1 /*- 2 * Copyright (c) 2016-2018 3 * Netflix Inc. All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice, this list of conditions and the following disclaimer. 10 * 2. Redistributions in binary form must reproduce the above copyright 11 * notice, this list of conditions and the following disclaimer in the 12 * documentation and/or other materials provided with the distribution. 13 * 14 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 17 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 24 * SUCH DAMAGE. 25 * 26 */ 27 28 #include <sys/cdefs.h> 29 __FBSDID("$FreeBSD$"); 30 31 #include "opt_inet.h" 32 #include "opt_inet6.h" 33 #include "opt_ipsec.h" 34 #include "opt_tcpdebug.h" 35 36 #include <sys/param.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 NETFLIX_STATS 53 #include <sys/stats.h> 54 #endif 55 #include <sys/refcount.h> 56 #include <sys/queue.h> 57 #include <sys/smp.h> 58 #include <sys/kthread.h> 59 #include <sys/kern_prefetch.h> 60 61 #include <vm/uma.h> 62 63 #include <net/route.h> 64 #include <net/vnet.h> 65 66 #define TCPSTATES /* for logging */ 67 68 #include <netinet/in.h> 69 #include <netinet/in_kdtrace.h> 70 #include <netinet/in_pcb.h> 71 #include <netinet/ip.h> 72 #include <netinet/ip_icmp.h> /* required for icmp_var.h */ 73 #include <netinet/icmp_var.h> /* for ICMP_BANDLIM */ 74 #include <netinet/ip_var.h> 75 #include <netinet/ip6.h> 76 #include <netinet6/in6_pcb.h> 77 #include <netinet6/ip6_var.h> 78 #include <netinet/tcp.h> 79 #define TCPOUTFLAGS 80 #include <netinet/tcp_fsm.h> 81 #include <netinet/tcp_log_buf.h> 82 #include <netinet/tcp_seq.h> 83 #include <netinet/tcp_timer.h> 84 #include <netinet/tcp_var.h> 85 #include <netinet/tcp_hpts.h> 86 #include <netinet/tcpip.h> 87 #include <netinet/cc/cc.h> 88 #ifdef NETFLIX_CWV 89 #include <netinet/tcp_newcwv.h> 90 #endif 91 #include <netinet/tcp_fastopen.h> 92 #ifdef TCPDEBUG 93 #include <netinet/tcp_debug.h> 94 #endif /* TCPDEBUG */ 95 #ifdef TCP_OFFLOAD 96 #include <netinet/tcp_offload.h> 97 #endif 98 #ifdef INET6 99 #include <netinet6/tcp6_var.h> 100 #endif 101 102 #include <netipsec/ipsec_support.h> 103 104 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 105 #include <netipsec/ipsec.h> 106 #include <netipsec/ipsec6.h> 107 #endif /* IPSEC */ 108 109 #include <netinet/udp.h> 110 #include <netinet/udp_var.h> 111 #include <machine/in_cksum.h> 112 113 #ifdef MAC 114 #include <security/mac/mac_framework.h> 115 #endif 116 #include "sack_filter.h" 117 #include "tcp_rack.h" 118 #include "rack_bbr_common.h" 119 120 uma_zone_t rack_zone; 121 uma_zone_t rack_pcb_zone; 122 123 #ifndef TICKS2SBT 124 #define TICKS2SBT(__t) (tick_sbt * ((sbintime_t)(__t))) 125 #endif 126 127 struct sysctl_ctx_list rack_sysctl_ctx; 128 struct sysctl_oid *rack_sysctl_root; 129 130 #define CUM_ACKED 1 131 #define SACKED 2 132 133 /* 134 * The RACK module incorporates a number of 135 * TCP ideas that have been put out into the IETF 136 * over the last few years: 137 * - Matt Mathis's Rate Halving which slowly drops 138 * the congestion window so that the ack clock can 139 * be maintained during a recovery. 140 * - Yuchung Cheng's RACK TCP (for which its named) that 141 * will stop us using the number of dup acks and instead 142 * use time as the gage of when we retransmit. 143 * - Reorder Detection of RFC4737 and the Tail-Loss probe draft 144 * of Dukkipati et.al. 145 * RACK depends on SACK, so if an endpoint arrives that 146 * cannot do SACK the state machine below will shuttle the 147 * connection back to using the "default" TCP stack that is 148 * in FreeBSD. 149 * 150 * To implement RACK the original TCP stack was first decomposed 151 * into a functional state machine with individual states 152 * for each of the possible TCP connection states. The do_segement 153 * functions role in life is to mandate the connection supports SACK 154 * initially and then assure that the RACK state matches the conenction 155 * state before calling the states do_segment function. Each 156 * state is simplified due to the fact that the original do_segment 157 * has been decomposed and we *know* what state we are in (no 158 * switches on the state) and all tests for SACK are gone. This 159 * greatly simplifies what each state does. 160 * 161 * TCP output is also over-written with a new version since it 162 * must maintain the new rack scoreboard. 163 * 164 */ 165 static int32_t rack_precache = 1; 166 static int32_t rack_tlp_thresh = 1; 167 static int32_t rack_reorder_thresh = 2; 168 static int32_t rack_reorder_fade = 60000; /* 0 - never fade, def 60,000 169 * - 60 seconds */ 170 static int32_t rack_pkt_delay = 1; 171 static int32_t rack_inc_var = 0;/* For TLP */ 172 static int32_t rack_reduce_largest_on_idle = 0; 173 static int32_t rack_min_pace_time = 0; 174 static int32_t rack_min_pace_time_seg_req=6; 175 static int32_t rack_early_recovery = 1; 176 static int32_t rack_early_recovery_max_seg = 6; 177 static int32_t rack_send_a_lot_in_prr = 1; 178 static int32_t rack_min_to = 1; /* Number of ms minimum timeout */ 179 static int32_t rack_tlp_in_recovery = 1; /* Can we do TLP in recovery? */ 180 static int32_t rack_verbose_logging = 0; 181 static int32_t rack_ignore_data_after_close = 1; 182 /* 183 * Currently regular tcp has a rto_min of 30ms 184 * the backoff goes 12 times so that ends up 185 * being a total of 122.850 seconds before a 186 * connection is killed. 187 */ 188 static int32_t rack_tlp_min = 10; 189 static int32_t rack_rto_min = 30; /* 30ms same as main freebsd */ 190 static int32_t rack_rto_max = 30000; /* 30 seconds */ 191 static const int32_t rack_free_cache = 2; 192 static int32_t rack_hptsi_segments = 40; 193 static int32_t rack_rate_sample_method = USE_RTT_LOW; 194 static int32_t rack_pace_every_seg = 1; 195 static int32_t rack_delayed_ack_time = 200; /* 200ms */ 196 static int32_t rack_slot_reduction = 4; 197 static int32_t rack_lower_cwnd_at_tlp = 0; 198 static int32_t rack_use_proportional_reduce = 0; 199 static int32_t rack_proportional_rate = 10; 200 static int32_t rack_tlp_max_resend = 2; 201 static int32_t rack_limited_retran = 0; 202 static int32_t rack_always_send_oldest = 0; 203 static int32_t rack_sack_block_limit = 128; 204 static int32_t rack_use_sack_filter = 1; 205 static int32_t rack_tlp_threshold_use = TLP_USE_TWO_ONE; 206 207 /* Rack specific counters */ 208 counter_u64_t rack_badfr; 209 counter_u64_t rack_badfr_bytes; 210 counter_u64_t rack_rtm_prr_retran; 211 counter_u64_t rack_rtm_prr_newdata; 212 counter_u64_t rack_timestamp_mismatch; 213 counter_u64_t rack_reorder_seen; 214 counter_u64_t rack_paced_segments; 215 counter_u64_t rack_unpaced_segments; 216 counter_u64_t rack_saw_enobuf; 217 counter_u64_t rack_saw_enetunreach; 218 219 /* Tail loss probe counters */ 220 counter_u64_t rack_tlp_tot; 221 counter_u64_t rack_tlp_newdata; 222 counter_u64_t rack_tlp_retran; 223 counter_u64_t rack_tlp_retran_bytes; 224 counter_u64_t rack_tlp_retran_fail; 225 counter_u64_t rack_to_tot; 226 counter_u64_t rack_to_arm_rack; 227 counter_u64_t rack_to_arm_tlp; 228 counter_u64_t rack_to_alloc; 229 counter_u64_t rack_to_alloc_hard; 230 counter_u64_t rack_to_alloc_emerg; 231 232 counter_u64_t rack_sack_proc_all; 233 counter_u64_t rack_sack_proc_short; 234 counter_u64_t rack_sack_proc_restart; 235 counter_u64_t rack_runt_sacks; 236 counter_u64_t rack_used_tlpmethod; 237 counter_u64_t rack_used_tlpmethod2; 238 counter_u64_t rack_enter_tlp_calc; 239 counter_u64_t rack_input_idle_reduces; 240 counter_u64_t rack_tlp_does_nada; 241 242 /* Temp CPU counters */ 243 counter_u64_t rack_find_high; 244 245 counter_u64_t rack_progress_drops; 246 counter_u64_t rack_out_size[TCP_MSS_ACCT_SIZE]; 247 counter_u64_t rack_opts_arry[RACK_OPTS_SIZE]; 248 249 static void 250 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick, int event, int line); 251 252 static int 253 rack_process_ack(struct mbuf *m, struct tcphdr *th, 254 struct socket *so, struct tcpcb *tp, struct tcpopt *to, 255 uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t * ret_val); 256 static int 257 rack_process_data(struct mbuf *m, struct tcphdr *th, 258 struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, 259 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); 260 static void 261 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, 262 struct tcphdr *th, uint16_t nsegs, uint16_t type, int32_t recovery); 263 static struct rack_sendmap *rack_alloc(struct tcp_rack *rack); 264 static struct rack_sendmap * 265 rack_check_recovery_mode(struct tcpcb *tp, 266 uint32_t tsused); 267 static void 268 rack_cong_signal(struct tcpcb *tp, struct tcphdr *th, 269 uint32_t type); 270 static void rack_counter_destroy(void); 271 static int 272 rack_ctloutput(struct socket *so, struct sockopt *sopt, 273 struct inpcb *inp, struct tcpcb *tp); 274 static int32_t rack_ctor(void *mem, int32_t size, void *arg, int32_t how); 275 static void 276 rack_do_segment(struct mbuf *m, struct tcphdr *th, 277 struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, 278 uint8_t iptos); 279 static void rack_dtor(void *mem, int32_t size, void *arg); 280 static void 281 rack_earlier_retran(struct tcpcb *tp, struct rack_sendmap *rsm, 282 uint32_t t, uint32_t cts); 283 static struct rack_sendmap * 284 rack_find_high_nonack(struct tcp_rack *rack, 285 struct rack_sendmap *rsm); 286 static struct rack_sendmap *rack_find_lowest_rsm(struct tcp_rack *rack); 287 static void rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm); 288 static void rack_fini(struct tcpcb *tp, int32_t tcb_is_purged); 289 static int 290 rack_get_sockopt(struct socket *so, struct sockopt *sopt, 291 struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack); 292 static int32_t rack_handoff_ok(struct tcpcb *tp); 293 static int32_t rack_init(struct tcpcb *tp); 294 static void rack_init_sysctls(void); 295 static void 296 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, 297 struct tcphdr *th); 298 static void 299 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len, 300 uint32_t seq_out, uint8_t th_flags, int32_t err, uint32_t ts, 301 uint8_t pass, struct rack_sendmap *hintrsm); 302 static void 303 rack_log_sack_passed(struct tcpcb *tp, struct tcp_rack *rack, 304 struct rack_sendmap *rsm); 305 static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num); 306 static int32_t rack_output(struct tcpcb *tp); 307 static void 308 rack_hpts_do_segment(struct mbuf *m, struct tcphdr *th, 309 struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, 310 uint8_t iptos, int32_t nxt_pkt, struct timeval *tv); 311 312 static uint32_t 313 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, 314 struct sackblk *sack, struct tcpopt *to, struct rack_sendmap **prsm, 315 uint32_t cts); 316 static void rack_post_recovery(struct tcpcb *tp, struct tcphdr *th); 317 static void rack_remxt_tmr(struct tcpcb *tp); 318 static int 319 rack_set_sockopt(struct socket *so, struct sockopt *sopt, 320 struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack); 321 static void rack_set_state(struct tcpcb *tp, struct tcp_rack *rack); 322 static int32_t rack_stopall(struct tcpcb *tp); 323 static void 324 rack_timer_activate(struct tcpcb *tp, uint32_t timer_type, 325 uint32_t delta); 326 static int32_t rack_timer_active(struct tcpcb *tp, uint32_t timer_type); 327 static void rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line); 328 static void rack_timer_stop(struct tcpcb *tp, uint32_t timer_type); 329 static uint32_t 330 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack, 331 struct rack_sendmap *rsm, uint32_t ts, int32_t * lenp); 332 static void 333 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack, 334 struct rack_sendmap *rsm, uint32_t ts); 335 static int 336 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack, 337 struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type); 338 static int32_t tcp_addrack(module_t mod, int32_t type, void *data); 339 static void 340 rack_challenge_ack(struct mbuf *m, struct tcphdr *th, 341 struct tcpcb *tp, int32_t * ret_val); 342 static int 343 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, 344 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, 345 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); 346 static int 347 rack_do_closing(struct mbuf *m, struct tcphdr *th, 348 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, 349 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); 350 static void 351 rack_do_drop(struct mbuf *m, struct tcpcb *tp); 352 static void 353 rack_do_dropafterack(struct mbuf *m, struct tcpcb *tp, 354 struct tcphdr *th, int32_t thflags, int32_t tlen, int32_t * ret_val); 355 static void 356 rack_do_dropwithreset(struct mbuf *m, struct tcpcb *tp, 357 struct tcphdr *th, int32_t rstreason, int32_t tlen); 358 static int 359 rack_do_established(struct mbuf *m, struct tcphdr *th, 360 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, 361 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); 362 static int 363 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, 364 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, 365 int32_t tlen, uint32_t tiwin, int32_t nxt_pkt); 366 static int 367 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, 368 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, 369 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); 370 static int 371 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, 372 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, 373 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); 374 static int 375 rack_do_lastack(struct mbuf *m, struct tcphdr *th, 376 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, 377 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); 378 static int 379 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, 380 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, 381 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); 382 static int 383 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, 384 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, 385 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); 386 static int 387 rack_drop_checks(struct tcpopt *to, struct mbuf *m, 388 struct tcphdr *th, struct tcpcb *tp, int32_t * tlenp, int32_t * thf, 389 int32_t * drop_hdrlen, int32_t * ret_val); 390 static int 391 rack_process_rst(struct mbuf *m, struct tcphdr *th, 392 struct socket *so, struct tcpcb *tp); 393 struct rack_sendmap * 394 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, 395 uint32_t tsused); 396 static void tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt); 397 static void 398 tcp_rack_partialack(struct tcpcb *tp, struct tcphdr *th); 399 400 static int 401 rack_ts_check(struct mbuf *m, struct tcphdr *th, 402 struct tcpcb *tp, int32_t tlen, int32_t thflags, int32_t * ret_val); 403 404 int32_t rack_clear_counter=0; 405 406 407 static int 408 sysctl_rack_clear(SYSCTL_HANDLER_ARGS) 409 { 410 uint32_t stat; 411 int32_t error; 412 413 error = SYSCTL_OUT(req, &rack_clear_counter, sizeof(uint32_t)); 414 if (error || req->newptr == NULL) 415 return error; 416 417 error = SYSCTL_IN(req, &stat, sizeof(uint32_t)); 418 if (error) 419 return (error); 420 if (stat == 1) { 421 #ifdef INVARIANTS 422 printf("Clearing RACK counters\n"); 423 #endif 424 counter_u64_zero(rack_badfr); 425 counter_u64_zero(rack_badfr_bytes); 426 counter_u64_zero(rack_rtm_prr_retran); 427 counter_u64_zero(rack_rtm_prr_newdata); 428 counter_u64_zero(rack_timestamp_mismatch); 429 counter_u64_zero(rack_reorder_seen); 430 counter_u64_zero(rack_tlp_tot); 431 counter_u64_zero(rack_tlp_newdata); 432 counter_u64_zero(rack_tlp_retran); 433 counter_u64_zero(rack_tlp_retran_bytes); 434 counter_u64_zero(rack_tlp_retran_fail); 435 counter_u64_zero(rack_to_tot); 436 counter_u64_zero(rack_to_arm_rack); 437 counter_u64_zero(rack_to_arm_tlp); 438 counter_u64_zero(rack_paced_segments); 439 counter_u64_zero(rack_unpaced_segments); 440 counter_u64_zero(rack_saw_enobuf); 441 counter_u64_zero(rack_saw_enetunreach); 442 counter_u64_zero(rack_to_alloc_hard); 443 counter_u64_zero(rack_to_alloc_emerg); 444 counter_u64_zero(rack_sack_proc_all); 445 counter_u64_zero(rack_sack_proc_short); 446 counter_u64_zero(rack_sack_proc_restart); 447 counter_u64_zero(rack_to_alloc); 448 counter_u64_zero(rack_find_high); 449 counter_u64_zero(rack_runt_sacks); 450 counter_u64_zero(rack_used_tlpmethod); 451 counter_u64_zero(rack_used_tlpmethod2); 452 counter_u64_zero(rack_enter_tlp_calc); 453 counter_u64_zero(rack_progress_drops); 454 counter_u64_zero(rack_tlp_does_nada); 455 } 456 rack_clear_counter = 0; 457 return (0); 458 } 459 460 461 462 static void 463 rack_init_sysctls() 464 { 465 SYSCTL_ADD_S32(&rack_sysctl_ctx, 466 SYSCTL_CHILDREN(rack_sysctl_root), 467 OID_AUTO, "rate_sample_method", CTLFLAG_RW, 468 &rack_rate_sample_method , USE_RTT_LOW, 469 "What method should we use for rate sampling 0=high, 1=low "); 470 SYSCTL_ADD_S32(&rack_sysctl_ctx, 471 SYSCTL_CHILDREN(rack_sysctl_root), 472 OID_AUTO, "data_after_close", CTLFLAG_RW, 473 &rack_ignore_data_after_close, 0, 474 "Do we hold off sending a RST until all pending data is ack'd"); 475 SYSCTL_ADD_S32(&rack_sysctl_ctx, 476 SYSCTL_CHILDREN(rack_sysctl_root), 477 OID_AUTO, "tlpmethod", CTLFLAG_RW, 478 &rack_tlp_threshold_use, TLP_USE_TWO_ONE, 479 "What method do we do for TLP time calc 0=no-de-ack-comp, 1=ID, 2=2.1, 3=2.2"); 480 SYSCTL_ADD_S32(&rack_sysctl_ctx, 481 SYSCTL_CHILDREN(rack_sysctl_root), 482 OID_AUTO, "min_pace_time", CTLFLAG_RW, 483 &rack_min_pace_time, 0, 484 "Should we enforce a minimum pace time of 1ms"); 485 SYSCTL_ADD_S32(&rack_sysctl_ctx, 486 SYSCTL_CHILDREN(rack_sysctl_root), 487 OID_AUTO, "min_pace_segs", CTLFLAG_RW, 488 &rack_min_pace_time_seg_req, 6, 489 "How many segments have to be in the len to enforce min-pace-time"); 490 SYSCTL_ADD_S32(&rack_sysctl_ctx, 491 SYSCTL_CHILDREN(rack_sysctl_root), 492 OID_AUTO, "idle_reduce_high", CTLFLAG_RW, 493 &rack_reduce_largest_on_idle, 0, 494 "Should we reduce the largest cwnd seen to IW on idle reduction"); 495 SYSCTL_ADD_S32(&rack_sysctl_ctx, 496 SYSCTL_CHILDREN(rack_sysctl_root), 497 OID_AUTO, "bb_verbose", CTLFLAG_RW, 498 &rack_verbose_logging, 0, 499 "Should RACK black box logging be verbose"); 500 SYSCTL_ADD_S32(&rack_sysctl_ctx, 501 SYSCTL_CHILDREN(rack_sysctl_root), 502 OID_AUTO, "sackfiltering", CTLFLAG_RW, 503 &rack_use_sack_filter, 1, 504 "Do we use sack filtering?"); 505 SYSCTL_ADD_S32(&rack_sysctl_ctx, 506 SYSCTL_CHILDREN(rack_sysctl_root), 507 OID_AUTO, "delayed_ack", CTLFLAG_RW, 508 &rack_delayed_ack_time, 200, 509 "Delayed ack time (200ms)"); 510 SYSCTL_ADD_S32(&rack_sysctl_ctx, 511 SYSCTL_CHILDREN(rack_sysctl_root), 512 OID_AUTO, "tlpminto", CTLFLAG_RW, 513 &rack_tlp_min, 10, 514 "TLP minimum timeout per the specification (10ms)"); 515 SYSCTL_ADD_S32(&rack_sysctl_ctx, 516 SYSCTL_CHILDREN(rack_sysctl_root), 517 OID_AUTO, "precache", CTLFLAG_RW, 518 &rack_precache, 0, 519 "Where should we precache the mcopy (0 is not at all)"); 520 SYSCTL_ADD_S32(&rack_sysctl_ctx, 521 SYSCTL_CHILDREN(rack_sysctl_root), 522 OID_AUTO, "sblklimit", CTLFLAG_RW, 523 &rack_sack_block_limit, 128, 524 "When do we start paying attention to small sack blocks"); 525 SYSCTL_ADD_S32(&rack_sysctl_ctx, 526 SYSCTL_CHILDREN(rack_sysctl_root), 527 OID_AUTO, "send_oldest", CTLFLAG_RW, 528 &rack_always_send_oldest, 1, 529 "Should we always send the oldest TLP and RACK-TLP"); 530 SYSCTL_ADD_S32(&rack_sysctl_ctx, 531 SYSCTL_CHILDREN(rack_sysctl_root), 532 OID_AUTO, "rack_tlp_in_recovery", CTLFLAG_RW, 533 &rack_tlp_in_recovery, 1, 534 "Can we do a TLP during recovery?"); 535 SYSCTL_ADD_S32(&rack_sysctl_ctx, 536 SYSCTL_CHILDREN(rack_sysctl_root), 537 OID_AUTO, "rack_tlimit", CTLFLAG_RW, 538 &rack_limited_retran, 0, 539 "How many times can a rack timeout drive out sends"); 540 SYSCTL_ADD_S32(&rack_sysctl_ctx, 541 SYSCTL_CHILDREN(rack_sysctl_root), 542 OID_AUTO, "minrto", CTLFLAG_RW, 543 &rack_rto_min, 0, 544 "Minimum RTO in ms -- set with caution below 1000 due to TLP"); 545 SYSCTL_ADD_S32(&rack_sysctl_ctx, 546 SYSCTL_CHILDREN(rack_sysctl_root), 547 OID_AUTO, "maxrto", CTLFLAG_RW, 548 &rack_rto_max, 0, 549 "Maxiumum RTO in ms -- should be at least as large as min_rto"); 550 SYSCTL_ADD_S32(&rack_sysctl_ctx, 551 SYSCTL_CHILDREN(rack_sysctl_root), 552 OID_AUTO, "tlp_retry", CTLFLAG_RW, 553 &rack_tlp_max_resend, 2, 554 "How many times does TLP retry a single segment or multiple with no ACK"); 555 SYSCTL_ADD_S32(&rack_sysctl_ctx, 556 SYSCTL_CHILDREN(rack_sysctl_root), 557 OID_AUTO, "recovery_loss_prop", CTLFLAG_RW, 558 &rack_use_proportional_reduce, 0, 559 "Should we proportionaly reduce cwnd based on the number of losses "); 560 SYSCTL_ADD_S32(&rack_sysctl_ctx, 561 SYSCTL_CHILDREN(rack_sysctl_root), 562 OID_AUTO, "recovery_prop", CTLFLAG_RW, 563 &rack_proportional_rate, 10, 564 "What percent reduction per loss"); 565 SYSCTL_ADD_S32(&rack_sysctl_ctx, 566 SYSCTL_CHILDREN(rack_sysctl_root), 567 OID_AUTO, "tlp_cwnd_flag", CTLFLAG_RW, 568 &rack_lower_cwnd_at_tlp, 0, 569 "When a TLP completes a retran should we enter recovery?"); 570 SYSCTL_ADD_S32(&rack_sysctl_ctx, 571 SYSCTL_CHILDREN(rack_sysctl_root), 572 OID_AUTO, "hptsi_reduces", CTLFLAG_RW, 573 &rack_slot_reduction, 4, 574 "When setting a slot should we reduce by divisor"); 575 SYSCTL_ADD_S32(&rack_sysctl_ctx, 576 SYSCTL_CHILDREN(rack_sysctl_root), 577 OID_AUTO, "hptsi_every_seg", CTLFLAG_RW, 578 &rack_pace_every_seg, 1, 579 "Should we pace out every segment hptsi"); 580 SYSCTL_ADD_S32(&rack_sysctl_ctx, 581 SYSCTL_CHILDREN(rack_sysctl_root), 582 OID_AUTO, "hptsi_seg_max", CTLFLAG_RW, 583 &rack_hptsi_segments, 6, 584 "Should we pace out only a limited size of segments"); 585 SYSCTL_ADD_S32(&rack_sysctl_ctx, 586 SYSCTL_CHILDREN(rack_sysctl_root), 587 OID_AUTO, "prr_sendalot", CTLFLAG_RW, 588 &rack_send_a_lot_in_prr, 1, 589 "Send a lot in prr"); 590 SYSCTL_ADD_S32(&rack_sysctl_ctx, 591 SYSCTL_CHILDREN(rack_sysctl_root), 592 OID_AUTO, "minto", CTLFLAG_RW, 593 &rack_min_to, 1, 594 "Minimum rack timeout in milliseconds"); 595 SYSCTL_ADD_S32(&rack_sysctl_ctx, 596 SYSCTL_CHILDREN(rack_sysctl_root), 597 OID_AUTO, "earlyrecoveryseg", CTLFLAG_RW, 598 &rack_early_recovery_max_seg, 6, 599 "Max segments in early recovery"); 600 SYSCTL_ADD_S32(&rack_sysctl_ctx, 601 SYSCTL_CHILDREN(rack_sysctl_root), 602 OID_AUTO, "earlyrecovery", CTLFLAG_RW, 603 &rack_early_recovery, 1, 604 "Do we do early recovery with rack"); 605 SYSCTL_ADD_S32(&rack_sysctl_ctx, 606 SYSCTL_CHILDREN(rack_sysctl_root), 607 OID_AUTO, "reorder_thresh", CTLFLAG_RW, 608 &rack_reorder_thresh, 2, 609 "What factor for rack will be added when seeing reordering (shift right)"); 610 SYSCTL_ADD_S32(&rack_sysctl_ctx, 611 SYSCTL_CHILDREN(rack_sysctl_root), 612 OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW, 613 &rack_tlp_thresh, 1, 614 "what divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)"); 615 SYSCTL_ADD_S32(&rack_sysctl_ctx, 616 SYSCTL_CHILDREN(rack_sysctl_root), 617 OID_AUTO, "reorder_fade", CTLFLAG_RW, 618 &rack_reorder_fade, 0, 619 "Does reorder detection fade, if so how many ms (0 means never)"); 620 SYSCTL_ADD_S32(&rack_sysctl_ctx, 621 SYSCTL_CHILDREN(rack_sysctl_root), 622 OID_AUTO, "pktdelay", CTLFLAG_RW, 623 &rack_pkt_delay, 1, 624 "Extra RACK time (in ms) besides reordering thresh"); 625 SYSCTL_ADD_S32(&rack_sysctl_ctx, 626 SYSCTL_CHILDREN(rack_sysctl_root), 627 OID_AUTO, "inc_var", CTLFLAG_RW, 628 &rack_inc_var, 0, 629 "Should rack add to the TLP timer the variance in rtt calculation"); 630 rack_badfr = counter_u64_alloc(M_WAITOK); 631 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 632 SYSCTL_CHILDREN(rack_sysctl_root), 633 OID_AUTO, "badfr", CTLFLAG_RD, 634 &rack_badfr, "Total number of bad FRs"); 635 rack_badfr_bytes = counter_u64_alloc(M_WAITOK); 636 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 637 SYSCTL_CHILDREN(rack_sysctl_root), 638 OID_AUTO, "badfr_bytes", CTLFLAG_RD, 639 &rack_badfr_bytes, "Total number of bad FRs"); 640 rack_rtm_prr_retran = counter_u64_alloc(M_WAITOK); 641 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 642 SYSCTL_CHILDREN(rack_sysctl_root), 643 OID_AUTO, "prrsndret", CTLFLAG_RD, 644 &rack_rtm_prr_retran, 645 "Total number of prr based retransmits"); 646 rack_rtm_prr_newdata = counter_u64_alloc(M_WAITOK); 647 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 648 SYSCTL_CHILDREN(rack_sysctl_root), 649 OID_AUTO, "prrsndnew", CTLFLAG_RD, 650 &rack_rtm_prr_newdata, 651 "Total number of prr based new transmits"); 652 rack_timestamp_mismatch = counter_u64_alloc(M_WAITOK); 653 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 654 SYSCTL_CHILDREN(rack_sysctl_root), 655 OID_AUTO, "tsnf", CTLFLAG_RD, 656 &rack_timestamp_mismatch, 657 "Total number of timestamps that we could not find the reported ts"); 658 rack_find_high = counter_u64_alloc(M_WAITOK); 659 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 660 SYSCTL_CHILDREN(rack_sysctl_root), 661 OID_AUTO, "findhigh", CTLFLAG_RD, 662 &rack_find_high, 663 "Total number of FIN causing find-high"); 664 rack_reorder_seen = counter_u64_alloc(M_WAITOK); 665 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 666 SYSCTL_CHILDREN(rack_sysctl_root), 667 OID_AUTO, "reordering", CTLFLAG_RD, 668 &rack_reorder_seen, 669 "Total number of times we added delay due to reordering"); 670 rack_tlp_tot = counter_u64_alloc(M_WAITOK); 671 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 672 SYSCTL_CHILDREN(rack_sysctl_root), 673 OID_AUTO, "tlp_to_total", CTLFLAG_RD, 674 &rack_tlp_tot, 675 "Total number of tail loss probe expirations"); 676 rack_tlp_newdata = counter_u64_alloc(M_WAITOK); 677 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 678 SYSCTL_CHILDREN(rack_sysctl_root), 679 OID_AUTO, "tlp_new", CTLFLAG_RD, 680 &rack_tlp_newdata, 681 "Total number of tail loss probe sending new data"); 682 683 rack_tlp_retran = counter_u64_alloc(M_WAITOK); 684 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 685 SYSCTL_CHILDREN(rack_sysctl_root), 686 OID_AUTO, "tlp_retran", CTLFLAG_RD, 687 &rack_tlp_retran, 688 "Total number of tail loss probe sending retransmitted data"); 689 rack_tlp_retran_bytes = counter_u64_alloc(M_WAITOK); 690 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 691 SYSCTL_CHILDREN(rack_sysctl_root), 692 OID_AUTO, "tlp_retran_bytes", CTLFLAG_RD, 693 &rack_tlp_retran_bytes, 694 "Total bytes of tail loss probe sending retransmitted data"); 695 rack_tlp_retran_fail = counter_u64_alloc(M_WAITOK); 696 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 697 SYSCTL_CHILDREN(rack_sysctl_root), 698 OID_AUTO, "tlp_retran_fail", CTLFLAG_RD, 699 &rack_tlp_retran_fail, 700 "Total number of tail loss probe sending retransmitted data that failed (wait for t3)"); 701 rack_to_tot = counter_u64_alloc(M_WAITOK); 702 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 703 SYSCTL_CHILDREN(rack_sysctl_root), 704 OID_AUTO, "rack_to_tot", CTLFLAG_RD, 705 &rack_to_tot, 706 "Total number of times the rack to expired?"); 707 rack_to_arm_rack = counter_u64_alloc(M_WAITOK); 708 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 709 SYSCTL_CHILDREN(rack_sysctl_root), 710 OID_AUTO, "arm_rack", CTLFLAG_RD, 711 &rack_to_arm_rack, 712 "Total number of times the rack timer armed?"); 713 rack_to_arm_tlp = counter_u64_alloc(M_WAITOK); 714 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 715 SYSCTL_CHILDREN(rack_sysctl_root), 716 OID_AUTO, "arm_tlp", CTLFLAG_RD, 717 &rack_to_arm_tlp, 718 "Total number of times the tlp timer armed?"); 719 rack_paced_segments = counter_u64_alloc(M_WAITOK); 720 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 721 SYSCTL_CHILDREN(rack_sysctl_root), 722 OID_AUTO, "paced", CTLFLAG_RD, 723 &rack_paced_segments, 724 "Total number of times a segment send caused hptsi"); 725 rack_unpaced_segments = counter_u64_alloc(M_WAITOK); 726 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 727 SYSCTL_CHILDREN(rack_sysctl_root), 728 OID_AUTO, "unpaced", CTLFLAG_RD, 729 &rack_unpaced_segments, 730 "Total number of times a segment did not cause hptsi"); 731 rack_saw_enobuf = counter_u64_alloc(M_WAITOK); 732 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 733 SYSCTL_CHILDREN(rack_sysctl_root), 734 OID_AUTO, "saw_enobufs", CTLFLAG_RD, 735 &rack_saw_enobuf, 736 "Total number of times a segment did not cause hptsi"); 737 rack_saw_enetunreach = counter_u64_alloc(M_WAITOK); 738 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 739 SYSCTL_CHILDREN(rack_sysctl_root), 740 OID_AUTO, "saw_enetunreach", CTLFLAG_RD, 741 &rack_saw_enetunreach, 742 "Total number of times a segment did not cause hptsi"); 743 rack_to_alloc = counter_u64_alloc(M_WAITOK); 744 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 745 SYSCTL_CHILDREN(rack_sysctl_root), 746 OID_AUTO, "allocs", CTLFLAG_RD, 747 &rack_to_alloc, 748 "Total allocations of tracking structures"); 749 rack_to_alloc_hard = counter_u64_alloc(M_WAITOK); 750 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 751 SYSCTL_CHILDREN(rack_sysctl_root), 752 OID_AUTO, "allochard", CTLFLAG_RD, 753 &rack_to_alloc_hard, 754 "Total allocations done with sleeping the hard way"); 755 rack_to_alloc_emerg = counter_u64_alloc(M_WAITOK); 756 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 757 SYSCTL_CHILDREN(rack_sysctl_root), 758 OID_AUTO, "allocemerg", CTLFLAG_RD, 759 &rack_to_alloc_emerg, 760 "Total alocations done from emergency cache"); 761 rack_sack_proc_all = counter_u64_alloc(M_WAITOK); 762 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 763 SYSCTL_CHILDREN(rack_sysctl_root), 764 OID_AUTO, "sack_long", CTLFLAG_RD, 765 &rack_sack_proc_all, 766 "Total times we had to walk whole list for sack processing"); 767 768 rack_sack_proc_restart = counter_u64_alloc(M_WAITOK); 769 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 770 SYSCTL_CHILDREN(rack_sysctl_root), 771 OID_AUTO, "sack_restart", CTLFLAG_RD, 772 &rack_sack_proc_restart, 773 "Total times we had to walk whole list due to a restart"); 774 rack_sack_proc_short = counter_u64_alloc(M_WAITOK); 775 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 776 SYSCTL_CHILDREN(rack_sysctl_root), 777 OID_AUTO, "sack_short", CTLFLAG_RD, 778 &rack_sack_proc_short, 779 "Total times we took shortcut for sack processing"); 780 rack_enter_tlp_calc = counter_u64_alloc(M_WAITOK); 781 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 782 SYSCTL_CHILDREN(rack_sysctl_root), 783 OID_AUTO, "tlp_calc_entered", CTLFLAG_RD, 784 &rack_enter_tlp_calc, 785 "Total times we called calc-tlp"); 786 rack_used_tlpmethod = counter_u64_alloc(M_WAITOK); 787 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 788 SYSCTL_CHILDREN(rack_sysctl_root), 789 OID_AUTO, "hit_tlp_method", CTLFLAG_RD, 790 &rack_used_tlpmethod, 791 "Total number of runt sacks"); 792 rack_used_tlpmethod2 = counter_u64_alloc(M_WAITOK); 793 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 794 SYSCTL_CHILDREN(rack_sysctl_root), 795 OID_AUTO, "hit_tlp_method2", CTLFLAG_RD, 796 &rack_used_tlpmethod2, 797 "Total number of runt sacks 2"); 798 rack_runt_sacks = counter_u64_alloc(M_WAITOK); 799 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 800 SYSCTL_CHILDREN(rack_sysctl_root), 801 OID_AUTO, "runtsacks", CTLFLAG_RD, 802 &rack_runt_sacks, 803 "Total number of runt sacks"); 804 rack_progress_drops = counter_u64_alloc(M_WAITOK); 805 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 806 SYSCTL_CHILDREN(rack_sysctl_root), 807 OID_AUTO, "prog_drops", CTLFLAG_RD, 808 &rack_progress_drops, 809 "Total number of progress drops"); 810 rack_input_idle_reduces = counter_u64_alloc(M_WAITOK); 811 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 812 SYSCTL_CHILDREN(rack_sysctl_root), 813 OID_AUTO, "idle_reduce_oninput", CTLFLAG_RD, 814 &rack_input_idle_reduces, 815 "Total number of idle reductions on input"); 816 rack_tlp_does_nada = counter_u64_alloc(M_WAITOK); 817 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 818 SYSCTL_CHILDREN(rack_sysctl_root), 819 OID_AUTO, "tlp_nada", CTLFLAG_RD, 820 &rack_tlp_does_nada, 821 "Total number of nada tlp calls"); 822 COUNTER_ARRAY_ALLOC(rack_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK); 823 SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), 824 OID_AUTO, "outsize", CTLFLAG_RD, 825 rack_out_size, TCP_MSS_ACCT_SIZE, "MSS send sizes"); 826 COUNTER_ARRAY_ALLOC(rack_opts_arry, RACK_OPTS_SIZE, M_WAITOK); 827 SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), 828 OID_AUTO, "opts", CTLFLAG_RD, 829 rack_opts_arry, RACK_OPTS_SIZE, "RACK Option Stats"); 830 SYSCTL_ADD_PROC(&rack_sysctl_ctx, 831 SYSCTL_CHILDREN(rack_sysctl_root), 832 OID_AUTO, "clear", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE, 833 &rack_clear_counter, 0, sysctl_rack_clear, "IU", "Clear counters"); 834 } 835 836 static inline int32_t 837 rack_progress_timeout_check(struct tcpcb *tp) 838 { 839 if (tp->t_maxunacktime && tp->t_acktime && TSTMP_GT(ticks, tp->t_acktime)) { 840 if ((ticks - tp->t_acktime) >= tp->t_maxunacktime) { 841 /* 842 * There is an assumption that the caller 843 * will drop the connection so we will 844 * increment the counters here. 845 */ 846 struct tcp_rack *rack; 847 rack = (struct tcp_rack *)tp->t_fb_ptr; 848 counter_u64_add(rack_progress_drops, 1); 849 #ifdef NETFLIX_STATS 850 TCPSTAT_INC(tcps_progdrops); 851 #endif 852 rack_log_progress_event(rack, tp, ticks, PROGRESS_DROP, __LINE__); 853 return (1); 854 } 855 } 856 return (0); 857 } 858 859 860 static void 861 rack_log_to_start(struct tcp_rack *rack, uint32_t cts, uint32_t to, int32_t slot, uint8_t which) 862 { 863 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 864 union tcp_log_stackspecific log; 865 866 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 867 log.u_bbr.flex1 = TICKS_2_MSEC(rack->rc_tp->t_srtt >> TCP_RTT_SHIFT); 868 log.u_bbr.flex2 = to; 869 log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags; 870 log.u_bbr.flex4 = slot; 871 log.u_bbr.flex5 = rack->rc_inp->inp_hptsslot; 872 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur; 873 log.u_bbr.flex8 = which; 874 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; 875 log.u_bbr.ininput = rack->rc_inp->inp_in_input; 876 TCP_LOG_EVENT(rack->rc_tp, NULL, 877 &rack->rc_inp->inp_socket->so_rcv, 878 &rack->rc_inp->inp_socket->so_snd, 879 BBR_LOG_TIMERSTAR, 0, 880 0, &log, false); 881 } 882 } 883 884 static void 885 rack_log_to_event(struct tcp_rack *rack, int32_t to_num) 886 { 887 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 888 union tcp_log_stackspecific log; 889 890 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 891 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; 892 log.u_bbr.ininput = rack->rc_inp->inp_in_input; 893 log.u_bbr.flex8 = to_num; 894 log.u_bbr.flex1 = rack->r_ctl.rc_rack_min_rtt; 895 log.u_bbr.flex2 = rack->rc_rack_rtt; 896 TCP_LOG_EVENT(rack->rc_tp, NULL, 897 &rack->rc_inp->inp_socket->so_rcv, 898 &rack->rc_inp->inp_socket->so_snd, 899 BBR_LOG_RTO, 0, 900 0, &log, false); 901 } 902 } 903 904 static void 905 rack_log_rtt_upd(struct tcpcb *tp, struct tcp_rack *rack, int32_t t, 906 uint32_t o_srtt, uint32_t o_var) 907 { 908 if (tp->t_logstate != TCP_LOG_STATE_OFF) { 909 union tcp_log_stackspecific log; 910 911 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 912 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; 913 log.u_bbr.ininput = rack->rc_inp->inp_in_input; 914 log.u_bbr.flex1 = t; 915 log.u_bbr.flex2 = o_srtt; 916 log.u_bbr.flex3 = o_var; 917 log.u_bbr.flex4 = rack->r_ctl.rack_rs.rs_rtt_lowest; 918 log.u_bbr.flex5 = rack->r_ctl.rack_rs.rs_rtt_highest; 919 log.u_bbr.flex6 = rack->r_ctl.rack_rs.rs_rtt_cnt; 920 log.u_bbr.rttProp = rack->r_ctl.rack_rs.rs_rtt_tot; 921 log.u_bbr.flex8 = rack->r_ctl.rc_rate_sample_method; 922 TCP_LOG_EVENT(tp, NULL, 923 &rack->rc_inp->inp_socket->so_rcv, 924 &rack->rc_inp->inp_socket->so_snd, 925 BBR_LOG_BBRRTT, 0, 926 0, &log, false); 927 } 928 } 929 930 static void 931 rack_log_rtt_sample(struct tcp_rack *rack, uint32_t rtt) 932 { 933 /* 934 * Log the rtt sample we are 935 * applying to the srtt algorithm in 936 * useconds. 