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 1429 i_cwnd = tcp_compute_initwnd(tcp_maxseg(tp)); 1430 1431 if (reduce_largest) { 1432 /* 1433 * Do we reduce the largest cwnd to make 1434 * rack play nice on restart hptsi wise? 1435 */ 1436 if (((struct tcp_rack *)tp->t_fb_ptr)->r_ctl.rc_rack_largest_cwnd > i_cwnd) 1437 ((struct tcp_rack *)tp->t_fb_ptr)->r_ctl.rc_rack_largest_cwnd = i_cwnd; 1438 } 1439 /* 1440 * Being idle is no differnt than the initial window. If the cc 1441 * clamps it down below the initial window raise it to the initial 1442 * window. 1443 */ 1444 if (tp->snd_cwnd < i_cwnd) { 1445 tp->snd_cwnd = i_cwnd; 1446 } 1447 } 1448 1449 1450 /* 1451 * Indicate whether this ack should be delayed. We can delay the ack if 1452 * following conditions are met: 1453 * - There is no delayed ack timer in progress. 1454 * - Our last ack wasn't a 0-sized window. We never want to delay 1455 * the ack that opens up a 0-sized window. 1456 * - LRO wasn't used for this segment. We make sure by checking that the 1457 * segment size is not larger than the MSS. 1458 * - Delayed acks are enabled or this is a half-synchronized T/TCP 1459 * connection. 1460 */ 1461 #define DELAY_ACK(tp, tlen) \ 1462 (((tp->t_flags & TF_RXWIN0SENT) == 0) && \ 1463 ((tp->t_flags & TF_DELACK) == 0) && \ 1464 (tlen <= tp->t_maxseg) && \ 1465 (tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN))) 1466 1467 static inline void 1468 rack_calc_rwin(struct socket *so, struct tcpcb *tp) 1469 { 1470 int32_t win; 1471 1472 /* 1473 * Calculate amount of space in receive window, and then do TCP 1474 * input processing. Receive window is amount of space in rcv queue, 1475 * but not less than advertised window. 1476 */ 1477 win = sbspace(&so->so_rcv); 1478 if (win < 0) 1479 win = 0; 1480 tp->rcv_wnd = imax(win, (int)(tp->rcv_adv - tp->rcv_nxt)); 1481 } 1482 1483 static void 1484 rack_do_drop(struct mbuf *m, struct tcpcb *tp) 1485 { 1486 /* 1487 * Drop space held by incoming segment and return. 1488 */ 1489 if (tp != NULL) 1490 INP_WUNLOCK(tp->t_inpcb); 1491 if (m) 1492 m_freem(m); 1493 } 1494 1495 static void 1496 rack_do_dropwithreset(struct mbuf *m, struct tcpcb *tp, struct tcphdr *th, 1497 int32_t rstreason, int32_t tlen) 1498 { 1499 if (tp != NULL) { 1500 tcp_dropwithreset(m, th, tp, tlen, rstreason); 1501 INP_WUNLOCK(tp->t_inpcb); 1502 } else 1503 tcp_dropwithreset(m, th, NULL, tlen, rstreason); 1504 } 1505 1506 /* 1507 * The value in ret_val informs the caller 1508 * if we dropped the tcb (and lock) or not. 1509 * 1 = we dropped it, 0 = the TCB is still locked 1510 * and valid. 1511 */ 1512 static void 1513 rack_do_dropafterack(struct mbuf *m, struct tcpcb *tp, struct tcphdr *th, int32_t thflags, int32_t tlen, int32_t * ret_val) 1514 { 1515 /* 1516 * Generate an ACK dropping incoming segment if it occupies sequence 1517 * space, where the ACK reflects our state. 1518 * 1519 * We can now skip the test for the RST flag since all paths to this 1520 * code happen after packets containing RST have been dropped. 1521 * 1522 * In the SYN-RECEIVED state, don't send an ACK unless the segment 1523 * we received passes the SYN-RECEIVED ACK test. If it fails send a 1524 * RST. This breaks the loop in the "LAND" DoS attack, and also 1525 * prevents an ACK storm between two listening ports that have been 1526 * sent forged SYN segments, each with the source address of the 1527 * other. 1528 */ 1529 struct tcp_rack *rack; 1530 1531 if (tp->t_state == TCPS_SYN_RECEIVED && (thflags & TH_ACK) && 1532 (SEQ_GT(tp->snd_una, th->th_ack) || 1533 SEQ_GT(th->th_ack, tp->snd_max))) { 1534 *ret_val = 1; 1535 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 1536 return; 1537 } else 1538 *ret_val = 0; 1539 rack = (struct tcp_rack *)tp->t_fb_ptr; 1540 rack->r_wanted_output++; 1541 tp->t_flags |= TF_ACKNOW; 1542 if (m) 1543 m_freem(m); 1544 } 1545 1546 1547 static int 1548 rack_process_rst(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp) 1549 { 1550 /* 1551 * RFC5961 Section 3.2 1552 * 1553 * - RST drops connection only if SEG.SEQ == RCV.NXT. - If RST is in 1554 * window, we send challenge ACK. 1555 * 1556 * Note: to take into account delayed ACKs, we should test against 1557 * last_ack_sent instead of rcv_nxt. Note 2: we handle special case 1558 * of closed window, not covered by the RFC. 1559 */ 1560 int dropped = 0; 1561 1562 if ((SEQ_GEQ(th->th_seq, (tp->last_ack_sent - 1)) && 1563 SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) || 1564 (tp->rcv_wnd == 0 && tp->last_ack_sent == th->th_seq)) { 1565 1566 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 1567 KASSERT(tp->t_state != TCPS_SYN_SENT, 1568 ("%s: TH_RST for TCPS_SYN_SENT th %p tp %p", 1569 __func__, th, tp)); 1570 1571 if (V_tcp_insecure_rst || 1572 (tp->last_ack_sent == th->th_seq) || 1573 (tp->rcv_nxt == th->th_seq) || 1574 ((tp->last_ack_sent - 1) == th->th_seq)) { 1575 TCPSTAT_INC(tcps_drops); 1576 /* Drop the connection. */ 1577 switch (tp->t_state) { 1578 case TCPS_SYN_RECEIVED: 1579 so->so_error = ECONNREFUSED; 1580 goto close; 1581 case TCPS_ESTABLISHED: 1582 case TCPS_FIN_WAIT_1: 1583 case TCPS_FIN_WAIT_2: 1584 case TCPS_CLOSE_WAIT: 1585 case TCPS_CLOSING: 1586 case TCPS_LAST_ACK: 1587 so->so_error = ECONNRESET; 1588 close: 1589 tcp_state_change(tp, TCPS_CLOSED); 1590 /* FALLTHROUGH */ 1591 default: 1592 tp = tcp_close(tp); 1593 } 1594 dropped = 1; 1595 rack_do_drop(m, tp); 1596 } else { 1597 TCPSTAT_INC(tcps_badrst); 1598 /* Send challenge ACK. */ 1599 tcp_respond(tp, mtod(m, void *), th, m, 1600 tp->rcv_nxt, tp->snd_nxt, TH_ACK); 1601 tp->last_ack_sent = tp->rcv_nxt; 1602 } 1603 } else { 1604 m_freem(m); 1605 } 1606 return (dropped); 1607 } 1608 1609 /* 1610 * The value in ret_val informs the caller 1611 * if we dropped the tcb (and lock) or not. 1612 * 1 = we dropped it, 0 = the TCB is still locked 1613 * and valid. 