stack_names[] = {
19858         __XSTRING(STACKNAME),
19859 #ifdef STACKALIAS
19860         __XSTRING(STACKALIAS),
19861 #endif
19862 };
19863
19864 static int
19865 rack_ctor(void *mem, int32_t size, void *arg, int32_t how)
19866 {
19867         memset(mem, 0, size);
19868         return (0);
19869 }
19870
19871 static void
19872 rack_dtor(void *mem, int32_t size, void *arg)
19873 {
19874
19875 }
19876
19877 static bool rack_mod_inited = false;
19878
19879 static int
19880 tcp_addrack(module_t mod, int32_t type, void *data)
19881 {
19882         int32_t err = 0;
19883         int num_stacks;
19884
19885         switch (type) {
19886         case MOD_LOAD:
19887                 rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
19888                     sizeof(struct rack_sendmap),
19889                     rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0);
19890
19891                 rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
19892                     sizeof(struct tcp_rack),
19893                     rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
19894
19895                 sysctl_ctx_init(&rack_sysctl_ctx);
19896                 rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
19897                     SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
19898                     OID_AUTO,
19899 #ifdef STACKALIAS
19900                     __XSTRING(STACKALIAS),
19901 #else
19902                     __XSTRING(STACKNAME),
19903 #endif
19904                     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
19905                     "");
19906                 if (rack_sysctl_root == NULL) {
19907                         printf("Failed to add sysctl node\n");
19908                         err = EFAULT;
19909                         goto free_uma;
19910                 }
19911                 rack_init_sysctls();
19912                 num_stacks = nitems(rack_stack_names);
19913                 err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK,
19914                     rack_stack_names, &num_stacks);
19915                 if (err) {
19916                         printf("Failed to register %s stack name for "
19917                             "%s module\n", rack_stack_names[num_stacks],
19918                             __XSTRING(MODNAME));
19919                         sysctl_ctx_free(&rack_sysctl_ctx);
19920 free_uma:
19921                         uma_zdestroy(rack_zone);
19922                         uma_zdestroy(rack_pcb_zone);
19923                         rack_counter_destroy();
19924                         printf("Failed to register rack module -- err:%d\n", err);
19925                         return (err);
19926                 }
19927                 tcp_lro_reg_mbufq();
19928                 rack_mod_inited = true;
19929                 break;
19930         case MOD_QUIESCE:
19931                 err = deregister_tcp_functions(&__tcp_rack, true, false);
19932                 break;
19933         case MOD_UNLOAD:
19934                 err = deregister_tcp_functions(&__tcp_rack, false, true);
19935                 if (err == EBUSY)
19936                         break;
19937                 if (rack_mod_inited) {
19938                         uma_zdestroy(rack_zone);
19939                         uma_zdestroy(rack_pcb_zone);
19940                         sysctl_ctx_free(&rack_sysctl_ctx);
19941                         rack_counter_destroy();
19942                         rack_mod_inited = false;
19943                 }
19944                 tcp_lro_dereg_mbufq();
19945                 err = 0;
19946                 break;
19947         default:
19948                 return (EOPNOTSUPP);
19949         }
19950         return (err);
19951 }
19952
19953 static moduledata_t tcp_rack = {
19954         .name = __XSTRING(MODNAME),
19955         .evhand = tcp_addrack,
19956         .priv = 0
19957 };
19958
19959 MODULE_VERSION(MODNAME, 1);
19960 DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
19961 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);