937 */ 938 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 939 union tcp_log_stackspecific log; 940 struct timeval tv; 941 942 /* Convert our ms to a microsecond */ 943 log.u_bbr.flex1 = rtt * 1000; 944 log.u_bbr.timeStamp = tcp_get_usecs(&tv); 945 TCP_LOG_EVENTP(rack->rc_tp, NULL, 946 &rack->rc_inp->inp_socket->so_rcv, 947 &rack->rc_inp->inp_socket->so_snd, 948 TCP_LOG_RTT, 0, 949 0, &log, false, &tv); 950 } 951 } 952 953 954 static inline void 955 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick, int event, int line) 956 { 957 if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) { 958 union tcp_log_stackspecific log; 959 960 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 961 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; 962 log.u_bbr.ininput = rack->rc_inp->inp_in_input; 963 log.u_bbr.flex1 = line; 964 log.u_bbr.flex2 = tick; 965 log.u_bbr.flex3 = tp->t_maxunacktime; 966 log.u_bbr.flex4 = tp->t_acktime; 967 log.u_bbr.flex8 = event; 968 TCP_LOG_EVENT(tp, NULL, 969 &rack->rc_inp->inp_socket->so_rcv, 970 &rack->rc_inp->inp_socket->so_snd, 971 BBR_LOG_PROGRESS, 0, 972 0, &log, false); 973 } 974 } 975 976 static void 977 rack_log_type_bbrsnd(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint32_t cts) 978 { 979 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 980 union tcp_log_stackspecific log; 981 982 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 983 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; 984 log.u_bbr.ininput = rack->rc_inp->inp_in_input; 985 log.u_bbr.flex1 = slot; 986 log.u_bbr.flex7 = (0x0000ffff & rack->r_ctl.rc_hpts_flags); 987 log.u_bbr.flex8 = rack->rc_in_persist; 988 TCP_LOG_EVENT(rack->rc_tp, NULL, 989 &rack->rc_inp->inp_socket->so_rcv, 990 &rack->rc_inp->inp_socket->so_snd, 991 BBR_LOG_BBRSND, 0, 992 0, &log, false); 993 } 994 } 995 996 static void 997 rack_log_doseg_done(struct tcp_rack *rack, uint32_t cts, int32_t nxt_pkt, int32_t did_out, int way_out) 998 { 999 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 1000 union tcp_log_stackspecific log; 1001 log.u_bbr.flex1 = did_out; 1002 log.u_bbr.flex2 = nxt_pkt; 1003 log.u_bbr.flex3 = way_out; 1004 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags; 1005 log.u_bbr.flex7 = rack->r_wanted_output; 1006 log.u_bbr.flex8 = rack->rc_in_persist; 1007 TCP_LOG_EVENT(rack->rc_tp, NULL, 1008 &rack->rc_inp->inp_socket->so_rcv, 1009 &rack->rc_inp->inp_socket->so_snd, 1010 BBR_LOG_DOSEG_DONE, 0, 1011 0, &log, false); 1012 } 1013 } 1014 1015 1016 static void 1017 rack_log_type_just_return(struct tcp_rack *rack, uint32_t cts, uint32_t tlen, uint32_t slot, uint8_t hpts_calling) 1018 { 1019 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 1020 union tcp_log_stackspecific log; 1021 1022 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 1023 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; 1024 log.u_bbr.ininput = rack->rc_inp->inp_in_input; 1025 log.u_bbr.flex1 = slot; 1026 log.u_bbr.flex2 = rack->r_ctl.rc_hpts_flags; 1027 log.u_bbr.flex7 = hpts_calling; 1028 log.u_bbr.flex8 = rack->rc_in_persist; 1029 TCP_LOG_EVENT(rack->rc_tp, NULL, 1030 &rack->rc_inp->inp_socket->so_rcv, 1031 &rack->rc_inp->inp_socket->so_snd, 1032 BBR_LOG_JUSTRET, 0, 1033 tlen, &log, false); 1034 } 1035 } 1036 1037 static void 1038 rack_log_to_cancel(struct tcp_rack *rack, int32_t hpts_removed, int line) 1039 { 1040 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 1041 union tcp_log_stackspecific log; 1042 1043 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 1044 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; 1045 log.u_bbr.ininput = rack->rc_inp->inp_in_input; 1046 log.u_bbr.flex1 = line; 1047 log.u_bbr.flex2 = 0; 1048 log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags; 1049 log.u_bbr.flex4 = 0; 1050 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur; 1051 log.u_bbr.flex8 = hpts_removed; 1052 TCP_LOG_EVENT(rack->rc_tp, NULL, 1053 &rack->rc_inp->inp_socket->so_rcv, 1054 &rack->rc_inp->inp_socket->so_snd, 1055 BBR_LOG_TIMERCANC, 0, 1056 0, &log, false); 1057 } 1058 } 1059 1060 static void 1061 rack_log_to_processing(struct tcp_rack *rack, uint32_t cts, int32_t ret, int32_t timers) 1062 { 1063 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 1064 union tcp_log_stackspecific log; 1065 1066 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 1067 log.u_bbr.flex1 = timers; 1068 log.u_bbr.flex2 = ret; 1069 log.u_bbr.flex3 = rack->r_ctl.rc_timer_exp; 1070 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags; 1071 log.u_bbr.flex5 = cts; 1072 TCP_LOG_EVENT(rack->rc_tp, NULL, 1073 &rack->rc_inp->inp_socket->so_rcv, 1074 &rack->rc_inp->inp_socket->so_snd, 1075 BBR_LOG_TO_PROCESS, 0, 1076 0, &log, false); 1077 } 1078 } 1079 1080 static void 1081 rack_counter_destroy() 1082 { 1083 counter_u64_free(rack_badfr); 1084 counter_u64_free(rack_badfr_bytes); 1085 counter_u64_free(rack_rtm_prr_retran); 1086 counter_u64_free(rack_rtm_prr_newdata); 1087 counter_u64_free(rack_timestamp_mismatch); 1088 counter_u64_free(rack_reorder_seen); 1089 counter_u64_free(rack_tlp_tot); 1090 counter_u64_free(rack_tlp_newdata); 1091 counter_u64_free(rack_tlp_retran); 1092 counter_u64_free(rack_tlp_retran_bytes); 1093 counter_u64_free(rack_tlp_retran_fail); 1094 counter_u64_free(rack_to_tot); 1095 counter_u64_free(rack_to_arm_rack); 1096 counter_u64_free(rack_to_arm_tlp); 1097 counter_u64_free(rack_paced_segments); 1098 counter_u64_free(rack_unpaced_segments); 1099 counter_u64_free(rack_saw_enobuf); 1100 counter_u64_free(rack_saw_enetunreach); 1101 counter_u64_free(rack_to_alloc_hard); 1102 counter_u64_free(rack_to_alloc_emerg); 1103 counter_u64_free(rack_sack_proc_all); 1104 counter_u64_free(rack_sack_proc_short); 1105 counter_u64_free(rack_sack_proc_restart); 1106 counter_u64_free(rack_to_alloc); 1107 counter_u64_free(rack_find_high); 1108 counter_u64_free(rack_runt_sacks); 1109 counter_u64_free(rack_enter_tlp_calc); 1110 counter_u64_free(rack_used_tlpmethod); 1111 counter_u64_free(rack_used_tlpmethod2); 1112 counter_u64_free(rack_progress_drops); 1113 counter_u64_free(rack_input_idle_reduces); 1114 counter_u64_free(rack_tlp_does_nada); 1115 COUNTER_ARRAY_FREE(rack_out_size, TCP_MSS_ACCT_SIZE); 1116 COUNTER_ARRAY_FREE(rack_opts_arry, RACK_OPTS_SIZE); 1117 } 1118 1119 static struct rack_sendmap * 1120 rack_alloc(struct tcp_rack *rack) 1121 { 1122 struct rack_sendmap *rsm; 1123 1124 counter_u64_add(rack_to_alloc, 1); 1125 rack->r_ctl.rc_num_maps_alloced++; 1126 rsm = uma_zalloc(rack_zone, M_NOWAIT); 1127 if (rsm) { 1128 return (rsm); 1129 } 1130 if (rack->rc_free_cnt) { 1131 counter_u64_add(rack_to_alloc_emerg, 1); 1132 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free); 1133 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_next); 1134 rack->rc_free_cnt--; 1135 return (rsm); 1136 } 1137 return (NULL); 1138 } 1139 1140 static void 1141 rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm) 1142 { 1143 rack->r_ctl.rc_num_maps_alloced--; 1144 if (rack->r_ctl.rc_tlpsend == rsm) 1145 rack->r_ctl.rc_tlpsend = NULL; 1146 if (rack->r_ctl.rc_next == rsm) 1147 rack->r_ctl.rc_next = NULL; 1148 if (rack->r_ctl.rc_sacklast == rsm) 1149 rack->r_ctl.rc_sacklast = NULL; 1150 if (rack->rc_free_cnt < rack_free_cache) { 1151 memset(rsm, 0, sizeof(struct rack_sendmap)); 1152 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_next); 1153 rack->rc_free_cnt++; 1154 return; 1155 } 1156 uma_zfree(rack_zone, rsm); 1157 } 1158 1159 /* 1160 * CC wrapper hook functions 1161 */ 1162 static void 1163 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, struct tcphdr *th, uint16_t nsegs, 1164 uint16_t type, int32_t recovery) 1165 { 1166 #ifdef NETFLIX_STATS 1167 int32_t gput; 1168 #endif 1169 #ifdef NETFLIX_CWV 1170 u_long old_cwnd = tp->snd_cwnd; 1171 #endif 1172 1173 INP_WLOCK_ASSERT(tp->t_inpcb); 1174 tp->ccv->nsegs = nsegs; 1175 tp->ccv->bytes_this_ack = BYTES_THIS_ACK(tp, th); 1176 if ((recovery) && (rack->r_ctl.rc_early_recovery_segs)) { 1177 uint32_t max; 1178 1179 max = rack->r_ctl.rc_early_recovery_segs * tp->t_maxseg; 1180 if (tp->ccv->bytes_this_ack > max) { 1181 tp->ccv->bytes_this_ack = max; 1182 } 1183 } 1184 if (tp->snd_cwnd <= tp->snd_wnd) 1185 tp->ccv->flags |= CCF_CWND_LIMITED; 1186 else 1187 tp->ccv->flags &= ~CCF_CWND_LIMITED; 1188 1189 if (type == CC_ACK) { 1190 #ifdef NETFLIX_STATS 1191 stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF, 1192 ((int32_t) tp->snd_cwnd) - tp->snd_wnd); 1193 if ((tp->t_flags & TF_GPUTINPROG) && 1194 SEQ_GEQ(th->th_ack, tp->gput_ack)) { 1195 gput = (((int64_t) (th->th_ack - tp->gput_seq)) << 3) / 1196 max(1, tcp_ts_getticks() - tp->gput_ts); 1197 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT, 1198 gput); 1199 /* 1200 * XXXLAS: This is a temporary hack, and should be 1201 * chained off VOI_TCP_GPUT when stats(9) grows an 1202 * API to deal with chained VOIs. 1203 */ 1204 if (tp->t_stats_gput_prev > 0) 1205 stats_voi_update_abs_s32(tp->t_stats, 1206 VOI_TCP_GPUT_ND, 1207 ((gput - tp->t_stats_gput_prev) * 100) / 1208 tp->t_stats_gput_prev); 1209 tp->t_flags &= ~TF_GPUTINPROG; 1210 tp->t_stats_gput_prev = gput; 1211 #ifdef NETFLIX_CWV 1212 if (tp->t_maxpeakrate) { 1213 /* 1214 * We update t_peakrate_thr. This gives us roughly 1215 * one update per round trip time. 1216 */ 1217 tcp_update_peakrate_thr(tp); 1218 } 1219 #endif 1220 } 1221 #endif 1222 if (tp->snd_cwnd > tp->snd_ssthresh) { 1223 tp->t_bytes_acked += min(tp->ccv->bytes_this_ack, 1224 nsegs * V_tcp_abc_l_var * tp->t_maxseg); 1225 if (tp->t_bytes_acked >= tp->snd_cwnd) { 1226 tp->t_bytes_acked -= tp->snd_cwnd; 1227 tp->ccv->flags |= CCF_ABC_SENTAWND; 1228 } 1229 } else { 1230 tp->ccv->flags &= ~CCF_ABC_SENTAWND; 1231 tp->t_bytes_acked = 0; 1232 } 1233 } 1234 if (CC_ALGO(tp)->ack_received != NULL) { 1235 /* XXXLAS: Find a way to live without this */ 1236 tp->ccv->curack = th->th_ack; 1237 CC_ALGO(tp)->ack_received(tp->ccv, type); 1238 } 1239 #ifdef NETFLIX_STATS 1240 stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, tp->snd_cwnd); 1241 #endif 1242 if (rack->r_ctl.rc_rack_largest_cwnd < tp->snd_cwnd) { 1243 rack->r_ctl.rc_rack_largest_cwnd = tp->snd_cwnd; 1244 } 1245 #ifdef NETFLIX_CWV 1246 if (tp->cwv_enabled) { 1247 /* 1248 * Per RFC 7661: The behaviour in the non-validated phase is 1249 * specified as: o A sender determines whether to increase 1250 * the cwnd based upon whether it is cwnd-limited (see 1251 * Section 4.5.3): * A sender that is cwnd-limited MAY use 1252 * the standard TCP method to increase cwnd (i.e., the 1253 * standard method permits a TCP sender that fully utilises 1254 * the cwnd to increase the cwnd each time it receives an 1255 * ACK). * A sender that is not cwnd-limited MUST NOT 1256 * increase the cwnd when ACK packets are received in this 1257 * phase (i.e., needs to avoid growing the cwnd when it has 1258 * not recently sent using the current size of cwnd). 1259 */ 1260 if ((tp->snd_cwnd > old_cwnd) && 1261 (tp->cwv_cwnd_valid == 0) && 1262 (!(tp->ccv->flags & CCF_CWND_LIMITED))) { 1263 tp->snd_cwnd = old_cwnd; 1264 } 1265 /* Try to update pipeAck and NCWV state */ 1266 if (TCPS_HAVEESTABLISHED(tp->t_state) && 1267 !IN_RECOVERY(tp->t_flags)) { 1268 uint32_t data = sbavail(&(tp->t_inpcb->inp_socket->so_snd)); 1269 1270 tcp_newcwv_update_pipeack(tp, data); 1271 } 1272 } 1273 /* we enforce max peak rate if it is set. */ 1274 if (tp->t_peakrate_thr && tp->snd_cwnd > tp->t_peakrate_thr) { 1275 tp->snd_cwnd = tp->t_peakrate_thr; 1276 } 1277 #endif 1278 } 1279 1280 static void 1281 tcp_rack_partialack(struct tcpcb *tp, struct tcphdr *th) 1282 { 1283 struct tcp_rack *rack; 1284 1285 rack = (struct tcp_rack *)tp->t_fb_ptr; 1286 INP_WLOCK_ASSERT(tp->t_inpcb); 1287 if (rack->r_ctl.rc_prr_sndcnt > 0) 1288 rack->r_wanted_output++; 1289 } 1290 1291 static void 1292 rack_post_recovery(struct tcpcb *tp, struct tcphdr *th) 1293 { 1294 struct tcp_rack *rack; 1295 1296 INP_WLOCK_ASSERT(tp->t_inpcb); 1297 rack = (struct tcp_rack *)tp->t_fb_ptr; 1298 if (CC_ALGO(tp)->post_recovery != NULL) { 1299 tp->ccv->curack = th->th_ack; 1300 CC_ALGO(tp)->post_recovery(tp->ccv); 1301 } 1302 /* 1303 * Here we can in theory adjust cwnd to be based on the number of 1304 * losses in the window (rack->r_ctl.rc_loss_count). This is done 1305 * based on the rack_use_proportional flag. 1306 */ 1307 if (rack->r_ctl.rc_prop_reduce && rack->r_ctl.rc_prop_rate) { 1308 int32_t reduce; 1309 1310 reduce = (rack->r_ctl.rc_loss_count * rack->r_ctl.rc_prop_rate); 1311 if (reduce > 50) { 1312 reduce = 50; 1313 } 1314 tp->snd_cwnd -= ((reduce * tp->snd_cwnd) / 100); 1315 } else { 1316 if (tp->snd_cwnd > tp->snd_ssthresh) { 1317 /* Drop us down to the ssthresh (1/2 cwnd at loss) */ 1318 tp->snd_cwnd = tp->snd_ssthresh; 1319 } 1320 } 1321 if (rack->r_ctl.rc_prr_sndcnt > 0) { 1322 /* Suck the next prr cnt back into cwnd */ 1323 tp->snd_cwnd += rack->r_ctl.rc_prr_sndcnt; 1324 rack->r_ctl.rc_prr_sndcnt = 0; 1325 } 1326 EXIT_RECOVERY(tp->t_flags); 1327 1328 1329 #ifdef NETFLIX_CWV 1330 if (tp->cwv_enabled) { 1331 if ((tp->cwv_cwnd_valid == 0) && 1332 (tp->snd_cwv.in_recovery)) 1333 tcp_newcwv_end_recovery(tp); 1334 } 1335 #endif 1336 } 1337 1338 static void 1339 rack_cong_signal(struct tcpcb *tp, struct tcphdr *th, uint32_t type) 1340 { 1341 struct tcp_rack *rack; 1342 1343 INP_WLOCK_ASSERT(tp->t_inpcb); 1344 1345 rack = (struct tcp_rack *)tp->t_fb_ptr; 1346 switch (type) { 1347 case CC_NDUPACK: 1348 /* rack->r_ctl.rc_ssthresh_set = 1;*/ 1349 if (!IN_FASTRECOVERY(tp->t_flags)) { 1350 rack->r_ctl.rc_tlp_rtx_out = 0; 1351 rack->r_ctl.rc_prr_delivered = 0; 1352 rack->r_ctl.rc_prr_out = 0; 1353 rack->r_ctl.rc_loss_count = 0; 1354 rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; 1355 rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una; 1356 tp->snd_recover = tp->snd_max; 1357 if (tp->t_flags & TF_ECN_PERMIT) 1358 tp->t_flags |= TF_ECN_SND_CWR; 1359 } 1360 break; 1361 case CC_ECN: 1362 if (!IN_CONGRECOVERY(tp->t_flags)) { 1363 TCPSTAT_INC(tcps_ecn_rcwnd); 1364 tp->snd_recover = tp->snd_max; 1365 if (tp->t_flags & TF_ECN_PERMIT) 1366 tp->t_flags |= TF_ECN_SND_CWR; 1367 } 1368 break; 1369 case CC_RTO: 1370 tp->t_dupacks = 0; 1371 tp->t_bytes_acked = 0; 1372 EXIT_RECOVERY(tp->t_flags); 1373 tp->snd_ssthresh = max(2, min(tp->snd_wnd, tp->snd_cwnd) / 2 / 1374 tp->t_maxseg) * tp->t_maxseg; 1375 tp->snd_cwnd = tp->t_maxseg; 1376 break; 1377 case CC_RTO_ERR: 1378 TCPSTAT_INC(tcps_sndrexmitbad); 1379 /* RTO was unnecessary, so reset everything. */ 1380 tp->snd_cwnd = tp->snd_cwnd_prev; 1381 tp->snd_ssthresh = tp->snd_ssthresh_prev; 1382 tp->snd_recover = tp->snd_recover_prev; 1383 if (tp->t_flags & TF_WASFRECOVERY) 1384 ENTER_FASTRECOVERY(tp->t_flags); 1385 if (tp->t_flags & TF_WASCRECOVERY) 1386 ENTER_CONGRECOVERY(tp->t_flags); 1387 tp->snd_nxt = tp->snd_max; 1388 tp->t_badrxtwin = 0; 1389 break; 1390 } 1391 1392 if (CC_ALGO(tp)->cong_signal != NULL) { 1393 if (th != NULL) 1394 tp->ccv->curack = th->th_ack; 1395 CC_ALGO(tp)->cong_signal(tp->ccv, type); 1396 } 1397 #ifdef NETFLIX_CWV 1398 if (tp->cwv_enabled) { 1399 if (tp->snd_cwv.in_recovery == 0 && IN_RECOVERY(tp->t_flags)) { 1400 tcp_newcwv_enter_recovery(tp); 1401 } 1402 if (type == CC_RTO) { 1403 tcp_newcwv_reset(tp); 1404 } 1405 } 1406 #endif 1407 } 1408 1409 1410 1411 static inline void 1412 rack_cc_after_idle(struct tcpcb *tp, int reduce_largest) 1413 { 1414 uint32_t i_cwnd; 1415 1416 INP_WLOCK_ASSERT(tp->t_inpcb); 1417 1418 #ifdef NETFLIX_STATS 1419 TCPSTAT_INC(tcps_idle_restarts); 1420 if (tp->t_state == TCPS_ESTABLISHED) 1421 TCPSTAT_INC(tcps_idle_estrestarts); 1422 #endif 1423 if (CC_ALGO(tp)->after_idle != NULL) 1424 CC_ALGO(tp)->after_idle(tp->ccv); 1425 1426 if (tp->snd_cwnd == 1) 1427 i_cwnd = tp->t_maxseg; /* SYN(-ACK) lost */ 1428 else if (V_tcp_initcwnd_segments) 1429 i_cwnd = min((V_tcp_initcwnd_segments * tp->t_maxseg), 1430 max(2 * tp->t_maxseg, V_tcp_initcwnd_segments * 1460)); 1431 else if (V_tcp_do_rfc3390) 1432 i_cwnd = min(4 * tp->t_maxseg, 1433 max(2 * tp->t_maxseg, 4380)); 1434 else { 1435 /* Per RFC5681 Section 3.1 */ 1436 if (tp->t_maxseg > 2190) 1437 i_cwnd = 2 * tp->t_maxseg; 1438 else if (tp->t_maxseg > 1095) 1439 i_cwnd = 3 * tp->t_maxseg; 1440 else 1441 i_cwnd = 4 * tp->t_maxseg; 1442 } 1443 if (reduce_largest) { 1444 /* 1445 * Do we reduce the largest cwnd to make 1446 * rack play nice on restart hptsi wise? 1447 */ 1448 if (((struct tcp_rack *)tp->t_fb_ptr)->r_ctl.rc_rack_largest_cwnd > i_cwnd) 1449 ((struct tcp_rack *)tp->t_fb_ptr)->r_ctl.rc_rack_largest_cwnd = i_cwnd; 1450 } 1451 /* 1452 * Being idle is no differnt than the initial window. If the cc 1453 * clamps it down below the initial window raise it to the initial 1454 * window. 1455 */ 1456 if (tp->snd_cwnd < i_cwnd) { 1457 tp->snd_cwnd = i_cwnd; 1458 } 1459 } 1460 1461 1462 /* 1463 * Indicate whether this ack should be delayed. We can delay the ack if 1464 * following conditions are met: 1465 * - There is no delayed ack timer in progress. 1466 * - Our last ack wasn't a 0-sized window. We never want to delay 1467 * the ack that opens up a 0-sized window. 1468 * - LRO wasn't used for this segment. We make sure by checking that the 1469 * segment size is not larger than the MSS. 1470 * - Delayed acks are enabled or this is a half-synchronized T/TCP 1471 * connection. 1472 */ 1473 #define DELAY_ACK(tp, tlen) \ 1474 (((tp->t_flags & TF_RXWIN0SENT) == 0) && \ 1475 ((tp->t_flags & TF_DELACK) == 0) && \ 1476 (tlen <= tp->t_maxseg) && \ 1477 (tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN))) 1478 1479 static inline void 1480 rack_calc_rwin(struct socket *so, struct tcpcb *tp) 1481 { 1482 int32_t win; 1483 1484 /* 1485 * Calculate amount of space in receive window, and then do TCP 1486 * input processing. Receive window is amount of space in rcv queue, 1487 * but not less than advertised window. 1488 */ 1489 win = sbspace(&so->so_rcv); 1490 if (win < 0) 1491 win = 0; 1492 tp->rcv_wnd = imax(win, (int)(tp->rcv_adv - tp->rcv_nxt)); 1493 } 1494 1495 static void 1496 rack_do_drop(struct mbuf *m, struct tcpcb *tp) 1497 { 1498 /* 1499 * Drop space held by incoming segment and return. 1500 */ 1501 if (tp != NULL) 1502 INP_WUNLOCK(tp->t_inpcb); 1503 if (m) 1504 m_freem(m); 1505 } 1506 1507 static void 1508 rack_do_dropwithreset(struct mbuf *m, struct tcpcb *tp, struct tcphdr *th, 1509 int32_t rstreason, int32_t tlen) 1510 { 1511 if (tp != NULL) { 1512 tcp_dropwithreset(m, th, tp, tlen, rstreason); 1513 INP_WUNLOCK(tp->t_inpcb); 1514 } else 1515 tcp_dropwithreset(m, th, NULL, tlen, rstreason); 1516 } 1517 1518 /* 1519 * The value in ret_val informs the caller 1520 * if we dropped the tcb (and lock) or not. 1521 * 1 = we dropped it, 0 = the TCB is still locked 1522 * and valid. 1523 */ 1524 static void 1525 rack_do_dropafterack(struct mbuf *m, struct tcpcb *tp, struct tcphdr *th, int32_t thflags, int32_t tlen, int32_t * ret_val) 1526 { 1527 /* 1528 * Generate an ACK dropping incoming segment if it occupies sequence 1529 * space, where the ACK reflects our state. 1530 * 1531 * We can now skip the test for the RST flag since all paths to this 1532 * code happen after packets containing RST have been dropped. 1533 * 1534 * In the SYN-RECEIVED state, don't send an ACK unless the segment 1535 * we received passes the SYN-RECEIVED ACK test. If it fails send a 1536 * RST. This breaks the loop in the "LAND" DoS attack, and also 1537 * prevents an ACK storm between two listening ports that have been 1538 * sent forged SYN segments, each with the source address of the 1539 * other. 1540 */ 1541 struct tcp_rack *rack; 1542 1543 if (tp->t_state == TCPS_SYN_RECEIVED && (thflags & TH_ACK) && 1544 (SEQ_GT(tp->snd_una, th->th_ack) || 1545 SEQ_GT(th->th_ack, tp->snd_max))) { 1546 *ret_val = 1; 1547 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 1548 return; 1549 } else 1550 *ret_val = 0; 1551 rack = (struct tcp_rack *)tp->t_fb_ptr; 1552 rack->r_wanted_output++; 1553 tp->t_flags |= TF_ACKNOW; 1554 if (m) 1555 m_freem(m); 1556 } 1557 1558 1559 static int 1560 rack_process_rst(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp) 1561 { 1562 /* 1563 * RFC5961 Section 3.2 1564 * 1565 * - RST drops connection only if SEG.SEQ == RCV.NXT. - If RST is in 1566 * window, we send challenge ACK. 1567 * 1568 * Note: to take into account delayed ACKs, we should test against 1569 * last_ack_sent instead of rcv_nxt. Note 2: we handle special case 1570 * of closed window, not covered by the RFC. 1571 */ 1572 int dropped = 0; 1573 1574 if ((SEQ_GEQ(th->th_seq, (tp->last_ack_sent - 1)) && 1575 SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) || 1576 (tp->rcv_wnd == 0 && tp->last_ack_sent == th->th_seq)) { 1577 1578 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 1579 KASSERT(tp->t_state != TCPS_SYN_SENT, 1580 ("%s: TH_RST for TCPS_SYN_SENT th %p tp %p", 1581 __func__, th, tp)); 1582 1583 if (V_tcp_insecure_rst || 1584 (tp->last_ack_sent == th->th_seq) || 1585 (tp->rcv_nxt == th->th_seq) || 1586 ((tp->last_ack_sent - 1) == th->th_seq)) { 1587 TCPSTAT_INC(tcps_drops); 1588 /* Drop the connection. */ 1589 switch (tp->t_state) { 1590 case TCPS_SYN_RECEIVED: 1591 so->so_error = ECONNREFUSED; 1592 goto close; 1593 case TCPS_ESTABLISHED: 1594 case TCPS_FIN_WAIT_1: 1595 case TCPS_FIN_WAIT_2: 1596 case TCPS_CLOSE_WAIT: 1597 case TCPS_CLOSING: 1598 case TCPS_LAST_ACK: 1599 so->so_error = ECONNRESET; 1600 close: 1601 tcp_state_change(tp, TCPS_CLOSED); 1602 /* FALLTHROUGH */ 1603 default: 1604 tp = tcp_close(tp); 1605 } 1606 dropped = 1; 1607 rack_do_drop(m, tp); 1608 } else { 1609 TCPSTAT_INC(tcps_badrst); 1610 /* Send challenge ACK. */ 1611 tcp_respond(tp, mtod(m, void *), th, m, 1612 tp->rcv_nxt, tp->snd_nxt, TH_ACK); 1613 tp->last_ack_sent = tp->rcv_nxt; 1614 } 1615 } else { 1616 m_freem(m); 1617 } 1618 return (dropped); 1619 } 1620 1621 /* 1622 * The value in ret_val informs the caller 1623 * if we dropped the tcb (and lock) or not. 1624 * 1 = we dropped it, 0 = the TCB is still locked 1625 * and valid. 1626 */ 1627 static void 1628 rack_challenge_ack(struct mbuf *m, struct tcphdr *th, struct tcpcb *tp, int32_t * ret_val) 1629 { 1630 1631 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 1632 1633 TCPSTAT_INC(tcps_badsyn); 1634 if (V_tcp_insecure_syn && 1635 SEQ_GEQ(th->th_seq, tp->last_ack_sent) && 1636 SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) { 1637 tp = tcp_drop(tp, ECONNRESET); 1638 *ret_val = 1; 1639 rack_do_drop(m, tp); 1640 } else { 1641 /* Send challenge ACK. */ 1642 tcp_respond(tp, mtod(m, void *), th, m, tp->rcv_nxt, 1643 tp->snd_nxt, TH_ACK); 1644 tp->last_ack_sent = tp->rcv_nxt; 1645 m = NULL; 1646 *ret_val = 0; 1647 rack_do_drop(m, NULL); 1648 } 1649 } 1650 1651 /* 1652 * rack_ts_check returns 1 for you should not proceed. It places 1653 * in ret_val what should be returned 1/0 by the caller. The 1 indicates 1654 * that the TCB is unlocked and probably dropped. The 0 indicates the 1655 * TCB is still valid and locked. 1656 */ 1657 static int 1658 rack_ts_check(struct mbuf *m, struct tcphdr *th, struct tcpcb *tp, int32_t tlen, int32_t thflags, int32_t * ret_val) 1659 { 1660 1661 /* Check to see if ts_recent is over 24 days old. */ 1662 if (tcp_ts_getticks() - tp->ts_recent_age > TCP_PAWS_IDLE) { 1663 /* 1664 * Invalidate ts_recent. If this segment updates ts_recent, 1665 * the age will be reset later and ts_recent will get a 1666 * valid value. If it does not, setting ts_recent to zero 1667 * will at least satisfy the requirement that zero be placed 1668 * in the timestamp echo reply when ts_recent isn't valid. 1669 * The age isn't reset until we get a valid ts_recent 1670 * because we don't want out-of-order segments to be dropped 1671 * when ts_recent is old. 1672 */ 1673 tp->ts_recent = 0; 1674 } else { 1675 TCPSTAT_INC(tcps_rcvduppack); 1676 TCPSTAT_ADD(tcps_rcvdupbyte, tlen); 1677 TCPSTAT_INC(tcps_pawsdrop); 1678 *ret_val = 0; 1679 if (tlen) { 1680 rack_do_dropafterack(m, tp, th, thflags, tlen, ret_val); 1681 } else { 1682 rack_do_drop(m, NULL); 1683 } 1684 return (1); 1685 } 1686 return (0); 1687 } 1688 1689 /* 1690 * rack_drop_checks returns 1 for you should not proceed. It places 1691 * in ret_val what should be returned 1/0 by the caller. The 1 indicates 1692 * that the TCB is unlocked and probably dropped. The 0 indicates the 1693 * TCB is still valid and locked. 1694 */ 1695 static int 1696 rack_drop_checks(struct tcpopt *to, struct mbuf *m, struct tcphdr *th, struct tcpcb *tp, int32_t * tlenp, int32_t * thf, int32_t * drop_hdrlen, int32_t * ret_val) 1697 { 1698 int32_t todrop; 1699 int32_t thflags; 1700 int32_t tlen; 1701 1702 thflags = *thf; 1703 tlen = *tlenp; 1704 todrop = tp->rcv_nxt - th->th_seq; 1705 if (todrop > 0) { 1706 if (thflags & TH_SYN) { 1707 thflags &= ~TH_SYN; 1708 th->th_seq++; 1709 if (th->th_urp > 1) 1710 th->th_urp--; 1711 else 1712 thflags &= ~TH_URG; 1713 todrop--; 1714 } 1715 /* 1716 * Following if statement from Stevens, vol. 2, p. 960. 1717 */ 1718 if (todrop > tlen 1719 || (todrop == tlen && (thflags & TH_FIN) == 0)) { 1720 /* 1721 * Any valid FIN must be to the left of the window. 1722 * At this point the FIN must be a duplicate or out 1723 * of sequence; drop it. 1724 */ 1725 thflags &= ~TH_FIN; 1726 /* 1727 * Send an ACK to resynchronize and drop any data. 1728 * But keep on processing for RST or ACK. 1729 */ 1730 tp->t_flags |= TF_ACKNOW; 1731 todrop = tlen; 1732 TCPSTAT_INC(tcps_rcvduppack); 1733 TCPSTAT_ADD(tcps_rcvdupbyte, todrop); 1734 } else { 1735 TCPSTAT_INC(tcps_rcvpartduppack); 1736 TCPSTAT_ADD(tcps_rcvpartdupbyte, todrop); 1737 } 1738 *drop_hdrlen += todrop; /* drop from the top afterwards */ 1739 th->th_seq += todrop; 1740 tlen -= todrop; 1741 if (th->th_urp > todrop) 1742 th->th_urp -= todrop; 1743 else { 1744 thflags &= ~TH_URG; 1745 th->th_urp = 0; 1746 } 1747 } 1748 /* 1749 * If segment ends after window, drop trailing data (and PUSH and 1750 * FIN); if nothing left, just ACK. 1751 */ 1752 todrop = (th->th_seq + tlen) - (tp->rcv_nxt + tp->rcv_wnd); 1753 if (todrop > 0) { 1754 TCPSTAT_INC(tcps_rcvpackafterwin); 1755 if (todrop >= tlen) { 1756 TCPSTAT_ADD(tcps_rcvbyteafterwin, tlen); 1757 /* 1758 * If window is closed can only take segments at 1759 * window edge, and have to drop data and PUSH from 1760 * incoming segments. Continue processing, but 1761 * remember to ack. Otherwise, drop segment and 1762 * ack. 1763 */ 1764 if (tp->rcv_wnd == 0 && th->th_seq == tp->rcv_nxt) { 1765 tp->t_flags |= TF_ACKNOW; 1766 TCPSTAT_INC(tcps_rcvwinprobe); 1767 } else { 1768 rack_do_dropafterack(m, tp, th, thflags, tlen, ret_val); 1769 return (1); 1770 } 1771 } else 1772 TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop); 1773 m_adj(m, -todrop); 1774 tlen -= todrop; 1775 thflags &= ~(TH_PUSH | TH_FIN); 1776 } 1777 *thf = thflags; 1778 *tlenp = tlen; 1779 return (0); 1780 } 1781 1782 static struct rack_sendmap * 1783 rack_find_lowest_rsm(struct tcp_rack *rack) 1784 { 1785 struct rack_sendmap *rsm; 1786 1787 /* 1788 * Walk the time-order transmitted list looking for an rsm that is 1789 * not acked. This will be the one that was sent the longest time 1790 * ago that is still outstanding. 1791 */ 1792 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) { 1793 if (rsm->r_flags & RACK_ACKED) { 1794 continue; 1795 } 1796 goto finish; 1797 } 1798 finish: 1799 return (rsm); 1800 } 1801 1802 static struct rack_sendmap * 1803 rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm) 1804 { 1805 struct rack_sendmap *prsm; 1806 1807 /* 1808 * Walk the sequence order list backward until we hit and arrive at 1809 * the highest seq not acked. In theory when this is called it 1810 * should be the last segment (which it was not). 1811 */ 1812 counter_u64_add(rack_find_high, 1); 1813 prsm = rsm; 1814 TAILQ_FOREACH_REVERSE_FROM(prsm, &rack->r_ctl.rc_map, rack_head, r_next) { 1815 if (prsm->r_flags & (RACK_ACKED | RACK_HAS_FIN)) { 1816 continue; 1817 } 1818 return (prsm); 1819 } 1820 return (NULL); 1821 } 1822 1823 1824 static uint32_t 1825 rack_calc_thresh_rack(struct tcp_rack *rack, uint32_t srtt, uint32_t cts) 1826 { 1827 int32_t lro; 1828 uint32_t thresh; 1829 1830 /* 1831 * lro is the flag we use to determine if we have seen reordering. 1832 * If it gets set we have seen reordering. The reorder logic either 1833 * works in one of two ways: 1834 * 1835 * If reorder-fade is configured, then we track the last time we saw 1836 * re-ordering occur. If we reach the point where enough time as 1837 * passed we no longer consider reordering has occuring. 1838 * 1839 * Or if reorder-face is 0, then once we see reordering we consider 1840 * the connection to alway be subject to reordering and just set lro 1841 * to 1. 1842 * 1843 * In the end if lro is non-zero we add the extra time for 1844 * reordering in. 1845 */ 1846 if (srtt == 0) 1847 srtt = 1; 1848 if (rack->r_ctl.rc_reorder_ts) { 1849 if (rack->r_ctl.rc_reorder_fade) { 1850 if (SEQ_GEQ(cts, rack->r_ctl.rc_reorder_ts)) { 1851 lro = cts - rack->r_ctl.rc_reorder_ts; 1852 if (lro == 0) { 1853 /* 1854 * No time as passed since the last 1855 * reorder, mark it as reordering. 1856 */ 1857 lro = 1; 1858 } 1859 } else { 1860 /* Negative time? */ 1861 lro = 0; 1862 } 1863 if (lro > rack->r_ctl.rc_reorder_fade) { 1864 /* Turn off reordering seen too */ 1865 rack->r_ctl.rc_reorder_ts = 0; 1866 lro = 0; 1867 } 1868 } else { 1869 /* Reodering does not fade */ 1870 lro = 1; 1871 } 1872 } else { 1873 lro = 0; 1874 } 1875 thresh = srtt + rack->r_ctl.rc_pkt_delay; 1876 if (lro) { 1877 /* It must be set, if not you get 1/4 rtt */ 1878 if (rack->r_ctl.rc_reorder_shift) 1879 thresh += (srtt >> rack->r_ctl.rc_reorder_shift); 1880 else 1881 thresh += (srtt >> 2); 1882 } else { 1883 thresh += 1; 1884 } 1885 /* We don't let the rack timeout be above a RTO */ 1886 1887 if (thresh > TICKS_2_MSEC(rack->rc_tp->t_rxtcur)) { 1888 thresh = TICKS_2_MSEC(rack->rc_tp->t_rxtcur); 1889 } 1890 /* And we don't want it above the RTO max either */ 1891 if (thresh > rack_rto_max) { 1892 thresh = rack_rto_max; 1893 } 1894 return (thresh); 1895 } 1896 1897 static uint32_t 1898 rack_calc_thresh_tlp(struct tcpcb *tp, struct tcp_rack *rack, 1899 struct rack_sendmap *rsm, uint32_t srtt) 1900 { 1901 struct rack_sendmap *prsm; 1902 uint32_t thresh, len; 1903 int maxseg; 1904 1905 if (srtt == 0) 1906 srtt = 1; 1907 if (rack->r_ctl.rc_tlp_threshold) 1908 thresh = srtt + (srtt / rack->r_ctl.rc_tlp_threshold); 1909 else 1910 thresh = (srtt * 2); 1911 1912 /* Get the previous sent packet, if any */ 1913 maxseg = tcp_maxseg(tp); 1914 counter_u64_add(rack_enter_tlp_calc, 1); 1915 len = rsm->r_end - rsm->r_start; 1916 if (rack->rack_tlp_threshold_use == TLP_USE_ID) { 1917 /* Exactly like the ID */ 1918 if (((tp->snd_max - tp->snd_una) - rack->r_ctl.rc_sacked + rack->r_ctl.rc_holes_rxt) <= maxseg) { 1919 uint32_t alt_thresh; 1920 /* 1921 * Compensate for delayed-ack with the d-ack time. 1922 */ 1923 counter_u64_add(rack_used_tlpmethod, 1); 1924 alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time; 1925 if (alt_thresh > thresh) 1926 thresh = alt_thresh; 1927 } 1928 } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_ONE) { 1929 /* 2.1 behavior */ 1930 prsm = TAILQ_PREV(rsm, rack_head, r_tnext); 1931 if (prsm && (len <= maxseg)) { 1932 /* 1933 * Two packets outstanding, thresh should be (2*srtt) + 1934 * possible inter-packet delay (if any). 1935 */ 1936 uint32_t inter_gap = 0; 1937 int idx, nidx; 1938 1939 counter_u64_add(rack_used_tlpmethod, 1); 1940 idx = rsm->r_rtr_cnt - 1; 1941 nidx = prsm->r_rtr_cnt - 1; 1942 if (TSTMP_GEQ(rsm->r_tim_lastsent[nidx], prsm->r_tim_lastsent[idx])) { 1943 /* Yes it was sent later (or at the same time) */ 1944 inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx]; 1945 } 1946 thresh += inter_gap; 1947 } else if (len <= maxseg) { 1948 /* 1949 * Possibly compensate for delayed-ack. 1950 */ 1951 uint32_t alt_thresh; 1952 1953 counter_u64_add(rack_used_tlpmethod2, 1); 1954 alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time; 1955 if (alt_thresh > thresh) 1956 thresh = alt_thresh; 1957 } 1958 } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_TWO) { 1959 /* 2.2 behavior */ 1960 if (len <= maxseg) { 1961 uint32_t alt_thresh; 1962 /* 1963 * Compensate for delayed-ack with the d-ack time. 1964 */ 1965 counter_u64_add(rack_used_tlpmethod, 1); 1966 alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time; 1967 if (alt_thresh > thresh) 1968 thresh = alt_thresh; 1969 } 1970 } 1971 /* Not above an RTO */ 1972 if (thresh > TICKS_2_MSEC(tp->t_rxtcur)) { 1973 thresh = TICKS_2_MSEC(tp->t_rxtcur); 1974 } 1975 /* Not above a RTO max */ 1976 if (thresh > rack_rto_max) { 1977 thresh = rack_rto_max; 1978 } 1979 /* Apply user supplied min TLP */ 1980 if (thresh < rack_tlp_min) { 1981 thresh = rack_tlp_min; 1982 } 1983 return (thresh); 1984 } 1985 1986 static struct rack_sendmap * 1987 rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused) 1988 { 1989 /* 1990 * Check to see that we don't need to fall into recovery. We will 1991 * need to do so if our oldest transmit is past the time we should 1992 * have had an ack. 1993 */ 1994 struct tcp_rack *rack; 1995 struct rack_sendmap *rsm; 1996 int32_t idx; 1997 uint32_t srtt_cur, srtt, thresh; 1998 1999 rack = (struct tcp_rack *)tp->t_fb_ptr; 2000 if (TAILQ_EMPTY(&rack->r_ctl.rc_map)) { 2001 return (NULL); 2002 } 2003 srtt_cur = tp->t_srtt >> TCP_RTT_SHIFT; 2004 srtt = TICKS_2_MSEC(srtt_cur); 2005 if (rack->rc_rack_rtt && (srtt > rack->rc_rack_rtt)) 2006 srtt = rack->rc_rack_rtt; 2007 2008 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); 2009 if (rsm == NULL) 2010 return (NULL); 2011 2012 if (rsm->r_flags & RACK_ACKED) { 2013 rsm = rack_find_lowest_rsm(rack); 2014 if (rsm == NULL) 2015 return (NULL); 2016 } 2017 idx = rsm->r_rtr_cnt - 1; 2018 thresh = rack_calc_thresh_rack(rack, srtt, tsused); 2019 if (tsused < rsm->r_tim_lastsent[idx]) { 2020 return (NULL); 2021 } 2022 if ((tsused - rsm->r_tim_lastsent[idx]) < thresh) { 2023 return (NULL); 2024 } 2025 /* Ok if we reach here we are over-due */ 2026 rack->r_ctl.rc_rsm_start = rsm->r_start; 2027 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd; 2028 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh; 2029 rack_cong_signal(tp, NULL, CC_NDUPACK); 2030 return (rsm); 2031 } 2032 2033 static uint32_t 2034 rack_get_persists_timer_val(struct tcpcb *tp, struct tcp_rack *rack) 2035 { 2036 int32_t t; 2037 int32_t tt; 2038 uint32_t ret_val; 2039 2040 t = TICKS_2_MSEC((tp->t_srtt >> TCP_RTT_SHIFT) + ((tp->t_rttvar * 4) >> TCP_RTT_SHIFT)); 2041 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], 2042 tcp_persmin, tcp_persmax); 2043 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 2044 tp->t_rxtshift++; 2045 rack->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT; 2046 ret_val = (uint32_t)tt; 2047 return (ret_val); 2048 } 2049 2050 static uint32_t 2051 rack_timer_start(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) 2052 { 2053 /* 2054 * Start the FR timer, we do this based on getting the first one in 2055 * the rc_tmap. Note that if its NULL we must stop the timer. in all 2056 * events we need to stop the running timer (if its running) before 2057 * starting the new one. 2058 */ 2059 uint32_t thresh, exp, to, srtt, time_since_sent; 2060 uint32_t srtt_cur; 2061 int32_t idx; 2062 int32_t is_tlp_timer = 0; 2063 struct rack_sendmap *rsm; 2064 2065 if (rack->t_timers_stopped) { 2066 /* All timers have been stopped none are to run */ 2067 return (0); 2068 } 2069 if (rack->rc_in_persist) { 2070 /* We can't start any timer in persists */ 2071 return (rack_get_persists_timer_val(tp, rack)); 2072 } 2073 if (tp->t_state < TCPS_ESTABLISHED) 2074 goto activate_rxt; 2075 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); 2076 if (rsm == NULL) { 2077 /* Nothing on the send map */ 2078 activate_rxt: 2079 if (SEQ_LT(tp->snd_una, tp->snd_max) || sbavail(&(tp->t_inpcb->inp_socket->so_snd))) { 2080 rack->r_ctl.rc_hpts_flags |= PACE_TMR_RXT; 2081 to = TICKS_2_MSEC(tp->t_rxtcur); 2082 if (to == 0) 2083 to = 1; 2084 return (to); 2085 } 2086 return (0); 2087 } 2088 if (rsm->r_flags & RACK_ACKED) { 2089 rsm = rack_find_lowest_rsm(rack); 2090 if (rsm == NULL) { 2091 /* No lowest? */ 2092 goto activate_rxt; 2093 } 2094 } 2095 /* Convert from ms to usecs */ 2096 if (rsm->r_flags & RACK_SACK_PASSED) { 2097 if ((tp->t_flags & TF_SENTFIN) && 2098 ((tp->snd_max - tp->snd_una) == 1) && 2099 (rsm->r_flags & RACK_HAS_FIN)) { 2100 /* 2101 * We don't start a rack timer if all we have is a 2102 * FIN outstanding. 2103 */ 2104 goto activate_rxt; 2105 } 2106 if (tp->t_srtt) { 2107 srtt_cur = (tp->t_srtt >> TCP_RTT_SHIFT); 2108 srtt = TICKS_2_MSEC(srtt_cur); 2109 } else 2110 srtt = RACK_INITIAL_RTO; 2111 2112 thresh = rack_calc_thresh_rack(rack, srtt, cts); 2113 idx = rsm->r_rtr_cnt - 1; 2114 exp = rsm->r_tim_lastsent[idx] + thresh; 2115 if (SEQ_GEQ(exp, cts)) { 2116 to = exp - cts; 2117 if (to < rack->r_ctl.rc_min_to) { 2118 to = rack->r_ctl.rc_min_to; 2119 } 2120 } else { 2121 to = rack->r_ctl.rc_min_to; 2122 } 2123 } else { 2124 /* Ok we need to do a TLP not RACK */ 2125 if ((rack->rc_tlp_in_progress != 0) || 2126 (rack->r_ctl.rc_tlp_rtx_out != 0)) { 2127 /* 2128 * The previous send was a TLP or a tlp_rtx is in 2129 * process. 2130 */ 2131 goto activate_rxt; 2132 } 2133 rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext); 2134 if (rsm == NULL) { 2135 /* We found no rsm to TLP with. */ 2136 goto activate_rxt; 2137 } 2138 if (rsm->r_flags & RACK_HAS_FIN) { 2139 /* If its a FIN we dont do TLP */ 2140 rsm = NULL; 2141 goto activate_rxt; 2142 } 2143 idx = rsm->r_rtr_cnt - 1; 2144 if (TSTMP_GT(cts, rsm->r_tim_lastsent[idx])) 2145 time_since_sent = cts - rsm->r_tim_lastsent[idx]; 2146 else 2147 time_since_sent = 0; 2148 is_tlp_timer = 1; 2149 if (tp->t_srtt) { 2150 srtt_cur = (tp->t_srtt >> TCP_RTT_SHIFT); 2151 srtt = TICKS_2_MSEC(srtt_cur); 2152 } else 2153 srtt = RACK_INITIAL_RTO; 2154 thresh = rack_calc_thresh_tlp(tp, rack, rsm, srtt); 2155 if (thresh > time_since_sent) 2156 to = thresh - time_since_sent; 2157 else 2158 to = rack->r_ctl.rc_min_to; 2159 if (to > TCPTV_REXMTMAX) { 2160 /* 2161 * If the TLP time works out to larger than the max 2162 * RTO lets not do TLP.. just RTO. 2163 */ 2164 goto activate_rxt; 2165 } 2166 if (rsm->r_start != rack->r_ctl.rc_last_tlp_seq) { 2167 /* 2168 * The tail is no longer the last one I did a probe 2169 * on 2170 */ 2171 rack->r_ctl.rc_tlp_seg_send_cnt = 0; 2172 rack->r_ctl.rc_last_tlp_seq = rsm->r_start; 2173 } 2174 } 2175 if (is_tlp_timer == 0) { 2176 rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK; 2177 } else { 2178 if ((rack->r_ctl.rc_tlp_send_cnt > rack_tlp_max_resend) || 2179 (rack->r_ctl.rc_tlp_seg_send_cnt > rack_tlp_max_resend)) { 2180 /* 2181 * We have exceeded how many times we can retran the 2182 * current TLP timer, switch to the RTO timer. 