1614 */ 1615 static void 1616 rack_challenge_ack(struct mbuf *m, struct tcphdr *th, struct tcpcb *tp, int32_t * ret_val) 1617 { 1618 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 1619 1620 TCPSTAT_INC(tcps_badsyn); 1621 if (V_tcp_insecure_syn && 1622 SEQ_GEQ(th->th_seq, tp->last_ack_sent) && 1623 SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) { 1624 tp = tcp_drop(tp, ECONNRESET); 1625 *ret_val = 1; 1626 rack_do_drop(m, tp); 1627 } else { 1628 /* Send challenge ACK. */ 1629 tcp_respond(tp, mtod(m, void *), th, m, tp->rcv_nxt, 1630 tp->snd_nxt, TH_ACK); 1631 tp->last_ack_sent = tp->rcv_nxt; 1632 m = NULL; 1633 *ret_val = 0; 1634 rack_do_drop(m, NULL); 1635 } 1636 } 1637 1638 /* 1639 * rack_ts_check returns 1 for you should not proceed. It places 1640 * in ret_val what should be returned 1/0 by the caller. The 1 indicates 1641 * that the TCB is unlocked and probably dropped. The 0 indicates the 1642 * TCB is still valid and locked. 1643 */ 1644 static int 1645 rack_ts_check(struct mbuf *m, struct tcphdr *th, struct tcpcb *tp, int32_t tlen, int32_t thflags, int32_t * ret_val) 1646 { 1647 1648 /* Check to see if ts_recent is over 24 days old. */ 1649 if (tcp_ts_getticks() - tp->ts_recent_age > TCP_PAWS_IDLE) { 1650 /* 1651 * Invalidate ts_recent. If this segment updates ts_recent, 1652 * the age will be reset later and ts_recent will get a 1653 * valid value. If it does not, setting ts_recent to zero 1654 * will at least satisfy the requirement that zero be placed 1655 * in the timestamp echo reply when ts_recent isn't valid. 1656 * The age isn't reset until we get a valid ts_recent 1657 * because we don't want out-of-order segments to be dropped 1658 * when ts_recent is old. 1659 */ 1660 tp->ts_recent = 0; 1661 } else { 1662 TCPSTAT_INC(tcps_rcvduppack); 1663 TCPSTAT_ADD(tcps_rcvdupbyte, tlen); 1664 TCPSTAT_INC(tcps_pawsdrop); 1665 *ret_val = 0; 1666 if (tlen) { 1667 rack_do_dropafterack(m, tp, th, thflags, tlen, ret_val); 1668 } else { 1669 rack_do_drop(m, NULL); 1670 } 1671 return (1); 1672 } 1673 return (0); 1674 } 1675 1676 /* 1677 * rack_drop_checks returns 1 for you should not proceed. It places 1678 * in ret_val what should be returned 1/0 by the caller. The 1 indicates 1679 * that the TCB is unlocked and probably dropped. The 0 indicates the 1680 * TCB is still valid and locked. 1681 */ 1682 static int 1683 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) 1684 { 1685 int32_t todrop; 1686 int32_t thflags; 1687 int32_t tlen; 1688 1689 thflags = *thf; 1690 tlen = *tlenp; 1691 todrop = tp->rcv_nxt - th->th_seq; 1692 if (todrop > 0) { 1693 if (thflags & TH_SYN) { 1694 thflags &= ~TH_SYN; 1695 th->th_seq++; 1696 if (th->th_urp > 1) 1697 th->th_urp--; 1698 else 1699 thflags &= ~TH_URG; 1700 todrop--; 1701 } 1702 /* 1703 * Following if statement from Stevens, vol. 2, p. 960. 1704 */ 1705 if (todrop > tlen 1706 || (todrop == tlen && (thflags & TH_FIN) == 0)) { 1707 /* 1708 * Any valid FIN must be to the left of the window. 1709 * At this point the FIN must be a duplicate or out 1710 * of sequence; drop it. 1711 */ 1712 thflags &= ~TH_FIN; 1713 /* 1714 * Send an ACK to resynchronize and drop any data. 1715 * But keep on processing for RST or ACK. 1716 */ 1717 tp->t_flags |= TF_ACKNOW; 1718 todrop = tlen; 1719 TCPSTAT_INC(tcps_rcvduppack); 1720 TCPSTAT_ADD(tcps_rcvdupbyte, todrop); 1721 } else { 1722 TCPSTAT_INC(tcps_rcvpartduppack); 1723 TCPSTAT_ADD(tcps_rcvpartdupbyte, todrop); 1724 } 1725 *drop_hdrlen += todrop; /* drop from the top afterwards */ 1726 th->th_seq += todrop; 1727 tlen -= todrop; 1728 if (th->th_urp > todrop) 1729 th->th_urp -= todrop; 1730 else { 1731 thflags &= ~TH_URG; 1732 th->th_urp = 0; 1733 } 1734 } 1735 /* 1736 * If segment ends after window, drop trailing data (and PUSH and 1737 * FIN); if nothing left, just ACK. 1738 */ 1739 todrop = (th->th_seq + tlen) - (tp->rcv_nxt + tp->rcv_wnd); 1740 if (todrop > 0) { 1741 TCPSTAT_INC(tcps_rcvpackafterwin); 1742 if (todrop >= tlen) { 1743 TCPSTAT_ADD(tcps_rcvbyteafterwin, tlen); 1744 /* 1745 * If window is closed can only take segments at 1746 * window edge, and have to drop data and PUSH from 1747 * incoming segments. Continue processing, but 1748 * remember to ack. Otherwise, drop segment and 1749 * ack. 1750 */ 1751 if (tp->rcv_wnd == 0 && th->th_seq == tp->rcv_nxt) { 1752 tp->t_flags |= TF_ACKNOW; 1753 TCPSTAT_INC(tcps_rcvwinprobe); 1754 } else { 1755 rack_do_dropafterack(m, tp, th, thflags, tlen, ret_val); 1756 return (1); 1757 } 1758 } else 1759 TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop); 1760 m_adj(m, -todrop); 1761 tlen -= todrop; 1762 thflags &= ~(TH_PUSH | TH_FIN); 1763 } 1764 *thf = thflags; 1765 *tlenp = tlen; 1766 return (0); 1767 } 1768 1769 static struct rack_sendmap * 1770 rack_find_lowest_rsm(struct tcp_rack *rack) 1771 { 1772 struct rack_sendmap *rsm; 1773 1774 /* 1775 * Walk the time-order transmitted list looking for an rsm that is 1776 * not acked. This will be the one that was sent the longest time 1777 * ago that is still outstanding. 1778 */ 1779 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) { 1780 if (rsm->r_flags & RACK_ACKED) { 1781 continue; 1782 } 1783 goto finish; 1784 } 1785 finish: 1786 return (rsm); 1787 } 1788 1789 static struct rack_sendmap * 1790 rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm) 1791 { 1792 struct rack_sendmap *prsm; 1793 1794 /* 1795 * Walk the sequence order list backward until we hit and arrive at 1796 * the highest seq not acked. In theory when this is called it 1797 * should be the last segment (which it was not). 1798 */ 1799 counter_u64_add(rack_find_high, 1); 1800 prsm = rsm; 1801 TAILQ_FOREACH_REVERSE_FROM(prsm, &rack->r_ctl.rc_map, rack_head, r_next) { 1802 if (prsm->r_flags & (RACK_ACKED | RACK_HAS_FIN)) { 1803 continue; 1804 } 1805 return (prsm); 1806 } 1807 return (NULL); 1808 } 1809 1810 1811 static uint32_t 1812 rack_calc_thresh_rack(struct tcp_rack *rack, uint32_t srtt, uint32_t cts) 1813 { 1814 int32_t lro; 1815 uint32_t thresh; 1816 1817 /* 1818 * lro is the flag we use to determine if we have seen reordering. 1819 * If it gets set we have seen reordering. The reorder logic either 1820 * works in one of two ways: 1821 * 1822 * If reorder-fade is configured, then we track the last time we saw 1823 * re-ordering occur. If we reach the point where enough time as 1824 * passed we no longer consider reordering has occuring. 1825 * 1826 * Or if reorder-face is 0, then once we see reordering we consider 1827 * the connection to alway be subject to reordering and just set lro 1828 * to 1. 