2183 */ 2184 goto activate_rxt; 2185 } else { 2186 rack->r_ctl.rc_hpts_flags |= PACE_TMR_TLP; 2187 } 2188 } 2189 if (to == 0) 2190 to = 1; 2191 return (to); 2192 } 2193 2194 static void 2195 rack_enter_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) 2196 { 2197 if (rack->rc_in_persist == 0) { 2198 if (((tp->t_flags & TF_SENTFIN) == 0) && 2199 (tp->snd_max - tp->snd_una) >= sbavail(&rack->rc_inp->inp_socket->so_snd)) 2200 /* Must need to send more data to enter persist */ 2201 return; 2202 rack->r_ctl.rc_went_idle_time = cts; 2203 rack_timer_cancel(tp, rack, cts, __LINE__); 2204 tp->t_rxtshift = 0; 2205 rack->rc_in_persist = 1; 2206 } 2207 } 2208 2209 static void 2210 rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack) 2211 { 2212 if (rack->rc_inp->inp_in_hpts) { 2213 tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT); 2214 rack->r_ctl.rc_hpts_flags = 0; 2215 } 2216 rack->rc_in_persist = 0; 2217 rack->r_ctl.rc_went_idle_time = 0; 2218 tp->t_flags &= ~TF_FORCEDATA; 2219 tp->t_rxtshift = 0; 2220 } 2221 2222 static void 2223 rack_start_hpts_timer(struct tcp_rack *rack, struct tcpcb *tp, uint32_t cts, int32_t line, 2224 int32_t slot, uint32_t tot_len_this_send, int32_t frm_out_sbavail) 2225 { 2226 struct inpcb *inp; 2227 uint32_t delayed_ack = 0; 2228 uint32_t hpts_timeout; 2229 uint8_t stopped; 2230 uint32_t left = 0; 2231 2232 inp = tp->t_inpcb; 2233 if (inp->inp_in_hpts) { 2234 /* A previous call is already set up */ 2235 return; 2236 } 2237 if (tp->t_state == TCPS_CLOSED) { 2238 return; 2239 } 2240 stopped = rack->rc_tmr_stopped; 2241 if (stopped && TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) { 2242 left = rack->r_ctl.rc_timer_exp - cts; 2243 } 2244 rack->r_ctl.rc_timer_exp = 0; 2245 if (rack->rc_inp->inp_in_hpts == 0) { 2246 rack->r_ctl.rc_hpts_flags = 0; 2247 } 2248 if (slot) { 2249 /* We are hptsi too */ 2250 rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT; 2251 } else if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) { 2252 /* 2253 * We are still left on the hpts when the to goes 2254 * it will be for output. 2255 */ 2256 if (TSTMP_GT(cts, rack->r_ctl.rc_last_output_to)) 2257 slot = cts - rack->r_ctl.rc_last_output_to; 2258 else 2259 slot = 1; 2260 } 2261 if ((tp->snd_wnd == 0) && TCPS_HAVEESTABLISHED(tp->t_state)) { 2262 /* No send window.. we must enter persist */ 2263 rack_enter_persist(tp, rack, cts); 2264 } else if ((frm_out_sbavail && 2265 (frm_out_sbavail > (tp->snd_max - tp->snd_una)) && 2266 (tp->snd_wnd < tp->t_maxseg)) && 2267 TCPS_HAVEESTABLISHED(tp->t_state)) { 2268 /* 2269 * If we have no window or we can't send a segment (and have 2270 * data to send.. we cheat here and frm_out_sbavail is 2271 * passed in with the sbavail(sb) only from bbr_output) and 2272 * we are established, then we must enter persits (if not 2273 * already in persits). 2274 */ 2275 rack_enter_persist(tp, rack, cts); 2276 } 2277 hpts_timeout = rack_timer_start(tp, rack, cts); 2278 if (tp->t_flags & TF_DELACK) { 2279 delayed_ack = tcp_delacktime; 2280 rack->r_ctl.rc_hpts_flags |= PACE_TMR_DELACK; 2281 } 2282 if (delayed_ack && ((hpts_timeout == 0) || 2283 (delayed_ack < hpts_timeout))) 2284 hpts_timeout = delayed_ack; 2285 else 2286 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK; 2287 /* 2288 * If no timers are going to run and we will fall off the hptsi 2289 * wheel, we resort to a keep-alive timer if its configured. 2290 */ 2291 if ((hpts_timeout == 0) && 2292 (slot == 0)) { 2293 if ((tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && 2294 (tp->t_state <= TCPS_CLOSING)) { 2295 /* 2296 * Ok we have no timer (persists, rack, tlp, rxt or 2297 * del-ack), we don't have segments being paced. So 2298 * all that is left is the keepalive timer. 2299 */ 2300 if (TCPS_HAVEESTABLISHED(tp->t_state)) { 2301 /* Get the established keep-alive time */ 2302 hpts_timeout = TP_KEEPIDLE(tp); 2303 } else { 2304 /* Get the initial setup keep-alive time */ 2305 hpts_timeout = TP_KEEPINIT(tp); 2306 } 2307 rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP; 2308 } 2309 } 2310 if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) == 2311 (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) { 2312 /* 2313 * RACK, TLP, persists and RXT timers all are restartable 2314 * based on actions input .. i.e we received a packet (ack 2315 * or sack) and that changes things (rw, or snd_una etc). 2316 * Thus we can restart them with a new value. For 2317 * keep-alive, delayed_ack we keep track of what was left 2318 * and restart the timer with a smaller value. 2319 */ 2320 if (left < hpts_timeout) 2321 hpts_timeout = left; 2322 } 2323 if (hpts_timeout) { 2324 /* 2325 * Hack alert for now we can't time-out over 2,147,483 2326 * seconds (a bit more than 596 hours), which is probably ok 2327 * :). 2328 */ 2329 if (hpts_timeout > 0x7ffffffe) 2330 hpts_timeout = 0x7ffffffe; 2331 rack->r_ctl.rc_timer_exp = cts + hpts_timeout; 2332 } 2333 if (slot) { 2334 rack->r_ctl.rc_last_output_to = cts + slot; 2335 if ((hpts_timeout == 0) || (hpts_timeout > slot)) { 2336 if (rack->rc_inp->inp_in_hpts == 0) 2337 tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(slot)); 2338 rack_log_to_start(rack, cts, hpts_timeout, slot, 1); 2339 } else { 2340 /* 2341 * Arrange for the hpts to kick back in after the 2342 * t-o if the t-o does not cause a send. 2343 */ 2344 if (rack->rc_inp->inp_in_hpts == 0) 2345 tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(hpts_timeout)); 2346 rack_log_to_start(rack, cts, hpts_timeout, slot, 0); 2347 } 2348 } else if (hpts_timeout) { 2349 if (rack->rc_inp->inp_in_hpts == 0) 2350 tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(hpts_timeout)); 2351 rack_log_to_start(rack, cts, hpts_timeout, slot, 0); 2352 } else { 2353 /* No timer starting */ 2354 #ifdef INVARIANTS 2355 if (SEQ_GT(tp->snd_max, tp->snd_una)) { 2356 panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?", 2357 tp, rack, tot_len_this_send, cts, slot, hpts_timeout); 2358 } 2359 #endif 2360 } 2361 rack->rc_tmr_stopped = 0; 2362 if (slot) 2363 rack_log_type_bbrsnd(rack, tot_len_this_send, slot, cts); 2364 } 2365 2366 /* 2367 * RACK Timer, here we simply do logging and house keeping. 2368 * the normal rack_output() function will call the 2369 * appropriate thing to check if we need to do a RACK retransmit. 2370 * We return 1, saying don't proceed with rack_output only 2371 * when all timers have been stopped (destroyed PCB?). 2372 */ 2373 static int 2374 rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) 2375 { 2376 /* 2377 * This timer simply provides an internal trigger to send out data. 2378 * The check_recovery_mode call will see if there are needed 2379 * retransmissions, if so we will enter fast-recovery. The output 2380 * call may or may not do the same thing depending on sysctl 2381 * settings. 2382 */ 2383 struct rack_sendmap *rsm; 2384 int32_t recovery; 2385 2386 if (tp->t_timers->tt_flags & TT_STOPPED) { 2387 return (1); 2388 } 2389 if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) { 2390 /* Its not time yet */ 2391 return (0); 2392 } 2393 rack_log_to_event(rack, RACK_TO_FRM_RACK); 2394 recovery = IN_RECOVERY(tp->t_flags); 2395 counter_u64_add(rack_to_tot, 1); 2396 if (rack->r_state && (rack->r_state != tp->t_state)) 2397 rack_set_state(tp, rack); 2398 rsm = rack_check_recovery_mode(tp, cts); 2399 if (rsm) { 2400 uint32_t rtt; 2401 2402 rtt = rack->rc_rack_rtt; 2403 if (rtt == 0) 2404 rtt = 1; 2405 if ((recovery == 0) && 2406 (rack->r_ctl.rc_prr_sndcnt < tp->t_maxseg)) { 2407 /* 2408 * The rack-timeout that enter's us into recovery 2409 * will force out one MSS and set us up so that we 2410 * can do one more send in 2*rtt (transitioning the 2411 * rack timeout into a rack-tlp). 2412 */ 2413 rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; 2414 } else if ((rack->r_ctl.rc_prr_sndcnt < tp->t_maxseg) && 2415 ((rsm->r_end - rsm->r_start) > rack->r_ctl.rc_prr_sndcnt)) { 2416 /* 2417 * When a rack timer goes, we have to send at 2418 * least one segment. They will be paced a min of 1ms 2419 * apart via the next rack timer (or further 2420 * if the rack timer dictates it). 2421 */ 2422 rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; 2423 } 2424 } else { 2425 /* This is a case that should happen rarely if ever */ 2426 counter_u64_add(rack_tlp_does_nada, 1); 2427 #ifdef TCP_BLACKBOX 2428 tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true); 2429 #endif 2430 rack->r_ctl.rc_resend = TAILQ_FIRST(&rack->r_ctl.rc_tmap); 2431 } 2432 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK; 2433 return (0); 2434 } 2435 2436 /* 2437 * TLP Timer, here we simply setup what segment we want to 2438 * have the TLP expire on, the normal rack_output() will then 2439 * send it out. 2440 * 2441 * We return 1, saying don't proceed with rack_output only 2442 * when all timers have been stopped (destroyed PCB?). 2443 */ 2444 static int 2445 rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) 2446 { 2447 /* 2448 * Tail Loss Probe. 2449 */ 2450 struct rack_sendmap *rsm = NULL; 2451 struct socket *so; 2452 uint32_t amm, old_prr_snd = 0; 2453 uint32_t out, avail; 2454 2455 if (tp->t_timers->tt_flags & TT_STOPPED) { 2456 return (1); 2457 } 2458 if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) { 2459 /* Its not time yet */ 2460 return (0); 2461 } 2462 if (rack_progress_timeout_check(tp)) { 2463 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); 2464 return (1); 2465 } 2466 /* 2467 * A TLP timer has expired. We have been idle for 2 rtts. So we now 2468 * need to figure out how to force a full MSS segment out. 2469 */ 2470 rack_log_to_event(rack, RACK_TO_FRM_TLP); 2471 counter_u64_add(rack_tlp_tot, 1); 2472 if (rack->r_state && (rack->r_state != tp->t_state)) 2473 rack_set_state(tp, rack); 2474 so = tp->t_inpcb->inp_socket; 2475 avail = sbavail(&so->so_snd); 2476 out = tp->snd_max - tp->snd_una; 2477 rack->rc_timer_up = 1; 2478 /* 2479 * If we are in recovery we can jazz out a segment if new data is 2480 * present simply by setting rc_prr_sndcnt to a segment. 2481 */ 2482 if ((avail > out) && 2483 ((rack_always_send_oldest == 0) || (TAILQ_EMPTY(&rack->r_ctl.rc_tmap)))) { 2484 /* New data is available */ 2485 amm = avail - out; 2486 if (amm > tp->t_maxseg) { 2487 amm = tp->t_maxseg; 2488 } else if ((amm < tp->t_maxseg) && ((tp->t_flags & TF_NODELAY) == 0)) { 2489 /* not enough to fill a MTU and no-delay is off */ 2490 goto need_retran; 2491 } 2492 if (IN_RECOVERY(tp->t_flags)) { 2493 /* Unlikely */ 2494 old_prr_snd = rack->r_ctl.rc_prr_sndcnt; 2495 if (out + amm <= tp->snd_wnd) 2496 rack->r_ctl.rc_prr_sndcnt = amm; 2497 else 2498 goto need_retran; 2499 } else { 2500 /* Set the send-new override */ 2501 if (out + amm <= tp->snd_wnd) 2502 rack->r_ctl.rc_tlp_new_data = amm; 2503 else 2504 goto need_retran; 2505 } 2506 rack->r_ctl.rc_tlp_seg_send_cnt = 0; 2507 rack->r_ctl.rc_last_tlp_seq = tp->snd_max; 2508 rack->r_ctl.rc_tlpsend = NULL; 2509 counter_u64_add(rack_tlp_newdata, 1); 2510 goto send; 2511 } 2512 need_retran: 2513 /* 2514 * Ok we need to arrange the last un-acked segment to be re-sent, or 2515 * optionally the first un-acked segment. 2516 */ 2517 if (rack_always_send_oldest) 2518 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); 2519 else { 2520 rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_map, rack_sendmap, r_next); 2521 if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) { 2522 rsm = rack_find_high_nonack(rack, rsm); 2523 } 2524 } 2525 if (rsm == NULL) { 2526 counter_u64_add(rack_tlp_does_nada, 1); 2527 #ifdef TCP_BLACKBOX 2528 tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true); 2529 #endif 2530 goto out; 2531 } 2532 if ((rsm->r_end - rsm->r_start) > tp->t_maxseg) { 2533 /* 2534 * We need to split this the last segment in two. 2535 */ 2536 int32_t idx; 2537 struct rack_sendmap *nrsm; 2538 2539 nrsm = rack_alloc(rack); 2540 if (nrsm == NULL) { 2541 /* 2542 * No memory to split, we will just exit and punt 2543 * off to the RXT timer. 2544 */ 2545 counter_u64_add(rack_tlp_does_nada, 1); 2546 goto out; 2547 } 2548 nrsm->r_start = (rsm->r_end - tp->t_maxseg); 2549 nrsm->r_end = rsm->r_end; 2550 nrsm->r_rtr_cnt = rsm->r_rtr_cnt; 2551 nrsm->r_flags = rsm->r_flags; 2552 nrsm->r_sndcnt = rsm->r_sndcnt; 2553 nrsm->r_rtr_bytes = 0; 2554 rsm->r_end = nrsm->r_start; 2555 for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { 2556 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; 2557 } 2558 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_map, rsm, nrsm, r_next); 2559 if (rsm->r_in_tmap) { 2560 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 2561 nrsm->r_in_tmap = 1; 2562 } 2563 rsm->r_flags &= (~RACK_HAS_FIN); 2564 rsm = nrsm; 2565 } 2566 rack->r_ctl.rc_tlpsend = rsm; 2567 rack->r_ctl.rc_tlp_rtx_out = 1; 2568 if (rsm->r_start == rack->r_ctl.rc_last_tlp_seq) { 2569 rack->r_ctl.rc_tlp_seg_send_cnt++; 2570 tp->t_rxtshift++; 2571 } else { 2572 rack->r_ctl.rc_last_tlp_seq = rsm->r_start; 2573 rack->r_ctl.rc_tlp_seg_send_cnt = 1; 2574 } 2575 send: 2576 rack->r_ctl.rc_tlp_send_cnt++; 2577 if (rack->r_ctl.rc_tlp_send_cnt > rack_tlp_max_resend) { 2578 /* 2579 * Can't [re]/transmit a segment we have not heard from the 2580 * peer in max times. We need the retransmit timer to take 2581 * over. 2582 */ 2583 restore: 2584 rack->r_ctl.rc_tlpsend = NULL; 2585 if (rsm) 2586 rsm->r_flags &= ~RACK_TLP; 2587 rack->r_ctl.rc_prr_sndcnt = old_prr_snd; 2588 counter_u64_add(rack_tlp_retran_fail, 1); 2589 goto out; 2590 } else if (rsm) { 2591 rsm->r_flags |= RACK_TLP; 2592 } 2593 if (rsm && (rsm->r_start == rack->r_ctl.rc_last_tlp_seq) && 2594 (rack->r_ctl.rc_tlp_seg_send_cnt > rack_tlp_max_resend)) { 2595 /* 2596 * We don't want to send a single segment more than the max 2597 * either. 2598 */ 2599 goto restore; 2600 } 2601 rack->r_timer_override = 1; 2602 rack->r_tlp_running = 1; 2603 rack->rc_tlp_in_progress = 1; 2604 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP; 2605 return (0); 2606 out: 2607 rack->rc_timer_up = 0; 2608 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP; 2609 return (0); 2610 } 2611 2612 /* 2613 * Delayed ack Timer, here we simply need to setup the 2614 * ACK_NOW flag and remove the DELACK flag. From there 2615 * the output routine will send the ack out. 2616 * 2617 * We only return 1, saying don't proceed, if all timers 2618 * are stopped (destroyed PCB?). 2619 */ 2620 static int 2621 rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) 2622 { 2623 if (tp->t_timers->tt_flags & TT_STOPPED) { 2624 return (1); 2625 } 2626 rack_log_to_event(rack, RACK_TO_FRM_DELACK); 2627 tp->t_flags &= ~TF_DELACK; 2628 tp->t_flags |= TF_ACKNOW; 2629 TCPSTAT_INC(tcps_delack); 2630 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK; 2631 return (0); 2632 } 2633 2634 /* 2635 * Persists timer, here we simply need to setup the 2636 * FORCE-DATA flag the output routine will send 2637 * the one byte send. 2638 * 2639 * We only return 1, saying don't proceed, if all timers 2640 * are stopped (destroyed PCB?). 2641 */ 2642 static int 2643 rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) 2644 { 2645 struct inpcb *inp; 2646 int32_t retval = 0; 2647 2648 inp = tp->t_inpcb; 2649 2650 if (tp->t_timers->tt_flags & TT_STOPPED) { 2651 return (1); 2652 } 2653 if (rack->rc_in_persist == 0) 2654 return (0); 2655 if (rack_progress_timeout_check(tp)) { 2656 tcp_set_inp_to_drop(inp, ETIMEDOUT); 2657 return (1); 2658 } 2659 KASSERT(inp != NULL, ("%s: tp %p tp->t_inpcb == NULL", __func__, tp)); 2660 /* 2661 * Persistence timer into zero window. Force a byte to be output, if 2662 * possible. 2663 */ 2664 TCPSTAT_INC(tcps_persisttimeo); 2665 /* 2666 * Hack: if the peer is dead/unreachable, we do not time out if the 2667 * window is closed. After a full backoff, drop the connection if 2668 * the idle time (no responses to probes) reaches the maximum 2669 * backoff that we would use if retransmitting. 2670 */ 2671 if (tp->t_rxtshift == TCP_MAXRXTSHIFT && 2672 (ticks - tp->t_rcvtime >= tcp_maxpersistidle || 2673 ticks - tp->t_rcvtime >= TCP_REXMTVAL(tp) * tcp_totbackoff)) { 2674 TCPSTAT_INC(tcps_persistdrop); 2675 retval = 1; 2676 tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT); 2677 goto out; 2678 } 2679 if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) && 2680 tp->snd_una == tp->snd_max) 2681 rack_exit_persist(tp, rack); 2682 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT; 2683 /* 2684 * If the user has closed the socket then drop a persisting 2685 * connection after a much reduced timeout. 2686 */ 2687 if (tp->t_state > TCPS_CLOSE_WAIT && 2688 (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) { 2689 retval = 1; 2690 TCPSTAT_INC(tcps_persistdrop); 2691 tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT); 2692 goto out; 2693 } 2694 tp->t_flags |= TF_FORCEDATA; 2695 out: 2696 rack_log_to_event(rack, RACK_TO_FRM_PERSIST); 2697 return (retval); 2698 } 2699 2700 /* 2701 * If a keepalive goes off, we had no other timers 2702 * happening. We always return 1 here since this 2703 * routine either drops the connection or sends 2704 * out a segment with respond. 2705 */ 2706 static int 2707 rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) 2708 { 2709 struct tcptemp *t_template; 2710 struct inpcb *inp; 2711 2712 if (tp->t_timers->tt_flags & TT_STOPPED) { 2713 return (1); 2714 } 2715 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP; 2716 inp = tp->t_inpcb; 2717 rack_log_to_event(rack, RACK_TO_FRM_KEEP); 2718 /* 2719 * Keep-alive timer went off; send something or drop connection if 2720 * idle for too long. 2721 */ 2722 TCPSTAT_INC(tcps_keeptimeo); 2723 if (tp->t_state < TCPS_ESTABLISHED) 2724 goto dropit; 2725 if ((tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && 2726 tp->t_state <= TCPS_CLOSING) { 2727 if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp)) 2728 goto dropit; 2729 /* 2730 * Send a packet designed to force a response if the peer is 2731 * up and reachable: either an ACK if the connection is 2732 * still alive, or an RST if the peer has closed the 2733 * connection due to timeout or reboot. Using sequence 2734 * number tp->snd_una-1 causes the transmitted zero-length 2735 * segment to lie outside the receive window; by the 2736 * protocol spec, this requires the correspondent TCP to 2737 * respond. 2738 */ 2739 TCPSTAT_INC(tcps_keepprobe); 2740 t_template = tcpip_maketemplate(inp); 2741 if (t_template) { 2742 tcp_respond(tp, t_template->tt_ipgen, 2743 &t_template->tt_t, (struct mbuf *)NULL, 2744 tp->rcv_nxt, tp->snd_una - 1, 0); 2745 free(t_template, M_TEMP); 2746 } 2747 } 2748 rack_start_hpts_timer(rack, tp, cts, __LINE__, 0, 0, 0); 2749 return (1); 2750 dropit: 2751 TCPSTAT_INC(tcps_keepdrops); 2752 tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT); 2753 return (1); 2754 } 2755 2756 /* 2757 * Retransmit helper function, clear up all the ack 2758 * flags and take care of important book keeping. 2759 */ 2760 static void 2761 rack_remxt_tmr(struct tcpcb *tp) 2762 { 2763 /* 2764 * The retransmit timer went off, all sack'd blocks must be 2765 * un-acked. 2766 */ 2767 struct rack_sendmap *rsm, *trsm = NULL; 2768 struct tcp_rack *rack; 2769 int32_t cnt = 0; 2770 2771 rack = (struct tcp_rack *)tp->t_fb_ptr; 2772 rack_timer_cancel(tp, rack, tcp_ts_getticks(), __LINE__); 2773 rack_log_to_event(rack, RACK_TO_FRM_TMR); 2774 if (rack->r_state && (rack->r_state != tp->t_state)) 2775 rack_set_state(tp, rack); 2776 /* 2777 * Ideally we would like to be able to 2778 * mark SACK-PASS on anything not acked here. 2779 * However, if we do that we would burst out 2780 * all that data 1ms apart. This would be unwise, 2781 * so for now we will just let the normal rxt timer 2782 * and tlp timer take care of it. 2783 */ 2784 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_map, r_next) { 2785 if (rsm->r_flags & RACK_ACKED) { 2786 cnt++; 2787 rsm->r_sndcnt = 0; 2788 if (rsm->r_in_tmap == 0) { 2789 /* We must re-add it back to the tlist */ 2790 if (trsm == NULL) { 2791 TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext); 2792 } else { 2793 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext); 2794 } 2795 rsm->r_in_tmap = 1; 2796 trsm = rsm; 2797 } 2798 } 2799 rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS); 2800 } 2801 /* Clear the count (we just un-acked them) */ 2802 rack->r_ctl.rc_sacked = 0; 2803 /* Clear the tlp rtx mark */ 2804 rack->r_ctl.rc_tlp_rtx_out = 0; 2805 rack->r_ctl.rc_tlp_seg_send_cnt = 0; 2806 rack->r_ctl.rc_resend = TAILQ_FIRST(&rack->r_ctl.rc_map); 2807 /* Setup so we send one segment */ 2808 if (rack->r_ctl.rc_prr_sndcnt < tp->t_maxseg) 2809 rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; 2810 rack->r_timer_override = 1; 2811 } 2812 2813 /* 2814 * Re-transmit timeout! If we drop the PCB we will return 1, otherwise 2815 * we will setup to retransmit the lowest seq number outstanding. 2816 */ 2817 static int 2818 rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) 2819 { 2820 int32_t rexmt; 2821 struct inpcb *inp; 2822 int32_t retval = 0; 2823 2824 inp = tp->t_inpcb; 2825 if (tp->t_timers->tt_flags & TT_STOPPED) { 2826 return (1); 2827 } 2828 if (rack_progress_timeout_check(tp)) { 2829 tcp_set_inp_to_drop(inp, ETIMEDOUT); 2830 return (1); 2831 } 2832 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT; 2833 if (TCPS_HAVEESTABLISHED(tp->t_state) && 2834 (tp->snd_una == tp->snd_max)) { 2835 /* Nothing outstanding .. nothing to do */ 2836 return (0); 2837 } 2838 /* 2839 * Retransmission timer went off. Message has not been acked within 2840 * retransmit interval. Back off to a longer retransmit interval 2841 * and retransmit one segment. 2842 */ 2843 if (++tp->t_rxtshift > TCP_MAXRXTSHIFT) { 2844 tp->t_rxtshift = TCP_MAXRXTSHIFT; 2845 TCPSTAT_INC(tcps_timeoutdrop); 2846 retval = 1; 2847 tcp_set_inp_to_drop(rack->rc_inp, 2848 (tp->t_softerror ? (uint16_t) tp->t_softerror : ETIMEDOUT)); 2849 goto out; 2850 } 2851 rack_remxt_tmr(tp); 2852 if (tp->t_state == TCPS_SYN_SENT) { 2853 /* 2854 * If the SYN was retransmitted, indicate CWND to be limited 2855 * to 1 segment in cc_conn_init(). 2856 */ 2857 tp->snd_cwnd = 1; 2858 } else if (tp->t_rxtshift == 1) { 2859 /* 2860 * first retransmit; record ssthresh and cwnd so they can be 2861 * recovered if this turns out to be a "bad" retransmit. A 2862 * retransmit is considered "bad" if an ACK for this segment 2863 * is received within RTT/2 interval; the assumption here is 2864 * that the ACK was already in flight. See "On Estimating 2865 * End-to-End Network Path Properties" by Allman and Paxson 2866 * for more details. 2867 */ 2868 tp->snd_cwnd_prev = tp->snd_cwnd; 2869 tp->snd_ssthresh_prev = tp->snd_ssthresh; 2870 tp->snd_recover_prev = tp->snd_recover; 2871 if (IN_FASTRECOVERY(tp->t_flags)) 2872 tp->t_flags |= TF_WASFRECOVERY; 2873 else 2874 tp->t_flags &= ~TF_WASFRECOVERY; 2875 if (IN_CONGRECOVERY(tp->t_flags)) 2876 tp->t_flags |= TF_WASCRECOVERY; 2877 else 2878 tp->t_flags &= ~TF_WASCRECOVERY; 2879 tp->t_badrxtwin = ticks + (tp->t_srtt >> (TCP_RTT_SHIFT + 1)); 2880 tp->t_flags |= TF_PREVVALID; 2881 } else 2882 tp->t_flags &= ~TF_PREVVALID; 2883 TCPSTAT_INC(tcps_rexmttimeo); 2884 if ((tp->t_state == TCPS_SYN_SENT) || 2885 (tp->t_state == TCPS_SYN_RECEIVED)) 2886 rexmt = MSEC_2_TICKS(RACK_INITIAL_RTO * tcp_syn_backoff[tp->t_rxtshift]); 2887 else 2888 rexmt = TCP_REXMTVAL(tp) * tcp_backoff[tp->t_rxtshift]; 2889 TCPT_RANGESET(tp->t_rxtcur, rexmt, 2890 max(MSEC_2_TICKS(rack_rto_min), rexmt), 2891 MSEC_2_TICKS(rack_rto_max)); 2892 /* 2893 * We enter the path for PLMTUD if connection is established or, if 2894 * connection is FIN_WAIT_1 status, reason for the last is that if 2895 * amount of data we send is very small, we could send it in couple 2896 * of packets and process straight to FIN. In that case we won't 2897 * catch ESTABLISHED state. 2898 */ 2899 if (V_tcp_pmtud_blackhole_detect && (((tp->t_state == TCPS_ESTABLISHED)) 2900 || (tp->t_state == TCPS_FIN_WAIT_1))) { 2901 #ifdef INET6 2902 int32_t isipv6; 2903 #endif 2904 2905 /* 2906 * Idea here is that at each stage of mtu probe (usually, 2907 * 1448 -> 1188 -> 524) should be given 2 chances to recover 2908 * before further clamping down. 'tp->t_rxtshift % 2 == 0' 2909 * should take care of that. 2910 */ 2911 if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) == 2912 (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) && 2913 (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 && 2914 tp->t_rxtshift % 2 == 0)) { 2915 /* 2916 * Enter Path MTU Black-hole Detection mechanism: - 2917 * Disable Path MTU Discovery (IP "DF" bit). - 2918 * Reduce MTU to lower value than what we negotiated 2919 * with peer. 2920 */ 2921 if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) { 2922 /* Record that we may have found a black hole. */ 2923 tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE; 2924 /* Keep track of previous MSS. */ 2925 tp->t_pmtud_saved_maxseg = tp->t_maxseg; 2926 } 2927 2928 /* 2929 * Reduce the MSS to blackhole value or to the 2930 * default in an attempt to retransmit. 2931 */ 2932 #ifdef INET6 2933 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) ? 1 : 0; 2934 if (isipv6 && 2935 tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) { 2936 /* Use the sysctl tuneable blackhole MSS. */ 2937 tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss; 2938 TCPSTAT_INC(tcps_pmtud_blackhole_activated); 2939 } else if (isipv6) { 2940 /* Use the default MSS. */ 2941 tp->t_maxseg = V_tcp_v6mssdflt; 2942 /* 2943 * Disable Path MTU Discovery when we switch 2944 * to minmss. 2945 */ 2946 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 2947 TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss); 2948 } 2949 #endif 2950 #if defined(INET6) && defined(INET) 2951 else 2952 #endif 2953 #ifdef INET 2954 if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) { 2955 /* Use the sysctl tuneable blackhole MSS. */ 2956 tp->t_maxseg = V_tcp_pmtud_blackhole_mss; 2957 TCPSTAT_INC(tcps_pmtud_blackhole_activated); 2958 } else { 2959 /* Use the default MSS. */ 2960 tp->t_maxseg = V_tcp_mssdflt; 2961 /* 2962 * Disable Path MTU Discovery when we switch 2963 * to minmss. 2964 */ 2965 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 2966 TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss); 2967 } 2968 #endif 2969 } else { 2970 /* 2971 * If further retransmissions are still unsuccessful 2972 * with a lowered MTU, maybe this isn't a blackhole 2973 * and we restore the previous MSS and blackhole 2974 * detection flags. The limit '6' is determined by 2975 * giving each probe stage (1448, 1188, 524) 2 2976 * chances to recover. 2977 */ 2978 if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) && 2979 (tp->t_rxtshift >= 6)) { 2980 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 2981 tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE; 2982 tp->t_maxseg = tp->t_pmtud_saved_maxseg; 2983 TCPSTAT_INC(tcps_pmtud_blackhole_failed); 2984 } 2985 } 2986 } 2987 /* 2988 * Disable RFC1323 and SACK if we haven't got any response to our 2989 * third SYN to work-around some broken terminal servers (most of 2990 * which have hopefully been retired) that have bad VJ header 2991 * compression code which trashes TCP segments containing 2992 * unknown-to-them TCP options. 2993 */ 2994 if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) && 2995 (tp->t_rxtshift == 3)) 2996 tp->t_flags &= ~(TF_REQ_SCALE | TF_REQ_TSTMP | TF_SACK_PERMIT); 2997 /* 2998 * If we backed off this far, our srtt estimate is probably bogus. 2999 * Clobber it so we'll take the next rtt measurement as our srtt; 3000 * move the current srtt into rttvar to keep the current retransmit 3001 * times until then. 3002 */ 3003 if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) { 3004 #ifdef INET6 3005 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0) 3006 in6_losing(tp->t_inpcb); 3007 else 3008 #endif 3009 in_losing(tp->t_inpcb); 3010 tp->t_rttvar += (tp->t_srtt >> TCP_RTT_SHIFT); 3011 tp->t_srtt = 0; 3012 } 3013 if (rack_use_sack_filter) 3014 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una); 3015 tp->snd_recover = tp->snd_max; 3016 tp->t_flags |= TF_ACKNOW; 3017 tp->t_rtttime = 0; 3018 rack_cong_signal(tp, NULL, CC_RTO); 3019 out: 3020 return (retval); 3021 } 3022 3023 static int 3024 rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling) 3025 { 3026 int32_t ret = 0; 3027 int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK); 3028 3029 if (timers == 0) { 3030 return (0); 3031 } 3032 if (tp->t_state == TCPS_LISTEN) { 3033 /* no timers on listen sockets */ 3034 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) 3035 return (0); 3036 return (1); 3037 } 3038 if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) { 3039 uint32_t left; 3040 3041 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) { 3042 ret = -1; 3043 rack_log_to_processing(rack, cts, ret, 0); 3044 return (0); 3045 } 3046 if (hpts_calling == 0) { 3047 ret = -2; 3048 rack_log_to_processing(rack, cts, ret, 0); 3049 return (0); 3050 } 3051 /* 3052 * Ok our timer went off early and we are not paced false 3053 * alarm, go back to sleep. 3054 */ 3055 ret = -3; 3056 left = rack->r_ctl.rc_timer_exp - cts; 3057 tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(left)); 3058 rack_log_to_processing(rack, cts, ret, left); 3059 rack->rc_last_pto_set = 0; 3060 return (1); 3061 } 3062 rack->rc_tmr_stopped = 0; 3063 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK; 3064 if (timers & PACE_TMR_DELACK) { 3065 ret = rack_timeout_delack(tp, rack, cts); 3066 } else if (timers & PACE_TMR_RACK) { 3067 ret = rack_timeout_rack(tp, rack, cts); 3068 } else if (timers & PACE_TMR_TLP) { 3069 ret = rack_timeout_tlp(tp, rack, cts); 3070 } else if (timers & PACE_TMR_RXT) { 3071 ret = rack_timeout_rxt(tp, rack, cts); 3072 } else if (timers & PACE_TMR_PERSIT) { 3073 ret = rack_timeout_persist(tp, rack, cts); 3074 } else if (timers & PACE_TMR_KEEP) { 3075 ret = rack_timeout_keepalive(tp, rack, cts); 3076 } 3077 rack_log_to_processing(rack, cts, ret, timers); 3078 return (ret); 3079 } 3080 3081 static void 3082 rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line) 3083 { 3084 uint8_t hpts_removed = 0; 3085 3086 if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) && 3087 TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to)) { 3088 tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT); 3089 hpts_removed = 1; 3090 } 3091 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) { 3092 rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK; 3093 if (rack->rc_inp->inp_in_hpts && 3094 ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) { 3095 /* 3096 * Canceling timer's when we have no output being 3097 * paced. We also must remove ourselves from the 3098 * hpts. 3099 */ 3100 tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT); 3101 hpts_removed = 1; 3102 } 3103 rack_log_to_cancel(rack, hpts_removed, line); 3104 rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK); 3105 } 3106 } 3107 3108 static void 3109 rack_timer_stop(struct tcpcb *tp, uint32_t timer_type) 3110 { 3111 return; 3112 } 3113 3114 static int 3115 rack_stopall(struct tcpcb *tp) 3116 { 3117 struct tcp_rack *rack; 3118 rack = (struct tcp_rack *)tp->t_fb_ptr; 3119 rack->t_timers_stopped = 1; 3120 return (0); 3121 } 3122 3123 static void 3124 rack_timer_activate(struct tcpcb *tp, uint32_t timer_type, uint32_t delta) 3125 { 3126 return; 3127 } 3128 3129 static int 3130 rack_timer_active(struct tcpcb *tp, uint32_t timer_type) 3131 { 3132 return (0); 3133 } 3134 3135 static void 3136 rack_stop_all_timers(struct tcpcb *tp) 3137 { 3138 struct tcp_rack *rack; 3139 3140 /* 3141 * Assure no timers are running. 3142 */ 3143 if (tcp_timer_active(tp, TT_PERSIST)) { 3144 /* We enter in persists, set the flag appropriately */ 3145 rack = (struct tcp_rack *)tp->t_fb_ptr; 3146 rack->rc_in_persist = 1; 3147 } 3148 tcp_timer_suspend(tp, TT_PERSIST); 3149 tcp_timer_suspend(tp, TT_REXMT); 3150 tcp_timer_suspend(tp, TT_KEEP); 3151 tcp_timer_suspend(tp, TT_DELACK); 3152 } 3153 3154 static void 3155 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack, 3156 struct rack_sendmap *rsm, uint32_t ts) 3157 { 3158 int32_t idx; 3159 3160 rsm->r_rtr_cnt++; 3161 rsm->r_sndcnt++; 3162 if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) { 3163 rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS; 3164 rsm->r_flags |= RACK_OVERMAX; 3165 } 3166 if ((rsm->r_rtr_cnt > 1) && (rack->r_tlp_running == 0)) { 3167 rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start); 3168 rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start); 3169 } 3170 idx = rsm->r_rtr_cnt - 1; 3171 rsm->r_tim_lastsent[idx] = ts; 3172 if (rsm->r_flags & RACK_ACKED) { 3173 /* Problably MTU discovery messing with us */ 3174 rsm->r_flags &= ~RACK_ACKED; 3175 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); 3176 } 3177 if (rsm->r_in_tmap) { 3178 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext); 3179 } 3180 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext); 3181 rsm->r_in_tmap = 1; 3182 if (rsm->r_flags & RACK_SACK_PASSED) { 3183 /* We have retransmitted due to the SACK pass */ 3184 rsm->r_flags &= ~RACK_SACK_PASSED; 3185 rsm->r_flags |= RACK_WAS_SACKPASS; 3186 } 3187 /* Update memory for next rtr */ 3188 rack->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next); 3189 } 3190 3191 3192 static uint32_t 3193 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack, 3194 struct rack_sendmap *rsm, uint32_t ts, int32_t * lenp) 3195 { 3196 /* 3197 * We (re-)transmitted starting at rsm->r_start for some length 3198 * (possibly less than r_end. 3199 */ 3200 struct rack_sendmap *nrsm; 3201 uint32_t c_end; 3202 int32_t len; 3203 int32_t idx; 3204 3205 len = *lenp; 3206 c_end = rsm->r_start + len; 3207 if (SEQ_GEQ(c_end, rsm->r_end)) { 3208 /* 3209 * We retransmitted the whole piece or more than the whole 3210 * slopping into the next rsm. 3211 */ 3212 rack_update_rsm(tp, rack, rsm, ts); 3213 if (c_end == rsm->r_end) { 3214 *lenp = 0; 3215 return (0); 3216 } else { 3217 int32_t act_len; 3218 3219 /* Hangs over the end return whats left */ 3220 act_len = rsm->r_end - rsm->r_start; 3221 *lenp = (len - act_len); 3222 return (rsm->r_end); 3223 } 3224 /* We don't get out of this block. */ 3225 } 3226 /* 3227 * Here we retransmitted less than the whole thing which means we 3228 * have to split this into what was transmitted and what was not. 3229 */ 3230 nrsm = rack_alloc(rack); 3231 if (nrsm == NULL) { 3232 /* 3233 * We can't get memory, so lets not proceed. 3234 */ 3235 *lenp = 0; 3236 return (0); 3237 } 3238 /* 3239 * So here we are going to take the original rsm and make it what we 3240 * retransmitted. nrsm will be the tail portion we did not 3241 * retransmit. For example say the chunk was 1, 11 (10 bytes). And 3242 * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to 3243 * 1, 6 and the new piece will be 6, 11. 3244 */ 3245 nrsm->r_start = c_end; 3246 nrsm->r_end = rsm->r_end; 3247 nrsm->r_rtr_cnt = rsm->r_rtr_cnt; 3248 nrsm->r_flags = rsm->r_flags; 3249 nrsm->r_sndcnt = rsm->r_sndcnt; 3250 nrsm->r_rtr_bytes = 0; 3251 rsm->r_end = c_end; 3252 for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { 3253 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; 3254 } 3255 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_map, rsm, nrsm, r_next); 3256 if (rsm->r_in_tmap) { 3257 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 3258 nrsm->r_in_tmap = 1; 3259 } 3260 rsm->r_flags &= (~RACK_HAS_FIN); 3261 rack_update_rsm(tp, rack, rsm, ts); 3262 *lenp = 0; 3263 return (0); 3264 } 3265 3266 3267 static void 3268 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len, 3269 uint32_t seq_out, uint8_t th_flags, int32_t err, uint32_t ts, 3270 uint8_t pass, struct rack_sendmap *hintrsm) 3271 { 3272 struct tcp_rack *rack; 3273 struct rack_sendmap *rsm, *nrsm; 3274 register uint32_t snd_max, snd_una; 3275 int32_t idx; 3276 3277 /* 3278 * Add to the RACK log of packets in flight or retransmitted. If 3279 * there is a TS option we will use the TS echoed, if not we will 3280 * grab a TS. 3281 * 3282 * Retransmissions will increment the count and move the ts to its 3283 * proper place. Note that if options do not include TS's then we 3284 * won't be able to effectively use the ACK for an RTT on a retran. 3285 * 3286 * Notes about r_start and r_end. Lets consider a send starting at 3287 * sequence 1 for 10 bytes. In such an example the r_start would be 3288 * 1 (starting sequence) but the r_end would be r_start+len i.e. 11. 3289 * This means that r_end is actually the first sequence for the next 3290 * slot (11). 3291 * 3292 */ 3293 /* 3294 * If err is set what do we do XXXrrs? should we not add the thing? 3295 * -- i.e. return if err != 0 or should we pretend we sent it? -- 3296 * i.e. proceed with add ** do this for now. 3297 */ 3298 INP_WLOCK_ASSERT(tp->t_inpcb); 3299 if (err) 3300 /* 3301 * We don't log errors -- we could but snd_max does not 3302 * advance in this case either. 