1829 * 1830 * In the end if lro is non-zero we add the extra time for 1831 * reordering in. 1832 */ 1833 if (srtt == 0) 1834 srtt = 1; 1835 if (rack->r_ctl.rc_reorder_ts) { 1836 if (rack->r_ctl.rc_reorder_fade) { 1837 if (SEQ_GEQ(cts, rack->r_ctl.rc_reorder_ts)) { 1838 lro = cts - rack->r_ctl.rc_reorder_ts; 1839 if (lro == 0) { 1840 /* 1841 * No time as passed since the last 1842 * reorder, mark it as reordering. 1843 */ 1844 lro = 1; 1845 } 1846 } else { 1847 /* Negative time? */ 1848 lro = 0; 1849 } 1850 if (lro > rack->r_ctl.rc_reorder_fade) { 1851 /* Turn off reordering seen too */ 1852 rack->r_ctl.rc_reorder_ts = 0; 1853 lro = 0; 1854 } 1855 } else { 1856 /* Reodering does not fade */ 1857 lro = 1; 1858 } 1859 } else { 1860 lro = 0; 1861 } 1862 thresh = srtt + rack->r_ctl.rc_pkt_delay; 1863 if (lro) { 1864 /* It must be set, if not you get 1/4 rtt */ 1865 if (rack->r_ctl.rc_reorder_shift) 1866 thresh += (srtt >> rack->r_ctl.rc_reorder_shift); 1867 else 1868 thresh += (srtt >> 2); 1869 } else { 1870 thresh += 1; 1871 } 1872 /* We don't let the rack timeout be above a RTO */ 1873 1874 if (thresh > TICKS_2_MSEC(rack->rc_tp->t_rxtcur)) { 1875 thresh = TICKS_2_MSEC(rack->rc_tp->t_rxtcur); 1876 } 1877 /* And we don't want it above the RTO max either */ 1878 if (thresh > rack_rto_max) { 1879 thresh = rack_rto_max; 1880 } 1881 return (thresh); 1882 } 1883 1884 static uint32_t 1885 rack_calc_thresh_tlp(struct tcpcb *tp, struct tcp_rack *rack, 1886 struct rack_sendmap *rsm, uint32_t srtt) 1887 { 1888 struct rack_sendmap *prsm; 1889 uint32_t thresh, len; 1890 int maxseg; 1891 1892 if (srtt == 0) 1893 srtt = 1; 1894 if (rack->r_ctl.rc_tlp_threshold) 1895 thresh = srtt + (srtt / rack->r_ctl.rc_tlp_threshold); 1896 else 1897 thresh = (srtt * 2); 1898 1899 /* Get the previous sent packet, if any */ 1900 maxseg = tcp_maxseg(tp); 1901 counter_u64_add(rack_enter_tlp_calc, 1); 1902 len = rsm->r_end - rsm->r_start; 1903 if (rack->rack_tlp_threshold_use == TLP_USE_ID) { 1904 /* Exactly like the ID */ 1905 if (((tp->snd_max - tp->snd_una) - rack->r_ctl.rc_sacked + rack->r_ctl.rc_holes_rxt) <= maxseg) { 1906 uint32_t alt_thresh; 1907 /* 1908 * Compensate for delayed-ack with the d-ack time. 1909 */ 1910 counter_u64_add(rack_used_tlpmethod, 1); 1911 alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time; 1912 if (alt_thresh > thresh) 1913 thresh = alt_thresh; 1914 } 1915 } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_ONE) { 1916 /* 2.1 behavior */ 1917 prsm = TAILQ_PREV(rsm, rack_head, r_tnext); 1918 if (prsm && (len <= maxseg)) { 1919 /* 1920 * Two packets outstanding, thresh should be (2*srtt) + 1921 * possible inter-packet delay (if any). 1922 */ 1923 uint32_t inter_gap = 0; 1924 int idx, nidx; 1925 1926 counter_u64_add(rack_used_tlpmethod, 1); 1927 idx = rsm->r_rtr_cnt - 1; 1928 nidx = prsm->r_rtr_cnt - 1; 1929 if (TSTMP_GEQ(rsm->r_tim_lastsent[nidx], prsm->r_tim_lastsent[idx])) { 1930 /* Yes it was sent later (or at the same time) */ 1931 inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx]; 1932 } 1933 thresh += inter_gap; 1934 } else if (len <= maxseg) { 1935 /* 1936 * Possibly compensate for delayed-ack. 1937 */ 1938 uint32_t alt_thresh; 1939 1940 counter_u64_add(rack_used_tlpmethod2, 1); 1941 alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time; 1942 if (alt_thresh > thresh) 1943 thresh = alt_thresh; 1944 } 1945 } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_TWO) { 1946 /* 2.2 behavior */ 1947 if (len <= maxseg) { 1948 uint32_t alt_thresh; 1949 /* 1950 * Compensate for delayed-ack with the d-ack time. 1951 */ 1952 counter_u64_add(rack_used_tlpmethod, 1); 1953 alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time; 1954 if (alt_thresh > thresh) 1955 thresh = alt_thresh; 1956 } 1957 } 1958 /* Not above an RTO */ 1959 if (thresh > TICKS_2_MSEC(tp->t_rxtcur)) { 1960 thresh = TICKS_2_MSEC(tp->t_rxtcur); 1961 } 1962 /* Not above a RTO max */ 1963 if (thresh > rack_rto_max) { 1964 thresh = rack_rto_max; 1965 } 1966 /* Apply user supplied min TLP */ 1967 if (thresh < rack_tlp_min) { 1968 thresh = rack_tlp_min; 1969 } 1970 return (thresh); 1971 } 1972 1973 static struct rack_sendmap * 1974 rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused) 1975 { 1976 /* 1977 * Check to see that we don't need to fall into recovery. We will 1978 * need to do so if our oldest transmit is past the time we should 1979 * have had an ack. 1980 */ 1981 struct tcp_rack *rack; 1982 struct rack_sendmap *rsm; 1983 int32_t idx; 1984 uint32_t srtt_cur, srtt, thresh; 1985 1986 rack = (struct tcp_rack *)tp->t_fb_ptr; 1987 if (TAILQ_EMPTY(&rack->r_ctl.rc_map)) { 1988 return (NULL); 1989 } 1990 srtt_cur = tp->t_srtt >> TCP_RTT_SHIFT; 1991 srtt = TICKS_2_MSEC(srtt_cur); 1992 if (rack->rc_rack_rtt && (srtt > rack->rc_rack_rtt)) 1993 srtt = rack->rc_rack_rtt; 1994 1995 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); 1996 if (rsm == NULL) 1997 return (NULL); 1998 1999 if (rsm->r_flags & RACK_ACKED) { 2000 rsm = rack_find_lowest_rsm(rack); 2001 if (rsm == NULL) 2002 return (NULL); 2003 } 2004 idx = rsm->r_rtr_cnt - 1; 2005 thresh = rack_calc_thresh_rack(rack, srtt, tsused); 2006 if (tsused < rsm->r_tim_lastsent[idx]) { 2007 return (NULL); 2008 } 2009 if ((tsused - rsm->r_tim_lastsent[idx]) < thresh) { 2010 return (NULL); 2011 } 2012 /* Ok if we reach here we are over-due */ 2013 rack->r_ctl.rc_rsm_start = rsm->r_start; 2014 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd; 2015 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh; 2016 rack_cong_signal(tp, NULL, CC_NDUPACK); 2017 return (rsm); 2018 } 2019 2020 static uint32_t 2021 rack_get_persists_timer_val(struct tcpcb *tp, struct tcp_rack *rack) 2022 { 2023 int32_t t; 2024 int32_t tt; 2025 uint32_t ret_val; 2026 2027 t = TICKS_2_MSEC((tp->t_srtt >> TCP_RTT_SHIFT) + ((tp->t_rttvar * 4) >> TCP_RTT_SHIFT)); 2028 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], 2029 tcp_persmin, tcp_persmax); 2030 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 2031 tp->t_rxtshift++; 2032 rack->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT; 2033 ret_val = (uint32_t)tt; 2034 return (ret_val); 2035 } 2036 2037 static uint32_t 2038 rack_timer_start(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) 2039 { 2040 /* 2041 * Start the FR timer, we do this based on getting the first one in 2042 * the rc_tmap. Note that if its NULL we must stop the timer. in all 2043 * events we need to stop the running timer (if its running) before 2044 * starting the new one. 2045 */ 2046 uint32_t thresh, exp, to, srtt, time_since_sent; 2047 uint32_t srtt_cur; 2048 int32_t idx; 2049 int32_t is_tlp_timer = 0; 2050 struct rack_sendmap *rsm; 2051 2052 if (rack->t_timers_stopped) { 2053 /* All timers have been stopped none are to run */ 2054 return (0); 2055 } 2056 if (rack->rc_in_persist) { 2057 /* We can't start any timer in persists */ 2058 return (rack_get_persists_timer_val(tp, rack)); 2059 } 2060 if (tp->t_state < TCPS_ESTABLISHED) 2061 goto activate_rxt; 2062 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); 2063 if (rsm == NULL) { 2064 /* Nothing on the send map */ 2065 activate_rxt: 2066 if (SEQ_LT(tp->snd_una, tp->snd_max) || sbavail(&(tp->t_inpcb->inp_socket->so_snd))) { 2067 rack->r_ctl.rc_hpts_flags |= PACE_TMR_RXT; 2068 to = TICKS_2_MSEC(tp->t_rxtcur); 2069 if (to == 0) 2070 to = 1; 2071 return (to); 2072 } 2073 return (0); 2074 } 2075 if (rsm->r_flags & RACK_ACKED) { 2076 rsm = rack_find_lowest_rsm(rack); 2077 if (rsm == NULL) { 2078 /* No lowest? */ 2079 goto activate_rxt; 2080 } 2081 } 2082 /* Convert from ms to usecs */ 2083 if (rsm->r_flags & RACK_SACK_PASSED) { 2084 if ((tp->t_flags & TF_SENTFIN) && 2085 ((tp->snd_max - tp->snd_una) == 1) && 2086 (rsm->r_flags & RACK_HAS_FIN)) { 2087 /* 2088 * We don't start a rack timer if all we have is a 2089 * FIN outstanding. 2090 */ 2091 goto activate_rxt; 2092 } 2093 if (tp->t_srtt) { 2094 srtt_cur = (tp->t_srtt >> TCP_RTT_SHIFT); 2095 srtt = TICKS_2_MSEC(srtt_cur); 2096 } else 2097 srtt = RACK_INITIAL_RTO; 2098 2099 thresh = rack_calc_thresh_rack(rack, srtt, cts); 2100 idx = rsm->r_rtr_cnt - 1; 2101 exp = rsm->r_tim_lastsent[idx] + thresh; 2102 if (SEQ_GEQ(exp, cts)) { 2103 to = exp - cts; 2104 if (to < rack->r_ctl.rc_min_to) { 2105 to = rack->r_ctl.rc_min_to; 2106 } 2107 } else { 2108 to = rack->r_ctl.rc_min_to; 2109 } 2110 } else { 2111 /* Ok we need to do a TLP not RACK */ 2112 if ((rack->rc_tlp_in_progress != 0) || 2113 (rack->r_ctl.rc_tlp_rtx_out != 0)) { 2114 /* 2115 * The previous send was a TLP or a tlp_rtx is in 2116 * process. 2117 */ 2118 goto activate_rxt; 2119 } 2120 rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext); 2121 if (rsm == NULL) { 2122 /* We found no rsm to TLP with. */ 2123 goto activate_rxt; 2124 } 2125 if (rsm->r_flags & RACK_HAS_FIN) { 2126 /* If its a FIN we dont do TLP */ 2127 rsm = NULL; 2128 goto activate_rxt; 2129 } 2130 idx = rsm->r_rtr_cnt - 1; 2131 if (TSTMP_GT(cts, rsm->r_tim_lastsent[idx])) 2132 time_since_sent = cts - rsm->r_tim_lastsent[idx]; 2133 else 2134 time_since_sent = 0; 2135 is_tlp_timer = 1; 2136 if (tp->t_srtt) { 2137 srtt_cur = (tp->t_srtt >> TCP_RTT_SHIFT); 2138 srtt = TICKS_2_MSEC(srtt_cur); 2139 } else 2140 srtt = RACK_INITIAL_RTO; 2141 thresh = rack_calc_thresh_tlp(tp, rack, rsm, srtt); 2142 if (thresh > time_since_sent) 2143 to = thresh - time_since_sent; 2144 else 2145 to = rack->r_ctl.rc_min_to; 2146 if (to > TCPTV_REXMTMAX) { 2147 /* 2148 * If the TLP time works out to larger than the max 2149 * RTO lets not do TLP.. just RTO. 2150 */ 2151 goto activate_rxt; 2152 } 2153 if (rsm->r_start != rack->r_ctl.rc_last_tlp_seq) { 2154 /* 2155 * The tail is no longer the last one I did a probe 2156 * on 2157 */ 2158 rack->r_ctl.rc_tlp_seg_send_cnt = 0; 2159 rack->r_ctl.rc_last_tlp_seq = rsm->r_start; 2160 } 2161 } 2162 if (is_tlp_timer == 0) { 2163 rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK; 2164 } else { 2165 if ((rack->r_ctl.rc_tlp_send_cnt > rack_tlp_max_resend) || 2166 (rack->r_ctl.rc_tlp_seg_send_cnt > rack_tlp_max_resend)) { 2167 /* 2168 * We have exceeded how many times we can retran the 2169 * current TLP timer, switch to the RTO timer. 2170 */ 2171 goto activate_rxt; 2172 } else { 2173 rack->r_ctl.rc_hpts_flags |= PACE_TMR_TLP; 2174 } 2175 } 2176 if (to == 0) 2177 to = 1; 2178 return (to); 2179 } 2180 2181 static void 2182 rack_enter_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) 2183 { 2184 if (rack->rc_in_persist == 0) { 2185 if (((tp->t_flags & TF_SENTFIN) == 0) && 2186 (tp->snd_max - tp->snd_una) >= sbavail(&rack->rc_inp->inp_socket->so_snd)) 2187 /* Must need to send more data to enter persist */ 2188 return; 2189 rack->r_ctl.rc_went_idle_time = cts; 2190 rack_timer_cancel(tp, rack, cts, __LINE__); 2191 tp->t_rxtshift = 0; 2192 rack->rc_in_persist = 1; 2193 } 2194 } 2195 2196 static void 2197 rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack) 2198 { 2199 if (rack->rc_inp->inp_in_hpts) { 2200 tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT); 2201 rack->r_ctl.rc_hpts_flags = 0; 2202 } 2203 rack->rc_in_persist = 0; 2204 rack->r_ctl.rc_went_idle_time = 0; 2205 tp->t_flags &= ~TF_FORCEDATA; 2206 tp->t_rxtshift = 0; 2207 } 2208 2209 static void 2210 rack_start_hpts_timer(struct tcp_rack *rack, struct tcpcb *tp, uint32_t cts, int32_t line, 2211 int32_t slot, uint32_t tot_len_this_send, int32_t frm_out_sbavail) 2212 { 2213 struct inpcb *inp; 2214 uint32_t delayed_ack = 0; 2215 uint32_t hpts_timeout; 2216 uint8_t stopped; 2217 uint32_t left = 0; 2218 2219 inp = tp->t_inpcb; 2220 if (inp->inp_in_hpts) { 2221 /* A previous call is already set up */ 2222 return; 2223 } 2224 if (tp->t_state == TCPS_CLOSED) { 2225 return; 2226 } 2227 stopped = rack->rc_tmr_stopped; 2228 if (stopped && TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) { 2229 left = rack->r_ctl.rc_timer_exp - cts; 2230 } 2231 rack->r_ctl.rc_timer_exp = 0; 2232 if (rack->rc_inp->inp_in_hpts == 0) { 2233 rack->r_ctl.rc_hpts_flags = 0; 2234 } 2235 if (slot) { 2236 /* We are hptsi too */ 2237 rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT; 2238 } else if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) { 2239 /* 2240 * We are still left on the hpts when the to goes 2241 * it will be for output. 