3303 */ 3304 return; 3305 3306 if (th_flags & TH_RST) { 3307 /* 3308 * We don't log resets and we return immediately from 3309 * sending 3310 */ 3311 return; 3312 } 3313 rack = (struct tcp_rack *)tp->t_fb_ptr; 3314 snd_una = tp->snd_una; 3315 if (SEQ_LEQ((seq_out + len), snd_una)) { 3316 /* Are sending an old segment to induce an ack (keep-alive)? */ 3317 return; 3318 } 3319 if (SEQ_LT(seq_out, snd_una)) { 3320 /* huh? should we panic? */ 3321 uint32_t end; 3322 3323 end = seq_out + len; 3324 seq_out = snd_una; 3325 len = end - seq_out; 3326 } 3327 snd_max = tp->snd_max; 3328 if (th_flags & (TH_SYN | TH_FIN)) { 3329 /* 3330 * The call to rack_log_output is made before bumping 3331 * snd_max. This means we can record one extra byte on a SYN 3332 * or FIN if seq_out is adding more on and a FIN is present 3333 * (and we are not resending). 3334 */ 3335 if (th_flags & TH_SYN) 3336 len++; 3337 if (th_flags & TH_FIN) 3338 len++; 3339 if (SEQ_LT(snd_max, tp->snd_nxt)) { 3340 /* 3341 * The add/update as not been done for the FIN/SYN 3342 * yet. 3343 */ 3344 snd_max = tp->snd_nxt; 3345 } 3346 } 3347 if (len == 0) { 3348 /* We don't log zero window probes */ 3349 return; 3350 } 3351 rack->r_ctl.rc_time_last_sent = ts; 3352 if (IN_RECOVERY(tp->t_flags)) { 3353 rack->r_ctl.rc_prr_out += len; 3354 } 3355 /* First question is it a retransmission? */ 3356 if (seq_out == snd_max) { 3357 again: 3358 rsm = rack_alloc(rack); 3359 if (rsm == NULL) { 3360 /* 3361 * Hmm out of memory and the tcb got destroyed while 3362 * we tried to wait. 3363 */ 3364 #ifdef INVARIANTS 3365 panic("Out of memory when we should not be rack:%p", rack); 3366 #endif 3367 return; 3368 } 3369 if (th_flags & TH_FIN) { 3370 rsm->r_flags = RACK_HAS_FIN; 3371 } else { 3372 rsm->r_flags = 0; 3373 } 3374 rsm->r_tim_lastsent[0] = ts; 3375 rsm->r_rtr_cnt = 1; 3376 rsm->r_rtr_bytes = 0; 3377 if (th_flags & TH_SYN) { 3378 /* The data space is one beyond snd_una */ 3379 rsm->r_start = seq_out + 1; 3380 rsm->r_end = rsm->r_start + (len - 1); 3381 } else { 3382 /* Normal case */ 3383 rsm->r_start = seq_out; 3384 rsm->r_end = rsm->r_start + len; 3385 } 3386 rsm->r_sndcnt = 0; 3387 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_map, rsm, r_next); 3388 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext); 3389 rsm->r_in_tmap = 1; 3390 return; 3391 } 3392 /* 3393 * If we reach here its a retransmission and we need to find it. 3394 */ 3395 more: 3396 if (hintrsm && (hintrsm->r_start == seq_out)) { 3397 rsm = hintrsm; 3398 hintrsm = NULL; 3399 } else if (rack->r_ctl.rc_next) { 3400 /* We have a hint from a previous run */ 3401 rsm = rack->r_ctl.rc_next; 3402 } else { 3403 /* No hints sorry */ 3404 rsm = NULL; 3405 } 3406 if ((rsm) && (rsm->r_start == seq_out)) { 3407 /* 3408 * We used rc_next or hintrsm to retransmit, hopefully the 3409 * likely case. 3410 */ 3411 seq_out = rack_update_entry(tp, rack, rsm, ts, &len); 3412 if (len == 0) { 3413 return; 3414 } else { 3415 goto more; 3416 } 3417 } 3418 /* Ok it was not the last pointer go through it the hard way. */ 3419 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_map, r_next) { 3420 if (rsm->r_start == seq_out) { 3421 seq_out = rack_update_entry(tp, rack, rsm, ts, &len); 3422 rack->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next); 3423 if (len == 0) { 3424 return; 3425 } else { 3426 continue; 3427 } 3428 } 3429 if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) { 3430 /* Transmitted within this piece */ 3431 /* 3432 * Ok we must split off the front and then let the 3433 * update do the rest 3434 */ 3435 nrsm = rack_alloc(rack); 3436 if (nrsm == NULL) { 3437 #ifdef INVARIANTS 3438 panic("Ran out of memory that was preallocated? rack:%p", rack); 3439 #endif 3440 rack_update_rsm(tp, rack, rsm, ts); 3441 return; 3442 } 3443 /* 3444 * copy rsm to nrsm and then trim the front of rsm 3445 * to not include this part. 3446 */ 3447 nrsm->r_start = seq_out; 3448 nrsm->r_end = rsm->r_end; 3449 nrsm->r_rtr_cnt = rsm->r_rtr_cnt; 3450 nrsm->r_flags = rsm->r_flags; 3451 nrsm->r_sndcnt = rsm->r_sndcnt; 3452 nrsm->r_rtr_bytes = 0; 3453 for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { 3454 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; 3455 } 3456 rsm->r_end = nrsm->r_start; 3457 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_map, rsm, nrsm, r_next); 3458 if (rsm->r_in_tmap) { 3459 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 3460 nrsm->r_in_tmap = 1; 3461 } 3462 rsm->r_flags &= (~RACK_HAS_FIN); 3463 seq_out = rack_update_entry(tp, rack, nrsm, ts, &len); 3464 if (len == 0) { 3465 return; 3466 } 3467 } 3468 } 3469 /* 3470 * Hmm not found in map did they retransmit both old and on into the 3471 * new? 3472 */ 3473 if (seq_out == tp->snd_max) { 3474 goto again; 3475 } else if (SEQ_LT(seq_out, tp->snd_max)) { 3476 #ifdef INVARIANTS 3477 printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n", 3478 seq_out, len, tp->snd_una, tp->snd_max); 3479 printf("Starting Dump of all rack entries\n"); 3480 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_map, r_next) { 3481 printf("rsm:%p start:%u end:%u\n", 3482 rsm, rsm->r_start, rsm->r_end); 3483 } 3484 printf("Dump complete\n"); 3485 panic("seq_out not found rack:%p tp:%p", 3486 rack, tp); 3487 #endif 3488 } else { 3489 #ifdef INVARIANTS 3490 /* 3491 * Hmm beyond sndmax? (only if we are using the new rtt-pack 3492 * flag) 3493 */ 3494 panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p", 3495 seq_out, len, tp->snd_max, tp); 3496 #endif 3497 } 3498 } 3499 3500 /* 3501 * Record one of the RTT updates from an ack into 3502 * our sample structure. 3503 */ 3504 static void 3505 tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt) 3506 { 3507 if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) || 3508 (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) { 3509 rack->r_ctl.rack_rs.rs_rtt_lowest = rtt; 3510 } 3511 if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) || 3512 (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) { 3513 rack->r_ctl.rack_rs.rs_rtt_highest = rtt; 3514 } 3515 rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID; 3516 rack->r_ctl.rack_rs.rs_rtt_tot += rtt; 3517 rack->r_ctl.rack_rs.rs_rtt_cnt++; 3518 } 3519 3520 /* 3521 * Collect new round-trip time estimate 3522 * and update averages and current timeout. 3523 */ 3524 static void 3525 tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp) 3526 { 3527 int32_t delta; 3528 uint32_t o_srtt, o_var; 3529 int32_t rtt; 3530 3531 if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) 3532 /* No valid sample */ 3533 return; 3534 if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) { 3535 /* We are to use the lowest RTT seen in a single ack */ 3536 rtt = rack->r_ctl.rack_rs.rs_rtt_lowest; 3537 } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) { 3538 /* We are to use the highest RTT seen in a single ack */ 3539 rtt = rack->r_ctl.rack_rs.rs_rtt_highest; 3540 } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) { 3541 /* We are to use the average RTT seen in a single ack */ 3542 rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot / 3543 (uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt); 3544 } else { 3545 #ifdef INVARIANTS 3546 panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method); 3547 #endif 3548 return; 3549 } 3550 if (rtt == 0) 3551 rtt = 1; 3552 rack_log_rtt_sample(rack, rtt); 3553 o_srtt = tp->t_srtt; 3554 o_var = tp->t_rttvar; 3555 rack = (struct tcp_rack *)tp->t_fb_ptr; 3556 if (tp->t_srtt != 0) { 3557 /* 3558 * srtt is stored as fixed point with 5 bits after the 3559 * binary point (i.e., scaled by 8). The following magic is 3560 * equivalent to the smoothing algorithm in rfc793 with an 3561 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point). 3562 * Adjust rtt to origin 0. 3563 */ 3564 delta = ((rtt - 1) << TCP_DELTA_SHIFT) 3565 - (tp->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT)); 3566 3567 tp->t_srtt += delta; 3568 if (tp->t_srtt <= 0) 3569 tp->t_srtt = 1; 3570 3571 /* 3572 * We accumulate a smoothed rtt variance (actually, a 3573 * smoothed mean difference), then set the retransmit timer 3574 * to smoothed rtt + 4 times the smoothed variance. rttvar 3575 * is stored as fixed point with 4 bits after the binary 3576 * point (scaled by 16). The following is equivalent to 3577 * rfc793 smoothing with an alpha of .75 (rttvar = 3578 * rttvar*3/4 + |delta| / 4). This replaces rfc793's 3579 * wired-in beta. 3580 */ 3581 if (delta < 0) 3582 delta = -delta; 3583 delta -= tp->t_rttvar >> (TCP_RTTVAR_SHIFT - TCP_DELTA_SHIFT); 3584 tp->t_rttvar += delta; 3585 if (tp->t_rttvar <= 0) 3586 tp->t_rttvar = 1; 3587 if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar) 3588 tp->t_rttbest = tp->t_srtt + tp->t_rttvar; 3589 } else { 3590 /* 3591 * No rtt measurement yet - use the unsmoothed rtt. Set the 3592 * variance to half the rtt (so our first retransmit happens 3593 * at 3*rtt). 3594 */ 3595 tp->t_srtt = rtt << TCP_RTT_SHIFT; 3596 tp->t_rttvar = rtt << (TCP_RTTVAR_SHIFT - 1); 3597 tp->t_rttbest = tp->t_srtt + tp->t_rttvar; 3598 } 3599 TCPSTAT_INC(tcps_rttupdated); 3600 rack_log_rtt_upd(tp, rack, rtt, o_srtt, o_var); 3601 tp->t_rttupdated++; 3602 #ifdef NETFLIX_STATS 3603 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt)); 3604 #endif 3605 tp->t_rxtshift = 0; 3606 3607 /* 3608 * the retransmit should happen at rtt + 4 * rttvar. Because of the 3609 * way we do the smoothing, srtt and rttvar will each average +1/2 3610 * tick of bias. When we compute the retransmit timer, we want 1/2 3611 * tick of rounding and 1 extra tick because of +-1/2 tick 3612 * uncertainty in the firing of the timer. The bias will give us 3613 * exactly the 1.5 tick we need. But, because the bias is 3614 * statistical, we have to test that we don't drop below the minimum 3615 * feasible timer (which is 2 ticks). 3616 */ 3617 TCPT_RANGESET(tp->t_rxtcur, TCP_REXMTVAL(tp), 3618 max(MSEC_2_TICKS(rack_rto_min), rtt + 2), MSEC_2_TICKS(rack_rto_max)); 3619 tp->t_softerror = 0; 3620 } 3621 3622 static void 3623 rack_earlier_retran(struct tcpcb *tp, struct rack_sendmap *rsm, 3624 uint32_t t, uint32_t cts) 3625 { 3626 /* 3627 * For this RSM, we acknowledged the data from a previous 3628 * transmission, not the last one we made. This means we did a false 3629 * retransmit. 3630 */ 3631 struct tcp_rack *rack; 3632 3633 if (rsm->r_flags & RACK_HAS_FIN) { 3634 /* 3635 * The sending of the FIN often is multiple sent when we 3636 * have everything outstanding ack'd. We ignore this case 3637 * since its over now. 3638 */ 3639 return; 3640 } 3641 if (rsm->r_flags & RACK_TLP) { 3642 /* 3643 * We expect TLP's to have this occur. 3644 */ 3645 return; 3646 } 3647 rack = (struct tcp_rack *)tp->t_fb_ptr; 3648 /* should we undo cc changes and exit recovery? */ 3649 if (IN_RECOVERY(tp->t_flags)) { 3650 if (rack->r_ctl.rc_rsm_start == rsm->r_start) { 3651 /* 3652 * Undo what we ratched down and exit recovery if 3653 * possible 3654 */ 3655 EXIT_RECOVERY(tp->t_flags); 3656 tp->snd_recover = tp->snd_una; 3657 if (rack->r_ctl.rc_cwnd_at > tp->snd_cwnd) 3658 tp->snd_cwnd = rack->r_ctl.rc_cwnd_at; 3659 if (rack->r_ctl.rc_ssthresh_at > tp->snd_ssthresh) 3660 tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at; 3661 } 3662 } 3663 if (rsm->r_flags & RACK_WAS_SACKPASS) { 3664 /* 3665 * We retransmitted based on a sack and the earlier 3666 * retransmission ack'd it - re-ordering is occuring. 3667 */ 3668 counter_u64_add(rack_reorder_seen, 1); 3669 rack->r_ctl.rc_reorder_ts = cts; 3670 } 3671 counter_u64_add(rack_badfr, 1); 3672 counter_u64_add(rack_badfr_bytes, (rsm->r_end - rsm->r_start)); 3673 } 3674 3675 3676 static int 3677 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack, 3678 struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type) 3679 { 3680 int32_t i; 3681 uint32_t t; 3682 3683 if (rsm->r_flags & RACK_ACKED) 3684 /* Already done */ 3685 return (0); 3686 3687 3688 if ((rsm->r_rtr_cnt == 1) || 3689 ((ack_type == CUM_ACKED) && 3690 (to->to_flags & TOF_TS) && 3691 (to->to_tsecr) && 3692 (rsm->r_tim_lastsent[rsm->r_rtr_cnt - 1] == to->to_tsecr)) 3693 ) { 3694 /* 3695 * We will only find a matching timestamp if its cum-acked. 3696 * But if its only one retransmission its for-sure matching 3697 * :-) 3698 */ 3699 t = cts - rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]; 3700 if ((int)t <= 0) 3701 t = 1; 3702 if (!tp->t_rttlow || tp->t_rttlow > t) 3703 tp->t_rttlow = t; 3704 if (!rack->r_ctl.rc_rack_min_rtt || 3705 SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) { 3706 rack->r_ctl.rc_rack_min_rtt = t; 3707 if (rack->r_ctl.rc_rack_min_rtt == 0) { 3708 rack->r_ctl.rc_rack_min_rtt = 1; 3709 } 3710 } 3711 tcp_rack_xmit_timer(rack, TCP_TS_TO_TICKS(t) + 1); 3712 if ((rsm->r_flags & RACK_TLP) && 3713 (!IN_RECOVERY(tp->t_flags))) { 3714 /* Segment was a TLP and our retrans matched */ 3715 if (rack->r_ctl.rc_tlp_cwnd_reduce) { 3716 rack->r_ctl.rc_rsm_start = tp->snd_max; 3717 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd; 3718 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh; 3719 rack_cong_signal(tp, NULL, CC_NDUPACK); 3720 /* 3721 * When we enter recovery we need to assure 3722 * we send one packet. 3723 */ 3724 rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; 3725 } else 3726 rack->r_ctl.rc_tlp_rtx_out = 0; 3727 } 3728 if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) { 3729 /* New more recent rack_tmit_time */ 3730 rack->r_ctl.rc_rack_tmit_time = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]; 3731 rack->rc_rack_rtt = t; 3732 } 3733 return (1); 3734 } 3735 /* 3736 * We clear the soft/rxtshift since we got an ack. 3737 * There is no assurance we will call the commit() function 3738 * so we need to clear these to avoid incorrect handling. 3739 */ 3740 tp->t_rxtshift = 0; 3741 tp->t_softerror = 0; 3742 if ((to->to_flags & TOF_TS) && 3743 (ack_type == CUM_ACKED) && 3744 (to->to_tsecr) && 3745 ((rsm->r_flags & (RACK_DEFERRED | RACK_OVERMAX)) == 0)) { 3746 /* 3747 * Now which timestamp does it match? In this block the ACK 3748 * must be coming from a previous transmission. 3749 */ 3750 for (i = 0; i < rsm->r_rtr_cnt; i++) { 3751 if (rsm->r_tim_lastsent[i] == to->to_tsecr) { 3752 t = cts - rsm->r_tim_lastsent[i]; 3753 if ((int)t <= 0) 3754 t = 1; 3755 if ((i + 1) < rsm->r_rtr_cnt) { 3756 /* Likely */ 3757 rack_earlier_retran(tp, rsm, t, cts); 3758 } 3759 if (!tp->t_rttlow || tp->t_rttlow > t) 3760 tp->t_rttlow = t; 3761 if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) { 3762 rack->r_ctl.rc_rack_min_rtt = t; 3763 if (rack->r_ctl.rc_rack_min_rtt == 0) { 3764 rack->r_ctl.rc_rack_min_rtt = 1; 3765 } 3766 } 3767 /* 3768 * Note the following calls to 3769 * tcp_rack_xmit_timer() are being commented 3770 * out for now. They give us no more accuracy 3771 * and often lead to a wrong choice. We have 3772 * enough samples that have not been 3773 * retransmitted. I leave the commented out 3774 * code in here in case in the future we 3775 * decide to add it back (though I can't forsee 3776 * doing that). That way we will easily see 3777 * where they need to be placed. 3778 */ 3779 if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, 3780 rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) { 3781 /* New more recent rack_tmit_time */ 3782 rack->r_ctl.rc_rack_tmit_time = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]; 3783 rack->rc_rack_rtt = t; 3784 } 3785 return (1); 3786 } 3787 } 3788 goto ts_not_found; 3789 } else { 3790 /* 3791 * Ok its a SACK block that we retransmitted. or a windows 3792 * machine without timestamps. We can tell nothing from the 3793 * time-stamp since its not there or the time the peer last 3794 * recieved a segment that moved forward its cum-ack point. 3795 */ 3796 ts_not_found: 3797 i = rsm->r_rtr_cnt - 1; 3798 t = cts - rsm->r_tim_lastsent[i]; 3799 if ((int)t <= 0) 3800 t = 1; 3801 if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) { 3802 /* 3803 * We retransmitted and the ack came back in less 3804 * than the smallest rtt we have observed. We most 3805 * likey did an improper retransmit as outlined in 3806 * 4.2 Step 3 point 2 in the rack-draft. 3807 */ 3808 i = rsm->r_rtr_cnt - 2; 3809 t = cts - rsm->r_tim_lastsent[i]; 3810 rack_earlier_retran(tp, rsm, t, cts); 3811 } else if (rack->r_ctl.rc_rack_min_rtt) { 3812 /* 3813 * We retransmitted it and the retransmit did the 3814 * job. 3815 */ 3816 if (!rack->r_ctl.rc_rack_min_rtt || 3817 SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) { 3818 rack->r_ctl.rc_rack_min_rtt = t; 3819 if (rack->r_ctl.rc_rack_min_rtt == 0) { 3820 rack->r_ctl.rc_rack_min_rtt = 1; 3821 } 3822 } 3823 if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, rsm->r_tim_lastsent[i])) { 3824 /* New more recent rack_tmit_time */ 3825 rack->r_ctl.rc_rack_tmit_time = rsm->r_tim_lastsent[i]; 3826 rack->rc_rack_rtt = t; 3827 } 3828 return (1); 3829 } 3830 } 3831 return (0); 3832 } 3833 3834 /* 3835 * Mark the SACK_PASSED flag on all entries prior to rsm send wise. 3836 */ 3837 static void 3838 rack_log_sack_passed(struct tcpcb *tp, 3839 struct tcp_rack *rack, struct rack_sendmap *rsm) 3840 { 3841 struct rack_sendmap *nrsm; 3842 uint32_t ts; 3843 int32_t idx; 3844 3845 idx = rsm->r_rtr_cnt - 1; 3846 ts = rsm->r_tim_lastsent[idx]; 3847 nrsm = rsm; 3848 TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap, 3849 rack_head, r_tnext) { 3850 if (nrsm == rsm) { 3851 /* Skip orginal segment he is acked */ 3852 continue; 3853 } 3854 if (nrsm->r_flags & RACK_ACKED) { 3855 /* Skip ack'd segments */ 3856 continue; 3857 } 3858 idx = nrsm->r_rtr_cnt - 1; 3859 if (ts == nrsm->r_tim_lastsent[idx]) { 3860 /* 3861 * For this case lets use seq no, if we sent in a 3862 * big block (TSO) we would have a bunch of segments 3863 * sent at the same time. 3864 * 3865 * We would only get a report if its SEQ is earlier. 3866 * If we have done multiple retransmits the times 3867 * would not be equal. 3868 */ 3869 if (SEQ_LT(nrsm->r_start, rsm->r_start)) { 3870 nrsm->r_flags |= RACK_SACK_PASSED; 3871 nrsm->r_flags &= ~RACK_WAS_SACKPASS; 3872 } 3873 } else { 3874 /* 3875 * Here they were sent at different times, not a big 3876 * block. Since we transmitted this one later and 3877 * see it sack'd then this must also be missing (or 3878 * we would have gotten a sack block for it) 3879 */ 3880 nrsm->r_flags |= RACK_SACK_PASSED; 3881 nrsm->r_flags &= ~RACK_WAS_SACKPASS; 3882 } 3883 } 3884 } 3885 3886 static uint32_t 3887 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack, 3888 struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts) 3889 { 3890 int32_t idx; 3891 int32_t times = 0; 3892 uint32_t start, end, changed = 0; 3893 struct rack_sendmap *rsm, *nrsm; 3894 int32_t used_ref = 1; 3895 3896 start = sack->start; 3897 end = sack->end; 3898 rsm = *prsm; 3899 if (rsm && SEQ_LT(start, rsm->r_start)) { 3900 TAILQ_FOREACH_REVERSE_FROM(rsm, &rack->r_ctl.rc_map, rack_head, r_next) { 3901 if (SEQ_GEQ(start, rsm->r_start) && 3902 SEQ_LT(start, rsm->r_end)) { 3903 goto do_rest_ofb; 3904 } 3905 } 3906 } 3907 if (rsm == NULL) { 3908 start_at_beginning: 3909 rsm = NULL; 3910 used_ref = 0; 3911 } 3912 /* First lets locate the block where this guy is */ 3913 TAILQ_FOREACH_FROM(rsm, &rack->r_ctl.rc_map, r_next) { 3914 if (SEQ_GEQ(start, rsm->r_start) && 3915 SEQ_LT(start, rsm->r_end)) { 3916 break; 3917 } 3918 } 3919 do_rest_ofb: 3920 if (rsm == NULL) { 3921 /* 3922 * This happens when we get duplicate sack blocks with the 3923 * same end. For example SACK 4: 100 SACK 3: 100 The sort 3924 * will not change there location so we would just start at 3925 * the end of the first one and get lost. 3926 */ 3927 if (tp->t_flags & TF_SENTFIN) { 3928 /* 3929 * Check to see if we have not logged the FIN that 3930 * went out. 3931 */ 3932 nrsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_map, rack_sendmap, r_next); 3933 if (nrsm && (nrsm->r_end + 1) == tp->snd_max) { 3934 /* 3935 * Ok we did not get the FIN logged. 3936 */ 3937 nrsm->r_end++; 3938 rsm = nrsm; 3939 goto do_rest_ofb; 3940 } 3941 } 3942 if (times == 1) { 3943 #ifdef INVARIANTS 3944 panic("tp:%p rack:%p sack:%p to:%p prsm:%p", 3945 tp, rack, sack, to, prsm); 3946 #else 3947 goto out; 3948 #endif 3949 } 3950 times++; 3951 counter_u64_add(rack_sack_proc_restart, 1); 3952 goto start_at_beginning; 3953 } 3954 /* Ok we have an ACK for some piece of rsm */ 3955 if (rsm->r_start != start) { 3956 /* 3957 * Need to split this in two pieces the before and after. 3958 */ 3959 nrsm = rack_alloc(rack); 3960 if (nrsm == NULL) { 3961 /* 3962 * failed XXXrrs what can we do but loose the sack 3963 * info? 3964 */ 3965 goto out; 3966 } 3967 nrsm->r_start = start; 3968 nrsm->r_rtr_bytes = 0; 3969 nrsm->r_end = rsm->r_end; 3970 nrsm->r_rtr_cnt = rsm->r_rtr_cnt; 3971 nrsm->r_flags = rsm->r_flags; 3972 nrsm->r_sndcnt = rsm->r_sndcnt; 3973 for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { 3974 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; 3975 } 3976 rsm->r_end = nrsm->r_start; 3977 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_map, rsm, nrsm, r_next); 3978 if (rsm->r_in_tmap) { 3979 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 3980 nrsm->r_in_tmap = 1; 3981 } 3982 rsm->r_flags &= (~RACK_HAS_FIN); 3983 rsm = nrsm; 3984 } 3985 if (SEQ_GEQ(end, rsm->r_end)) { 3986 /* 3987 * The end of this block is either beyond this guy or right 3988 * at this guy. 3989 */ 3990 3991 if ((rsm->r_flags & RACK_ACKED) == 0) { 3992 rack_update_rtt(tp, rack, rsm, to, cts, SACKED); 3993 changed += (rsm->r_end - rsm->r_start); 3994 rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start); 3995 rack_log_sack_passed(tp, rack, rsm); 3996 /* Is Reordering occuring? */ 3997 if (rsm->r_flags & RACK_SACK_PASSED) { 3998 counter_u64_add(rack_reorder_seen, 1); 3999 rack->r_ctl.rc_reorder_ts = cts; 4000 } 4001 rsm->r_flags |= RACK_ACKED; 4002 rsm->r_flags &= ~RACK_TLP; 4003 if (rsm->r_in_tmap) { 4004 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext); 4005 rsm->r_in_tmap = 0; 4006 } 4007 } 4008 if (end == rsm->r_end) { 4009 /* This block only - done */ 4010 goto out; 4011 } 4012 /* There is more not coverend by this rsm move on */ 4013 start = rsm->r_end; 4014 nrsm = TAILQ_NEXT(rsm, r_next); 4015 rsm = nrsm; 4016 times = 0; 4017 goto do_rest_ofb; 4018 } 4019 /* Ok we need to split off this one at the tail */ 4020 nrsm = rack_alloc(rack); 4021 if (nrsm == NULL) { 4022 /* failed rrs what can we do but loose the sack info? */ 4023 goto out; 4024 } 4025 /* Clone it */ 4026 nrsm->r_start = end; 4027 nrsm->r_end = rsm->r_end; 4028 nrsm->r_rtr_bytes = 0; 4029 nrsm->r_rtr_cnt = rsm->r_rtr_cnt; 4030 nrsm->r_flags = rsm->r_flags; 4031 nrsm->r_sndcnt = rsm->r_sndcnt; 4032 for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { 4033 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; 4034 } 4035 /* The sack block does not cover this guy fully */ 4036 rsm->r_flags &= (~RACK_HAS_FIN); 4037 rsm->r_end = end; 4038 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_map, rsm, nrsm, r_next); 4039 if (rsm->r_in_tmap) { 4040 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 4041 nrsm->r_in_tmap = 1; 4042 } 4043 if (rsm->r_flags & RACK_ACKED) { 4044 /* Been here done that */ 4045 goto out; 4046 } 4047 rack_update_rtt(tp, rack, rsm, to, cts, SACKED); 4048 changed += (rsm->r_end - rsm->r_start); 4049 rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start); 4050 rack_log_sack_passed(tp, rack, rsm); 4051 /* Is Reordering occuring? */ 4052 if (rsm->r_flags & RACK_SACK_PASSED) { 4053 counter_u64_add(rack_reorder_seen, 1); 4054 rack->r_ctl.rc_reorder_ts = cts; 4055 } 4056 rsm->r_flags |= RACK_ACKED; 4057 rsm->r_flags &= ~RACK_TLP; 4058 if (rsm->r_in_tmap) { 4059 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext); 4060 rsm->r_in_tmap = 0; 4061 } 4062 out: 4063 if (used_ref == 0) { 4064 counter_u64_add(rack_sack_proc_all, 1); 4065 } else { 4066 counter_u64_add(rack_sack_proc_short, 1); 4067 } 4068 /* Save off where we last were */ 4069 if (rsm) 4070 rack->r_ctl.rc_sacklast = TAILQ_NEXT(rsm, r_next); 4071 else 4072 rack->r_ctl.rc_sacklast = NULL; 4073 *prsm = rsm; 4074 return (changed); 4075 } 4076 4077 static void inline 4078 rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack) 4079 { 4080 struct rack_sendmap *tmap; 4081 4082 tmap = NULL; 4083 while (rsm && (rsm->r_flags & RACK_ACKED)) { 4084 /* Its no longer sacked, mark it so */ 4085 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); 4086 #ifdef INVARIANTS 4087 if (rsm->r_in_tmap) { 4088 panic("rack:%p rsm:%p flags:0x%x in tmap?", 4089 rack, rsm, rsm->r_flags); 4090 } 4091 #endif 4092 rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS); 4093 /* Rebuild it into our tmap */ 4094 if (tmap == NULL) { 4095 TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext); 4096 tmap = rsm; 4097 } else { 4098 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext); 4099 tmap = rsm; 4100 } 4101 tmap->r_in_tmap = 1; 4102 rsm = TAILQ_NEXT(rsm, r_next); 4103 } 4104 /* 4105 * Now lets possibly clear the sack filter so we start 4106 * recognizing sacks that cover this area. 4107 */ 4108 if (rack_use_sack_filter) 4109 sack_filter_clear(&rack->r_ctl.rack_sf, th_ack); 4110 4111 } 4112 4113 static void 4114 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th) 4115 { 4116 uint32_t changed, last_seq, entered_recovery = 0; 4117 struct tcp_rack *rack; 4118 struct rack_sendmap *rsm; 4119 struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1]; 4120 register uint32_t th_ack; 4121 int32_t i, j, k, num_sack_blks = 0; 4122 uint32_t cts, acked, ack_point, sack_changed = 0; 4123 4124 INP_WLOCK_ASSERT(tp->t_inpcb); 4125 if (th->th_flags & TH_RST) { 4126 /* We don't log resets */ 4127 return; 4128 } 4129 rack = (struct tcp_rack *)tp->t_fb_ptr; 4130 cts = tcp_ts_getticks(); 4131 rsm = TAILQ_FIRST(&rack->r_ctl.rc_map); 4132 changed = 0; 4133 th_ack = th->th_ack; 4134 4135 if (SEQ_GT(th_ack, tp->snd_una)) { 4136 rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__); 4137 tp->t_acktime = ticks; 4138 } 4139 if (rsm && SEQ_GT(th_ack, rsm->r_start)) 4140 changed = th_ack - rsm->r_start; 4141 if (changed) { 4142 /* 4143 * The ACK point is advancing to th_ack, we must drop off 4144 * the packets in the rack log and calculate any eligble 4145 * RTT's. 4146 */ 4147 rack->r_wanted_output++; 4148 more: 4149 rsm = TAILQ_FIRST(&rack->r_ctl.rc_map); 4150 if (rsm == NULL) { 4151 if ((th_ack - 1) == tp->iss) { 4152 /* 4153 * For the SYN incoming case we will not 4154 * have called tcp_output for the sending of 4155 * the SYN, so there will be no map. All 4156 * other cases should probably be a panic. 4157 */ 4158 goto proc_sack; 4159 } 4160 if (tp->t_flags & TF_SENTFIN) { 4161 /* if we send a FIN we will not hav a map */ 4162 goto proc_sack; 4163 } 4164 #ifdef INVARIANTS 4165 panic("No rack map tp:%p for th:%p state:%d rack:%p snd_una:%u snd_max:%u snd_nxt:%u chg:%d\n", 4166 tp, 4167 th, tp->t_state, rack, 4168 tp->snd_una, tp->snd_max, tp->snd_nxt, changed); 4169 #endif 4170 goto proc_sack; 4171 } 4172 if (SEQ_LT(th_ack, rsm->r_start)) { 4173 /* Huh map is missing this */ 4174 #ifdef INVARIANTS 4175 printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n", 4176 rsm->r_start, 4177 th_ack, tp->t_state, rack->r_state); 4178 #endif 4179 goto proc_sack; 4180 } 4181 rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED); 4182 /* Now do we consume the whole thing? */ 4183 if (SEQ_GEQ(th_ack, rsm->r_end)) { 4184 /* Its all consumed. */ 4185 uint32_t left; 4186 4187 rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; 4188 rsm->r_rtr_bytes = 0; 4189 TAILQ_REMOVE(&rack->r_ctl.rc_map, rsm, r_next); 4190 if (rsm->r_in_tmap) { 4191 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext); 4192 rsm->r_in_tmap = 0; 4193 } 4194 if (rack->r_ctl.rc_next == rsm) { 4195 /* scoot along the marker */ 4196 rack->r_ctl.rc_next = TAILQ_FIRST(&rack->r_ctl.rc_map); 4197 } 4198 if (rsm->r_flags & RACK_ACKED) { 4199 /* 4200 * It was acked on the scoreboard -- remove 4201 * it from total 4202 */ 4203 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); 4204 } else if (rsm->r_flags & RACK_SACK_PASSED) { 4205 /* 4206 * There are acked segments ACKED on the 4207 * scoreboard further up. We are seeing 4208 * reordering. 4209 */ 4210 counter_u64_add(rack_reorder_seen, 1); 4211 rsm->r_flags |= RACK_ACKED; 4212 rack->r_ctl.rc_reorder_ts = cts; 4213 } 4214 left = th_ack - rsm->r_end; 4215 if (rsm->r_rtr_cnt > 1) { 4216 /* 4217 * Technically we should make r_rtr_cnt be 4218 * monotonicly increasing and just mod it to 4219 * the timestamp it is replacing.. that way 4220 * we would have the last 3 retransmits. Now 4221 * rc_loss_count will be wrong if we 4222 * retransmit something more than 2 times in 4223 * recovery :( 4224 */ 4225 rack->r_ctl.rc_loss_count += (rsm->r_rtr_cnt - 1); 4226 } 4227 /* Free back to zone */ 4228 rack_free(rack, rsm); 4229 if (left) { 4230 goto more; 4231 } 4232 goto proc_sack; 4233 } 4234 if (rsm->r_flags & RACK_ACKED) { 4235 /* 4236 * It was acked on the scoreboard -- remove it from 4237 * total for the part being cum-acked. 4238 */ 4239 rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start); 4240 } 4241 rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; 4242 rsm->r_rtr_bytes = 0; 4243 rsm->r_start = th_ack; 4244 } 4245 proc_sack: 4246 /* Check for reneging */ 4247 rsm = TAILQ_FIRST(&rack->r_ctl.rc_map); 4248 if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) { 4249 /* 4250 * The peer has moved snd_una up to 4251 * the edge of this send, i.e. one 4252 * that it had previously acked. The only 4253 * way that can be true if the peer threw 4254 * away data (space issues) that it had 4255 * previously sacked (else it would have 4256 * given us snd_una up to (rsm->r_end). 4257 * We need to undo the acked markings here. 4258 * 4259 * Note we have to look to make sure th_ack is 4260 * our rsm->r_start in case we get an old ack 4261 * where th_ack is behind snd_una. 4262 */ 4263 rack_peer_reneges(rack, rsm, th->th_ack); 4264 } 4265 if ((to->to_flags & TOF_SACK) == 0) { 4266 /* We are done nothing left to log */ 4267 goto out; 4268 } 4269 rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_map, rack_sendmap, r_next); 4270 if (rsm) { 4271 last_seq = rsm->r_end; 4272 } else { 4273 last_seq = tp->snd_max; 4274 } 4275 /* Sack block processing */ 4276 if (SEQ_GT(th_ack, tp->snd_una)) 4277 ack_point = th_ack; 4278 else 4279 ack_point = tp->snd_una; 4280 for (i = 0; i < to->to_nsacks; i++) { 4281 bcopy((to->to_sacks + i * TCPOLEN_SACK), 4282 &sack, sizeof(sack)); 4283 sack.start = ntohl(sack.start); 4284 sack.end = ntohl(sack.end); 4285 if (SEQ_GT(sack.end, sack.start) && 4286 SEQ_GT(sack.start, ack_point) && 4287 SEQ_LT(sack.start, tp->snd_max) && 4288 SEQ_GT(sack.end, ack_point) && 4289 SEQ_LEQ(sack.end, tp->snd_max)) { 4290 if ((rack->r_ctl.rc_num_maps_alloced > rack_sack_block_limit) && 4291 (SEQ_LT(sack.end, last_seq)) && 4292 ((sack.end - sack.start) < (tp->t_maxseg / 8))) { 4293 /* 4294 * Not the last piece and its smaller than 4295 * 1/8th of a MSS. We ignore this. 4296 */ 4297 counter_u64_add(rack_runt_sacks, 1); 4298 continue; 4299 } 4300 sack_blocks[num_sack_blks] = sack; 4301 num_sack_blks++; 4302 #ifdef NETFLIX_STATS 4303 } else if (SEQ_LEQ(sack.start, th_ack) && 4304 SEQ_LEQ(sack.end, th_ack)) { 4305 /* 4306 * Its a D-SACK block. 4307 */ 4308 tcp_record_dsack(sack.start, sack.end); 4309 #endif 4310 } 4311 4312 } 4313 if (num_sack_blks == 0) 4314 goto out; 4315 /* 4316 * Sort the SACK blocks so we can update the rack scoreboard with 4317 * just one pass. 4318 */ 4319 if (rack_use_sack_filter) { 4320 num_sack_blks = sack_filter_blks(&rack->r_ctl.rack_sf, sack_blocks, num_sack_blks, th->th_ack); 4321 } 4322 if (num_sack_blks < 2) { 4323 goto do_sack_work; 4324 } 4325 /* Sort the sacks */ 4326 for (i = 0; i < num_sack_blks; i++) { 4327 for (j = i + 1; j < num_sack_blks; j++) { 4328 if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) { 4329 sack = sack_blocks[i]; 4330 sack_blocks[i] = sack_blocks[j]; 4331 sack_blocks[j] = sack; 4332 } 4333 } 4334 } 4335 /* 4336 * Now are any of the sack block ends the same (yes some 4337 * implememtations send these)? 4338 */ 4339 again: 4340 if (num_sack_blks > 1) { 4341 for (i = 0; i < num_sack_blks; i++) { 4342 for (j = i + 1; j < num_sack_blks; j++) { 4343 if (sack_blocks[i].end == sack_blocks[j].end) { 4344 /* 4345 * Ok these two have the same end we 4346 * want the smallest end and then 4347 * throw away the larger and start 4348 * again. 4349 */ 4350 if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) { 4351 /* 4352 * The second block covers 4353 * more area use that 4354 */ 4355 sack_blocks[i].start = sack_blocks[j].start; 4356 } 4357 /* 4358 * Now collapse out the dup-sack and 4359 * lower the count 4360 */ 4361 for (k = (j + 1); k < num_sack_blks; k++) { 4362 sack_blocks[j].start = sack_blocks[k].start; 4363 sack_blocks[j].end = sack_blocks[k].end; 4364 j++; 4365 } 4366 num_sack_blks--; 4367 goto again; 4368 } 4369 } 4370 } 4371 } 4372 do_sack_work: 4373 rsm = rack->r_ctl.rc_sacklast; 4374 for (i = 0; i < num_sack_blks; i++) { 4375 acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts); 4376 if (acked) { 4377 rack->r_wanted_output++; 4378 changed += acked; 4379 sack_changed += acked; 4380 } 4381 } 4382 out: 4383 if (changed) { 4384 /* Something changed cancel the rack timer */ 4385 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 4386 } 4387 if ((sack_changed) && (!IN_RECOVERY(tp->t_flags))) { 4388 /* 4389 * Ok we have a high probability that we need to go in to 4390 * recovery since we have data sack'd 4391 */ 4392 struct rack_sendmap *rsm; 4393 uint32_t tsused; 4394 4395 tsused = tcp_ts_getticks(); 4396 rsm = tcp_rack_output(tp, rack, tsused); 4397 if (rsm) { 4398 /* Enter recovery */ 4399 rack->r_ctl.rc_rsm_start = rsm->r_start; 4400 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd; 4401 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh; 4402 entered_recovery = 1; 4403 rack_cong_signal(tp, NULL, CC_NDUPACK); 4404 /* 4405 * When we enter recovery we need to assure we send 4406 * one packet. 4407 */ 4408 rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; 4409 rack->r_timer_override = 1; 4410 } 4411 } 4412 if (IN_RECOVERY(tp->t_flags) && (entered_recovery == 0)) { 4413 /* Deal with changed an PRR here (in recovery only) */ 4414 uint32_t pipe, snd_una; 4415 4416 rack->r_ctl.rc_prr_delivered += changed; 4417 /* Compute prr_sndcnt */ 4418 if (SEQ_GT(tp->snd_una, th_ack)) { 4419 snd_una = tp->snd_una; 4420 } else { 4421 snd_una = th_ack; 4422 } 4423 pipe = ((tp->snd_max - snd_una) - rack->r_ctl.rc_sacked) + rack->r_ctl.rc_holes_rxt; 4424 if (pipe > tp->snd_ssthresh) { 4425 long sndcnt; 4426 4427 sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh; 4428 if (rack->r_ctl.rc_prr_recovery_fs > 0) 4429 sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs; 4430 else { 4431 rack->r_ctl.rc_prr_sndcnt = 0; 4432 sndcnt = 0; 4433 } 4434 sndcnt++; 4435 if (sndcnt > (long)rack->r_ctl.rc_prr_out) 4436 sndcnt -= rack->r_ctl.rc_prr_out; 4437 else 4438 sndcnt = 0; 4439 rack->r_ctl.rc_prr_sndcnt = sndcnt; 4440 } else { 4441 uint32_t limit; 4442 4443 if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out) 4444 limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out); 4445 else 4446 limit = 0; 4447 if (changed > limit) 4448 limit = changed; 4449 limit += tp->t_maxseg; 4450 if (tp->snd_ssthresh > pipe) { 4451 rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit); 4452 } else { 4453 rack->r_ctl.rc_prr_sndcnt = min(0, limit); 4454 } 4455 } 4456 if (rack->r_ctl.rc_prr_sndcnt >= tp->t_maxseg) { 4457 rack->r_timer_override = 1; 4458 } 4459 } 4460 } 4461 4462 /* 4463 * Return value of 1, we do not need to call rack_process_data(). 4464 * return value of 0, rack_process_data can be called. 4465 * For ret_val if its 0 the TCP is locked, if its non-zero 4466 * its unlocked and probably unsafe to touch the TCB. 4467 */ 4468 static int 4469 rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so, 4470 struct tcpcb *tp, struct tcpopt *to, 4471 uint32_t tiwin, int32_t tlen, 4472 int32_t * ofia, int32_t thflags, int32_t * ret_val) 4473 { 4474 int32_t ourfinisacked = 0; 4475 int32_t nsegs, acked_amount; 4476 int32_t acked; 4477 struct mbuf *mfree; 4478 struct tcp_rack *rack; 4479 int32_t recovery = 0; 4480 4481 rack = (struct tcp_rack *)tp->t_fb_ptr; 4482 if (SEQ_GT(th->th_ack, tp->snd_max)) { 4483 rack_do_dropafterack(m, tp, th, thflags, tlen, ret_val); 4484 return (1); 4485 } 4486 if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) { 4487 rack_log_ack(tp, to, th); 4488 } 4489 if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) { 4490 /* 4491 * Old ack, behind (or duplicate to) the last one rcv'd 4492 * Note: Should mark reordering is occuring! We should also 4493 * look for sack blocks arriving e.g. ack 1, 4-4 then ack 1, 4494 * 3-3, 4-4 would be reording. As well as ack 1, 3-3 <no 4495 * retran and> ack 3 4496 */ 4497 return (0); 4498 } 4499 /* 4500 * If we reach this point, ACK is not a duplicate, i.e., it ACKs 4501 * something we sent. 4502 */ 4503 if (tp->t_flags & TF_NEEDSYN) { 4504 /* 4505 * T/TCP: Connection was half-synchronized, and our SYN has 4506 * been ACK'd (so connection is now fully synchronized). Go 4507 * to non-starred state, increment snd_una for ACK of SYN, 4508 * and check if we can do window scaling. 