2242 */ 2243 if (TSTMP_GT(cts, rack->r_ctl.rc_last_output_to)) 2244 slot = cts - rack->r_ctl.rc_last_output_to; 2245 else 2246 slot = 1; 2247 } 2248 if ((tp->snd_wnd == 0) && TCPS_HAVEESTABLISHED(tp->t_state)) { 2249 /* No send window.. we must enter persist */ 2250 rack_enter_persist(tp, rack, cts); 2251 } else if ((frm_out_sbavail && 2252 (frm_out_sbavail > (tp->snd_max - tp->snd_una)) && 2253 (tp->snd_wnd < tp->t_maxseg)) && 2254 TCPS_HAVEESTABLISHED(tp->t_state)) { 2255 /* 2256 * If we have no window or we can't send a segment (and have 2257 * data to send.. we cheat here and frm_out_sbavail is 2258 * passed in with the sbavail(sb) only from bbr_output) and 2259 * we are established, then we must enter persits (if not 2260 * already in persits). 2261 */ 2262 rack_enter_persist(tp, rack, cts); 2263 } 2264 hpts_timeout = rack_timer_start(tp, rack, cts); 2265 if (tp->t_flags & TF_DELACK) { 2266 delayed_ack = TICKS_2_MSEC(tcp_delacktime); 2267 rack->r_ctl.rc_hpts_flags |= PACE_TMR_DELACK; 2268 } 2269 if (delayed_ack && ((hpts_timeout == 0) || 2270 (delayed_ack < hpts_timeout))) 2271 hpts_timeout = delayed_ack; 2272 else 2273 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK; 2274 /* 2275 * If no timers are going to run and we will fall off the hptsi 2276 * wheel, we resort to a keep-alive timer if its configured. 2277 */ 2278 if ((hpts_timeout == 0) && 2279 (slot == 0)) { 2280 if ((tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && 2281 (tp->t_state <= TCPS_CLOSING)) { 2282 /* 2283 * Ok we have no timer (persists, rack, tlp, rxt or 2284 * del-ack), we don't have segments being paced. So 2285 * all that is left is the keepalive timer. 2286 */ 2287 if (TCPS_HAVEESTABLISHED(tp->t_state)) { 2288 /* Get the established keep-alive time */ 2289 hpts_timeout = TP_KEEPIDLE(tp); 2290 } else { 2291 /* Get the initial setup keep-alive time */ 2292 hpts_timeout = TP_KEEPINIT(tp); 2293 } 2294 rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP; 2295 } 2296 } 2297 if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) == 2298 (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) { 2299 /* 2300 * RACK, TLP, persists and RXT timers all are restartable 2301 * based on actions input .. i.e we received a packet (ack 2302 * or sack) and that changes things (rw, or snd_una etc). 2303 * Thus we can restart them with a new value. For 2304 * keep-alive, delayed_ack we keep track of what was left 2305 * and restart the timer with a smaller value. 2306 */ 2307 if (left < hpts_timeout) 2308 hpts_timeout = left; 2309 } 2310 if (hpts_timeout) { 2311 /* 2312 * Hack alert for now we can't time-out over 2,147,483 2313 * seconds (a bit more than 596 hours), which is probably ok 2314 * :). 2315 */ 2316 if (hpts_timeout > 0x7ffffffe) 2317 hpts_timeout = 0x7ffffffe; 2318 rack->r_ctl.rc_timer_exp = cts + hpts_timeout; 2319 } 2320 if (slot) { 2321 rack->r_ctl.rc_last_output_to = cts + slot; 2322 if ((hpts_timeout == 0) || (hpts_timeout > slot)) { 2323 if (rack->rc_inp->inp_in_hpts == 0) 2324 tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(slot)); 2325 rack_log_to_start(rack, cts, hpts_timeout, slot, 1); 2326 } else { 2327 /* 2328 * Arrange for the hpts to kick back in after the 2329 * t-o if the t-o does not cause a send. 2330 */ 2331 if (rack->rc_inp->inp_in_hpts == 0) 2332 tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(hpts_timeout)); 2333 rack_log_to_start(rack, cts, hpts_timeout, slot, 0); 2334 } 2335 } else if (hpts_timeout) { 2336 if (rack->rc_inp->inp_in_hpts == 0) 2337 tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(hpts_timeout)); 2338 rack_log_to_start(rack, cts, hpts_timeout, slot, 0); 2339 } else { 2340 /* No timer starting */ 2341 #ifdef INVARIANTS 2342 if (SEQ_GT(tp->snd_max, tp->snd_una)) { 2343 panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?", 2344 tp, rack, tot_len_this_send, cts, slot, hpts_timeout); 2345 } 2346 #endif 2347 } 2348 rack->rc_tmr_stopped = 0; 2349 if (slot) 2350 rack_log_type_bbrsnd(rack, tot_len_this_send, slot, cts); 2351 } 2352 2353 /* 2354 * RACK Timer, here we simply do logging and house keeping. 2355 * the normal rack_output() function will call the 2356 * appropriate thing to check if we need to do a RACK retransmit. 2357 * We return 1, saying don't proceed with rack_output only 2358 * when all timers have been stopped (destroyed PCB?). 2359 */ 2360 static int 2361 rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) 2362 { 2363 /* 2364 * This timer simply provides an internal trigger to send out data. 2365 * The check_recovery_mode call will see if there are needed 2366 * retransmissions, if so we will enter fast-recovery. The output 2367 * call may or may not do the same thing depending on sysctl 2368 * settings. 2369 */ 2370 struct rack_sendmap *rsm; 2371 int32_t recovery; 2372 2373 if (tp->t_timers->tt_flags & TT_STOPPED) { 2374 return (1); 2375 } 2376 if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) { 2377 /* Its not time yet */ 2378 return (0); 2379 } 2380 rack_log_to_event(rack, RACK_TO_FRM_RACK); 2381 recovery = IN_RECOVERY(tp->t_flags); 2382 counter_u64_add(rack_to_tot, 1); 2383 if (rack->r_state && (rack->r_state != tp->t_state)) 2384 rack_set_state(tp, rack); 2385 rsm = rack_check_recovery_mode(tp, cts); 2386 if (rsm) { 2387 uint32_t rtt; 2388 2389 rtt = rack->rc_rack_rtt; 2390 if (rtt == 0) 2391 rtt = 1; 2392 if ((recovery == 0) && 2393 (rack->r_ctl.rc_prr_sndcnt < tp->t_maxseg)) { 2394 /* 2395 * The rack-timeout that enter's us into recovery 2396 * will force out one MSS and set us up so that we 2397 * can do one more send in 2*rtt (transitioning the 2398 * rack timeout into a rack-tlp). 2399 */ 2400 rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; 2401 } else if ((rack->r_ctl.rc_prr_sndcnt < tp->t_maxseg) && 2402 ((rsm->r_end - rsm->r_start) > rack->r_ctl.rc_prr_sndcnt)) { 2403 /* 2404 * When a rack timer goes, we have to send at 2405 * least one segment. They will be paced a min of 1ms 2406 * apart via the next rack timer (or further 2407 * if the rack timer dictates it). 