4509 */ 4510 tp->t_flags &= ~TF_NEEDSYN; 4511 tp->snd_una++; 4512 /* Do window scaling? */ 4513 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == 4514 (TF_RCVD_SCALE | TF_REQ_SCALE)) { 4515 tp->rcv_scale = tp->request_r_scale; 4516 /* Send window already scaled. */ 4517 } 4518 } 4519 nsegs = max(1, m->m_pkthdr.lro_nsegs); 4520 INP_WLOCK_ASSERT(tp->t_inpcb); 4521 4522 acked = BYTES_THIS_ACK(tp, th); 4523 TCPSTAT_ADD(tcps_rcvackpack, nsegs); 4524 TCPSTAT_ADD(tcps_rcvackbyte, acked); 4525 4526 /* 4527 * If we just performed our first retransmit, and the ACK arrives 4528 * within our recovery window, then it was a mistake to do the 4529 * retransmit in the first place. Recover our original cwnd and 4530 * ssthresh, and proceed to transmit where we left off. 4531 */ 4532 if (tp->t_flags & TF_PREVVALID) { 4533 tp->t_flags &= ~TF_PREVVALID; 4534 if (tp->t_rxtshift == 1 && 4535 (int)(ticks - tp->t_badrxtwin) < 0) 4536 rack_cong_signal(tp, th, CC_RTO_ERR); 4537 } 4538 /* 4539 * If we have a timestamp reply, update smoothed round trip time. If 4540 * no timestamp is present but transmit timer is running and timed 4541 * sequence number was acked, update smoothed round trip time. Since 4542 * we now have an rtt measurement, cancel the timer backoff (cf., 4543 * Phil Karn's retransmit alg.). Recompute the initial retransmit 4544 * timer. 4545 * 4546 * Some boxes send broken timestamp replies during the SYN+ACK 4547 * phase, ignore timestamps of 0 or we could calculate a huge RTT 4548 * and blow up the retransmit timer. 4549 */ 4550 /* 4551 * If all outstanding data is acked, stop retransmit timer and 4552 * remember to restart (more output or persist). If there is more 4553 * data to be acked, restart retransmit timer, using current 4554 * (possibly backed-off) value. 4555 */ 4556 if (th->th_ack == tp->snd_max) { 4557 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 4558 rack->r_wanted_output++; 4559 } 4560 /* 4561 * If no data (only SYN) was ACK'd, skip rest of ACK processing. 4562 */ 4563 if (acked == 0) { 4564 if (ofia) 4565 *ofia = ourfinisacked; 4566 return (0); 4567 } 4568 if (rack->r_ctl.rc_early_recovery) { 4569 if (IN_FASTRECOVERY(tp->t_flags)) { 4570 if (SEQ_LT(th->th_ack, tp->snd_recover)) { 4571 tcp_rack_partialack(tp, th); 4572 } else { 4573 rack_post_recovery(tp, th); 4574 recovery = 1; 4575 } 4576 } 4577 } 4578 /* 4579 * Let the congestion control algorithm update congestion control 4580 * related information. This typically means increasing the 4581 * congestion window. 4582 */ 4583 rack_ack_received(tp, rack, th, nsegs, CC_ACK, recovery); 4584 SOCKBUF_LOCK(&so->so_snd); 4585 acked_amount = min(acked, (int)sbavail(&so->so_snd)); 4586 tp->snd_wnd -= acked_amount; 4587 mfree = sbcut_locked(&so->so_snd, acked_amount); 4588 if ((sbused(&so->so_snd) == 0) && 4589 (acked > acked_amount) && 4590 (tp->t_state >= TCPS_FIN_WAIT_1)) { 4591 ourfinisacked = 1; 4592 } 4593 /* NB: sowwakeup_locked() does an implicit unlock. */ 4594 sowwakeup_locked(so); 4595 m_freem(mfree); 4596 if (rack->r_ctl.rc_early_recovery == 0) { 4597 if (IN_FASTRECOVERY(tp->t_flags)) { 4598 if (SEQ_LT(th->th_ack, tp->snd_recover)) { 4599 tcp_rack_partialack(tp, th); 4600 } else { 4601 rack_post_recovery(tp, th); 4602 } 4603 } 4604 } 4605 tp->snd_una = th->th_ack; 4606 if (SEQ_GT(tp->snd_una, tp->snd_recover)) 4607 tp->snd_recover = tp->snd_una; 4608 4609 if (SEQ_LT(tp->snd_nxt, tp->snd_una)) { 4610 tp->snd_nxt = tp->snd_una; 4611 } 4612 if (tp->snd_una == tp->snd_max) { 4613 /* Nothing left outstanding */ 4614 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__); 4615 tp->t_acktime = 0; 4616 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 4617 /* Set need output so persist might get set */ 4618 rack->r_wanted_output++; 4619 if (rack_use_sack_filter) 4620 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una); 4621 if ((tp->t_state >= TCPS_FIN_WAIT_1) && 4622 (sbavail(&so->so_snd) == 0) && 4623 (tp->t_flags2 & TF2_DROP_AF_DATA)) { 4624 /* 4625 * The socket was gone and the 4626 * peer sent data, time to 4627 * reset him. 4628 */ 4629 *ret_val = 1; 4630 tp = tcp_close(tp); 4631 rack_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen); 4632 return (1); 4633 } 4634 } 4635 if (ofia) 4636 *ofia = ourfinisacked; 4637 return (0); 4638 } 4639 4640 4641 /* 4642 * Return value of 1, the TCB is unlocked and most 4643 * likely gone, return value of 0, the TCP is still 4644 * locked. 4645 */ 4646 static int 4647 rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so, 4648 struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, 4649 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 4650 { 4651 /* 4652 * Update window information. Don't look at window if no ACK: TAC's 4653 * send garbage on first SYN. 4654 */ 4655 int32_t nsegs; 4656 int32_t tfo_syn; 4657 struct tcp_rack *rack; 4658 4659 rack = (struct tcp_rack *)tp->t_fb_ptr; 4660 INP_WLOCK_ASSERT(tp->t_inpcb); 4661 4662 nsegs = max(1, m->m_pkthdr.lro_nsegs); 4663 if ((thflags & TH_ACK) && 4664 (SEQ_LT(tp->snd_wl1, th->th_seq) || 4665 (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) || 4666 (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) { 4667 /* keep track of pure window updates */ 4668 if (tlen == 0 && 4669 tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd) 4670 TCPSTAT_INC(tcps_rcvwinupd); 4671 tp->snd_wnd = tiwin; 4672 tp->snd_wl1 = th->th_seq; 4673 tp->snd_wl2 = th->th_ack; 4674 if (tp->snd_wnd > tp->max_sndwnd) 4675 tp->max_sndwnd = tp->snd_wnd; 4676 rack->r_wanted_output++; 4677 } else if (thflags & TH_ACK) { 4678 if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) { 4679 tp->snd_wnd = tiwin; 4680 tp->snd_wl1 = th->th_seq; 4681 tp->snd_wl2 = th->th_ack; 4682 } 4683 } 4684 /* Was persist timer active and now we have window space? */ 4685 if ((rack->rc_in_persist != 0) && tp->snd_wnd) { 4686 rack_exit_persist(tp, rack); 4687 tp->snd_nxt = tp->snd_max; 4688 /* Make sure we output to start the timer */ 4689 rack->r_wanted_output++; 4690 } 4691 /* 4692 * Process segments with URG. 4693 */ 4694 if ((thflags & TH_URG) && th->th_urp && 4695 TCPS_HAVERCVDFIN(tp->t_state) == 0) { 4696 /* 4697 * This is a kludge, but if we receive and accept random 4698 * urgent pointers, we'll crash in soreceive. It's hard to 4699 * imagine someone actually wanting to send this much urgent 4700 * data. 4701 */ 4702 SOCKBUF_LOCK(&so->so_rcv); 4703 if (th->th_urp + sbavail(&so->so_rcv) > sb_max) { 4704 th->th_urp = 0; /* XXX */ 4705 thflags &= ~TH_URG; /* XXX */ 4706 SOCKBUF_UNLOCK(&so->so_rcv); /* XXX */ 4707 goto dodata; /* XXX */ 4708 } 4709 /* 4710 * If this segment advances the known urgent pointer, then 4711 * mark the data stream. This should not happen in 4712 * CLOSE_WAIT, CLOSING, LAST_ACK or TIME_WAIT STATES since a 4713 * FIN has been received from the remote side. In these 4714 * states we ignore the URG. 4715 * 4716 * According to RFC961 (Assigned Protocols), the urgent 4717 * pointer points to the last octet of urgent data. We 4718 * continue, however, to consider it to indicate the first 4719 * octet of data past the urgent section as the original 4720 * spec states (in one of two places). 4721 */ 4722 if (SEQ_GT(th->th_seq + th->th_urp, tp->rcv_up)) { 4723 tp->rcv_up = th->th_seq + th->th_urp; 4724 so->so_oobmark = sbavail(&so->so_rcv) + 4725 (tp->rcv_up - tp->rcv_nxt) - 1; 4726 if (so->so_oobmark == 0) 4727 so->so_rcv.sb_state |= SBS_RCVATMARK; 4728 sohasoutofband(so); 4729 tp->t_oobflags &= ~(TCPOOB_HAVEDATA | TCPOOB_HADDATA); 4730 } 4731 SOCKBUF_UNLOCK(&so->so_rcv); 4732 /* 4733 * Remove out of band data so doesn't get presented to user. 4734 * This can happen independent of advancing the URG pointer, 4735 * but if two URG's are pending at once, some out-of-band 4736 * data may creep in... ick. 4737 */ 4738 if (th->th_urp <= (uint32_t) tlen && 4739 !(so->so_options & SO_OOBINLINE)) { 4740 /* hdr drop is delayed */ 4741 tcp_pulloutofband(so, th, m, drop_hdrlen); 4742 } 4743 } else { 4744 /* 4745 * If no out of band data is expected, pull receive urgent 4746 * pointer along with the receive window. 4747 */ 4748 if (SEQ_GT(tp->rcv_nxt, tp->rcv_up)) 4749 tp->rcv_up = tp->rcv_nxt; 4750 } 4751 dodata: /* XXX */ 4752 INP_WLOCK_ASSERT(tp->t_inpcb); 4753 4754 /* 4755 * Process the segment text, merging it into the TCP sequencing 4756 * queue, and arranging for acknowledgment of receipt if necessary. 4757 * This process logically involves adjusting tp->rcv_wnd as data is 4758 * presented to the user (this happens in tcp_usrreq.c, case 4759 * PRU_RCVD). If a FIN has already been received on this connection 4760 * then we just ignore the text. 4761 */ 4762 tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) && 4763 IS_FASTOPEN(tp->t_flags)); 4764 if ((tlen || (thflags & TH_FIN) || tfo_syn) && 4765 TCPS_HAVERCVDFIN(tp->t_state) == 0) { 4766 tcp_seq save_start = th->th_seq; 4767 4768 m_adj(m, drop_hdrlen); /* delayed header drop */ 4769 /* 4770 * Insert segment which includes th into TCP reassembly 4771 * queue with control block tp. Set thflags to whether 4772 * reassembly now includes a segment with FIN. This handles 4773 * the common case inline (segment is the next to be 4774 * received on an established connection, and the queue is 4775 * empty), avoiding linkage into and removal from the queue 4776 * and repetition of various conversions. Set DELACK for 4777 * segments received in order, but ack immediately when 4778 * segments are out of order (so fast retransmit can work). 4779 */ 4780 if (th->th_seq == tp->rcv_nxt && 4781 LIST_EMPTY(&tp->t_segq) && 4782 (TCPS_HAVEESTABLISHED(tp->t_state) || 4783 tfo_syn)) { 4784 if (DELAY_ACK(tp, tlen) || tfo_syn) { 4785 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 4786 tp->t_flags |= TF_DELACK; 4787 } else { 4788 rack->r_wanted_output++; 4789 tp->t_flags |= TF_ACKNOW; 4790 } 4791 tp->rcv_nxt += tlen; 4792 thflags = th->th_flags & TH_FIN; 4793 TCPSTAT_ADD(tcps_rcvpack, nsegs); 4794 TCPSTAT_ADD(tcps_rcvbyte, tlen); 4795 SOCKBUF_LOCK(&so->so_rcv); 4796 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) 4797 m_freem(m); 4798 else 4799 sbappendstream_locked(&so->so_rcv, m, 0); 4800 /* NB: sorwakeup_locked() does an implicit unlock. */ 4801 sorwakeup_locked(so); 4802 } else { 4803 /* 4804 * XXX: Due to the header drop above "th" is 4805 * theoretically invalid by now. Fortunately 4806 * m_adj() doesn't actually frees any mbufs when 4807 * trimming from the head. 4808 */ 4809 thflags = tcp_reass(tp, th, &tlen, m); 4810 tp->t_flags |= TF_ACKNOW; 4811 } 4812 if (tlen > 0) 4813 tcp_update_sack_list(tp, save_start, save_start + tlen); 4814 } else { 4815 m_freem(m); 4816 thflags &= ~TH_FIN; 4817 } 4818 4819 /* 4820 * If FIN is received ACK the FIN and let the user know that the 4821 * connection is closing. 4822 */ 4823 if (thflags & TH_FIN) { 4824 if (TCPS_HAVERCVDFIN(tp->t_state) == 0) { 4825 socantrcvmore(so); 4826 /* 4827 * If connection is half-synchronized (ie NEEDSYN 4828 * flag on) then delay ACK, so it may be piggybacked 4829 * when SYN is sent. Otherwise, since we received a 4830 * FIN then no more input can be expected, send ACK 4831 * now. 4832 */ 4833 if (tp->t_flags & TF_NEEDSYN) { 4834 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 4835 tp->t_flags |= TF_DELACK; 4836 } else { 4837 tp->t_flags |= TF_ACKNOW; 4838 } 4839 tp->rcv_nxt++; 4840 } 4841 switch (tp->t_state) { 4842 4843 /* 4844 * In SYN_RECEIVED and ESTABLISHED STATES enter the 4845 * CLOSE_WAIT state. 4846 */ 4847 case TCPS_SYN_RECEIVED: 4848 tp->t_starttime = ticks; 4849 /* FALLTHROUGH */ 4850 case TCPS_ESTABLISHED: 4851 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 4852 tcp_state_change(tp, TCPS_CLOSE_WAIT); 4853 break; 4854 4855 /* 4856 * If still in FIN_WAIT_1 STATE FIN has not been 4857 * acked so enter the CLOSING state. 4858 */ 4859 case TCPS_FIN_WAIT_1: 4860 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 4861 tcp_state_change(tp, TCPS_CLOSING); 4862 break; 4863 4864 /* 4865 * In FIN_WAIT_2 state enter the TIME_WAIT state, 4866 * starting the time-wait timer, turning off the 4867 * other standard timers. 4868 */ 4869 case TCPS_FIN_WAIT_2: 4870 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 4871 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 4872 tcp_twstart(tp); 4873 return (1); 4874 } 4875 } 4876 /* 4877 * Return any desired output. 4878 */ 4879 if ((tp->t_flags & TF_ACKNOW) || (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) { 4880 rack->r_wanted_output++; 4881 } 4882 INP_WLOCK_ASSERT(tp->t_inpcb); 4883 return (0); 4884 } 4885 4886 /* 4887 * Here nothing is really faster, its just that we 4888 * have broken out the fast-data path also just like 4889 * the fast-ack. 4890 */ 4891 static int 4892 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so, 4893 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 4894 uint32_t tiwin, int32_t nxt_pkt) 4895 { 4896 int32_t nsegs; 4897 int32_t newsize = 0; /* automatic sockbuf scaling */ 4898 struct tcp_rack *rack; 4899 #ifdef TCPDEBUG 4900 /* 4901 * The size of tcp_saveipgen must be the size of the max ip header, 4902 * now IPv6. 4903 */ 4904 u_char tcp_saveipgen[IP6_HDR_LEN]; 4905 struct tcphdr tcp_savetcp; 4906 short ostate = 0; 4907 4908 #endif 4909 /* 4910 * If last ACK falls within this segment's sequence numbers, record 4911 * the timestamp. NOTE that the test is modified according to the 4912 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26). 4913 */ 4914 if (__predict_false(th->th_seq != tp->rcv_nxt)) { 4915 return (0); 4916 } 4917 if (__predict_false(tp->snd_nxt != tp->snd_max)) { 4918 return (0); 4919 } 4920 if (tiwin && tiwin != tp->snd_wnd) { 4921 return (0); 4922 } 4923 if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) { 4924 return (0); 4925 } 4926 if (__predict_false((to->to_flags & TOF_TS) && 4927 (TSTMP_LT(to->to_tsval, tp->ts_recent)))) { 4928 return (0); 4929 } 4930 if (__predict_false((th->th_ack != tp->snd_una))) { 4931 return (0); 4932 } 4933 if (__predict_false(tlen > sbspace(&so->so_rcv))) { 4934 return (0); 4935 } 4936 if ((to->to_flags & TOF_TS) != 0 && 4937 SEQ_LEQ(th->th_seq, tp->last_ack_sent)) { 4938 tp->ts_recent_age = tcp_ts_getticks(); 4939 tp->ts_recent = to->to_tsval; 4940 } 4941 rack = (struct tcp_rack *)tp->t_fb_ptr; 4942 /* 4943 * This is a pure, in-sequence data packet with nothing on the 4944 * reassembly queue and we have enough buffer space to take it. 4945 */ 4946 nsegs = max(1, m->m_pkthdr.lro_nsegs); 4947 4948 4949 /* Clean receiver SACK report if present */ 4950 if (tp->rcv_numsacks) 4951 tcp_clean_sackreport(tp); 4952 TCPSTAT_INC(tcps_preddat); 4953 tp->rcv_nxt += tlen; 4954 /* 4955 * Pull snd_wl1 up to prevent seq wrap relative to th_seq. 4956 */ 4957 tp->snd_wl1 = th->th_seq; 4958 /* 4959 * Pull rcv_up up to prevent seq wrap relative to rcv_nxt. 4960 */ 4961 tp->rcv_up = tp->rcv_nxt; 4962 TCPSTAT_ADD(tcps_rcvpack, nsegs); 4963 TCPSTAT_ADD(tcps_rcvbyte, tlen); 4964 #ifdef TCPDEBUG 4965 if (so->so_options & SO_DEBUG) 4966 tcp_trace(TA_INPUT, ostate, tp, 4967 (void *)tcp_saveipgen, &tcp_savetcp, 0); 4968 #endif 4969 newsize = tcp_autorcvbuf(m, th, so, tp, tlen); 4970 4971 /* Add data to socket buffer. */ 4972 SOCKBUF_LOCK(&so->so_rcv); 4973 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 4974 m_freem(m); 4975 } else { 4976 /* 4977 * Set new socket buffer size. Give up when limit is 4978 * reached. 4979 */ 4980 if (newsize) 4981 if (!sbreserve_locked(&so->so_rcv, 4982 newsize, so, NULL)) 4983 so->so_rcv.sb_flags &= ~SB_AUTOSIZE; 4984 m_adj(m, drop_hdrlen); /* delayed header drop */ 4985 sbappendstream_locked(&so->so_rcv, m, 0); 4986 rack_calc_rwin(so, tp); 4987 } 4988 /* NB: sorwakeup_locked() does an implicit unlock. */ 4989 sorwakeup_locked(so); 4990 if (DELAY_ACK(tp, tlen)) { 4991 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 4992 tp->t_flags |= TF_DELACK; 4993 } else { 4994 tp->t_flags |= TF_ACKNOW; 4995 rack->r_wanted_output++; 4996 } 4997 if ((tp->snd_una == tp->snd_max) && rack_use_sack_filter) 4998 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una); 4999 return (1); 5000 } 5001 5002 /* 5003 * This subfunction is used to try to highly optimize the 5004 * fast path. We again allow window updates that are 5005 * in sequence to remain in the fast-path. We also add 5006 * in the __predict's to attempt to help the compiler. 5007 * Note that if we return a 0, then we can *not* process 5008 * it and the caller should push the packet into the 5009 * slow-path. 5010 */ 5011 static int 5012 rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so, 5013 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 5014 uint32_t tiwin, int32_t nxt_pkt, uint32_t cts) 5015 { 5016 int32_t acked; 5017 int32_t nsegs; 5018 5019 #ifdef TCPDEBUG 5020 /* 5021 * The size of tcp_saveipgen must be the size of the max ip header, 5022 * now IPv6. 5023 */ 5024 u_char tcp_saveipgen[IP6_HDR_LEN]; 5025 struct tcphdr tcp_savetcp; 5026 short ostate = 0; 5027 5028 #endif 5029 struct tcp_rack *rack; 5030 5031 if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) { 5032 /* Old ack, behind (or duplicate to) the last one rcv'd */ 5033 return (0); 5034 } 5035 if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) { 5036 /* Above what we have sent? */ 5037 return (0); 5038 } 5039 if (__predict_false(tp->snd_nxt != tp->snd_max)) { 5040 /* We are retransmitting */ 5041 return (0); 5042 } 5043 if (__predict_false(tiwin == 0)) { 5044 /* zero window */ 5045 return (0); 5046 } 5047 if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) { 5048 /* We need a SYN or a FIN, unlikely.. */ 5049 return (0); 5050 } 5051 if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) { 5052 /* Timestamp is behind .. old ack with seq wrap? */ 5053 return (0); 5054 } 5055 if (__predict_false(IN_RECOVERY(tp->t_flags))) { 5056 /* Still recovering */ 5057 return (0); 5058 } 5059 rack = (struct tcp_rack *)tp->t_fb_ptr; 5060 if (rack->r_ctl.rc_sacked) { 5061 /* We have sack holes on our scoreboard */ 5062 return (0); 5063 } 5064 /* Ok if we reach here, we can process a fast-ack */ 5065 nsegs = max(1, m->m_pkthdr.lro_nsegs); 5066 rack_log_ack(tp, to, th); 5067 /* Did the window get updated? */ 5068 if (tiwin != tp->snd_wnd) { 5069 tp->snd_wnd = tiwin; 5070 tp->snd_wl1 = th->th_seq; 5071 if (tp->snd_wnd > tp->max_sndwnd) 5072 tp->max_sndwnd = tp->snd_wnd; 5073 } 5074 if ((rack->rc_in_persist != 0) && (tp->snd_wnd >= tp->t_maxseg)) { 5075 rack_exit_persist(tp, rack); 5076 } 5077 /* 5078 * If last ACK falls within this segment's sequence numbers, record 5079 * the timestamp. NOTE that the test is modified according to the 5080 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26). 5081 */ 5082 if ((to->to_flags & TOF_TS) != 0 && 5083 SEQ_LEQ(th->th_seq, tp->last_ack_sent)) { 5084 tp->ts_recent_age = tcp_ts_getticks(); 5085 tp->ts_recent = to->to_tsval; 5086 } 5087 /* 5088 * This is a pure ack for outstanding data. 5089 */ 5090 TCPSTAT_INC(tcps_predack); 5091 5092 /* 5093 * "bad retransmit" recovery. 5094 */ 5095 if (tp->t_flags & TF_PREVVALID) { 5096 tp->t_flags &= ~TF_PREVVALID; 5097 if (tp->t_rxtshift == 1 && 5098 (int)(ticks - tp->t_badrxtwin) < 0) 5099 rack_cong_signal(tp, th, CC_RTO_ERR); 5100 } 5101 /* 5102 * Recalculate the transmit timer / rtt. 5103 * 5104 * Some boxes send broken timestamp replies during the SYN+ACK 5105 * phase, ignore timestamps of 0 or we could calculate a huge RTT 5106 * and blow up the retransmit timer. 5107 */ 5108 acked = BYTES_THIS_ACK(tp, th); 5109 5110 #ifdef TCP_HHOOK 5111 /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */ 5112 hhook_run_tcp_est_in(tp, th, to); 5113 #endif 5114 5115 TCPSTAT_ADD(tcps_rcvackpack, nsegs); 5116 TCPSTAT_ADD(tcps_rcvackbyte, acked); 5117 sbdrop(&so->so_snd, acked); 5118 /* 5119 * Let the congestion control algorithm update congestion control 5120 * related information. This typically means increasing the 5121 * congestion window. 5122 */ 5123 rack_ack_received(tp, rack, th, nsegs, CC_ACK, 0); 5124 5125 tp->snd_una = th->th_ack; 5126 /* 5127 * Pull snd_wl2 up to prevent seq wrap relative to th_ack. 5128 */ 5129 tp->snd_wl2 = th->th_ack; 5130 tp->t_dupacks = 0; 5131 m_freem(m); 5132 /* ND6_HINT(tp); *//* Some progress has been made. */ 5133 5134 /* 5135 * If all outstanding data are acked, stop retransmit timer, 5136 * otherwise restart timer using current (possibly backed-off) 5137 * value. If process is waiting for space, wakeup/selwakeup/signal. 5138 * If data are ready to send, let tcp_output decide between more 5139 * output or persist. 5140 */ 5141 #ifdef TCPDEBUG 5142 if (so->so_options & SO_DEBUG) 5143 tcp_trace(TA_INPUT, ostate, tp, 5144 (void *)tcp_saveipgen, 5145 &tcp_savetcp, 0); 5146 #endif 5147 if (tp->snd_una == tp->snd_max) { 5148 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__); 5149 tp->t_acktime = 0; 5150 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 5151 } 5152 /* Wake up the socket if we have room to write more */ 5153 sowwakeup(so); 5154 if (sbavail(&so->so_snd)) { 5155 rack->r_wanted_output++; 5156 } 5157 return (1); 5158 } 5159 5160 /* 5161 * Return value of 1, the TCB is unlocked and most 5162 * likely gone, return value of 0, the TCP is still 5163 * locked. 5164 */ 5165 static int 5166 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so, 5167 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 5168 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 5169 { 5170 int32_t ret_val = 0; 5171 int32_t todrop; 5172 int32_t ourfinisacked = 0; 5173 5174 rack_calc_rwin(so, tp); 5175 /* 5176 * If the state is SYN_SENT: if seg contains an ACK, but not for our 5177 * SYN, drop the input. if seg contains a RST, then drop the 5178 * connection. if seg does not contain SYN, then drop it. Otherwise 5179 * this is an acceptable SYN segment initialize tp->rcv_nxt and 5180 * tp->irs if seg contains ack then advance tp->snd_una if seg 5181 * contains an ECE and ECN support is enabled, the stream is ECN 5182 * capable. if SYN has been acked change to ESTABLISHED else 5183 * SYN_RCVD state arrange for segment to be acked (eventually) 5184 * continue processing rest of data/controls, beginning with URG 5185 */ 5186 if ((thflags & TH_ACK) && 5187 (SEQ_LEQ(th->th_ack, tp->iss) || 5188 SEQ_GT(th->th_ack, tp->snd_max))) { 5189 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 5190 return (1); 5191 } 5192 if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) { 5193 TCP_PROBE5(connect__refused, NULL, tp, 5194 mtod(m, const char *), tp, th); 5195 tp = tcp_drop(tp, ECONNREFUSED); 5196 rack_do_drop(m, tp); 5197 return (1); 5198 } 5199 if (thflags & TH_RST) { 5200 rack_do_drop(m, tp); 5201 return (1); 5202 } 5203 if (!(thflags & TH_SYN)) { 5204 rack_do_drop(m, tp); 5205 return (1); 5206 } 5207 tp->irs = th->th_seq; 5208 tcp_rcvseqinit(tp); 5209 if (thflags & TH_ACK) { 5210 int tfo_partial = 0; 5211 5212 TCPSTAT_INC(tcps_connects); 5213 soisconnected(so); 5214 #ifdef MAC 5215 mac_socketpeer_set_from_mbuf(m, so); 5216 #endif 5217 /* Do window scaling on this connection? */ 5218 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == 5219 (TF_RCVD_SCALE | TF_REQ_SCALE)) { 5220 tp->rcv_scale = tp->request_r_scale; 5221 } 5222 tp->rcv_adv += min(tp->rcv_wnd, 5223 TCP_MAXWIN << tp->rcv_scale); 5224 /* 5225 * If not all the data that was sent in the TFO SYN 5226 * has been acked, resend the remainder right away. 5227 */ 5228 if (IS_FASTOPEN(tp->t_flags) && 5229 (tp->snd_una != tp->snd_max)) { 5230 tp->snd_nxt = th->th_ack; 5231 tfo_partial = 1; 5232 } 5233 /* 5234 * If there's data, delay ACK; if there's also a FIN ACKNOW 5235 * will be turned on later. 5236 */ 5237 if (DELAY_ACK(tp, tlen) && tlen != 0 && (tfo_partial == 0)) { 5238 rack_timer_cancel(tp, (struct tcp_rack *)tp->t_fb_ptr, 5239 ((struct tcp_rack *)tp->t_fb_ptr)->r_ctl.rc_rcvtime, __LINE__); 5240 tp->t_flags |= TF_DELACK; 5241 } else { 5242 ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output++; 5243 tp->t_flags |= TF_ACKNOW; 5244 } 5245 5246 if ((thflags & TH_ECE) && V_tcp_do_ecn) { 5247 tp->t_flags |= TF_ECN_PERMIT; 5248 TCPSTAT_INC(tcps_ecn_shs); 5249 } 5250 if (SEQ_GT(th->th_ack, tp->snd_una)) { 5251 /* 5252 * We advance snd_una for the 5253 * fast open case. If th_ack is 5254 * acknowledging data beyond 5255 * snd_una we can't just call 5256 * ack-processing since the 5257 * data stream in our send-map 5258 * will start at snd_una + 1 (one 5259 * beyond the SYN). If its just 5260 * equal we don't need to do that 5261 * and there is no send_map. 5262 */ 5263 tp->snd_una++; 5264 } 5265 /* 5266 * Received <SYN,ACK> in SYN_SENT[*] state. Transitions: 5267 * SYN_SENT --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1 5268 */ 5269 tp->t_starttime = ticks; 5270 if (tp->t_flags & TF_NEEDFIN) { 5271 tcp_state_change(tp, TCPS_FIN_WAIT_1); 5272 tp->t_flags &= ~TF_NEEDFIN; 5273 thflags &= ~TH_SYN; 5274 } else { 5275 tcp_state_change(tp, TCPS_ESTABLISHED); 5276 TCP_PROBE5(connect__established, NULL, tp, 5277 mtod(m, const char *), tp, th); 5278 cc_conn_init(tp); 5279 } 5280 } else { 5281 /* 5282 * Received initial SYN in SYN-SENT[*] state => simultaneous 5283 * open. If segment contains CC option and there is a 5284 * cached CC, apply TAO test. If it succeeds, connection is * 5285 * half-synchronized. Otherwise, do 3-way handshake: 5286 * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If 5287 * there was no CC option, clear cached CC value. 5288 */ 5289 tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN); 5290 tcp_state_change(tp, TCPS_SYN_RECEIVED); 5291 } 5292 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 5293 INP_WLOCK_ASSERT(tp->t_inpcb); 5294 /* 5295 * Advance th->th_seq to correspond to first data byte. If data, 5296 * trim to stay within window, dropping FIN if necessary. 5297 */ 5298 th->th_seq++; 5299 if (tlen > tp->rcv_wnd) { 5300 todrop = tlen - tp->rcv_wnd; 5301 m_adj(m, -todrop); 5302 tlen = tp->rcv_wnd; 5303 thflags &= ~TH_FIN; 5304 TCPSTAT_INC(tcps_rcvpackafterwin); 5305 TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop); 5306 } 5307 tp->snd_wl1 = th->th_seq - 1; 5308 tp->rcv_up = th->th_seq; 5309 /* 5310 * Client side of transaction: already sent SYN and data. If the 5311 * remote host used T/TCP to validate the SYN, our data will be 5312 * ACK'd; if so, enter normal data segment processing in the middle 5313 * of step 5, ack processing. Otherwise, goto step 6. 5314 */ 5315 if (thflags & TH_ACK) { 5316 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) 5317 return (ret_val); 5318 /* We may have changed to FIN_WAIT_1 above */ 5319 if (tp->t_state == TCPS_FIN_WAIT_1) { 5320 /* 5321 * In FIN_WAIT_1 STATE in addition to the processing 5322 * for the ESTABLISHED state if our FIN is now 5323 * acknowledged then enter FIN_WAIT_2. 5324 */ 5325 if (ourfinisacked) { 5326 /* 5327 * If we can't receive any more data, then 5328 * closing user can proceed. Starting the 5329 * timer is contrary to the specification, 5330 * but if we don't get a FIN we'll hang 5331 * forever. 5332 * 5333 * XXXjl: we should release the tp also, and 5334 * use a compressed state. 5335 */ 5336 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 5337 soisdisconnected(so); 5338 tcp_timer_activate(tp, TT_2MSL, 5339 (tcp_fast_finwait2_recycle ? 5340 tcp_finwait2_timeout : 5341 TP_MAXIDLE(tp))); 5342 } 5343 tcp_state_change(tp, TCPS_FIN_WAIT_2); 5344 } 5345 } 5346 } 5347 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5348 tiwin, thflags, nxt_pkt)); 5349 } 5350 5351 /* 5352 * Return value of 1, the TCB is unlocked and most 5353 * likely gone, return value of 0, the TCP is still 5354 * locked. 5355 */ 5356 static int 5357 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so, 5358 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 5359 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 5360 { 5361 int32_t ret_val = 0; 5362 int32_t ourfinisacked = 0; 5363 5364 rack_calc_rwin(so, tp); 5365 5366 if ((thflags & TH_ACK) && 5367 (SEQ_LEQ(th->th_ack, tp->snd_una) || 5368 SEQ_GT(th->th_ack, tp->snd_max))) { 5369 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 5370 return (1); 5371 } 5372 if (IS_FASTOPEN(tp->t_flags)) { 5373 /* 5374 * When a TFO connection is in SYN_RECEIVED, the 5375 * only valid packets are the initial SYN, a 5376 * retransmit/copy of the initial SYN (possibly with 5377 * a subset of the original data), a valid ACK, a 5378 * FIN, or a RST. 5379 */ 5380 if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) { 5381 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 5382 return (1); 5383 } else if (thflags & TH_SYN) { 5384 /* non-initial SYN is ignored */ 5385 struct tcp_rack *rack; 5386 5387 rack = (struct tcp_rack *)tp->t_fb_ptr; 5388 if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) || 5389 (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) || 5390 (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) { 5391 rack_do_drop(m, NULL); 5392 return (0); 5393 } 5394 } else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) { 5395 rack_do_drop(m, NULL); 5396 return (0); 5397 } 5398 } 5399 if (thflags & TH_RST) 5400 return (rack_process_rst(m, th, so, tp)); 5401 /* 5402 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 5403 * synchronized state. 5404 */ 5405 if (thflags & TH_SYN) { 5406 rack_challenge_ack(m, th, tp, &ret_val); 5407 return (ret_val); 5408 } 5409 /* 5410 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 5411 * it's less than ts_recent, drop it. 5412 */ 5413 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 5414 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 5415 if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val)) 5416 return (ret_val); 5417 } 5418 /* 5419 * In the SYN-RECEIVED state, validate that the packet belongs to 5420 * this connection before trimming the data to fit the receive 5421 * window. Check the sequence number versus IRS since we know the 5422 * sequence numbers haven't wrapped. This is a partial fix for the 5423 * "LAND" DoS attack. 5424 */ 5425 if (SEQ_LT(th->th_seq, tp->irs)) { 5426 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 5427 return (1); 5428 } 5429 if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 5430 return (ret_val); 5431 } 5432 /* 5433 * If last ACK falls within this segment's sequence numbers, record 5434 * its timestamp. NOTE: 1) That the test incorporates suggestions 5435 * from the latest proposal of the tcplw@cray.com list (Braden 5436 * 1993/04/26). 2) That updating only on newer timestamps interferes 5437 * with our earlier PAWS tests, so this check should be solely 5438 * predicated on the sequence space of this segment. 3) That we 5439 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 5440 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 5441 * SEG.Len, This modified check allows us to overcome RFC1323's 5442 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 5443 * p.869. In such cases, we can still calculate the RTT correctly 5444 * when RCV.NXT == Last.ACK.Sent. 5445 */ 5446 if ((to->to_flags & TOF_TS) != 0 && 5447 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 5448 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 5449 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 5450 tp->ts_recent_age = tcp_ts_getticks(); 5451 tp->ts_recent = to->to_tsval; 5452 } 5453 /* 5454 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 5455 * is on (half-synchronized state), then queue data for later 5456 * processing; else drop segment and return. 5457 */ 5458 if ((thflags & TH_ACK) == 0) { 5459 if (IS_FASTOPEN(tp->t_flags)) { 5460 tp->snd_wnd = tiwin; 5461 cc_conn_init(tp); 5462 } 5463 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5464 tiwin, thflags, nxt_pkt)); 5465 } 5466 TCPSTAT_INC(tcps_connects); 5467 soisconnected(so); 5468 /* Do window scaling? */ 5469 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == 5470 (TF_RCVD_SCALE | TF_REQ_SCALE)) { 5471 tp->rcv_scale = tp->request_r_scale; 5472 tp->snd_wnd = tiwin; 5473 } 5474 /* 5475 * Make transitions: SYN-RECEIVED -> ESTABLISHED SYN-RECEIVED* -> 5476 * FIN-WAIT-1 5477 */ 5478 tp->t_starttime = ticks; 5479 if (tp->t_flags & TF_NEEDFIN) { 5480 tcp_state_change(tp, TCPS_FIN_WAIT_1); 5481 tp->t_flags &= ~TF_NEEDFIN; 5482 } else { 5483 tcp_state_change(tp, TCPS_ESTABLISHED); 5484 TCP_PROBE5(accept__established, NULL, tp, 5485 mtod(m, const char *), tp, th); 5486 if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) { 5487 tcp_fastopen_decrement_counter(tp->t_tfo_pending); 5488 tp->t_tfo_pending = NULL; 5489 5490 /* 5491 * Account for the ACK of our SYN prior to regular 5492 * ACK processing below. 5493 */ 5494 tp->snd_una++; 5495 } 5496 /* 5497 * TFO connections call cc_conn_init() during SYN 5498 * processing. Calling it again here for such connections 5499 * is not harmless as it would undo the snd_cwnd reduction 5500 * that occurs when a TFO SYN|ACK is retransmitted. 5501 */ 5502 if (!IS_FASTOPEN(tp->t_flags)) 5503 cc_conn_init(tp); 5504 } 5505 /* 5506 * If segment contains data or ACK, will call tcp_reass() later; if 5507 * not, do so now to pass queued data to user. 5508 */ 5509 if (tlen == 0 && (thflags & TH_FIN) == 0) 5510 (void)tcp_reass(tp, (struct tcphdr *)0, 0, 5511 (struct mbuf *)0); 5512 tp->snd_wl1 = th->th_seq - 1; 5513 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 5514 return (ret_val); 5515 } 5516 if (tp->t_state == TCPS_FIN_WAIT_1) { 5517 /* We could have went to FIN_WAIT_1 (or EST) above */ 5518 /* 5519 * In FIN_WAIT_1 STATE in addition to the processing for the 5520 * ESTABLISHED state if our FIN is now acknowledged then 5521 * enter FIN_WAIT_2. 5522 */ 5523 if (ourfinisacked) { 5524 /* 5525 * If we can't receive any more data, then closing 5526 * user can proceed. Starting the timer is contrary 5527 * to the specification, but if we don't get a FIN 5528 * we'll hang forever. 5529 * 5530 * XXXjl: we should release the tp also, and use a 5531 * compressed state. 5532 */ 5533 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 5534 soisdisconnected(so); 5535 tcp_timer_activate(tp, TT_2MSL, 5536 (tcp_fast_finwait2_recycle ? 5537 tcp_finwait2_timeout : 5538 TP_MAXIDLE(tp))); 5539 } 5540 tcp_state_change(tp, TCPS_FIN_WAIT_2); 5541 } 5542 } 5543 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5544 tiwin, thflags, nxt_pkt)); 5545 } 5546 5547 /* 5548 * Return value of 1, the TCB is unlocked and most 5549 * likely gone, return value of 0, the TCP is still 5550 * locked. 5551 */ 5552 static int 5553 rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so, 5554 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 5555 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 5556 { 5557 int32_t ret_val = 0; 5558 5559 /* 5560 * Header prediction: check for the two common cases of a 5561 * uni-directional data xfer. If the packet has no control flags, 5562 * is in-sequence, the window didn't change and we're not 5563 * retransmitting, it's a candidate. If the length is zero and the 5564 * ack moved forward, we're the sender side of the xfer. Just free 5565 * the data acked & wake any higher level process that was blocked 5566 * waiting for space. If the length is non-zero and the ack didn't 5567 * move, we're the receiver side. If we're getting packets in-order 5568 * (the reassembly queue is empty), add the data toc The socket 5569 * buffer and note that we need a delayed ack. Make sure that the 5570 * hidden state-flags are also off. Since we check for 5571 * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN. 5572 */ 5573 if (__predict_true(((to->to_flags & TOF_SACK) == 0)) && 5574 __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_URG | TH_ACK)) == TH_ACK) && 5575 __predict_true(LIST_EMPTY(&tp->t_segq)) && 5576 __predict_true(th->th_seq == tp->rcv_nxt)) { 5577 struct tcp_rack *rack; 5578 5579 rack = (struct tcp_rack *)tp->t_fb_ptr; 5580 if (tlen == 0) { 5581 if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen, 5582 tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) { 5583 return (0); 5584 } 5585 } else { 5586 if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen, 5587 tiwin, nxt_pkt)) { 5588 return (0); 5589 } 5590 } 5591 } 5592 rack_calc_rwin(so, tp); 5593 5594 if (thflags & TH_RST) 5595 return (rack_process_rst(m, th, so, tp)); 5596 5597 /* 5598 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 5599 * synchronized state. 5600 */ 5601 if (thflags & TH_SYN) { 5602 rack_challenge_ack(m, th, tp, &ret_val); 5603 return (ret_val); 5604 } 5605 /* 5606 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 5607 * it's less than ts_recent, drop it. 5608 */ 5609 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 5610 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 5611 if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val)) 5612 return (ret_val); 5613 } 5614 if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 5615 return (ret_val); 5616 } 5617 /* 5618 * If last ACK falls within this segment's sequence numbers, record 5619 * its timestamp. NOTE: 1) That the test incorporates suggestions 5620 * from the latest proposal of the tcplw@cray.com list (Braden 5621 * 1993/04/26). 