2408 */ 2409 rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; 2410 } 2411 } else { 2412 /* This is a case that should happen rarely if ever */ 2413 counter_u64_add(rack_tlp_does_nada, 1); 2414 #ifdef TCP_BLACKBOX 2415 tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true); 2416 #endif 2417 rack->r_ctl.rc_resend = TAILQ_FIRST(&rack->r_ctl.rc_tmap); 2418 } 2419 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK; 2420 return (0); 2421 } 2422 2423 /* 2424 * TLP Timer, here we simply setup what segment we want to 2425 * have the TLP expire on, the normal rack_output() will then 2426 * send it out. 2427 * 2428 * We return 1, saying don't proceed with rack_output only 2429 * when all timers have been stopped (destroyed PCB?). 2430 */ 2431 static int 2432 rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) 2433 { 2434 /* 2435 * Tail Loss Probe. 2436 */ 2437 struct rack_sendmap *rsm = NULL; 2438 struct socket *so; 2439 uint32_t amm, old_prr_snd = 0; 2440 uint32_t out, avail; 2441 2442 if (tp->t_timers->tt_flags & TT_STOPPED) { 2443 return (1); 2444 } 2445 if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) { 2446 /* Its not time yet */ 2447 return (0); 2448 } 2449 if (rack_progress_timeout_check(tp)) { 2450 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); 2451 return (1); 2452 } 2453 /* 2454 * A TLP timer has expired. We have been idle for 2 rtts. So we now 2455 * need to figure out how to force a full MSS segment out. 2456 */ 2457 rack_log_to_event(rack, RACK_TO_FRM_TLP); 2458 counter_u64_add(rack_tlp_tot, 1); 2459 if (rack->r_state && (rack->r_state != tp->t_state)) 2460 rack_set_state(tp, rack); 2461 so = tp->t_inpcb->inp_socket; 2462 avail = sbavail(&so->so_snd); 2463 out = tp->snd_max - tp->snd_una; 2464 rack->rc_timer_up = 1; 2465 /* 2466 * If we are in recovery we can jazz out a segment if new data is 2467 * present simply by setting rc_prr_sndcnt to a segment. 2468 */ 2469 if ((avail > out) && 2470 ((rack_always_send_oldest == 0) || (TAILQ_EMPTY(&rack->r_ctl.rc_tmap)))) { 2471 /* New data is available */ 2472 amm = avail - out; 2473 if (amm > tp->t_maxseg) { 2474 amm = tp->t_maxseg; 2475 } else if ((amm < tp->t_maxseg) && ((tp->t_flags & TF_NODELAY) == 0)) { 2476 /* not enough to fill a MTU and no-delay is off */ 2477 goto need_retran; 2478 } 2479 if (IN_RECOVERY(tp->t_flags)) { 2480 /* Unlikely */ 2481 old_prr_snd = rack->r_ctl.rc_prr_sndcnt; 2482 if (out + amm <= tp->snd_wnd) 2483 rack->r_ctl.rc_prr_sndcnt = amm; 2484 else 2485 goto need_retran; 2486 } else { 2487 /* Set the send-new override */ 2488 if (out + amm <= tp->snd_wnd) 2489 rack->r_ctl.rc_tlp_new_data = amm; 2490 else 2491 goto need_retran; 2492 } 2493 rack->r_ctl.rc_tlp_seg_send_cnt = 0; 2494 rack->r_ctl.rc_last_tlp_seq = tp->snd_max; 2495 rack->r_ctl.rc_tlpsend = NULL; 2496 counter_u64_add(rack_tlp_newdata, 1); 2497 goto send; 2498 } 2499 need_retran: 2500 /* 2501 * Ok we need to arrange the last un-acked segment to be re-sent, or 2502 * optionally the first un-acked segment. 2503 */ 2504 if (rack_always_send_oldest) 2505 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); 2506 else { 2507 rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_map, rack_sendmap, r_next); 2508 if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) { 2509 rsm = rack_find_high_nonack(rack, rsm); 2510 } 2511 } 2512 if (rsm == NULL) { 2513 counter_u64_add(rack_tlp_does_nada, 1); 2514 #ifdef TCP_BLACKBOX 2515 tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true); 2516 #endif 2517 goto out; 2518 } 2519 if ((rsm->r_end - rsm->r_start) > tp->t_maxseg) { 2520 /* 2521 * We need to split this the last segment in two. 2522 */ 2523 int32_t idx; 2524 struct rack_sendmap *nrsm; 2525 2526 nrsm = rack_alloc(rack); 2527 if (nrsm == NULL) { 2528 /* 2529 * No memory to split, we will just exit and punt 2530 * off to the RXT timer. 2531 */ 2532 counter_u64_add(rack_tlp_does_nada, 1); 2533 goto out; 2534 } 2535 nrsm->r_start = (rsm->r_end - tp->t_maxseg); 2536 nrsm->r_end = rsm->r_end; 2537 nrsm->r_rtr_cnt = rsm->r_rtr_cnt; 2538 nrsm->r_flags = rsm->r_flags; 2539 nrsm->r_sndcnt = rsm->r_sndcnt; 2540 nrsm->r_rtr_bytes = 0; 2541 rsm->r_end = nrsm->r_start; 2542 for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { 2543 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; 2544 } 2545 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_map, rsm, nrsm, r_next); 2546 if (rsm->r_in_tmap) { 2547 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 2548 nrsm->r_in_tmap = 1; 2549 } 2550 rsm->r_flags &= (~RACK_HAS_FIN); 2551 rsm = nrsm; 2552 } 2553 rack->r_ctl.rc_tlpsend = rsm; 2554 rack->r_ctl.rc_tlp_rtx_out = 1; 2555 if (rsm->r_start == rack->r_ctl.rc_last_tlp_seq) { 2556 rack->r_ctl.rc_tlp_seg_send_cnt++; 2557 tp->t_rxtshift++; 2558 } else { 2559 rack->r_ctl.rc_last_tlp_seq = rsm->r_start; 2560 rack->r_ctl.rc_tlp_seg_send_cnt = 1; 2561 } 2562 send: 2563 rack->r_ctl.rc_tlp_send_cnt++; 2564 if (rack->r_ctl.rc_tlp_send_cnt > rack_tlp_max_resend) { 2565 /* 2566 * Can't [re]/transmit a segment we have not heard from the 2567 * peer in max times. We need the retransmit timer to take 2568 * over. 2569 */ 2570 restore: 2571 rack->r_ctl.rc_tlpsend = NULL; 2572 if (rsm) 2573 rsm->r_flags &= ~RACK_TLP; 2574 rack->r_ctl.rc_prr_sndcnt = old_prr_snd; 2575 counter_u64_add(rack_tlp_retran_fail, 1); 2576 goto out; 2577 } else if (rsm) { 2578 rsm->r_flags |= RACK_TLP; 2579 } 2580 if (rsm && (rsm->r_start == rack->r_ctl.rc_last_tlp_seq) && 2581 (rack->r_ctl.rc_tlp_seg_send_cnt > rack_tlp_max_resend)) { 2582 /* 2583 * We don't want to send a single segment more than the max 2584 * either. 2585 */ 2586 goto restore; 2587 } 2588 rack->r_timer_override = 1; 2589 rack->r_tlp_running = 1; 2590 rack->rc_tlp_in_progress = 1; 2591 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP; 2592 return (0); 2593 out: 2594 rack->rc_timer_up = 0; 2595 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP; 2596 return (0); 2597 } 2598 2599 /* 2600 * Delayed ack Timer, here we simply need to setup the 2601 * ACK_NOW flag and remove the DELACK flag. From there 2602 * the output routine will send the ack out. 2603 * 2604 * We only return 1, saying don't proceed, if all timers 2605 * are stopped (destroyed PCB?). 2606 */ 2607 static int 2608 rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) 2609 { 2610 if (tp->t_timers->tt_flags & TT_STOPPED) { 2611 return (1); 2612 } 2613 rack_log_to_event(rack, RACK_TO_FRM_DELACK); 2614 tp->t_flags &= ~TF_DELACK; 2615 tp->t_flags |= TF_ACKNOW; 2616 TCPSTAT_INC(tcps_delack); 2617 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK; 2618 return (0); 2619 } 2620 2621 /* 2622 * Persists timer, here we simply need to setup the 2623 * FORCE-DATA flag the output routine will send 2624 * the one byte send. 2625 * 2626 * We only return 1, saying don't proceed, if all timers 2627 * are stopped (destroyed PCB?). 2628 */ 2629 static int 2630 rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) 2631 { 2632 struct inpcb *inp; 2633 int32_t retval = 0; 2634 2635 inp = tp->t_inpcb; 2636 2637 if (tp->t_timers->tt_flags & TT_STOPPED) { 2638 return (1); 2639 } 2640 if (rack->rc_in_persist == 0) 2641 return (0); 2642 if (rack_progress_timeout_check(tp)) { 2643 tcp_set_inp_to_drop(inp, ETIMEDOUT); 2644 return (1); 2645 } 2646 KASSERT(inp != NULL, ("%s: tp %p tp->t_inpcb == NULL", __func__, tp)); 2647 /* 2648 * Persistence timer into zero window. Force a byte to be output, if 2649 * possible. 2650 */ 2651 TCPSTAT_INC(tcps_persisttimeo); 2652 /* 2653 * Hack: if the peer is dead/unreachable, we do not time out if the 2654 * window is closed. After a full backoff, drop the connection if 2655 * the idle time (no responses to probes) reaches the maximum 2656 * backoff that we would use if retransmitting. 2657 */ 2658 if (tp->t_rxtshift == TCP_MAXRXTSHIFT && 2659 (ticks - tp->t_rcvtime >= tcp_maxpersistidle || 2660 ticks - tp->t_rcvtime >= TCP_REXMTVAL(tp) * tcp_totbackoff)) { 2661 TCPSTAT_INC(tcps_persistdrop); 2662 retval = 1; 2663 tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT); 2664 goto out; 2665 } 2666 if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) && 2667 tp->snd_una == tp->snd_max) 2668 rack_exit_persist(tp, rack); 2669 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT; 2670 /* 2671 * If the user has closed the socket then drop a persisting 2672 * connection after a much reduced timeout. 2673 */ 2674 if (tp->t_state > TCPS_CLOSE_WAIT && 2675 (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) { 2676 retval = 1; 2677 TCPSTAT_INC(tcps_persistdrop); 2678 tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT); 2679 goto out; 2680 } 2681 tp->t_flags |= TF_FORCEDATA; 2682 out: 2683 rack_log_to_event(rack, RACK_TO_FRM_PERSIST); 2684 return (retval); 2685 } 2686 2687 /* 2688 * If a keepalive goes off, we had no other timers 2689 * happening. We always return 1 here since this 2690 * routine either drops the connection or sends 2691 * out a segment with respond. 2692 */ 2693 static int 2694 rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) 2695 { 2696 struct tcptemp *t_template; 2697 struct inpcb *inp; 2698 2699 if (tp->t_timers->tt_flags & TT_STOPPED) { 2700 return (1); 2701 } 2702 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP; 2703 inp = tp->t_inpcb; 2704 rack_log_to_event(rack, RACK_TO_FRM_KEEP); 2705 /* 2706 * Keep-alive timer went off; send something or drop connection if 2707 * idle for too long. 2708 */ 2709 TCPSTAT_INC(tcps_keeptimeo); 2710 if (tp->t_state < TCPS_ESTABLISHED) 2711 goto dropit; 2712 if ((tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && 2713 tp->t_state <= TCPS_CLOSING) { 2714 if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp)) 2715 goto dropit; 2716 /* 2717 * Send a packet designed to force a response if the peer is 2718 * up and reachable: either an ACK if the connection is 2719 * still alive, or an RST if the peer has closed the 2720 * connection due to timeout or reboot. Using sequence 2721 * number tp->snd_una-1 causes the transmitted zero-length 2722 * segment to lie outside the receive window; by the 2723 * protocol spec, this requires the correspondent TCP to 2724 * respond. 2725 */ 2726 TCPSTAT_INC(tcps_keepprobe); 2727 t_template = tcpip_maketemplate(inp); 2728 if (t_template) { 2729 tcp_respond(tp, t_template->tt_ipgen, 2730 &t_template->tt_t, (struct mbuf *)NULL, 2731 tp->rcv_nxt, tp->snd_una - 1, 0); 2732 free(t_template, M_TEMP); 2733 } 2734 } 2735 rack_start_hpts_timer(rack, tp, cts, __LINE__, 0, 0, 0); 2736 return (1); 2737 dropit: 2738 TCPSTAT_INC(tcps_keepdrops); 2739 tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT); 2740 return (1); 2741 } 2742 2743 /* 2744 * Retransmit helper function, clear up all the ack 2745 * flags and take care of important book keeping. 2746 */ 2747 static void 2748 rack_remxt_tmr(struct tcpcb *tp) 2749 { 2750 /* 2751 * The retransmit timer went off, all sack'd blocks must be 2752 * un-acked. 2753 */ 2754 struct rack_sendmap *rsm, *trsm = NULL; 2755 struct tcp_rack *rack; 2756 int32_t cnt = 0; 2757 2758 rack = (struct tcp_rack *)tp->t_fb_ptr; 2759 rack_timer_cancel(tp, rack, tcp_ts_getticks(), __LINE__); 2760 rack_log_to_event(rack, RACK_TO_FRM_TMR); 2761 if (rack->r_state && (rack->r_state != tp->t_state)) 2762 rack_set_state(tp, rack); 2763 /* 2764 * Ideally we would like to be able to 2765 * mark SACK-PASS on anything not acked here. 2766 * However, if we do that we would burst out 2767 * all that data 1ms apart. This would be unwise, 2768 * so for now we will just let the normal rxt timer 2769 * and tlp timer take care of it. 2770 */ 2771 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_map, r_next) { 2772 if (rsm->r_flags & RACK_ACKED) { 2773 cnt++; 2774 rsm->r_sndcnt = 0; 2775 if (rsm->r_in_tmap == 0) { 2776 /* We must re-add it back to the tlist */ 2777 if (trsm == NULL) { 2778 TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext); 2779 } else { 2780 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext); 2781 } 2782 rsm->r_in_tmap = 1; 2783 trsm = rsm; 2784 } 2785 } 2786 rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS); 2787 } 2788 /* Clear the count (we just un-acked them) */ 2789 rack->r_ctl.rc_sacked = 0; 2790 /* Clear the tlp rtx mark */ 2791 rack->r_ctl.rc_tlp_rtx_out = 0; 2792 rack->r_ctl.rc_tlp_seg_send_cnt = 0; 2793 rack->r_ctl.rc_resend = TAILQ_FIRST(&rack->r_ctl.rc_map); 2794 /* Setup so we send one segment */ 2795 if (rack->r_ctl.rc_prr_sndcnt < tp->t_maxseg) 2796 rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; 2797 rack->r_timer_override = 1; 2798 } 2799 2800 /* 2801 * Re-transmit timeout! If we drop the PCB we will return 1, otherwise 2802 * we will setup to retransmit the lowest seq number outstanding. 