2) That updating only on newer timestamps interferes 5622 * with our earlier PAWS tests, so this check should be solely 5623 * predicated on the sequence space of this segment. 3) That we 5624 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 5625 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 5626 * SEG.Len, This modified check allows us to overcome RFC1323's 5627 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 5628 * p.869. In such cases, we can still calculate the RTT correctly 5629 * when RCV.NXT == Last.ACK.Sent. 5630 */ 5631 if ((to->to_flags & TOF_TS) != 0 && 5632 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 5633 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 5634 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 5635 tp->ts_recent_age = tcp_ts_getticks(); 5636 tp->ts_recent = to->to_tsval; 5637 } 5638 /* 5639 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 5640 * is on (half-synchronized state), then queue data for later 5641 * processing; else drop segment and return. 5642 */ 5643 if ((thflags & TH_ACK) == 0) { 5644 if (tp->t_flags & TF_NEEDSYN) { 5645 5646 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5647 tiwin, thflags, nxt_pkt)); 5648 5649 } else if (tp->t_flags & TF_ACKNOW) { 5650 rack_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 5651 return (ret_val); 5652 } else { 5653 rack_do_drop(m, NULL); 5654 return (0); 5655 } 5656 } 5657 /* 5658 * Ack processing. 5659 */ 5660 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) { 5661 return (ret_val); 5662 } 5663 if (sbavail(&so->so_snd)) { 5664 if (rack_progress_timeout_check(tp)) { 5665 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); 5666 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 5667 return (1); 5668 } 5669 } 5670 /* State changes only happen in rack_process_data() */ 5671 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5672 tiwin, thflags, nxt_pkt)); 5673 } 5674 5675 /* 5676 * Return value of 1, the TCB is unlocked and most 5677 * likely gone, return value of 0, the TCP is still 5678 * locked. 5679 */ 5680 static int 5681 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so, 5682 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 5683 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 5684 { 5685 int32_t ret_val = 0; 5686 5687 rack_calc_rwin(so, tp); 5688 if (thflags & TH_RST) 5689 return (rack_process_rst(m, th, so, tp)); 5690 /* 5691 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 5692 * synchronized state. 5693 */ 5694 if (thflags & TH_SYN) { 5695 rack_challenge_ack(m, th, tp, &ret_val); 5696 return (ret_val); 5697 } 5698 /* 5699 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 5700 * it's less than ts_recent, drop it. 5701 */ 5702 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 5703 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 5704 if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val)) 5705 return (ret_val); 5706 } 5707 if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 5708 return (ret_val); 5709 } 5710 /* 5711 * If last ACK falls within this segment's sequence numbers, record 5712 * its timestamp. NOTE: 1) That the test incorporates suggestions 5713 * from the latest proposal of the tcplw@cray.com list (Braden 5714 * 1993/04/26). 2) That updating only on newer timestamps interferes 5715 * with our earlier PAWS tests, so this check should be solely 5716 * predicated on the sequence space of this segment. 3) That we 5717 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 5718 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 5719 * SEG.Len, This modified check allows us to overcome RFC1323's 5720 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 5721 * p.869. In such cases, we can still calculate the RTT correctly 5722 * when RCV.NXT == Last.ACK.Sent. 5723 */ 5724 if ((to->to_flags & TOF_TS) != 0 && 5725 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 5726 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 5727 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 5728 tp->ts_recent_age = tcp_ts_getticks(); 5729 tp->ts_recent = to->to_tsval; 5730 } 5731 /* 5732 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 5733 * is on (half-synchronized state), then queue data for later 5734 * processing; else drop segment and return. 5735 */ 5736 if ((thflags & TH_ACK) == 0) { 5737 if (tp->t_flags & TF_NEEDSYN) { 5738 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5739 tiwin, thflags, nxt_pkt)); 5740 5741 } else if (tp->t_flags & TF_ACKNOW) { 5742 rack_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 5743 return (ret_val); 5744 } else { 5745 rack_do_drop(m, NULL); 5746 return (0); 5747 } 5748 } 5749 /* 5750 * Ack processing. 5751 */ 5752 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) { 5753 return (ret_val); 5754 } 5755 if (sbavail(&so->so_snd)) { 5756 if (rack_progress_timeout_check(tp)) { 5757 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); 5758 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 5759 return (1); 5760 } 5761 } 5762 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5763 tiwin, thflags, nxt_pkt)); 5764 } 5765 5766 static int 5767 rack_check_data_after_close(struct mbuf *m, 5768 struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so) 5769 { 5770 struct tcp_rack *rack; 5771 5772 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 5773 rack = (struct tcp_rack *)tp->t_fb_ptr; 5774 if (rack->rc_allow_data_af_clo == 0) { 5775 close_now: 5776 tp = tcp_close(tp); 5777 TCPSTAT_INC(tcps_rcvafterclose); 5778 rack_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen)); 5779 return (1); 5780 } 5781 if (sbavail(&so->so_snd) == 0) 5782 goto close_now; 5783 /* Ok we allow data that is ignored and a followup reset */ 5784 tp->rcv_nxt = th->th_seq + *tlen; 5785 tp->t_flags2 |= TF2_DROP_AF_DATA; 5786 rack->r_wanted_output = 1; 5787 *tlen = 0; 5788 return (0); 5789 } 5790 5791 /* 5792 * Return value of 1, the TCB is unlocked and most 5793 * likely gone, return value of 0, the TCP is still 5794 * locked. 5795 */ 5796 static int 5797 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so, 5798 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 5799 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 5800 { 5801 int32_t ret_val = 0; 5802 int32_t ourfinisacked = 0; 5803 5804 rack_calc_rwin(so, tp); 5805 5806 if (thflags & TH_RST) 5807 return (rack_process_rst(m, th, so, tp)); 5808 /* 5809 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 5810 * synchronized state. 5811 */ 5812 if (thflags & TH_SYN) { 5813 rack_challenge_ack(m, th, tp, &ret_val); 5814 return (ret_val); 5815 } 5816 /* 5817 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 5818 * it's less than ts_recent, drop it. 5819 */ 5820 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 5821 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 5822 if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val)) 5823 return (ret_val); 5824 } 5825 if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 5826 return (ret_val); 5827 } 5828 /* 5829 * If new data are received on a connection after the user processes 5830 * are gone, then RST the other end. 5831 */ 5832 if ((so->so_state & SS_NOFDREF) && tlen) { 5833 if (rack_check_data_after_close(m, tp, &tlen, th, so)) 5834 return (1); 5835 } 5836 /* 5837 * If last ACK falls within this segment's sequence numbers, record 5838 * its timestamp. NOTE: 1) That the test incorporates suggestions 5839 * from the latest proposal of the tcplw@cray.com list (Braden 5840 * 1993/04/26). 2) That updating only on newer timestamps interferes 5841 * with our earlier PAWS tests, so this check should be solely 5842 * predicated on the sequence space of this segment. 3) That we 5843 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 5844 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 5845 * SEG.Len, This modified check allows us to overcome RFC1323's 5846 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 5847 * p.869. In such cases, we can still calculate the RTT correctly 5848 * when RCV.NXT == Last.ACK.Sent. 5849 */ 5850 if ((to->to_flags & TOF_TS) != 0 && 5851 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 5852 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 5853 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 5854 tp->ts_recent_age = tcp_ts_getticks(); 5855 tp->ts_recent = to->to_tsval; 5856 } 5857 /* 5858 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 5859 * is on (half-synchronized state), then queue data for later 5860 * processing; else drop segment and return. 5861 */ 5862 if ((thflags & TH_ACK) == 0) { 5863 if (tp->t_flags & TF_NEEDSYN) { 5864 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5865 tiwin, thflags, nxt_pkt)); 5866 } else if (tp->t_flags & TF_ACKNOW) { 5867 rack_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 5868 return (ret_val); 5869 } else { 5870 rack_do_drop(m, NULL); 5871 return (0); 5872 } 5873 } 5874 /* 5875 * Ack processing. 5876 */ 5877 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 5878 return (ret_val); 5879 } 5880 if (ourfinisacked) { 5881 /* 5882 * If we can't receive any more data, then closing user can 5883 * proceed. Starting the timer is contrary to the 5884 * specification, but if we don't get a FIN we'll hang 5885 * forever. 5886 * 5887 * XXXjl: we should release the tp also, and use a 5888 * compressed state. 5889 */ 5890 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 5891 soisdisconnected(so); 5892 tcp_timer_activate(tp, TT_2MSL, 5893 (tcp_fast_finwait2_recycle ? 5894 tcp_finwait2_timeout : 5895 TP_MAXIDLE(tp))); 5896 } 5897 tcp_state_change(tp, TCPS_FIN_WAIT_2); 5898 } 5899 if (sbavail(&so->so_snd)) { 5900 if (rack_progress_timeout_check(tp)) { 5901 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); 5902 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 5903 return (1); 5904 } 5905 } 5906 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5907 tiwin, thflags, nxt_pkt)); 5908 } 5909 5910 /* 5911 * Return value of 1, the TCB is unlocked and most 5912 * likely gone, return value of 0, the TCP is still 5913 * locked. 5914 */ 5915 static int 5916 rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so, 5917 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 5918 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 5919 { 5920 int32_t ret_val = 0; 5921 int32_t ourfinisacked = 0; 5922 5923 rack_calc_rwin(so, tp); 5924 5925 if (thflags & TH_RST) 5926 return (rack_process_rst(m, th, so, tp)); 5927 /* 5928 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 5929 * synchronized state. 5930 */ 5931 if (thflags & TH_SYN) { 5932 rack_challenge_ack(m, th, tp, &ret_val); 5933 return (ret_val); 5934 } 5935 /* 5936 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 5937 * it's less than ts_recent, drop it. 5938 */ 5939 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 5940 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 5941 if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val)) 5942 return (ret_val); 5943 } 5944 if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 5945 return (ret_val); 5946 } 5947 /* 5948 * If new data are received on a connection after the user processes 5949 * are gone, then RST the other end. 5950 */ 5951 if ((so->so_state & SS_NOFDREF) && tlen) { 5952 if (rack_check_data_after_close(m, tp, &tlen, th, so)) 5953 return (1); 5954 } 5955 /* 5956 * If last ACK falls within this segment's sequence numbers, record 5957 * its timestamp. NOTE: 1) That the test incorporates suggestions 5958 * from the latest proposal of the tcplw@cray.com list (Braden 5959 * 1993/04/26). 2) That updating only on newer timestamps interferes 5960 * with our earlier PAWS tests, so this check should be solely 5961 * predicated on the sequence space of this segment. 3) That we 5962 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 5963 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 5964 * SEG.Len, This modified check allows us to overcome RFC1323's 5965 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 5966 * p.869. In such cases, we can still calculate the RTT correctly 5967 * when RCV.NXT == Last.ACK.Sent. 5968 */ 5969 if ((to->to_flags & TOF_TS) != 0 && 5970 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 5971 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 5972 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 5973 tp->ts_recent_age = tcp_ts_getticks(); 5974 tp->ts_recent = to->to_tsval; 5975 } 5976 /* 5977 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 5978 * is on (half-synchronized state), then queue data for later 5979 * processing; else drop segment and return. 5980 */ 5981 if ((thflags & TH_ACK) == 0) { 5982 if (tp->t_flags & TF_NEEDSYN) { 5983 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5984 tiwin, thflags, nxt_pkt)); 5985 } else if (tp->t_flags & TF_ACKNOW) { 5986 rack_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 5987 return (ret_val); 5988 } else { 5989 rack_do_drop(m, NULL); 5990 return (0); 5991 } 5992 } 5993 /* 5994 * Ack processing. 5995 */ 5996 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 5997 return (ret_val); 5998 } 5999 if (ourfinisacked) { 6000 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 6001 tcp_twstart(tp); 6002 m_freem(m); 6003 return (1); 6004 } 6005 if (sbavail(&so->so_snd)) { 6006 if (rack_progress_timeout_check(tp)) { 6007 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); 6008 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 6009 return (1); 6010 } 6011 } 6012 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 6013 tiwin, thflags, nxt_pkt)); 6014 } 6015 6016 /* 6017 * Return value of 1, the TCB is unlocked and most 6018 * likely gone, return value of 0, the TCP is still 6019 * locked. 6020 */ 6021 static int 6022 rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so, 6023 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 6024 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 6025 { 6026 int32_t ret_val = 0; 6027 int32_t ourfinisacked = 0; 6028 6029 rack_calc_rwin(so, tp); 6030 6031 if (thflags & TH_RST) 6032 return (rack_process_rst(m, th, so, tp)); 6033 /* 6034 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 6035 * synchronized state. 6036 */ 6037 if (thflags & TH_SYN) { 6038 rack_challenge_ack(m, th, tp, &ret_val); 6039 return (ret_val); 6040 } 6041 /* 6042 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 6043 * it's less than ts_recent, drop it. 6044 */ 6045 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 6046 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 6047 if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val)) 6048 return (ret_val); 6049 } 6050 if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 6051 return (ret_val); 6052 } 6053 /* 6054 * If new data are received on a connection after the user processes 6055 * are gone, then RST the other end. 6056 */ 6057 if ((so->so_state & SS_NOFDREF) && tlen) { 6058 if (rack_check_data_after_close(m, tp, &tlen, th, so)) 6059 return (1); 6060 } 6061 /* 6062 * If last ACK falls within this segment's sequence numbers, record 6063 * its timestamp. NOTE: 1) That the test incorporates suggestions 6064 * from the latest proposal of the tcplw@cray.com list (Braden 6065 * 1993/04/26). 2) That updating only on newer timestamps interferes 6066 * with our earlier PAWS tests, so this check should be solely 6067 * predicated on the sequence space of this segment. 3) That we 6068 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 6069 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 6070 * SEG.Len, This modified check allows us to overcome RFC1323's 6071 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 6072 * p.869. In such cases, we can still calculate the RTT correctly 6073 * when RCV.NXT == Last.ACK.Sent. 6074 */ 6075 if ((to->to_flags & TOF_TS) != 0 && 6076 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 6077 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 6078 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 6079 tp->ts_recent_age = tcp_ts_getticks(); 6080 tp->ts_recent = to->to_tsval; 6081 } 6082 /* 6083 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 6084 * is on (half-synchronized state), then queue data for later 6085 * processing; else drop segment and return. 6086 */ 6087 if ((thflags & TH_ACK) == 0) { 6088 if (tp->t_flags & TF_NEEDSYN) { 6089 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 6090 tiwin, thflags, nxt_pkt)); 6091 } else if (tp->t_flags & TF_ACKNOW) { 6092 rack_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 6093 return (ret_val); 6094 } else { 6095 rack_do_drop(m, NULL); 6096 return (0); 6097 } 6098 } 6099 /* 6100 * case TCPS_LAST_ACK: Ack processing. 6101 */ 6102 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 6103 return (ret_val); 6104 } 6105 if (ourfinisacked) { 6106 6107 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 6108 tp = tcp_close(tp); 6109 rack_do_drop(m, tp); 6110 return (1); 6111 } 6112 if (sbavail(&so->so_snd)) { 6113 if (rack_progress_timeout_check(tp)) { 6114 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); 6115 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 6116 return (1); 6117 } 6118 } 6119 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 6120 tiwin, thflags, nxt_pkt)); 6121 } 6122 6123 6124 /* 6125 * Return value of 1, the TCB is unlocked and most 6126 * likely gone, return value of 0, the TCP is still 6127 * locked. 6128 */ 6129 static int 6130 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so, 6131 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 6132 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 6133 { 6134 int32_t ret_val = 0; 6135 int32_t ourfinisacked = 0; 6136 6137 rack_calc_rwin(so, tp); 6138 6139 /* Reset receive buffer auto scaling when not in bulk receive mode. */ 6140 if (thflags & TH_RST) 6141 return (rack_process_rst(m, th, so, tp)); 6142 /* 6143 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 6144 * synchronized state. 6145 */ 6146 if (thflags & TH_SYN) { 6147 rack_challenge_ack(m, th, tp, &ret_val); 6148 return (ret_val); 6149 } 6150 /* 6151 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 6152 * it's less than ts_recent, drop it. 6153 */ 6154 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 6155 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 6156 if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val)) 6157 return (ret_val); 6158 } 6159 if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 6160 return (ret_val); 6161 } 6162 /* 6163 * If new data are received on a connection after the user processes 6164 * are gone, then RST the other end. 6165 */ 6166 if ((so->so_state & SS_NOFDREF) && 6167 tlen) { 6168 if (rack_check_data_after_close(m, tp, &tlen, th, so)) 6169 return (1); 6170 } 6171 /* 6172 * If last ACK falls within this segment's sequence numbers, record 6173 * its timestamp. NOTE: 1) That the test incorporates suggestions 6174 * from the latest proposal of the tcplw@cray.com list (Braden 6175 * 1993/04/26). 2) That updating only on newer timestamps interferes 6176 * with our earlier PAWS tests, so this check should be solely 6177 * predicated on the sequence space of this segment. 3) That we 6178 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 6179 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 6180 * SEG.Len, This modified check allows us to overcome RFC1323's 6181 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 6182 * p.869. In such cases, we can still calculate the RTT correctly 6183 * when RCV.NXT == Last.ACK.Sent. 6184 */ 6185 if ((to->to_flags & TOF_TS) != 0 && 6186 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 6187 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 6188 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 6189 tp->ts_recent_age = tcp_ts_getticks(); 6190 tp->ts_recent = to->to_tsval; 6191 } 6192 /* 6193 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 6194 * is on (half-synchronized state), then queue data for later 6195 * processing; else drop segment and return. 6196 */ 6197 if ((thflags & TH_ACK) == 0) { 6198 if (tp->t_flags & TF_NEEDSYN) { 6199 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 6200 tiwin, thflags, nxt_pkt)); 6201 } else if (tp->t_flags & TF_ACKNOW) { 6202 rack_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 6203 return (ret_val); 6204 } else { 6205 rack_do_drop(m, NULL); 6206 return (0); 6207 } 6208 } 6209 /* 6210 * Ack processing. 6211 */ 6212 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 6213 return (ret_val); 6214 } 6215 if (sbavail(&so->so_snd)) { 6216 if (rack_progress_timeout_check(tp)) { 6217 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); 6218 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 6219 return (1); 6220 } 6221 } 6222 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 6223 tiwin, thflags, nxt_pkt)); 6224 } 6225 6226 6227 static void inline 6228 rack_clear_rate_sample(struct tcp_rack *rack) 6229 { 6230 rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY; 6231 rack->r_ctl.rack_rs.rs_rtt_cnt = 0; 6232 rack->r_ctl.rack_rs.rs_rtt_tot = 0; 6233 } 6234 6235 static int 6236 rack_init(struct tcpcb *tp) 6237 { 6238 struct tcp_rack *rack = NULL; 6239 6240 tp->t_fb_ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT); 6241 if (tp->t_fb_ptr == NULL) { 6242 /* 6243 * We need to allocate memory but cant. The INP and INP_INFO 6244 * locks and they are recusive (happens during setup. So a 6245 * scheme to drop the locks fails :( 6246 * 6247 */ 6248 return (ENOMEM); 6249 } 6250 memset(tp->t_fb_ptr, 0, sizeof(struct tcp_rack)); 6251 6252 rack = (struct tcp_rack *)tp->t_fb_ptr; 6253 TAILQ_INIT(&rack->r_ctl.rc_map); 6254 TAILQ_INIT(&rack->r_ctl.rc_free); 6255 TAILQ_INIT(&rack->r_ctl.rc_tmap); 6256 rack->rc_tp = tp; 6257 if (tp->t_inpcb) { 6258 rack->rc_inp = tp->t_inpcb; 6259 } 6260 /* Probably not needed but lets be sure */ 6261 rack_clear_rate_sample(rack); 6262 rack->r_cpu = 0; 6263 rack->r_ctl.rc_reorder_fade = rack_reorder_fade; 6264 rack->rc_allow_data_af_clo = rack_ignore_data_after_close; 6265 rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh; 6266 rack->rc_pace_reduce = rack_slot_reduction; 6267 if (V_tcp_delack_enabled) 6268 tp->t_delayed_ack = 1; 6269 else 6270 tp->t_delayed_ack = 0; 6271 rack->rc_pace_max_segs = rack_hptsi_segments; 6272 rack->r_ctl.rc_early_recovery_segs = rack_early_recovery_max_seg; 6273 rack->r_ctl.rc_reorder_shift = rack_reorder_thresh; 6274 rack->r_ctl.rc_pkt_delay = rack_pkt_delay; 6275 rack->r_ctl.rc_prop_reduce = rack_use_proportional_reduce; 6276 rack->r_idle_reduce_largest = rack_reduce_largest_on_idle; 6277 rack->r_enforce_min_pace = rack_min_pace_time; 6278 rack->r_min_pace_seg_thresh = rack_min_pace_time_seg_req; 6279 rack->r_ctl.rc_prop_rate = rack_proportional_rate; 6280 rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp; 6281 rack->r_ctl.rc_early_recovery = rack_early_recovery; 6282 rack->rc_always_pace = rack_pace_every_seg; 6283 rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method; 6284 rack->rack_tlp_threshold_use = rack_tlp_threshold_use; 6285 rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr; 6286 rack->r_ctl.rc_min_to = rack_min_to; 6287 rack->r_ctl.rc_prr_inc_var = rack_inc_var; 6288 rack_start_hpts_timer(rack, tp, tcp_ts_getticks(), __LINE__, 0, 0, 0); 6289 if (tp->snd_una != tp->snd_max) { 6290 /* Create a send map for the current outstanding data */ 6291 struct rack_sendmap *rsm; 6292 6293 rsm = rack_alloc(rack); 6294 if (rsm == NULL) { 6295 uma_zfree(rack_pcb_zone, tp->t_fb_ptr); 6296 tp->t_fb_ptr = NULL; 6297 return (ENOMEM); 6298 } 6299 rsm->r_flags = RACK_OVERMAX; 6300 rsm->r_tim_lastsent[0] = tcp_ts_getticks(); 6301 rsm->r_rtr_cnt = 1; 6302 rsm->r_rtr_bytes = 0; 6303 rsm->r_start = tp->snd_una; 6304 rsm->r_end = tp->snd_max; 6305 rsm->r_sndcnt = 0; 6306 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_map, rsm, r_next); 6307 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext); 6308 rsm->r_in_tmap = 1; 6309 } 6310 return (0); 6311 } 6312 6313 static int 6314 rack_handoff_ok(struct tcpcb *tp) 6315 { 6316 if ((tp->t_state == TCPS_CLOSED) || 6317 (tp->t_state == TCPS_LISTEN)) { 6318 /* Sure no problem though it may not stick */ 6319 return (0); 6320 } 6321 if ((tp->t_state == TCPS_SYN_SENT) || 6322 (tp->t_state == TCPS_SYN_RECEIVED)) { 6323 /* 6324 * We really don't know you have to get to ESTAB or beyond 6325 * to tell. 6326 */ 6327 return (EAGAIN); 6328 } 6329 if (tp->t_flags & TF_SACK_PERMIT) { 6330 return (0); 6331 } 6332 /* 6333 * If we reach here we don't do SACK on this connection so we can 6334 * never do rack. 6335 */ 6336 return (EINVAL); 6337 } 6338 6339 static void 6340 rack_fini(struct tcpcb *tp, int32_t tcb_is_purged) 6341 { 6342 if (tp->t_fb_ptr) { 6343 struct tcp_rack *rack; 6344 struct rack_sendmap *rsm; 6345 6346 rack = (struct tcp_rack *)tp->t_fb_ptr; 6347 #ifdef TCP_BLACKBOX 6348 tcp_log_flowend(tp); 6349 #endif 6350 rsm = TAILQ_FIRST(&rack->r_ctl.rc_map); 6351 while (rsm) { 6352 TAILQ_REMOVE(&rack->r_ctl.rc_map, rsm, r_next); 6353 uma_zfree(rack_zone, rsm); 6354 rsm = TAILQ_FIRST(&rack->r_ctl.rc_map); 6355 } 6356 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free); 6357 while (rsm) { 6358 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_next); 6359 uma_zfree(rack_zone, rsm); 6360 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free); 6361 } 6362 rack->rc_free_cnt = 0; 6363 uma_zfree(rack_pcb_zone, tp->t_fb_ptr); 6364 tp->t_fb_ptr = NULL; 6365 } 6366 } 6367 6368 static void 6369 rack_set_state(struct tcpcb *tp, struct tcp_rack *rack) 6370 { 6371 switch (tp->t_state) { 6372 case TCPS_SYN_SENT: 6373 rack->r_state = TCPS_SYN_SENT; 6374 rack->r_substate = rack_do_syn_sent; 6375 break; 6376 case TCPS_SYN_RECEIVED: 6377 rack->r_state = TCPS_SYN_RECEIVED; 6378 rack->r_substate = rack_do_syn_recv; 6379 break; 6380 case TCPS_ESTABLISHED: 6381 rack->r_state = TCPS_ESTABLISHED; 6382 rack->r_substate = rack_do_established; 6383 break; 6384 case TCPS_CLOSE_WAIT: 6385 rack->r_state = TCPS_CLOSE_WAIT; 6386 rack->r_substate = rack_do_close_wait; 6387 break; 6388 case TCPS_FIN_WAIT_1: 6389 rack->r_state = TCPS_FIN_WAIT_1; 6390 rack->r_substate = rack_do_fin_wait_1; 6391 break; 6392 case TCPS_CLOSING: 6393 rack->r_state = TCPS_CLOSING; 6394 rack->r_substate = rack_do_closing; 6395 break; 6396 case TCPS_LAST_ACK: 6397 rack->r_state = TCPS_LAST_ACK; 6398 rack->r_substate = rack_do_lastack; 6399 break; 6400 case TCPS_FIN_WAIT_2: 6401 rack->r_state = TCPS_FIN_WAIT_2; 6402 rack->r_substate = rack_do_fin_wait_2; 6403 break; 6404 case TCPS_LISTEN: 6405 case TCPS_CLOSED: 6406 case TCPS_TIME_WAIT: 6407 default: 6408 #ifdef INVARIANTS 6409 panic("tcp tp:%p state:%d sees impossible state?", tp, tp->t_state); 6410 #endif 6411 break; 6412 }; 6413 } 6414 6415 6416 static void 6417 rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb) 6418 { 6419 /* 6420 * We received an ack, and then did not 6421 * call send or were bounced out due to the 6422 * hpts was running. Now a timer is up as well, is 6423 * it the right timer? 6424 */ 6425 struct rack_sendmap *rsm; 6426 int tmr_up; 6427 6428 tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK; 6429 if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT)) 6430 return; 6431 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); 6432 if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) && 6433 (tmr_up == PACE_TMR_RXT)) { 6434 /* Should be an RXT */ 6435 return; 6436 } 6437 if (rsm == NULL) { 6438 /* Nothing outstanding? */ 6439 if (tp->t_flags & TF_DELACK) { 6440 if (tmr_up == PACE_TMR_DELACK) 6441 /* We are supposed to have delayed ack up and we do */ 6442 return; 6443 } else if (sbavail(&tp->t_inpcb->inp_socket->so_snd) && (tmr_up == PACE_TMR_RXT)) { 6444 /* 6445 * if we hit enobufs then we would expect the possiblity 6446 * of nothing outstanding and the RXT up (and the hptsi timer). 6447 */ 6448 return; 6449 } else if (((tcp_always_keepalive || 6450 rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) && 6451 (tp->t_state <= TCPS_CLOSING)) && 6452 (tmr_up == PACE_TMR_KEEP) && 6453 (tp->snd_max == tp->snd_una)) { 6454 /* We should have keep alive up and we do */ 6455 return; 6456 } 6457 } 6458 if (rsm && (rsm->r_flags & RACK_SACK_PASSED)) { 6459 if ((tp->t_flags & TF_SENTFIN) && 6460 ((tp->snd_max - tp->snd_una) == 1) && 6461 (rsm->r_flags & RACK_HAS_FIN)) { 6462 /* needs to be a RXT */ 6463 if (tmr_up == PACE_TMR_RXT) 6464 return; 6465 } else if (tmr_up == PACE_TMR_RACK) 6466 return; 6467 } else if (SEQ_GT(tp->snd_max,tp->snd_una) && 6468 ((tmr_up == PACE_TMR_TLP) || 6469 (tmr_up == PACE_TMR_RXT))) { 6470 /* 6471 * Either a TLP or RXT is fine if no sack-passed 6472 * is in place and data is outstanding. 6473 */ 6474 return; 6475 } else if (tmr_up == PACE_TMR_DELACK) { 6476 /* 6477 * If the delayed ack was going to go off 6478 * before the rtx/tlp/rack timer were going to 6479 * expire, then that would be the timer in control. 6480 * Note we don't check the time here trusting the 6481 * code is correct. 6482 */ 6483 return; 6484 } 6485 /* 6486 * Ok the timer originally started is not what we want now. 6487 * We will force the hpts to be stopped if any, and restart 6488 * with the slot set to what was in the saved slot. 6489 */ 6490 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 6491 rack_start_hpts_timer(rack, tp, tcp_ts_getticks(), __LINE__, 0, 0, 0); 6492 } 6493 6494 static void 6495 rack_hpts_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so, 6496 struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos, 6497 int32_t nxt_pkt, struct timeval *tv) 6498 { 6499 int32_t thflags, retval, did_out = 0; 6500 int32_t way_out = 0; 6501 uint32_t cts; 6502 uint32_t tiwin; 6503 struct tcpopt to; 6504 struct tcp_rack *rack; 6505 struct rack_sendmap *rsm; 6506 int32_t prev_state = 0; 6507 6508 cts = tcp_tv_to_mssectick(tv); 6509 rack = (struct tcp_rack *)tp->t_fb_ptr; 6510 6511 kern_prefetch(rack, &prev_state); 6512 prev_state = 0; 6513 thflags = th->th_flags; 6514 /* 6515 * If this is either a state-changing packet or current state isn't 6516 * established, we require a read lock on tcbinfo. Otherwise, we 6517 * allow the tcbinfo to be in either locked or unlocked, as the 6518 * caller may have unnecessarily acquired a lock due to a race. 6519 */ 6520 if ((thflags & (TH_SYN | TH_FIN | TH_RST)) != 0 || 6521 tp->t_state != TCPS_ESTABLISHED) { 6522 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 6523 } 6524 INP_WLOCK_ASSERT(tp->t_inpcb); 6525 KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN", 6526 __func__)); 6527 KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT", 6528 __func__)); 6529 { 6530 union tcp_log_stackspecific log; 6531 6532 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 6533 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; 6534 log.u_bbr.ininput = rack->rc_inp->inp_in_input; 6535 TCP_LOG_EVENT(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0, 6536 tlen, &log, true); 6537 } 6538 /* 6539 * Segment received on connection. Reset idle time and keep-alive 6540 * timer. XXX: This should be done after segment validation to 6541 * ignore broken/spoofed segs. 6542 */ 6543 if (tp->t_idle_reduce && (tp->snd_max == tp->snd_una)) { 6544 #ifdef NETFLIX_CWV 6545 if ((tp->cwv_enabled) && 6546 ((tp->cwv_cwnd_valid == 0) && 6547 TCPS_HAVEESTABLISHED(tp->t_state) && 6548 (tp->snd_cwnd > tp->snd_cwv.init_cwnd))) { 6549 tcp_newcwv_nvp_closedown(tp); 6550 } else 6551 #endif 6552 if ((ticks - tp->t_rcvtime) >= tp->t_rxtcur) { 6553 counter_u64_add(rack_input_idle_reduces, 1); 6554 rack_cc_after_idle(tp, 6555 (rack->r_idle_reduce_largest ? 1 :0)); 6556 } 6557 } 6558 rack->r_ctl.rc_rcvtime = cts; 6559 tp->t_rcvtime = ticks; 6560 6561 #ifdef NETFLIX_CWV 6562 if (tp->cwv_enabled) { 6563 if ((tp->cwv_cwnd_valid == 0) && 6564 TCPS_HAVEESTABLISHED(tp->t_state) && 6565 (tp->snd_cwnd > tp->snd_cwv.init_cwnd)) 6566 tcp_newcwv_nvp_closedown(tp); 6567 } 6568 #endif 6569 /* 6570 * Unscale the window into a 32-bit value. For the SYN_SENT state 6571 * the scale is zero. 6572 */ 6573 tiwin = th->th_win << tp->snd_scale; 6574 #ifdef NETFLIX_STATS 6575 stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin); 6576 #endif 6577 /* 6578 * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move 6579 * this to occur after we've validated the segment. 6580 */ 6581 if (tp->t_flags & TF_ECN_PERMIT) { 6582 if (thflags & TH_CWR) 6583 tp->t_flags &= ~TF_ECN_SND_ECE; 6584 switch (iptos & IPTOS_ECN_MASK) { 6585 case IPTOS_ECN_CE: 6586 tp->t_flags |= TF_ECN_SND_ECE; 6587 TCPSTAT_INC(tcps_ecn_ce); 6588 break; 6589 case IPTOS_ECN_ECT0: 6590 TCPSTAT_INC(tcps_ecn_ect0); 6591 break; 6592 case IPTOS_ECN_ECT1: 6593 TCPSTAT_INC(tcps_ecn_ect1); 6594 break; 6595 } 6596 /* Congestion experienced. */ 6597 if (thflags & TH_ECE) { 6598 rack_cong_signal(tp, th, CC_ECN); 6599 } 6600 } 6601 /* 6602 * Parse options on any incoming segment. 6603 */ 6604 tcp_dooptions(&to, (u_char *)(th + 1), 6605 (th->th_off << 2) - sizeof(struct tcphdr), 6606 (thflags & TH_SYN) ? TO_SYN : 0); 6607 6608 /* 6609 * If echoed timestamp is later than the current time, fall back to 6610 * non RFC1323 RTT calculation. Normalize timestamp if syncookies 6611 * were used when this connection was established. 6612 */ 6613 if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) { 6614 to.to_tsecr -= tp->ts_offset; 6615 if (TSTMP_GT(to.to_tsecr, cts)) 6616 to.to_tsecr = 0; 6617 } 6618 /* 6619 * If its the first time in we need to take care of options and 6620 * verify we can do SACK for rack! 6621 */ 6622 if (rack->r_state == 0) { 6623 /* Should be init'd by rack_init() */ 6624 KASSERT(rack->rc_inp != NULL, 6625 ("%s: rack->rc_inp unexpectedly NULL", __func__)); 6626 if (rack->rc_inp == NULL) { 6627 rack->rc_inp = tp->t_inpcb; 6628 } 6629 6630 /* 6631 * Process options only when we get SYN/ACK back. The SYN 6632 * case for incoming connections is handled in tcp_syncache. 6633 * According to RFC1323 the window field in a SYN (i.e., a 6634 * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX 6635 * this is traditional behavior, may need to be cleaned up. 6636 */ 6637 rack->r_cpu = inp_to_cpuid(tp->t_inpcb); 6638 if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) { 6639 if ((to.to_flags & TOF_SCALE) && 6640 (tp->t_flags & TF_REQ_SCALE)) { 6641 tp->t_flags |= TF_RCVD_SCALE; 6642 tp->snd_scale = to.to_wscale; 6643 } 6644 /* 6645 * Initial send window. It will be updated with the 6646 * next incoming segment to the scaled value. 6647 */ 6648 tp->snd_wnd = th->th_win; 6649 if (to.to_flags & TOF_TS) { 6650 tp->t_flags |= TF_RCVD_TSTMP; 6651 tp->ts_recent = to.to_tsval; 6652 tp->ts_recent_age = cts; 6653 } 6654 if (to.to_flags & TOF_MSS) 6655 tcp_mss(tp, to.to_mss); 6656 if ((tp->t_flags & TF_SACK_PERMIT) && 6657 (to.to_flags & TOF_SACKPERM) == 0) 6658 tp->t_flags &= ~TF_SACK_PERMIT; 6659 if (IS_FASTOPEN(tp->t_flags)) { 6660 if (to.to_flags & TOF_FASTOPEN) { 6661 uint16_t mss; 6662 6663 if (to.to_flags & TOF_MSS) 6664 mss = to.to_mss; 6665 else 6666 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0) 6667 mss = TCP6_MSS; 6668 else 6669 mss = TCP_MSS; 6670 tcp_fastopen_update_cache(tp, mss, 6671 to.to_tfo_len, to.to_tfo_cookie); 6672 } else 6673 tcp_fastopen_disable_path(tp); 6674 } 6675 } 6676 /* 6677 * At this point we are at the initial call. Here we decide 6678 * if we are doing RACK or not. We do this by seeing if 6679 * TF_SACK_PERMIT is set, if not rack is *not* possible and 6680 * we switch to the default code. 6681 */ 6682 if ((tp->t_flags & TF_SACK_PERMIT) == 0) { 6683 tcp_switch_back_to_default(tp); 6684 (*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen, 6685 tlen, iptos); 6686 return; 6687 } 6688 /* Set the flag */ 6689 rack->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0; 6690 tcp_set_hpts(tp->t_inpcb); 6691 rack_stop_all_timers(tp); 6692 sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack); 6693 } 6694 /* 6695 * This is the one exception case where we set the rack state 6696 * always. All other times (timers etc) we must have a rack-state 6697 * set (so we assure we have done the checks above for SACK). 6698 */ 6699 if (rack->r_state != tp->t_state) 6700 rack_set_state(tp, rack); 6701 if (SEQ_GT(th->th_ack, tp->snd_una) && (rsm = TAILQ_FIRST(&rack->r_ctl.rc_map)) != NULL) 6702 kern_prefetch(rsm, &prev_state); 6703 prev_state = rack->r_state; 6704 rack->r_ctl.rc_tlp_send_cnt = 0; 6705 rack_clear_rate_sample(rack); 6706 retval = (*rack->r_substate) (m, th, so, 6707 tp, &to, drop_hdrlen, 6708 tlen, tiwin, thflags, nxt_pkt); 6709 #ifdef INVARIANTS 6710 if ((retval == 0) && 6711 (tp->t_inpcb == NULL)) { 6712 panic("retval:%d tp:%p t_inpcb:NULL state:%d", 6713 retval, tp, prev_state); 6714 } 6715 #endif 6716 if (retval == 0) { 6717 /* 6718 * If retval is 1 the tcb is unlocked and most likely the tp 6719 * is gone. 6720 */ 6721 INP_WLOCK_ASSERT(tp->t_inpcb); 6722 tcp_rack_xmit_timer_commit(rack, tp); 6723 if (((tp->snd_max - tp->snd_una) > tp->snd_wnd) && 6724 (rack->rc_in_persist == 0)){ 6725 /* 6726 * The peer shrunk its window on us to the point 6727 * where we have sent too much. The only thing 6728 * we can do here is stop any timers and 6729 * enter persist. We most likely lost the last 6730 * bytes we sent but oh well, we will have to 6731 * retransmit them after the peer is caught up. 6732 */ 6733 if (rack->rc_inp->inp_in_hpts) 6734 tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT); 6735 rack_timer_cancel(tp, rack, cts, __LINE__); 6736 rack_enter_persist(tp, rack, cts); 6737 rack_start_hpts_timer(rack, tp, tcp_ts_getticks(), __LINE__, 0, 0, 0); 6738 way_out = 3; 6739 goto done_with_input; 6740 } 6741 if (nxt_pkt == 0) { 6742 if (rack->r_wanted_output != 0) { 6743 did_out = 1; 6744 (void)tp->t_fb->tfb_tcp_output(tp); 6745 } 6746 rack_start_hpts_timer(rack, tp, cts, __LINE__, 0, 0, 0); 6747 } 6748 if (((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) && 6749 (SEQ_GT(tp->snd_max, tp->snd_una) || 6750 (tp->t_flags & TF_DELACK) || 6751 ((tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) && 6752 (tp->t_state <= TCPS_CLOSING)))) { 6753 /* We could not send (probably in the hpts but stopped the timer earlier)? */ 6754 if ((tp->snd_max == tp->snd_una) && 6755 ((tp->t_flags & TF_DELACK) == 0) && 6756 (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) { 6757 /* keep alive not needed if we are hptsi output yet */ 6758 ; 6759 } else { 6760 if (rack->rc_inp->inp_in_hpts) 6761 tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT); 6762 rack_start_hpts_timer(rack, tp, tcp_ts_getticks(), __LINE__, 0, 0, 0); 6763 } 6764 way_out = 1; 6765 } else { 6766 /* Do we have the correct timer running? */ 6767 rack_timer_audit(tp, rack, &so->so_snd); 6768 way_out = 2; 6769 } 6770 done_with_input: 6771 rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out); 6772 if (did_out) 6773 rack->r_wanted_output = 0; 6774 #ifdef INVARIANTS 6775 if (tp->t_inpcb == NULL) { 6776 panic("OP:%d retval:%d tp:%p t_inpcb:NULL state:%d", 6777 did_out, 6778 retval, tp, prev_state); 6779 } 6780 #endif 6781 INP_WUNLOCK(tp->t_inpcb); 6782 } 6783 } 6784 6785 void 6786 rack_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so, 6787 struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos) 6788 { 6789 struct timeval tv; 6790 #ifdef RSS 6791 struct tcp_function_block *tfb; 6792 struct tcp_rack *rack; 6793 struct epoch_tracker et; 6794 6795 rack = (struct tcp_rack *)tp->t_fb_ptr; 6796 if (rack->r_state == 0) { 6797 /* 6798 * Initial input (ACK to SYN-ACK etc)lets go ahead and get 6799 * it processed 6800 */ 6801 INP_INFO_RLOCK_ET(&V_tcbinfo, et); 6802 tcp_get_usecs(&tv); 6803 rack_hpts_do_segment(m, th, so, tp, drop_hdrlen, 6804 tlen, iptos, 0, &tv); 6805 INP_INFO_RUNLOCK_ET(&V_tcbinfo, et); 6806 return; 6807 } 6808 tcp_queue_to_input(tp, m, th, tlen, drop_hdrlen, iptos); 6809 INP_WUNLOCK(tp->t_inpcb); 6810 #else 6811 tcp_get_usecs(&tv); 6812 rack_hpts_do_segment(m, th, so, tp, drop_hdrlen, 6813 tlen, iptos, 0, &tv); 6814 #endif 6815 } 6816 6817 struct rack_sendmap * 6818 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused) 6819 { 6820 struct rack_sendmap *rsm = NULL; 6821 int32_t idx; 6822 uint32_t srtt_cur, srtt = 0, thresh = 0, ts_low = 0; 6823 6824 /* Return the next guy to be re-transmitted */ 6825 if (TAILQ_EMPTY(&rack->r_ctl.rc_map)) { 6826 return (NULL); 6827 } 6828 if (tp->t_flags & TF_SENTFIN) { 6829 /* retran the end FIN? */ 6830 return (NULL); 6831 } 6832 /* ok lets look at this one */ 6833 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); 6834 if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) { 6835 goto check_it; 6836 } 6837 rsm = rack_find_lowest_rsm(rack); 6838 if (rsm == NULL) { 6839 return (NULL); 6840 } 6841 check_it: 6842 srtt_cur = tp->t_srtt >> TCP_RTT_SHIFT; 6843 srtt = TICKS_2_MSEC(srtt_cur); 6844 if (rack->rc_rack_rtt && (srtt > rack->rc_rack_rtt)) 6845 srtt = rack->rc_rack_rtt; 6846 if (rsm->r_flags & RACK_ACKED) { 6847 return (NULL); 6848 } 6849 if ((rsm->r_flags & RACK_SACK_PASSED) == 0) { 6850 /* Its not yet ready */ 6851 return (NULL); 6852 } 6853 idx = rsm->r_rtr_cnt - 1; 6854 ts_low = rsm->r_tim_lastsent[idx]; 6855 thresh = rack_calc_thresh_rack(rack, srtt, tsused); 6856 if (tsused <= ts_low) { 6857 return (NULL); 6858 } 6859 if ((tsused - ts_low) >= thresh) { 6860 return (rsm); 6861 } 6862 return (NULL); 6863 } 6864 6865 static int 6866 rack_output(struct tcpcb *tp) 6867 { 6868 struct socket *so; 6869 uint32_t recwin, sendwin; 6870 uint32_t sb_offset; 6871 int32_t len, flags, error = 0; 6872 struct mbuf *m; 6873 struct mbuf *mb; 6874 uint32_t if_hw_tsomaxsegcount = 0; 6875 uint32_t if_hw_tsomaxsegsize; 6876 long tot_len_this_send = 0; 6877 struct ip *ip = NULL; 6878 #ifdef TCPDEBUG 6879 struct ipovly *ipov = NULL; 6880 #endif 6881 struct udphdr *udp = NULL; 6882 struct tcp_rack *rack; 6883 struct tcphdr *th; 6884 uint8_t pass = 0; 6885 uint8_t wanted_cookie = 0; 6886 u_char opt[TCP_MAXOLEN]; 6887 unsigned ipoptlen, optlen, hdrlen, ulen=0; 6888 uint32_t rack_seq; 6889 6890 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 6891 unsigned ipsec_optlen = 0; 6892 6893 #endif 6894 int32_t idle, sendalot; 6895 int32_t sub_from_prr = 0; 6896 volatile int32_t sack_rxmit; 6897 struct rack_sendmap *rsm = NULL; 6898 int32_t tso, mtu, would_have_fin = 0; 6899 struct tcpopt to; 6900 int32_t slot = 0; 6901 uint32_t cts; 6902 uint8_t hpts_calling, doing_tlp = 0; 6903 int32_t do_a_prefetch; 6904 int32_t prefetch_rsm = 0; 6905 int32_t prefetch_so_done = 0; 6906 struct tcp_log_buffer *lgb = NULL; 6907 struct inpcb *inp; 6908 struct sockbuf *sb; 6909 #ifdef INET6 6910 struct ip6_hdr *ip6 = NULL; 6911 int32_t isipv6; 6912 #endif 6913 /* setup and take the cache hits here */ 6914 rack = (struct tcp_rack *)tp->t_fb_ptr; 6915 inp = rack->rc_inp; 6916 so = inp->inp_socket; 6917 sb = &so->so_snd; 6918 kern_prefetch(sb, &do_a_prefetch); 6919 do_a_prefetch = 1; 6920 6921 INP_WLOCK_ASSERT(inp); 6922 #ifdef TCP_OFFLOAD 6923 if (tp->t_flags & TF_TOE) 6924 return (tcp_offload_output(tp)); 6925 #endif 6926 6927 /* 6928 * For TFO connections in SYN_RECEIVED, only allow the initial 6929 * SYN|ACK and those sent by the retransmit timer. 6930 */ 6931 if (IS_FASTOPEN(tp->t_flags) && 6932 (tp->t_state == TCPS_SYN_RECEIVED) && 6933 SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN|ACK sent */ 6934 (rack->r_ctl.rc_resend == NULL)) /* not a retransmit */ 6935 return (0); 6936 #ifdef INET6 6937 if (rack->r_state) { 6938 /* Use the cache line loaded if possible */ 6939 isipv6 = rack->r_is_v6; 6940 } else { 6941 isipv6 = (inp->inp_vflag & INP_IPV6) != 0; 6942 } 6943 #endif 6944 cts = tcp_ts_getticks(); 6945 if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) && 6946 inp->inp_in_hpts) { 6947 /* 6948 * We are on the hpts for some timer but not hptsi output. 6949 * Remove from the hpts unconditionally. 6950 */ 6951 rack_timer_cancel(tp, rack, cts, __LINE__); 6952 } 6953 /* Mark that we have called rack_output(). */ 6954 if ((rack->r_timer_override) || 6955 (tp->t_flags & TF_FORCEDATA) || 6956 (tp->t_state < TCPS_ESTABLISHED)) { 6957 if (tp->t_inpcb->inp_in_hpts) 6958 tcp_hpts_remove(tp->t_inpcb, HPTS_REMOVE_OUTPUT); 6959 } else if (tp->t_inpcb->inp_in_hpts) { 6960 /* 6961 * On the hpts you can't pass even if ACKNOW is on, we will 6962 * when the hpts fires. 6963 */ 6964 counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1); 6965 return (0); 6966 } 6967 hpts_calling = inp->inp_hpts_calls; 6968 inp->inp_hpts_calls = 0; 6969 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) { 6970 if (rack_process_timers(tp, rack, cts, hpts_calling)) { 6971 counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1); 6972 return (0); 6973 } 6974 } 6975 rack->r_wanted_output = 0; 6976 rack->r_timer_override = 0; 6977 /* 6978 * Determine length of data that should be transmitted, and flags 6979 * that will be used. If there is some data or critical controls 6980 * (SYN, RST) to send, then transmit; otherwise, investigate 6981 * further. 6982 */ 6983 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una); 6984 #ifdef NETFLIX_CWV 6985 if (tp->cwv_enabled) { 6986 if ((tp->cwv_cwnd_valid == 0) && 6987 TCPS_HAVEESTABLISHED(tp->t_state) && 6988 (tp->snd_cwnd > tp->snd_cwv.init_cwnd)) 6989 tcp_newcwv_nvp_closedown(tp); 6990 } else 6991 #endif 6992 if (tp->t_idle_reduce) { 6993 if (idle && ((ticks - tp->t_rcvtime) >= tp->t_rxtcur)) 6994 rack_cc_after_idle(tp, 6995 (rack->r_idle_reduce_largest ? 1 :0)); 6996 } 6997 tp->t_flags &= ~TF_LASTIDLE; 6998 if (idle) { 6999 if (tp->t_flags & TF_MORETOCOME) { 7000 tp->t_flags |= TF_LASTIDLE; 7001 idle = 0; 7002 } 7003 } 7004 again: 7005 /* 7006 * If we've recently taken a timeout, snd_max will be greater than 7007 * snd_nxt. There may be SACK information that allows us to avoid 7008 * resending already delivered data. Adjust snd_nxt accordingly. 7009 */ 7010 sendalot = 0; 7011 cts = tcp_ts_getticks(); 7012 tso = 0; 7013 mtu = 0; 7014 sb_offset = tp->snd_max - tp->snd_una; 7015 sendwin = min(tp->snd_wnd, tp->snd_cwnd); 7016 7017 flags = tcp_outflags[tp->t_state]; 7018 /* 7019 * Send any SACK-generated retransmissions. If we're explicitly 7020 * trying to send out new data (when sendalot is 1), bypass this 7021 * function. If we retransmit in fast recovery mode, decrement 7022 * snd_cwnd, since we're replacing a (future) new transmission with 7023 * a retransmission now, and we previously incremented snd_cwnd in 7024 * tcp_input(). 7025 */ 7026 /* 7027 * Still in sack recovery , reset rxmit flag to zero. 7028 */ 7029 while (rack->rc_free_cnt < rack_free_cache) { 7030 rsm = rack_alloc(rack); 7031 if (rsm == NULL) { 7032 if (inp->inp_hpts_calls) 7033 /* Retry in a ms */ 7034 slot = 1; 7035 goto just_return_nolock; 7036 } 7037 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_next); 7038 rack->rc_free_cnt++; 7039 rsm = NULL; 7040 } 7041 if (inp->inp_hpts_calls) 7042 inp->inp_hpts_calls = 0; 7043 sack_rxmit = 0; 7044 len = 0; 7045 rsm = NULL; 7046 if (flags & TH_RST) { 7047 SOCKBUF_LOCK(sb); 7048 goto send; 7049 } 7050 if (rack->r_ctl.rc_tlpsend) { 7051 /* Tail loss probe */ 7052 long cwin; 7053 long tlen; 7054 7055 doing_tlp = 1; 7056 rsm = rack->r_ctl.rc_tlpsend; 7057 rack->r_ctl.rc_tlpsend = NULL; 7058 sack_rxmit = 1; 7059 tlen = rsm->r_end - rsm->r_start; 7060 if (tlen > tp->t_maxseg) 7061 tlen = tp->t_maxseg; 7062 #ifdef INVARIANTS 7063 if (SEQ_GT(tp->snd_una, rsm->r_start)) { 7064 panic("tp:%p rack:%p snd_una:%u rsm:%p r_start:%u", 7065 tp, rack, tp->snd_una, rsm, rsm->r_start); 7066 } 7067 #endif 7068 sb_offset = rsm->r_start - tp->snd_una; 7069 cwin = min(tp->snd_wnd, tlen); 7070 len = cwin; 7071 } else if (rack->r_ctl.rc_resend) { 7072 /* Retransmit timer */ 7073 rsm = rack->r_ctl.rc_resend; 7074 rack->r_ctl.rc_resend = NULL; 7075 len = rsm->r_end - rsm->r_start; 7076 sack_rxmit = 1; 7077 sendalot = 0; 7078 sb_offset = rsm->r_start - tp->snd_una; 7079 if (len >= tp->t_maxseg) { 7080 len = tp->t_maxseg; 7081 } 7082 KASSERT(sb_offset >= 0, ("%s: sack block to the left of una : %d", 7083 __func__, sb_offset)); 7084 } else if ((rack->rc_in_persist == 0) && 7085 ((rsm = tcp_rack_output(tp, rack, cts)) != NULL)) { 7086 long tlen; 7087 7088 if ((!IN_RECOVERY(tp->t_flags)) && 7089 ((tp->t_flags & (TF_WASFRECOVERY | TF_WASCRECOVERY)) == 0)) { 7090 /* Enter recovery if not induced by a time-out */ 7091 rack->r_ctl.rc_rsm_start = rsm->r_start; 7092 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd; 7093 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh; 7094 rack_cong_signal(tp, NULL, CC_NDUPACK); 7095 /* 7096 * When we enter recovery we need to assure we send 7097 * one packet. 7098 */ 7099 rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; 7100 } 7101 #ifdef INVARIANTS 7102 if (SEQ_LT(rsm->r_start, tp->snd_una)) { 7103 panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n", 7104 tp, rack, rsm, rsm->r_start, tp->snd_una); 7105 } 7106 #endif 7107 tlen = rsm->r_end - rsm->r_start; 7108 sb_offset = rsm->r_start - tp->snd_una; 7109 if (tlen > rack->r_ctl.rc_prr_sndcnt) { 7110 len = rack->r_ctl.rc_prr_sndcnt; 7111 } else { 7112 len = tlen; 7113 } 7114 if (len >= tp->t_maxseg) { 7115 sendalot = 1; 7116 len = tp->t_maxseg; 7117 } else { 7118 sendalot = 0; 7119 if ((rack->rc_timer_up == 0) && 7120 (len < tlen)) { 7121 /* 7122 * If its not a timer don't send a partial 7123 * segment. 7124 */ 7125 len = 0; 7126 goto just_return_nolock; 7127 } 7128 } 7129 KASSERT(sb_offset >= 0, ("%s: sack block to the left of una : %d", 7130 __func__, sb_offset)); 7131 if (len > 0) { 7132 sub_from_prr = 1; 7133 sack_rxmit = 1; 7134 TCPSTAT_INC(tcps_sack_rexmits); 7135 TCPSTAT_ADD(tcps_sack_rexmit_bytes, 7136 min(len, tp->t_maxseg)); 7137 counter_u64_add(rack_rtm_prr_retran, 1); 7138 } 7139 } 7140 if (rsm && (rsm->r_flags & RACK_HAS_FIN)) { 7141 /* we are retransmitting the fin */ 7142 len--; 7143 if (len) { 7144 /* 7145 * When retransmitting data do *not* include the 7146 * FIN. This could happen from a TLP probe. 7147 */ 7148 flags &= ~TH_FIN; 7149 } 7150 } 7151 #ifdef INVARIANTS 7152 /* For debugging */ 7153 rack->r_ctl.rc_rsm_at_retran = rsm; 7154 #endif 7155 /* 7156 * Get standard flags, and add SYN or FIN if requested by 'hidden' 7157 * state flags. 7158 */ 7159 if (tp->t_flags & TF_NEEDFIN) 7160 flags |= TH_FIN; 7161 if (tp->t_flags & TF_NEEDSYN) 7162 flags |= TH_SYN; 7163 if ((sack_rxmit == 0) && (prefetch_rsm == 0)) { 7164 void *end_rsm; 7165 end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext); 7166 if (end_rsm) 7167 kern_prefetch(end_rsm, &prefetch_rsm); 7168 prefetch_rsm = 1; 7169 } 7170 SOCKBUF_LOCK(sb); 7171 /* 7172 * If in persist timeout with window of 0, send 1 byte. Otherwise, 7173 * if window is small but nonzero and time TF_SENTFIN expired, we 7174 * will send what we can and go to transmit state. 7175 */ 7176 if (tp->t_flags & TF_FORCEDATA) { 7177 if (sendwin == 0) { 7178 /* 7179 * If we still have some data to send, then clear 7180 * the FIN bit. Usually this would happen below 7181 * when it realizes that we aren't sending all the 7182 * data. However, if we have exactly 1 byte of 7183 * unsent data, then it won't clear the FIN bit 7184 * below, and if we are in persist state, we wind up 7185 * sending the packet without recording that we sent 7186 * the FIN bit. 7187 * 7188 * We can't just blindly clear the FIN bit, because 7189 * if we don't have any more data to send then the 7190 * probe will be the FIN itself. 7191 */ 7192 if (sb_offset < sbused(sb)) 7193 flags &= ~TH_FIN; 7194 sendwin = 1; 7195 } else { 7196 if (rack->rc_in_persist) 7197 rack_exit_persist(tp, rack); 7198 /* 7199 * If we are dropping persist mode then we need to 7200 * correct snd_nxt/snd_max and off. 7201 */ 7202 tp->snd_nxt = tp->snd_max; 7203 sb_offset = tp->snd_nxt - tp->snd_una; 7204 } 7205 } 7206 /* 7207 * If snd_nxt == snd_max and we have transmitted a FIN, the 7208 * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a 7209 * negative length. This can also occur when TCP opens up its 7210 * congestion window while receiving additional duplicate acks after 7211 * fast-retransmit because TCP will reset snd_nxt to snd_max after 7212 * the fast-retransmit. 7213 * 7214 * In the normal retransmit-FIN-only case, however, snd_nxt will be 7215 * set to snd_una, the sb_offset will be 0, and the length may wind 7216 * up 0. 7217 * 7218 * If sack_rxmit is true we are retransmitting from the scoreboard 7219 * in which case len is already set. 7220 */ 7221 if (sack_rxmit == 0) { 7222 uint32_t avail; 7223 7224 avail = sbavail(sb); 7225 if (SEQ_GT(tp->snd_nxt, tp->snd_una) && avail) 7226 sb_offset = tp->snd_nxt - tp->snd_una; 7227 else 7228 sb_offset = 0; 7229 if (IN_RECOVERY(tp->t_flags) == 0) { 7230 if (rack->r_ctl.rc_tlp_new_data) { 7231 /* TLP is forcing out new data */ 7232 if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) { 7233 rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset); 7234 } 7235 if (rack->r_ctl.rc_tlp_new_data > tp->snd_wnd) 7236 len = tp->snd_wnd; 7237 else 7238 len = rack->r_ctl.rc_tlp_new_data; 7239 rack->r_ctl.rc_tlp_new_data = 0; 7240 doing_tlp = 1; 7241 } else { 7242 if (sendwin > avail) { 7243 /* use the available */ 7244 if (avail > sb_offset) { 7245 len = (int32_t)(avail - sb_offset); 7246 } else { 7247 len = 0; 7248 } 7249 } else { 7250 if (sendwin > sb_offset) { 7251 len = (int32_t)(sendwin - sb_offset); 7252 } else { 7253 len = 0; 7254 } 7255 } 7256 } 7257 } else { 7258 uint32_t outstanding; 7259 7260 /* 7261 * We are inside of a SACK recovery episode and are 7262 * sending new data, having retransmitted all the 7263 * data possible so far in the scoreboard. 7264 */ 7265 outstanding = tp->snd_max - tp->snd_una; 7266 if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) 7267 len = 0; 7268 else if (avail > sb_offset) 7269 len = avail - sb_offset; 7270 else 7271 len = 0; 7272 if (len > 0) { 7273 if (len > rack->r_ctl.rc_prr_sndcnt) 7274 len = rack->r_ctl.rc_prr_sndcnt; 7275 7276 if (len > 0) { 7277 sub_from_prr = 1; 7278 counter_u64_add(rack_rtm_prr_newdata, 1); 7279 } 7280 } 7281 if (len > tp->t_maxseg) { 7282 /* 7283 * We should never send more than a MSS when 7284 * retransmitting or sending new data in prr 7285 * mode unless the override flag is on. Most 7286 * likely the PRR algorithm is not going to 7287 * let us send a lot as well :-) 7288 */ 7289 if (rack->r_ctl.rc_prr_sendalot == 0) 7290 len = tp->t_maxseg; 7291 } else if (len < tp->t_maxseg) { 7292 /* 7293 * Do we send any? The idea here is if the 7294 * send empty's the socket buffer we want to 7295 * do it. However if not then lets just wait 7296 * for our prr_sndcnt to get bigger. 7297 */ 7298 long leftinsb; 7299 7300 leftinsb = sbavail(sb) - sb_offset; 7301 if (leftinsb > len) { 7302 /* This send does not empty the sb */ 7303 len = 0; 7304 } 7305 } 7306 } 7307 } 7308 if (prefetch_so_done == 0) { 7309 kern_prefetch(so, &prefetch_so_done); 7310 prefetch_so_done = 1; 7311 } 7312 /* 7313 * Lop off SYN bit if it has already been sent. However, if this is 7314 * SYN-SENT state and if segment contains data and if we don't know 7315 * that foreign host supports TAO, suppress sending segment. 7316 */ 7317 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una) && 7318 ((sack_rxmit == 0) && (tp->t_rxtshift == 0))) { 7319 if (tp->t_state != TCPS_SYN_RECEIVED) 7320 flags &= ~TH_SYN; 7321 /* 7322 * When sending additional segments following a TFO SYN|ACK, 7323 * do not include the SYN bit. 7324 */ 7325 if (IS_FASTOPEN(tp->t_flags) && 7326 (tp->t_state == TCPS_SYN_RECEIVED)) 7327 flags &= ~TH_SYN; 7328 sb_offset--, len++; 7329 } 7330 /* 7331 * Be careful not to send data and/or FIN on SYN segments. This 7332 * measure is needed to prevent interoperability problems with not 7333 * fully conformant TCP implementations. 7334 */ 7335 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) { 7336 len = 0; 7337 flags &= ~TH_FIN; 7338 } 7339 /* 7340 * On TFO sockets, ensure no data is sent in the following cases: 7341 * 7342 * - When retransmitting SYN|ACK on a passively-created socket 7343 * 7344 * - When retransmitting SYN on an actively created socket 7345 * 7346 * - When sending a zero-length cookie (cookie request) on an 7347 * actively created socket 7348 * 7349 * - When the socket is in the CLOSED state (RST is being sent) 7350 */ 7351 if (IS_FASTOPEN(tp->t_flags) && 7352 (((flags & TH_SYN) && (tp->t_rxtshift > 0)) || 7353 ((tp->t_state == TCPS_SYN_SENT) && 7354 (tp->t_tfo_client_cookie_len == 0)) || 7355 (flags & TH_RST))) 7356 len = 0; 7357 /* Without fast-open there should never be data sent on a SYN */ 7358 if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags))) 7359 len = 0; 7360 if (len <= 0) { 7361 /* 7362 * If FIN has been sent but not acked, but we haven't been 7363 * called to retransmit, len will be < 0. Otherwise, window 7364 * shrank after we sent into it. If window shrank to 0, 7365 * cancel pending retransmit, pull snd_nxt back to (closed) 7366 * window, and set the persist timer if it isn't already 7367 * going. If the window didn't close completely, just wait 7368 * for an ACK. 7369 * 7370 * We also do a general check here to ensure that we will 7371 * set the persist timer when we have data to send, but a 7372 * 0-byte window. This makes sure the persist timer is set 7373 * even if the packet hits one of the "goto send" lines 7374 * below. 7375 */ 7376 len = 0; 7377 if ((tp->snd_wnd == 0) && 7378 (TCPS_HAVEESTABLISHED(tp->t_state)) && 7379 (sb_offset < (int)sbavail(sb))) { 7380 tp->snd_nxt = tp->snd_una; 7381 rack_enter_persist(tp, rack, cts); 7382 } 7383 } 7384 /* len will be >= 0 after this point. */ 7385 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 7386 tcp_sndbuf_autoscale(tp, so, sendwin); 7387 /* 7388 * Decide if we can use TCP Segmentation Offloading (if supported by 7389 * hardware). 7390 * 7391 * TSO may only be used if we are in a pure bulk sending state. The 7392 * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP 7393 * options prevent using TSO. With TSO the TCP header is the same 7394 * (except for the sequence number) for all generated packets. This 7395 * makes it impossible to transmit any options which vary per 7396 * generated segment or packet. 7397 * 7398 * IPv4 handling has a clear separation of ip options and ip header 7399 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does 7400 * the right thing below to provide length of just ip options and thus 7401 * checking for ipoptlen is enough to decide if ip options are present. 7402 */ 7403 7404 #ifdef INET6 7405 if (isipv6) 7406 ipoptlen = ip6_optlen(tp->t_inpcb); 7407 else 7408 #endif 7409 if (tp->t_inpcb->inp_options) 7410 ipoptlen = tp->t_inpcb->inp_options->m_len - 7411 offsetof(struct ipoption, ipopt_list); 7412 else 7413 ipoptlen = 0; 7414 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 7415 /* 7416 * Pre-calculate here as we save another lookup into the darknesses 7417 * of IPsec that way and can actually decide if TSO is ok. 7418 */ 7419 #ifdef INET6 7420 if (isipv6 && IPSEC_ENABLED(ipv6)) 7421 ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb); 7422 #ifdef INET 7423 else 7424 #endif 7425 #endif /* INET6 */ 7426 #ifdef INET 7427 if (IPSEC_ENABLED(ipv4)) 7428 ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb); 7429 #endif /* INET */ 7430 #endif 7431 7432 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 7433 ipoptlen += ipsec_optlen; 7434 #endif 7435 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg && 7436 (tp->t_port == 0) && 7437 ((tp->t_flags & TF_SIGNATURE) == 0) && 7438 tp->rcv_numsacks == 0 && sack_rxmit == 0 && 7439 ipoptlen == 0) 7440 tso = 1; 7441 { 7442 uint32_t outstanding; 7443 7444 outstanding = tp->snd_max - tp->snd_una; 7445 if (tp->t_flags & TF_SENTFIN) { 7446 /* 7447 * If we sent a fin, snd_max is 1 higher than 7448 * snd_una 7449 */ 7450 outstanding--; 7451 } 7452 if (outstanding > 0) { 7453 /* 7454 * This is sub-optimal. We only send a stand alone 7455 * FIN on its own segment. 7456 */ 7457 if (flags & TH_FIN) { 7458 flags &= ~TH_FIN; 7459 would_have_fin = 1; 7460 } 7461 } else if (sack_rxmit) { 7462 if ((rsm->r_flags & RACK_HAS_FIN) == 0) 7463 flags &= ~TH_FIN; 7464 } else { 7465 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + 7466 sbused(sb))) 7467 flags &= ~TH_FIN; 7468 } 7469 } 7470 recwin = sbspace(&so->so_rcv); 7471 7472 /* 7473 * Sender silly window avoidance. We transmit under the following 7474 * conditions when len is non-zero: 7475 * 7476 * - We have a full segment (or more with TSO) - This is the last 7477 * buffer in a write()/send() and we are either idle or running 7478 * NODELAY - we've timed out (e.g. persist timer) - we have more 7479 * then 1/2 the maximum send window's worth of data (receiver may be 7480 * limited the window size) - we need to retransmit 7481 */ 7482 if (len) { 7483 if (len >= tp->t_maxseg) { 7484 pass = 1; 7485 goto send; 7486 } 7487 /* 7488 * NOTE! on localhost connections an 'ack' from the remote 7489 * end may occur synchronously with the output and cause us 7490 * to flush a buffer queued with moretocome. XXX 7491 * 7492 */ 7493 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */ 7494 (idle || (tp->t_flags & TF_NODELAY)) && 7495 ((uint32_t)len + (uint32_t)sb_offset >= sbavail(&so->so_snd)) && 7496 (tp->t_flags & TF_NOPUSH) == 0) { 7497 pass = 2; 7498 goto send; 7499 } 7500 if (tp->t_flags & TF_FORCEDATA) { /* typ. timeout case */ 7501 pass = 3; 7502 goto send; 7503 } 7504 if ((tp->snd_una == tp->snd_max) && len) { /* Nothing outstanding */ 7505 goto send; 7506 } 7507 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) { 7508 pass = 4; 7509 goto send; 7510 } 7511 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { /* retransmit case */ 7512 pass = 5; 7513 goto send; 7514 } 7515 if (sack_rxmit) { 7516 pass = 6; 7517 goto send; 7518 } 7519 } 7520 /* 7521 * Sending of standalone window updates. 7522 * 7523 * Window updates are important when we close our window due to a 7524 * full socket buffer and are opening it again after the application 7525 * reads data from it. Once the window has opened again and the 7526 * remote end starts to send again the ACK clock takes over and 7527 * provides the most current window information. 7528 * 7529 * We must avoid the silly window syndrome whereas every read from 7530 * the receive buffer, no matter how small, causes a window update 7531 * to be sent. We also should avoid sending a flurry of window 7532 * updates when the socket buffer had queued a lot of data and the 7533 * application is doing small reads. 7534 * 7535 * Prevent a flurry of pointless window updates by only sending an 7536 * update when we can increase the advertized window by more than 7537 * 1/4th of the socket buffer capacity. When the buffer is getting 7538 * full or is very small be more aggressive and send an update 7539 * whenever we can increase by two mss sized segments. In all other 7540 * situations the ACK's to new incoming data will carry further 7541 * window increases. 7542 * 7543 * Don't send an independent window update if a delayed ACK is 7544 * pending (it will get piggy-backed on it) or the remote side 7545 * already has done a half-close and won't send more data. Skip 7546 * this if the connection is in T/TCP half-open state. 7547 */ 7548 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) && 7549 !(tp->t_flags & TF_DELACK) && 7550 !TCPS_HAVERCVDFIN(tp->t_state)) { 7551 /* 7552 * "adv" is the amount we could increase the window, taking 7553 * into account that we are limited by TCP_MAXWIN << 7554 * tp->rcv_scale. 7555 */ 7556 int32_t adv; 7557 int oldwin; 7558 7559 adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale); 7560 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) { 7561 oldwin = (tp->rcv_adv - tp->rcv_nxt); 7562 adv -= oldwin; 7563 } else 7564 oldwin = 0; 7565 7566 /* 7567 * If the new window size ends up being the same as the old 7568 * size when it is scaled, then don't force a window update. 7569 */ 7570 if (oldwin >> tp->rcv_scale == (adv + oldwin) >> tp->rcv_scale) 7571 goto dontupdate; 7572 7573 if (adv >= (int32_t)(2 * tp->t_maxseg) && 7574 (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) || 7575 recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) || 7576 so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg)) { 7577 pass = 7; 7578 goto send; 7579 } 7580 if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) 7581 goto send; 7582 } 7583 dontupdate: 7584 7585 /* 7586 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW 7587 * is also a catch-all for the retransmit timer timeout case. 7588 */ 7589 if (tp->t_flags & TF_ACKNOW) { 7590 pass = 8; 7591 goto send; 7592 } 7593 if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) { 7594 pass = 9; 7595 goto send; 7596 } 7597 if (SEQ_GT(tp->snd_up, tp->snd_una)) { 7598 pass = 10; 7599 goto send; 7600 } 7601 /* 7602 * If our state indicates that FIN should be sent and we have not 7603 * yet done so, then we need to send. 7604 */ 7605 if (flags & TH_FIN) { 7606 if ((tp->t_flags & TF_SENTFIN) || 7607 (((tp->t_flags & TF_SENTFIN) == 0) && 7608 (tp->snd_nxt == tp->snd_una))) { 7609 pass = 11; 7610 goto send; 7611 } 7612 } 7613 /* 7614 * No reason to send a segment, just return. 7615 */ 7616 just_return: 7617 SOCKBUF_UNLOCK(sb); 7618 just_return_nolock: 7619 if (tot_len_this_send == 0) 7620 counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1); 7621 rack_start_hpts_timer(rack, tp, cts, __LINE__, slot, tot_len_this_send, 1); 7622 rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling); 7623 tp->t_flags &= ~TF_FORCEDATA; 7624 return (0); 7625 7626 send: 7627 if (doing_tlp == 0) { 7628 /* 7629 * Data not a TLP, and its not the rxt firing. If it is the 7630 * rxt firing, we want to leave the tlp_in_progress flag on 7631 * so we don't send another TLP. It has to be a rack timer 7632 * or normal send (response to acked data) to clear the tlp 7633 * in progress flag. 7634 */ 7635 rack->rc_tlp_in_progress = 0; 7636 } 7637 SOCKBUF_LOCK_ASSERT(sb); 7638 if (len > 0) { 7639 if (len >= tp->t_maxseg) 7640 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT; 7641 else 7642 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT; 7643 } 7644 /* 7645 * Before ESTABLISHED, force sending of initial options unless TCP 7646 * set not to do any options. NOTE: we assume that the IP/TCP header 7647 * plus TCP options always fit in a single mbuf, leaving room for a 7648 * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr) 7649 * + optlen <= MCLBYTES 7650 */ 7651 optlen = 0; 7652 #ifdef INET6 7653 if (isipv6) 7654 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 7655 else 7656 #endif 7657 hdrlen = sizeof(struct tcpiphdr); 7658 7659 /* 7660 * Compute options for segment. We only have to care about SYN and 7661 * established connection segments. Options for SYN-ACK segments 7662 * are handled in TCP syncache. 7663 */ 7664 to.to_flags = 0; 7665 if ((tp->t_flags & TF_NOOPT) == 0) { 7666 /* Maximum segment size. */ 7667 if (flags & TH_SYN) { 7668 tp->snd_nxt = tp->iss; 7669 to.to_mss = tcp_mssopt(&inp->inp_inc); 7670 #ifdef NETFLIX_TCPOUDP 7671 if (tp->t_port) 7672 to.to_mss -= V_tcp_udp_tunneling_overhead; 7673 #endif 7674 to.to_flags |= TOF_MSS; 7675 7676 /* 7677 * On SYN or SYN|ACK transmits on TFO connections, 7678 * only include the TFO option if it is not a 7679 * retransmit, as the presence of the TFO option may 7680 * have caused the original SYN or SYN|ACK to have 7681 * been dropped by a middlebox. 7682 */ 7683 if (IS_FASTOPEN(tp->t_flags) && 7684 (tp->t_rxtshift == 0)) { 7685 if (tp->t_state == TCPS_SYN_RECEIVED) { 7686 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; 7687 to.to_tfo_cookie = 7688 (u_int8_t *)&tp->t_tfo_cookie.server; 7689 to.to_flags |= TOF_FASTOPEN; 7690 wanted_cookie = 1; 7691 } else if (tp->t_state == TCPS_SYN_SENT) { 7692 to.to_tfo_len = 7693 tp->t_tfo_client_cookie_len; 7694 to.to_tfo_cookie = 7695 tp->t_tfo_cookie.client; 7696 to.to_flags |= TOF_FASTOPEN; 7697 wanted_cookie = 1; 7698 /* 7699 * If we wind up having more data to 7700 * send with the SYN than can fit in 7701 * one segment, don't send any more 7702 * until the SYN|ACK comes back from 7703 * the other end. 7704 */ 7705 sendalot = 0; 7706 } 7707 } 7708 } 7709 /* Window scaling. */ 7710 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) { 7711 to.to_wscale = tp->request_r_scale; 7712 to.to_flags |= TOF_SCALE; 7713 } 7714 /* Timestamps. */ 7715 if ((tp->t_flags & TF_RCVD_TSTMP) || 7716 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) { 7717 to.to_tsval = cts + tp->ts_offset; 7718 to.to_tsecr = tp->ts_recent; 7719 to.to_flags |= TOF_TS; 7720 } 7721 /* Set receive buffer autosizing timestamp. */ 7722 if (tp->rfbuf_ts == 0 && 7723 (so->so_rcv.sb_flags & SB_AUTOSIZE)) 7724 tp->rfbuf_ts = tcp_ts_getticks(); 7725 /* Selective ACK's. */ 7726 if (flags & TH_SYN) 7727 to.to_flags |= TOF_SACKPERM; 7728 else if (TCPS_HAVEESTABLISHED(tp->t_state) && 7729 tp->rcv_numsacks > 0) { 7730 to.to_flags |= TOF_SACK; 7731 to.to_nsacks = tp->rcv_numsacks; 7732 to.to_sacks = (u_char *)tp->sackblks; 7733 } 7734 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 7735 /* TCP-MD5 (RFC2385). */ 7736 if (tp->t_flags & TF_SIGNATURE) 7737 to.to_flags |= TOF_SIGNATURE; 7738 #endif /* TCP_SIGNATURE */ 7739 7740 /* Processing the options. */ 7741 hdrlen += optlen = tcp_addoptions(&to, opt); 7742 /* 7743 * If we wanted a TFO option to be added, but it was unable 7744 * to fit, ensure no data is sent. 7745 */ 7746 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie && 7747 !(to.to_flags & TOF_FASTOPEN)) 7748 len = 0; 7749 } 7750 #ifdef NETFLIX_TCPOUDP 7751 if (tp->t_port) { 7752 if (V_tcp_udp_tunneling_port == 0) { 7753 /* The port was removed?? */ 7754 SOCKBUF_UNLOCK(&so->so_snd); 7755 return (EHOSTUNREACH); 7756 } 7757 hdrlen += sizeof(struct udphdr); 7758 } 7759 #endif 7760 ipoptlen = 0; 7761 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 7762 ipoptlen += ipsec_optlen; 7763 #endif 7764 7765 /* 7766 * Adjust data length if insertion of options will bump the packet 7767 * length beyond the t_maxseg length. Clear the FIN bit because we 7768 * cut off the tail of the segment. 7769 */ 7770 if (len + optlen + ipoptlen > tp->t_maxseg) { 7771 if (flags & TH_FIN) { 7772 would_have_fin = 1; 7773 flags &= ~TH_FIN; 7774 } 7775 if (tso) { 7776 uint32_t if_hw_tsomax; 7777 uint32_t moff; 7778 int32_t max_len; 7779 7780 /* extract TSO information */ 7781 if_hw_tsomax = tp->t_tsomax; 7782 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount; 7783 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize; 7784 KASSERT(ipoptlen == 0, 7785 ("%s: TSO can't do IP options", __func__)); 7786 7787 /* 7788 * Check if we should limit by maximum payload 7789 * length: 7790 */ 7791 if (if_hw_tsomax != 0) { 7792 /* compute maximum TSO length */ 7793 max_len = (if_hw_tsomax - hdrlen - 7794 max_linkhdr); 7795 if (max_len <= 0) { 7796 len = 0; 7797 } else if (len > max_len) { 7798 sendalot = 1; 7799 len = max_len; 7800 } 7801 } 7802 /* 7803 * Prevent the last segment from being fractional 7804 * unless the send sockbuf can be emptied: 7805 */ 7806 max_len = (tp->t_maxseg - optlen); 7807 if ((sb_offset + len) < sbavail(sb)) { 7808 moff = len % (u_int)max_len; 7809 if (moff != 0) { 7810 len -= moff; 7811 sendalot = 1; 7812 } 7813 } 7814 /* 7815 * In case there are too many small fragments don't 7816 * use TSO: 7817 */ 7818 if (len <= max_len) { 7819 len = max_len; 7820 sendalot = 1; 7821 tso = 0; 7822 } 7823 /* 7824 * Send the FIN in a separate segment after the bulk 7825 * sending is done. We don't trust the TSO 7826 * implementations to clear the FIN flag on all but 7827 * the last segment. 7828 */ 7829 if (tp->t_flags & TF_NEEDFIN) 7830 sendalot = 1; 7831 7832 } else { 7833 len = tp->t_maxseg - optlen - ipoptlen; 7834 sendalot = 1; 7835 } 7836 } else 7837 tso = 0; 7838 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET, 7839 ("%s: len > IP_MAXPACKET", __func__)); 7840 #ifdef DIAGNOSTIC 7841 #ifdef INET6 7842 if (max_linkhdr + hdrlen > MCLBYTES) 7843 #else 7844 if (max_linkhdr + hdrlen > MHLEN) 7845 #endif 7846 panic("tcphdr too big"); 7847 #endif 7848 7849 /* 7850 * This KASSERT is here to catch edge cases at a well defined place. 7851 * Before, those had triggered (random) panic conditions further 7852 * down. 7853 */ 7854 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 7855 if ((len == 0) && 7856 (flags & TH_FIN) && 7857 (sbused(sb))) { 7858 /* 7859 * We have outstanding data, don't send a fin by itself!. 7860 */ 7861 goto just_return; 7862 } 7863 /* 7864 * Grab a header mbuf, attaching a copy of data to be transmitted, 7865 * and initialize the header from the template for sends on this 7866 * connection. 7867 */ 7868 if (len) { 7869 uint32_t max_val; 7870 uint32_t moff; 7871 7872 if (rack->rc_pace_max_segs) 7873 max_val = rack->rc_pace_max_segs * tp->t_maxseg; 7874 else 7875 max_val = len; 7876 /* 7877 * We allow a limit on sending with hptsi. 7878 */ 7879 if (len > max_val) { 7880 len = max_val; 7881 } 7882 #ifdef INET6 7883 if (MHLEN < hdrlen + max_linkhdr) 7884 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); 7885 else 7886 #endif 7887 m = m_gethdr(M_NOWAIT, MT_DATA); 7888 7889 if (m == NULL) { 7890 SOCKBUF_UNLOCK(sb); 7891 error = ENOBUFS; 7892 sack_rxmit = 0; 7893 goto out; 7894 } 7895 m->m_data += max_linkhdr; 7896 m->m_len = hdrlen; 7897 7898 /* 7899 * Start the m_copy functions from the closest mbuf to the 7900 * sb_offset in the socket buffer chain. 7901 */ 7902 mb = sbsndptr_noadv(sb, sb_offset, &moff); 7903 if (len <= MHLEN - hdrlen - max_linkhdr) { 7904 m_copydata(mb, moff, (int)len, 7905 mtod(m, caddr_t)+hdrlen); 7906 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 7907 sbsndptr_adv(sb, mb, len); 7908 m->m_len += len; 7909 } else { 7910 struct sockbuf *msb; 7911 7912 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 7913 msb = NULL; 7914 else 7915 msb = sb; 7916 m->m_next = tcp_m_copym(mb, moff, &len, 7917 if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb); 7918 if (len <= (tp->t_maxseg - optlen)) { 7919 /* 7920 * Must have ran out of mbufs for the copy 7921 * shorten it to no longer need tso. Lets 7922 * not put on sendalot since we are low on 7923 * mbufs. 7924 */ 7925 tso = 0; 7926 } 7927 if (m->m_next == NULL) { 7928 SOCKBUF_UNLOCK(sb); 7929 (void)m_free(m); 7930 error = ENOBUFS; 7931 sack_rxmit = 0; 7932 goto out; 7933 } 7934 } 7935 if ((tp->t_flags & TF_FORCEDATA) && len == 1) { 7936 TCPSTAT_INC(tcps_sndprobe); 7937 #ifdef NETFLIX_STATS 7938 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 7939 stats_voi_update_abs_u32(tp->t_stats, 7940 VOI_TCP_RETXPB, len); 7941 else 7942 stats_voi_update_abs_u64(tp->t_stats, 7943 VOI_TCP_TXPB, len); 7944 #endif 7945 } else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) { 7946 if (rsm && (rsm->r_flags & RACK_TLP)) { 7947 /* 7948 * TLP should not count in retran count, but 7949 * in its own bin 7950 */ 7951 counter_u64_add(rack_tlp_retran, 1); 7952 counter_u64_add(rack_tlp_retran_bytes, len); 7953 } else { 7954 tp->t_sndrexmitpack++; 7955 TCPSTAT_INC(tcps_sndrexmitpack); 7956 TCPSTAT_ADD(tcps_sndrexmitbyte, len); 7957 } 7958 #ifdef NETFLIX_STATS 7959 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB, 7960 len); 7961 #endif 7962 } else { 7963 TCPSTAT_INC(tcps_sndpack); 7964 TCPSTAT_ADD(tcps_sndbyte, len); 7965 #ifdef NETFLIX_STATS 7966 stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB, 7967 len); 7968 #endif 7969 } 7970 /* 7971 * If we're sending everything we've got, set PUSH. (This 7972 * will keep happy those implementations which only give 7973 * data to the user when a buffer fills or a PUSH comes in.) 7974 */ 7975 if (sb_offset + len == sbused(sb) && 7976 sbused(sb) && 7977 !(flags & TH_SYN)) 7978 flags |= TH_PUSH; 7979 7980 /* 7981 * Are we doing hptsi, if so we must calculate the slot. We 7982 * only do hptsi in ESTABLISHED and with no RESET being 7983 * sent where we have data to send. 7984 */ 7985 if (((tp->t_state == TCPS_ESTABLISHED) || 7986 (tp->t_state == TCPS_CLOSE_WAIT) || 7987 ((tp->t_state == TCPS_FIN_WAIT_1) && 7988 ((tp->t_flags & TF_SENTFIN) == 0) && 7989 ((flags & TH_FIN) == 0))) && 7990 ((flags & TH_RST) == 0) && 7991 (rack->rc_always_pace)) { 7992 /* 7993 * We use the most optimistic possible cwnd/srtt for 7994 * sending calculations. This will make our 7995 * calculation anticipate getting more through 7996 * quicker then possible. But thats ok we don't want 7997 * the peer to have a gap in data sending. 7998 */ 7999 uint32_t srtt, cwnd, tr_perms = 0; 8000 8001 if (rack->r_ctl.rc_rack_min_rtt) 8002 srtt = rack->r_ctl.rc_rack_min_rtt; 8003 else 8004 srtt = TICKS_2_MSEC((tp->t_srtt >> TCP_RTT_SHIFT)); 8005 if (rack->r_ctl.rc_rack_largest_cwnd) 8006 cwnd = rack->r_ctl.rc_rack_largest_cwnd; 8007 else 8008 cwnd = tp->snd_cwnd; 8009 tr_perms = cwnd / srtt; 8010 if (tr_perms == 0) { 8011 tr_perms = tp->t_maxseg; 8012 } 8013 tot_len_this_send += len; 8014 /* 8015 * Calculate how long this will take to drain, if 8016 * the calculation comes out to zero, thats ok we 8017 * will use send_a_lot to possibly spin around for 8018 * more increasing tot_len_this_send to the point 8019 * that its going to require a pace, or we hit the 8020 * cwnd. Which in that case we are just waiting for 8021 * a ACK. 8022 */ 8023 slot = tot_len_this_send / tr_perms; 8024 /* Now do we reduce the time so we don't run dry? */ 8025 if (slot && rack->rc_pace_reduce) { 8026 int32_t reduce; 8027 8028 reduce = (slot / rack->rc_pace_reduce); 8029 if (reduce < slot) { 8030 slot -= reduce; 8031 } else 8032 slot = 0; 8033 } 8034 if (rack->r_enforce_min_pace && 8035 (slot == 0) && 8036 (tot_len_this_send >= (rack->r_min_pace_seg_thresh * tp->t_maxseg))) { 8037 /* We are enforcing a minimum pace time of 1ms */ 8038 slot = rack->r_enforce_min_pace; 8039 } 8040 } 8041 SOCKBUF_UNLOCK(sb); 8042 } else { 8043 SOCKBUF_UNLOCK(sb); 8044 if (tp->t_flags & TF_ACKNOW) 8045 TCPSTAT_INC(tcps_sndacks); 8046 else if (flags & (TH_SYN | TH_FIN | TH_RST)) 8047 TCPSTAT_INC(tcps_sndctrl); 8048 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 8049 TCPSTAT_INC(tcps_sndurg); 8050 else 8051 TCPSTAT_INC(tcps_sndwinup); 8052 8053 m = m_gethdr(M_NOWAIT, MT_DATA); 8054 if (m == NULL) { 8055 error = ENOBUFS; 8056 sack_rxmit = 0; 8057 goto out; 8058 } 8059 #ifdef INET6 8060 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && 8061 MHLEN >= hdrlen) { 8062 M_ALIGN(m, hdrlen); 8063 } else 8064 #endif 8065 m->m_data += max_linkhdr; 8066 m->m_len = hdrlen; 8067 } 8068 SOCKBUF_UNLOCK_ASSERT(sb); 8069 m->m_pkthdr.rcvif = (struct ifnet *)0; 8070 #ifdef MAC 8071 mac_inpcb_create_mbuf(inp, m); 8072 #endif 8073 #ifdef INET6 8074 if (isipv6) { 8075 ip6 = mtod(m, struct ip6_hdr *); 8076 #ifdef NETFLIX_TCPOUDP 8077 if (tp->t_port) { 8078 udp = (struct udphdr *)((caddr_t)ip6 + ipoptlen + sizeof(struct ip6_hdr)); 8079 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 8080 udp->uh_dport = tp->t_port; 8081 ulen = hdrlen + len - sizeof(struct ip6_hdr); 8082 udp->uh_ulen = htons(ulen); 8083 th = (struct tcphdr *)(udp + 1); 8084 } else 8085 #endif 8086 th = (struct tcphdr *)(ip6 + 1); 8087 tcpip_fillheaders(inp, ip6, th); 8088 } else 8089 #endif /* INET6 */ 8090 { 8091 ip = mtod(m, struct ip *); 8092 #ifdef TCPDEBUG 8093 ipov = (struct ipovly *)ip; 8094 #endif 8095 #ifdef NETFLIX_TCPOUDP 8096 if (tp->t_port) { 8097 udp = (struct udphdr *)((caddr_t)ip + ipoptlen + sizeof(struct ip)); 8098 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 8099 udp->uh_dport = tp->t_port; 8100 ulen = hdrlen + len - sizeof(struct ip); 8101 udp->uh_ulen = htons(ulen); 8102 th = (struct tcphdr *)(udp + 1); 8103 } else 8104 #endif 8105 th = (struct tcphdr *)(ip + 1); 8106 tcpip_fillheaders(inp, ip, th); 8107 } 8108 /* 8109 * Fill in fields, remembering maximum advertised window for use in 8110 * delaying messages about window sizes. If resending a FIN, be sure 8111 * not to use a new sequence number. 8112 */ 8113 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && 8114 tp->snd_nxt == tp->snd_max) 8115 tp->snd_nxt--; 8116 /* 8117 * If we are starting a connection, send ECN setup SYN packet. If we 8118 * are on a retransmit, we may resend those bits a number of times 8119 * as per RFC 3168. 8120 */ 8121 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) { 8122 if (tp->t_rxtshift >= 1) { 8123 if (tp->t_rxtshift <= V_tcp_ecn_maxretries) 8124 flags |= TH_ECE | TH_CWR; 8125 } else 8126 flags |= TH_ECE | TH_CWR; 8127 } 8128 if (tp->t_state == TCPS_ESTABLISHED && 8129 (tp->t_flags & TF_ECN_PERMIT)) { 8130 /* 8131 * If the peer has ECN, mark data packets with ECN capable 8132 * transmission (ECT). Ignore pure ack packets, 8133 * retransmissions and window probes. 8134 */ 8135 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) && 8136 !((tp->t_flags & TF_FORCEDATA) && len == 1)) { 8137 #ifdef INET6 8138 if (isipv6) 8139 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20); 8140 else 8141 #endif 8142 ip->ip_tos |= IPTOS_ECN_ECT0; 8143 TCPSTAT_INC(tcps_ecn_ect0); 8144 } 8145 /* 8146 * Reply with proper ECN notifications. 8147 */ 8148 if (tp->t_flags & TF_ECN_SND_CWR) { 8149 flags |= TH_CWR; 8150 tp->t_flags &= ~TF_ECN_SND_CWR; 8151 } 8152 if (tp->t_flags & TF_ECN_SND_ECE) 8153 flags |= TH_ECE; 8154 } 8155 /* 8156 * If we are doing retransmissions, then snd_nxt will not reflect 8157 * the first unsent octet. For ACK only packets, we do not want the 8158 * sequence number of the retransmitted packet, we want the sequence 8159 * number of the next unsent octet. So, if there is no data (and no 8160 * SYN or FIN), use snd_max instead of snd_nxt when filling in 8161 * ti_seq. But if we are in persist state, snd_max might reflect 8162 * one byte beyond the right edge of the window, so use snd_nxt in 8163 * that case, since we know we aren't doing a retransmission. 8164 * (retransmit and persist are mutually exclusive...) 8165 */ 8166 if (sack_rxmit == 0) { 8167 if (len || (flags & (TH_SYN | TH_FIN)) || 8168 rack->rc_in_persist) { 8169 th->th_seq = htonl(tp->snd_nxt); 8170 rack_seq = tp->snd_nxt; 8171 } else if (flags & TH_RST) { 8172 /* 8173 * For a Reset send the last cum ack in sequence 8174 * (this like any other choice may still generate a 8175 * challenge ack, if a ack-update packet is in 8176 * flight). 8177 */ 8178 th->th_seq = htonl(tp->snd_una); 8179 rack_seq = tp->snd_una; 8180 } else { 8181 th->th_seq = htonl(tp->snd_max); 8182 rack_seq = tp->snd_max; 8183 } 8184 } else { 8185 th->th_seq = htonl(rsm->r_start); 8186 rack_seq = rsm->r_start; 8187 } 8188 th->th_ack = htonl(tp->rcv_nxt); 8189 if (optlen) { 8190 bcopy(opt, th + 1, optlen); 8191 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 8192 } 8193 th->th_flags = flags; 8194 /* 8195 * Calculate receive window. Don't shrink window, but avoid silly 8196 * window syndrome. 8197 */ 8198 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) && 8199 recwin < (long)tp->t_maxseg) 8200 recwin = 0; 8201 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) && 8202 recwin < (long)(tp->rcv_adv - tp->rcv_nxt)) 8203 recwin = (long)(tp->rcv_adv - tp->rcv_nxt); 8204 if (recwin > (long)TCP_MAXWIN << tp->rcv_scale) 8205 recwin = (long)TCP_MAXWIN << tp->rcv_scale; 8206 8207 /* 8208 * According to RFC1323 the window field in a SYN (i.e., a <SYN> or 8209 * <SYN,ACK>) segment itself is never scaled. The <SYN,ACK> case is 8210 * handled in syncache. 8211 */ 8212 if (flags & TH_SYN) 8213 th->th_win = htons((u_short) 8214 (min(sbspace(&so->so_rcv), TCP_MAXWIN))); 8215 else 8216 th->th_win = htons((u_short)(recwin >> tp->rcv_scale)); 8217 /* 8218 * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0 8219 * window. This may cause the remote transmitter to stall. This 8220 * flag tells soreceive() to disable delayed acknowledgements when 8221 * draining the buffer. This can occur if the receiver is 8222 * attempting to read more data than can be buffered prior to 8223 * transmitting on the connection. 8224 */ 8225 if (th->th_win == 0) { 8226 tp->t_sndzerowin++; 8227 tp->t_flags |= TF_RXWIN0SENT; 8228 } else 8229 tp->t_flags &= ~TF_RXWIN0SENT; 8230 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 8231 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); 8232 th->th_flags |= TH_URG; 8233 } else 8234 /* 8235 * If no urgent pointer to send, then we pull the urgent 8236 * pointer to the left edge of the send window so that it 8237 * doesn't drift into the send window on sequence number 8238 * wraparound. 8239 */ 8240 tp->snd_up = tp->snd_una; /* drag it along */ 8241 8242 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 8243 if (to.to_flags & TOF_SIGNATURE) { 8244 /* 8245 * Calculate MD5 signature and put it into the place 8246 * determined before. 8247 * NOTE: since TCP options buffer doesn't point into 8248 * mbuf's data, calculate offset and use it. 8249 */ 8250 if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th, 8251 (u_char *)(th + 1) + (to.to_signature - opt)) != 0) { 8252 /* 8253 * Do not send segment if the calculation of MD5 8254 * digest has failed. 8255 */ 8256 goto out; 8257 } 8258 } 8259 #endif 8260 8261 /* 8262 * Put TCP length in extended header, and then checksum extended 8263 * header and data. 8264 */ 8265 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 8266 #ifdef INET6 8267 if (isipv6) { 8268 /* 8269 * ip6_plen is not need to be filled now, and will be filled 8270 * in ip6_output. 8271 */ 8272 if (tp->t_port) { 8273 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6; 8274 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 8275 udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0); 8276 th->th_sum = htons(0); 8277 } else { 8278 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 8279 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 8280 th->th_sum = in6_cksum_pseudo(ip6, 8281 sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP, 8282 0); 8283 } 8284 } 8285 #endif 8286 #if defined(INET6) && defined(INET) 8287 else 8288 #endif 8289 #ifdef INET 8290 { 8291 if (tp->t_port) { 8292 m->m_pkthdr.csum_flags = CSUM_UDP; 8293 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 8294 udp->uh_sum = in_pseudo(ip->ip_src.s_addr, 8295 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP)); 8296 th->th_sum = htons(0); 8297 } else { 8298 m->m_pkthdr.csum_flags = CSUM_TCP; 8299 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 8300 th->th_sum = in_pseudo(ip->ip_src.s_addr, 8301 ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) + 8302 IPPROTO_TCP + len + optlen)); 8303 } 8304 /* IP version must be set here for ipv4/ipv6 checking later */ 8305 KASSERT(ip->ip_v == IPVERSION, 8306 ("%s: IP version incorrect: %d", __func__, ip->ip_v)); 8307 } 8308 #endif 8309 8310 /* 8311 * Enable TSO and specify the size of the segments. The TCP pseudo 8312 * header checksum is always provided. XXX: Fixme: This is currently 8313 * not the case for IPv6. 8314 */ 8315 if (tso) { 8316 KASSERT(len > tp->t_maxseg - optlen, 8317 ("%s: len <= tso_segsz", __func__)); 8318 m->m_pkthdr.csum_flags |= CSUM_TSO; 8319 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen; 8320 } 8321 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 8322 KASSERT(len + hdrlen + ipoptlen - ipsec_optlen == m_length(m, NULL), 8323 ("%s: mbuf chain shorter than expected: %d + %u + %u - %u != %u", 8324 __func__, len, hdrlen, ipoptlen, ipsec_optlen, m_length(m, NULL))); 8325 #else 8326 KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL), 8327 ("%s: mbuf chain shorter than expected: %d + %u + %u != %u", 8328 __func__, len, hdrlen, ipoptlen, m_length(m, NULL))); 8329 #endif 8330 8331 #ifdef TCP_HHOOK 8332 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */ 8333 hhook_run_tcp_est_out(tp, th, &to, len, tso); 8334 #endif 8335 8336 #ifdef TCPDEBUG 8337 /* 8338 * Trace. 8339 */ 8340 if (so->so_options & SO_DEBUG) { 8341 u_short save = 0; 8342 8343 #ifdef INET6 8344 if (!isipv6) 8345 #endif 8346 { 8347 save = ipov->ih_len; 8348 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + 8349 * (th->th_off << 2) */ ); 8350 } 8351 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0); 8352 #ifdef INET6 8353 if (!isipv6) 8354 #endif 8355 ipov->ih_len = save; 8356 } 8357 #endif /* TCPDEBUG */ 8358 8359 /* We're getting ready to send; log now. */ 8360 if (tp->t_logstate != TCP_LOG_STATE_OFF) { 8361 union tcp_log_stackspecific log; 8362 8363 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 8364 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; 8365 log.u_bbr.ininput = rack->rc_inp->inp_in_input; 8366 log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt; 8367 if (rsm || sack_rxmit) { 8368 log.u_bbr.flex8 = 1; 8369 } else { 8370 log.u_bbr.flex8 = 0; 8371 } 8372 lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK, 8373 len, &log, false, NULL, NULL, 0, NULL); 8374 } else 8375 lgb = NULL; 8376 8377 /* 8378 * Fill in IP length and desired time to live and send to IP level. 8379 * There should be a better way to handle ttl and tos; we could keep 8380 * them in the template, but need a way to checksum without them. 8381 */ 8382 /* 8383 * m->m_pkthdr.len should have been set before cksum calcuration, 8384 * because in6_cksum() need it. 8385 */ 8386 #ifdef INET6 8387 if (isipv6) { 8388 /* 8389 * we separately set hoplimit for every segment, since the 8390 * user might want to change the value via setsockopt. Also, 8391 * desired default hop limit might be changed via Neighbor 8392 * Discovery. 8393 */ 8394 ip6->ip6_hlim = in6_selecthlim(inp, NULL); 8395 8396 /* 8397 * Set the packet size here for the benefit of DTrace 8398 * probes. ip6_output() will set it properly; it's supposed 8399 * to include the option header lengths as well. 8400 */ 8401 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6)); 8402 8403 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) 8404 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 8405 else 8406 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 8407 8408 if (tp->t_state == TCPS_SYN_SENT) 8409 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th); 8410 8411 TCP_PROBE5(send, NULL, tp, ip6, tp, th); 8412 /* TODO: IPv6 IP6TOS_ECT bit on */ 8413 error = ip6_output(m, tp->t_inpcb->in6p_outputopts, 8414 &inp->inp_route6, 8415 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 8416 NULL, NULL, inp); 8417 8418 if (error == EMSGSIZE && inp->inp_route6.ro_rt != NULL) 8419 mtu = inp->inp_route6.ro_rt->rt_mtu; 8420 } 8421 #endif /* INET6 */ 8422 #if defined(INET) && defined(INET6) 8423 else 8424 #endif 8425 #ifdef INET 8426 { 8427 ip->ip_len = htons(m->m_pkthdr.len); 8428 #ifdef INET6 8429 if (inp->inp_vflag & INP_IPV6PROTO) 8430 ip->ip_ttl = in6_selecthlim(inp, NULL); 8431 #endif /* INET6 */ 8432 /* 8433 * If we do path MTU discovery, then we set DF on every 8434 * packet. This might not be the best thing to do according 8435 * to RFC3390 Section 2. However the tcp hostcache migitates 8436 * the problem so it affects only the first tcp connection 8437 * with a host. 8438 * 8439 * NB: Don't set DF on small MTU/MSS to have a safe 8440 * fallback. 8441 */ 8442 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) { 8443 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 8444 if (tp->t_port == 0 || len < V_tcp_minmss) { 8445 ip->ip_off |= htons(IP_DF); 8446 } 8447 } else { 8448 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 8449 } 8450 8451 if (tp->t_state == TCPS_SYN_SENT) 8452 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th); 8453 8454 TCP_PROBE5(send, NULL, tp, ip, tp, th); 8455 8456 error = ip_output(m, tp->t_inpcb->inp_options, &inp->inp_route, 8457 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, 8458 inp); 8459 if (error == EMSGSIZE && inp->inp_route.ro_rt != NULL) 8460 mtu = inp->inp_route.ro_rt->rt_mtu; 8461 } 8462 #endif /* INET */ 8463 8464 out: 8465 if (lgb) { 8466 lgb->tlb_errno = error; 8467 lgb = NULL; 8468 } 8469 /* 8470 * In transmit state, time the transmission and arrange for the 8471 * retransmit. In persist state, just set snd_max. 8472 */ 8473 if (error == 0) { 8474 if (len == 0) 8475 counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1); 8476 else if (len == 1) { 8477 counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1); 8478 } else if (len > 1) { 8479 int idx; 8480 8481 idx = (len / tp->t_maxseg) + 3; 8482 if (idx >= TCP_MSS_ACCT_ATIMER) 8483 counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1); 8484 else 8485 counter_u64_add(rack_out_size[idx], 1); 8486 } 8487 } 8488 if (sub_from_prr && (error == 0)) { 8489 rack->r_ctl.rc_prr_sndcnt -= len; 8490 } 8491 sub_from_prr = 0; 8492 rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error, cts, 8493 pass, rsm); 8494 if ((tp->t_flags & TF_FORCEDATA) == 0 || 8495 (rack->rc_in_persist == 0)) { 8496 #ifdef NETFLIX_STATS 8497 tcp_seq startseq = tp->snd_nxt; 8498 #endif 8499 8500 /* 8501 * Advance snd_nxt over sequence space of this segment. 8502 */ 8503 if (error) 8504 /* We don't log or do anything with errors */ 8505 goto timer; 8506 8507 if (flags & (TH_SYN | TH_FIN)) { 8508 if (flags & TH_SYN) 8509 tp->snd_nxt++; 8510 if (flags & TH_FIN) { 8511 tp->snd_nxt++; 8512 tp->t_flags |= TF_SENTFIN; 8513 } 8514 } 8515 /* In the ENOBUFS case we do *not* update snd_max */ 8516 if (sack_rxmit) 8517 goto timer; 8518 8519 tp->snd_nxt += len; 8520 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 8521 if (tp->snd_una == tp->snd_max) { 8522 /* 8523 * Update the time we just added data since 8524 * none was outstanding. 8525 */ 8526 rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__); 8527 tp->t_acktime = ticks; 8528 } 8529 tp->snd_max = tp->snd_nxt; 8530 #ifdef NETFLIX_STATS 8531 if (!(tp->t_flags & TF_GPUTINPROG) && len) { 8532 tp->t_flags |= TF_GPUTINPROG; 8533 tp->gput_seq = startseq; 8534 tp->gput_ack = startseq + 8535 ulmin(sbavail(sb) - sb_offset, sendwin); 8536 tp->gput_ts = tcp_ts_getticks(); 8537 } 8538 #endif 8539 } 8540 /* 8541 * Set retransmit timer if not currently set, and not doing 8542 * a pure ack or a keep-alive probe. Initial value for 8543 * retransmit timer is smoothed round-trip time + 2 * 8544 * round-trip time variance. Initialize shift counter which 8545 * is used for backoff of retransmit time. 8546 */ 8547 timer: 8548 if ((tp->snd_wnd == 0) && 8549 TCPS_HAVEESTABLISHED(tp->t_state)) { 8550 /* 8551 * If the persists timer was set above (right before 8552 * the goto send), and still needs to be on. Lets 8553 * make sure all is canceled. If the persist timer 8554 * is not running, we want to get it up. 8555 */ 8556 if (rack->rc_in_persist == 0) { 8557 rack_enter_persist(tp, rack, cts); 8558 } 8559 } 8560 } else { 8561 /* 8562 * Persist case, update snd_max but since we are in persist 8563 * mode (no window) we do not update snd_nxt. 8564 */ 8565 int32_t xlen = len; 8566 8567 if (error) 8568 goto nomore; 8569 8570 if (flags & TH_SYN) 8571 ++xlen; 8572 if (flags & TH_FIN) { 8573 ++xlen; 8574 tp->t_flags |= TF_SENTFIN; 8575 } 8576 /* In the ENOBUFS case we do *not* update snd_max */ 8577 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) { 8578 if (tp->snd_una == tp->snd_max) { 8579 /* 8580 * Update the time we just added data since 8581 * none was outstanding. 8582 */ 8583 rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__); 8584 tp->t_acktime = ticks; 8585 } 8586 tp->snd_max = tp->snd_nxt + len; 8587 } 8588 } 8589 nomore: 8590 if (error) { 8591 SOCKBUF_UNLOCK_ASSERT(sb); /* Check gotos. */ 8592 /* 8593 * Failures do not advance the seq counter above. For the 8594 * case of ENOBUFS we will fall out and retry in 1ms with 8595 * the hpts. Everything else will just have to retransmit 8596 * with the timer. 8597 * 8598 * In any case, we do not want to loop around for another 8599 * send without a good reason. 8600 */ 8601 sendalot = 0; 8602 switch (error) { 8603 case EPERM: 8604 tp->t_flags &= ~TF_FORCEDATA; 8605 tp->t_softerror = error; 8606 return (error); 8607 case ENOBUFS: 8608 if (slot == 0) { 8609 /* 8610 * Pace us right away to retry in a some 8611 * time 8612 */ 8613 slot = 1 + rack->rc_enobuf; 8614 if (rack->rc_enobuf < 255) 8615 rack->rc_enobuf++; 8616 if (slot > (rack->rc_rack_rtt / 2)) { 8617 slot = rack->rc_rack_rtt / 2; 8618 } 8619 if (slot < 10) 8620 slot = 10; 8621 } 8622 counter_u64_add(rack_saw_enobuf, 1); 8623 error = 0; 8624 goto enobufs; 8625 case EMSGSIZE: 8626 /* 8627 * For some reason the interface we used initially 8628 * to send segments changed to another or lowered 8629 * its MTU. If TSO was active we either got an 8630 * interface without TSO capabilits or TSO was 8631 * turned off. If we obtained mtu from ip_output() 8632 * then update it and try again. 8633 */ 8634 if (tso) 8635 tp->t_flags &= ~TF_TSO; 8636 if (mtu != 0) { 8637 tcp_mss_update(tp, -1, mtu, NULL, NULL); 8638 goto again; 8639 } 8640 slot = 10; 8641 rack_start_hpts_timer(rack, tp, cts, __LINE__, slot, 0, 1); 8642 tp->t_flags &= ~TF_FORCEDATA; 8643 return (error); 8644 case ENETUNREACH: 8645 counter_u64_add(rack_saw_enetunreach, 1); 8646 case EHOSTDOWN: 8647 case EHOSTUNREACH: 8648 case ENETDOWN: 8649 if (TCPS_HAVERCVDSYN(tp->t_state)) { 8650 tp->t_softerror = error; 8651 } 8652 /* FALLTHROUGH */ 8653 default: 8654 slot = 10; 8655 rack_start_hpts_timer(rack, tp, cts, __LINE__, slot, 0, 1); 8656 tp->t_flags &= ~TF_FORCEDATA; 8657 return (error); 8658 } 8659 } else { 8660 rack->rc_enobuf = 0; 8661 } 8662 TCPSTAT_INC(tcps_sndtotal); 8663 8664 /* 8665 * Data sent (as far as we can tell). If this advertises a larger 8666 * window than any other segment, then remember the size of the 8667 * advertised window. Any pending ACK has now been sent. 8668 */ 8669 if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) 8670 tp->rcv_adv = tp->rcv_nxt + recwin; 8671 tp->last_ack_sent = tp->rcv_nxt; 8672 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); 8673 enobufs: 8674 rack->r_tlp_running = 0; 8675 if ((flags & TH_RST) || (would_have_fin == 1)) { 8676 /* 8677 * We don't send again after a RST. We also do *not* send 8678 * again if we would have had a find, but now have 8679 * outstanding data. 8680 */ 8681 slot = 0; 8682 sendalot = 0; 8683 } 8684 if (slot) { 8685 /* set the rack tcb into the slot N */ 8686 counter_u64_add(rack_paced_segments, 1); 8687 } else if (sendalot) { 8688 if (len) 8689 counter_u64_add(rack_unpaced_segments, 1); 8690 sack_rxmit = 0; 8691 tp->t_flags &= ~TF_FORCEDATA; 8692 goto again; 8693 } else if (len) { 8694 counter_u64_add(rack_unpaced_segments, 1); 8695 } 8696 tp->t_flags &= ~TF_FORCEDATA; 8697 rack_start_hpts_timer(rack, tp, cts, __LINE__, slot, tot_len_this_send, 1); 8698 return (error); 8699 } 8700 8701 /* 8702 * rack_ctloutput() must drop the inpcb lock before performing copyin on 8703 * socket option arguments. When it re-acquires the lock after the copy, it 8704 * has to revalidate that the connection is still valid for the socket 8705 * option. 8706 */ 8707 static int 8708 rack_set_sockopt(struct socket *so, struct sockopt *sopt, 8709 struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack) 8710 { 8711 int32_t error = 0, optval; 8712 8713 switch (sopt->sopt_name) { 8714 case TCP_RACK_PROP_RATE: 8715 case TCP_RACK_PROP: 8716 case TCP_RACK_TLP_REDUCE: 8717 case TCP_RACK_EARLY_RECOV: 8718 case TCP_RACK_PACE_ALWAYS: 8719 case TCP_DELACK: 8720 case TCP_RACK_PACE_REDUCE: 8721 case TCP_RACK_PACE_MAX_SEG: 8722 case TCP_RACK_PRR_SENDALOT: 8723 case TCP_RACK_MIN_TO: 8724 case TCP_RACK_EARLY_SEG: 8725 case TCP_RACK_REORD_THRESH: 8726 case TCP_RACK_REORD_FADE: 8727 case TCP_RACK_TLP_THRESH: 8728 case TCP_RACK_PKT_DELAY: 8729 case TCP_RACK_TLP_USE: 8730 case TCP_RACK_TLP_INC_VAR: 8731 case TCP_RACK_IDLE_REDUCE_HIGH: 8732 case TCP_RACK_MIN_PACE: 8733 case TCP_RACK_MIN_PACE_SEG: 8734 case TCP_BBR_RACK_RTT_USE: 8735 case TCP_DATA_AFTER_CLOSE: 8736 break; 8737 default: 8738 return (tcp_default_ctloutput(so, sopt, inp, tp)); 8739 break; 8740 } 8741 INP_WUNLOCK(inp); 8742 error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval)); 8743 if (error) 8744 return (error); 8745 INP_WLOCK(inp); 8746 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { 8747 INP_WUNLOCK(inp); 8748 return (ECONNRESET); 8749 } 8750 tp = intotcpcb(inp); 8751 rack = (struct tcp_rack *)tp->t_fb_ptr; 8752 switch (sopt->sopt_name) { 8753 case TCP_RACK_PROP_RATE: 8754 if ((optval <= 0) || (optval >= 100)) { 8755 error = EINVAL; 8756 break; 8757 } 8758 RACK_OPTS_INC(tcp_rack_prop_rate); 8759 rack->r_ctl.rc_prop_rate = optval; 8760 break; 8761 case TCP_RACK_TLP_USE: 8762 if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) { 8763 error = EINVAL; 8764 break; 8765 } 8766 RACK_OPTS_INC(tcp_tlp_use); 8767 rack->rack_tlp_threshold_use = optval; 8768 break; 8769 case TCP_RACK_PROP: 8770 /* RACK proportional rate reduction (bool) */ 8771 RACK_OPTS_INC(tcp_rack_prop); 8772 rack->r_ctl.rc_prop_reduce = optval; 8773 break; 8774 case TCP_RACK_TLP_REDUCE: 8775 /* RACK TLP cwnd reduction (bool) */ 8776 RACK_OPTS_INC(tcp_rack_tlp_reduce); 8777 rack->r_ctl.rc_tlp_cwnd_reduce = optval; 8778 break; 8779 case TCP_RACK_EARLY_RECOV: 8780 /* Should recovery happen early (bool) */ 8781 RACK_OPTS_INC(tcp_rack_early_recov); 8782 rack->r_ctl.rc_early_recovery = optval; 8783 break; 8784 case TCP_RACK_PACE_ALWAYS: 8785 /* Use the always pace method (bool) */ 8786 RACK_OPTS_INC(tcp_rack_pace_always); 8787 if (optval > 0) 8788 rack->rc_always_pace = 1; 8789 else 8790 rack->rc_always_pace = 0; 8791 break; 8792 case TCP_RACK_PACE_REDUCE: 8793 /* RACK Hptsi reduction factor (divisor) */ 8794 RACK_OPTS_INC(tcp_rack_pace_reduce); 8795 if (optval) 8796 /* Must be non-zero */ 8797 rack->rc_pace_reduce = optval; 8798 else 8799 error = EINVAL; 8800 break; 8801 case TCP_RACK_PACE_MAX_SEG: 8802 /* Max segments in a pace */ 8803 RACK_OPTS_INC(tcp_rack_max_seg); 8804 rack->rc_pace_max_segs = optval; 8805 break; 8806 case TCP_RACK_PRR_SENDALOT: 8807 /* Allow PRR to send more than one seg */ 8808 RACK_OPTS_INC(tcp_rack_prr_sendalot); 8809 rack->r_ctl.rc_prr_sendalot = optval; 8810 break; 8811 case TCP_RACK_MIN_TO: 8812 /* Minimum time between rack t-o's in ms */ 8813 RACK_OPTS_INC(tcp_rack_min_to); 8814 rack->r_ctl.rc_min_to = optval; 8815 break; 8816 case TCP_RACK_EARLY_SEG: 8817 /* If early recovery max segments */ 8818 RACK_OPTS_INC(tcp_rack_early_seg); 8819 rack->r_ctl.rc_early_recovery_segs = optval; 8820 break; 8821 case TCP_RACK_REORD_THRESH: 8822 /* RACK reorder threshold (shift amount) */ 8823 RACK_OPTS_INC(tcp_rack_reord_thresh); 8824 if ((optval > 0) && (optval < 31)) 8825 rack->r_ctl.rc_reorder_shift = optval; 8826 else 8827 error = EINVAL; 8828 break; 8829 case TCP_RACK_REORD_FADE: 8830 /* Does reordering fade after ms time */ 8831 RACK_OPTS_INC(tcp_rack_reord_fade); 8832 rack->r_ctl.rc_reorder_fade = optval; 8833 break; 8834 case TCP_RACK_TLP_THRESH: 8835 /* RACK TLP theshold i.e. srtt+(srtt/N) */ 8836 RACK_OPTS_INC(tcp_rack_tlp_thresh); 8837 if (optval) 8838 rack->r_ctl.rc_tlp_threshold = optval; 8839 else 8840 error = EINVAL; 8841 break; 8842 case TCP_RACK_PKT_DELAY: 8843 /* RACK added ms i.e. rack-rtt + reord + N */ 8844 RACK_OPTS_INC(tcp_rack_pkt_delay); 8845 rack->r_ctl.rc_pkt_delay = optval; 8846 break; 8847 case TCP_RACK_TLP_INC_VAR: 8848 /* Does TLP include rtt variance in t-o */ 8849 RACK_OPTS_INC(tcp_rack_tlp_inc_var); 8850 rack->r_ctl.rc_prr_inc_var = optval; 8851 break; 8852 case TCP_RACK_IDLE_REDUCE_HIGH: 8853 RACK_OPTS_INC(tcp_rack_idle_reduce_high); 8854 if (optval) 8855 rack->r_idle_reduce_largest = 1; 8856 else 8857 rack->r_idle_reduce_largest = 0; 8858 break; 8859 case TCP_DELACK: 8860 if (optval == 0) 8861 tp->t_delayed_ack = 0; 8862 else 8863 tp->t_delayed_ack = 1; 8864 if (tp->t_flags & TF_DELACK) { 8865 tp->t_flags &= ~TF_DELACK; 8866 tp->t_flags |= TF_ACKNOW; 8867 rack_output(tp); 8868 } 8869 break; 8870 case TCP_RACK_MIN_PACE: 8871 RACK_OPTS_INC(tcp_rack_min_pace); 8872 if (optval > 3) 8873 rack->r_enforce_min_pace = 3; 8874 else 8875 rack->r_enforce_min_pace = optval; 8876 break; 8877 case TCP_RACK_MIN_PACE_SEG: 8878 RACK_OPTS_INC(tcp_rack_min_pace_seg); 8879 if (optval >= 16) 8880 rack->r_min_pace_seg_thresh = 15; 8881 else 8882 rack->r_min_pace_seg_thresh = optval; 8883 break; 8884 case TCP_BBR_RACK_RTT_USE: 8885 if ((optval != USE_RTT_HIGH) && 8886 (optval != USE_RTT_LOW) && 8887 (optval != USE_RTT_AVG)) 8888 error = EINVAL; 8889 else 8890 rack->r_ctl.rc_rate_sample_method = optval; 8891 break; 8892 case TCP_DATA_AFTER_CLOSE: 8893 if (optval) 8894 rack->rc_allow_data_af_clo = 1; 8895 else 8896 rack->rc_allow_data_af_clo = 0; 8897 break; 8898 default: 8899 return (tcp_default_ctloutput(so, sopt, inp, tp)); 8900 break; 8901 } 8902 #ifdef NETFLIX_STATS 8903 tcp_log_socket_option(tp, sopt->sopt_name, optval, error); 8904 #endif 8905 INP_WUNLOCK(inp); 8906 return (error); 8907 } 8908 8909 static int 8910 rack_get_sockopt(struct socket *so, struct sockopt *sopt, 8911 struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack) 8912 { 8913 int32_t error, optval; 8914 8915 /* 8916 * Because all our options are either boolean or an int, we can just 8917 * pull everything into optval and then unlock and copy. If we ever 8918 * add a option that is not a int, then this will have quite an 8919 * impact to this routine. 8920 */ 8921 switch (sopt->sopt_name) { 8922 case TCP_RACK_PROP_RATE: 8923 optval = rack->r_ctl.rc_prop_rate; 8924 break; 8925 case TCP_RACK_PROP: 8926 /* RACK proportional rate reduction (bool) */ 8927 optval = rack->r_ctl.rc_prop_reduce; 8928 break; 8929 case TCP_RACK_TLP_REDUCE: 8930 /* RACK TLP cwnd reduction (bool) */ 8931 optval = rack->r_ctl.rc_tlp_cwnd_reduce; 8932 break; 8933 case TCP_RACK_EARLY_RECOV: 8934 /* Should recovery happen early (bool) */ 8935 optval = rack->r_ctl.rc_early_recovery; 8936 break; 8937 case TCP_RACK_PACE_REDUCE: 8938 /* RACK Hptsi reduction factor (divisor) */ 8939 optval = rack->rc_pace_reduce; 8940 break; 8941 case TCP_RACK_PACE_MAX_SEG: 8942 /* Max segments in a pace */ 8943 optval = rack->rc_pace_max_segs; 8944 break; 8945 case TCP_RACK_PACE_ALWAYS: 8946 /* Use the always pace method */ 8947 optval = rack->rc_always_pace; 8948 break; 8949 case TCP_RACK_PRR_SENDALOT: 8950 /* Allow PRR to send more than one seg */ 8951 optval = rack->r_ctl.rc_prr_sendalot; 8952 break; 8953 case TCP_RACK_MIN_TO: 8954 /* Minimum time between rack t-o's in ms */ 8955 optval = rack->r_ctl.rc_min_to; 8956 break; 8957 case TCP_RACK_EARLY_SEG: 8958 /* If early recovery max segments */ 8959 optval = rack->r_ctl.rc_early_recovery_segs; 8960 break; 8961 case TCP_RACK_REORD_THRESH: 8962 /* RACK reorder threshold (shift amount) */ 8963 optval = rack->r_ctl.rc_reorder_shift; 8964 break; 8965 case TCP_RACK_REORD_FADE: 8966 /* Does reordering fade after ms time */ 8967 optval = rack->r_ctl.rc_reorder_fade; 8968 break; 8969 case TCP_RACK_TLP_THRESH: 8970 /* RACK TLP theshold i.e. srtt+(srtt/N) */ 8971 optval = rack->r_ctl.rc_tlp_threshold; 8972 break; 8973 case TCP_RACK_PKT_DELAY: 8974 /* RACK added ms i.e. rack-rtt + reord + N */ 8975 optval = rack->r_ctl.rc_pkt_delay; 8976 break; 8977 case TCP_RACK_TLP_USE: 8978 optval = rack->rack_tlp_threshold_use; 8979 break; 8980 case TCP_RACK_TLP_INC_VAR: 8981 /* Does TLP include rtt variance in t-o */ 8982 optval = rack->r_ctl.rc_prr_inc_var; 8983 break; 8984 case TCP_RACK_IDLE_REDUCE_HIGH: 8985 optval = rack->r_idle_reduce_largest; 8986 break; 8987 case TCP_RACK_MIN_PACE: 8988 optval = rack->r_enforce_min_pace; 8989 break; 8990 case TCP_RACK_MIN_PACE_SEG: 8991 optval = rack->r_min_pace_seg_thresh; 8992 break; 8993 case TCP_BBR_RACK_RTT_USE: 8994 optval = rack->r_ctl.rc_rate_sample_method; 8995 break; 8996 case TCP_DELACK: 8997 optval = tp->t_delayed_ack; 8998 break; 8999 case TCP_DATA_AFTER_CLOSE: 9000 optval = rack->rc_allow_data_af_clo; 9001 break; 9002 default: 9003 return (tcp_default_ctloutput(so, sopt, inp, tp)); 9004 break; 9005 } 9006 INP_WUNLOCK(inp); 9007 error = sooptcopyout(sopt, &optval, sizeof optval); 9008 return (error); 9009 } 9010 9011 static int 9012 rack_ctloutput(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp) 9013 { 9014 int32_t error = EINVAL; 9015 struct tcp_rack *rack; 9016 9017 rack = (struct tcp_rack *)tp->t_fb_ptr; 9018 if (rack == NULL) { 9019 /* Huh? */ 9020 goto out; 9021 } 9022 if (sopt->sopt_dir == SOPT_SET) { 9023 return (rack_set_sockopt(so, sopt, inp, tp, rack)); 9024 } else if (sopt->sopt_dir == SOPT_GET) { 9025 return (rack_get_sockopt(so, sopt, inp, tp, rack)); 9026 } 9027 out: 9028 INP_WUNLOCK(inp); 9029 return (error); 9030 } 9031 9032 9033 struct tcp_function_block __tcp_rack = { 9034 .tfb_tcp_block_name = __XSTRING(STACKNAME), 9035 .tfb_tcp_output = rack_output, 9036 .tfb_tcp_do_segment = rack_do_segment, 9037 .tfb_tcp_hpts_do_segment = rack_hpts_do_segment, 9038 .tfb_tcp_ctloutput = rack_ctloutput, 9039 .tfb_tcp_fb_init = rack_init, 9040 .tfb_tcp_fb_fini = rack_fini, 9041 .tfb_tcp_timer_stop_all = rack_stopall, 9042 .tfb_tcp_timer_activate = rack_timer_activate, 9043 .tfb_tcp_timer_active = rack_timer_active, 9044 .tfb_tcp_timer_stop = rack_timer_stop, 9045 .tfb_tcp_rexmit_tmr = rack_remxt_tmr, 9046 .tfb_tcp_handoff_ok = rack_handoff_ok 9047 }; 9048 9049 static const char *rack_stack_names[] = { 9050 __XSTRING(STACKNAME), 9051 #ifdef STACKALIAS 9052 __XSTRING(STACKALIAS), 9053 #endif 9054 }; 9055 9056 static int 9057 rack_ctor(void *mem, int32_t size, void *arg, int32_t how) 9058 { 9059 memset(mem, 0, size); 9060 return (0); 9061 } 9062 9063 static void 9064 rack_dtor(void *mem, int32_t size, void *arg) 9065 { 9066 9067 } 9068 9069 static bool rack_mod_inited = false; 9070 9071 static int 9072 tcp_addrack(module_t mod, int32_t type, void *data) 9073 { 9074 int32_t err = 0; 9075 int num_stacks; 9076 9077 switch (type) { 9078 case MOD_LOAD: 9079 rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map", 9080 sizeof(struct rack_sendmap), 9081 rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0); 9082 9083 rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb", 9084 sizeof(struct tcp_rack), 9085 rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0); 9086 9087 sysctl_ctx_init(&rack_sysctl_ctx); 9088 rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx, 9089 SYSCTL_STATIC_CHILDREN(_net_inet_tcp), 9090 OID_AUTO, 9091 __XSTRING(STACKNAME), 9092 CTLFLAG_RW, 0, 9093 ""); 9094 if (rack_sysctl_root == NULL) { 9095 printf("Failed to add sysctl node\n"); 9096 err = EFAULT; 9097 goto free_uma; 9098 } 9099 rack_init_sysctls(); 9100 num_stacks = nitems(rack_stack_names); 9101 err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK, 9102 rack_stack_names, &num_stacks); 9103 if (err) { 9104 printf("Failed to register %s stack name for " 9105 "%s module\n", rack_stack_names[num_stacks], 9106 __XSTRING(MODNAME)); 9107 sysctl_ctx_free(&rack_sysctl_ctx); 9108 free_uma: 9109 uma_zdestroy(rack_zone); 9110 uma_zdestroy(rack_pcb_zone); 9111 rack_counter_destroy(); 9112 printf("Failed to register rack module -- err:%d\n", err); 9113 return (err); 9114 } 9115 rack_mod_inited = true; 9116 break; 9117 case MOD_QUIESCE: 9118 err = deregister_tcp_functions(&__tcp_rack, true, false); 9119 break; 9120 case MOD_UNLOAD: 9121 err = deregister_tcp_functions(&__tcp_rack, false, true); 9122 if (err == EBUSY) 9123 break; 9124 if (rack_mod_inited) { 9125 uma_zdestroy(rack_zone); 9126 uma_zdestroy(rack_pcb_zone); 9127 sysctl_ctx_free(&rack_sysctl_ctx); 9128 rack_counter_destroy(); 9129 rack_mod_inited = false; 9130 } 9131 err = 0; 9132 break; 9133 default: 9134 return (EOPNOTSUPP); 9135 } 9136 return (err); 9137 } 9138 9139 static moduledata_t tcp_rack = { 9140 .name = __XSTRING(MODNAME), 9141 .evhand = tcp_addrack, 9142 .priv = 0 9143 }; 9144 9145 MODULE_VERSION(MODNAME, 1); 9146 DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY); 9147 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1); 9148