2803 */ 2804 static int 2805 rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) 2806 { 2807 int32_t rexmt; 2808 struct inpcb *inp; 2809 int32_t retval = 0; 2810 2811 inp = tp->t_inpcb; 2812 if (tp->t_timers->tt_flags & TT_STOPPED) { 2813 return (1); 2814 } 2815 if (rack_progress_timeout_check(tp)) { 2816 tcp_set_inp_to_drop(inp, ETIMEDOUT); 2817 return (1); 2818 } 2819 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT; 2820 if (TCPS_HAVEESTABLISHED(tp->t_state) && 2821 (tp->snd_una == tp->snd_max)) { 2822 /* Nothing outstanding .. nothing to do */ 2823 return (0); 2824 } 2825 /* 2826 * Retransmission timer went off. Message has not been acked within 2827 * retransmit interval. Back off to a longer retransmit interval 2828 * and retransmit one segment. 2829 */ 2830 if (++tp->t_rxtshift > TCP_MAXRXTSHIFT) { 2831 tp->t_rxtshift = TCP_MAXRXTSHIFT; 2832 TCPSTAT_INC(tcps_timeoutdrop); 2833 retval = 1; 2834 tcp_set_inp_to_drop(rack->rc_inp, 2835 (tp->t_softerror ? (uint16_t) tp->t_softerror : ETIMEDOUT)); 2836 goto out; 2837 } 2838 rack_remxt_tmr(tp); 2839 if (tp->t_state == TCPS_SYN_SENT) { 2840 /* 2841 * If the SYN was retransmitted, indicate CWND to be limited 2842 * to 1 segment in cc_conn_init(). 2843 */ 2844 tp->snd_cwnd = 1; 2845 } else if (tp->t_rxtshift == 1) { 2846 /* 2847 * first retransmit; record ssthresh and cwnd so they can be 2848 * recovered if this turns out to be a "bad" retransmit. A 2849 * retransmit is considered "bad" if an ACK for this segment 2850 * is received within RTT/2 interval; the assumption here is 2851 * that the ACK was already in flight. See "On Estimating 2852 * End-to-End Network Path Properties" by Allman and Paxson 2853 * for more details. 2854 */ 2855 tp->snd_cwnd_prev = tp->snd_cwnd; 2856 tp->snd_ssthresh_prev = tp->snd_ssthresh; 2857 tp->snd_recover_prev = tp->snd_recover; 2858 if (IN_FASTRECOVERY(tp->t_flags)) 2859 tp->t_flags |= TF_WASFRECOVERY; 2860 else 2861 tp->t_flags &= ~TF_WASFRECOVERY; 2862 if (IN_CONGRECOVERY(tp->t_flags)) 2863 tp->t_flags |= TF_WASCRECOVERY; 2864 else 2865 tp->t_flags &= ~TF_WASCRECOVERY; 2866 tp->t_badrxtwin = ticks + (tp->t_srtt >> (TCP_RTT_SHIFT + 1)); 2867 tp->t_flags |= TF_PREVVALID; 2868 } else 2869 tp->t_flags &= ~TF_PREVVALID; 2870 TCPSTAT_INC(tcps_rexmttimeo); 2871 if ((tp->t_state == TCPS_SYN_SENT) || 2872 (tp->t_state == TCPS_SYN_RECEIVED)) 2873 rexmt = MSEC_2_TICKS(RACK_INITIAL_RTO * tcp_syn_backoff[tp->t_rxtshift]); 2874 else 2875 rexmt = TCP_REXMTVAL(tp) * tcp_backoff[tp->t_rxtshift]; 2876 TCPT_RANGESET(tp->t_rxtcur, rexmt, 2877 max(MSEC_2_TICKS(rack_rto_min), rexmt), 2878 MSEC_2_TICKS(rack_rto_max)); 2879 /* 2880 * We enter the path for PLMTUD if connection is established or, if 2881 * connection is FIN_WAIT_1 status, reason for the last is that if 2882 * amount of data we send is very small, we could send it in couple 2883 * of packets and process straight to FIN. In that case we won't 2884 * catch ESTABLISHED state. 2885 */ 2886 if (V_tcp_pmtud_blackhole_detect && (((tp->t_state == TCPS_ESTABLISHED)) 2887 || (tp->t_state == TCPS_FIN_WAIT_1))) { 2888 #ifdef INET6 2889 int32_t isipv6; 2890 #endif 2891 2892 /* 2893 * Idea here is that at each stage of mtu probe (usually, 2894 * 1448 -> 1188 -> 524) should be given 2 chances to recover 2895 * before further clamping down. 'tp->t_rxtshift % 2 == 0' 2896 * should take care of that. 2897 */ 2898 if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) == 2899 (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) && 2900 (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 && 2901 tp->t_rxtshift % 2 == 0)) { 2902 /* 2903 * Enter Path MTU Black-hole Detection mechanism: - 2904 * Disable Path MTU Discovery (IP "DF" bit). - 2905 * Reduce MTU to lower value than what we negotiated 2906 * with peer. 2907 */ 2908 if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) { 2909 /* Record that we may have found a black hole. */ 2910 tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE; 2911 /* Keep track of previous MSS. */ 2912 tp->t_pmtud_saved_maxseg = tp->t_maxseg; 2913 } 2914 2915 /* 2916 * Reduce the MSS to blackhole value or to the 2917 * default in an attempt to retransmit. 2918 */ 2919 #ifdef INET6 2920 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) ? 1 : 0; 2921 if (isipv6 && 2922 tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) { 2923 /* Use the sysctl tuneable blackhole MSS. */ 2924 tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss; 2925 TCPSTAT_INC(tcps_pmtud_blackhole_activated); 2926 } else if (isipv6) { 2927 /* Use the default MSS. */ 2928 tp->t_maxseg = V_tcp_v6mssdflt; 2929 /* 2930 * Disable Path MTU Discovery when we switch 2931 * to minmss. 2932 */ 2933 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 2934 TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss); 2935 } 2936 #endif 2937 #if defined(INET6) && defined(INET) 2938 else 2939 #endif 2940 #ifdef INET 2941 if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) { 2942 /* Use the sysctl tuneable blackhole MSS. */ 2943 tp->t_maxseg = V_tcp_pmtud_blackhole_mss; 2944 TCPSTAT_INC(tcps_pmtud_blackhole_activated); 2945 } else { 2946 /* Use the default MSS. */ 2947 tp->t_maxseg = V_tcp_mssdflt; 2948 /* 2949 * Disable Path MTU Discovery when we switch 2950 * to minmss. 2951 */ 2952 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 2953 TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss); 2954 } 2955 #endif 2956 } else { 2957 /* 2958 * If further retransmissions are still unsuccessful 2959 * with a lowered MTU, maybe this isn't a blackhole 2960 * and we restore the previous MSS and blackhole 2961 * detection flags. The limit '6' is determined by 2962 * giving each probe stage (1448, 1188, 524) 2 2963 * chances to recover. 2964 */ 2965 if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) && 2966 (tp->t_rxtshift >= 6)) { 2967 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 2968 tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE; 2969 tp->t_maxseg = tp->t_pmtud_saved_maxseg; 2970 TCPSTAT_INC(tcps_pmtud_blackhole_failed); 2971 } 2972 } 2973 } 2974 /* 2975 * Disable RFC1323 and SACK if we haven't got any response to our 2976 * third SYN to work-around some broken terminal servers (most of 2977 * which have hopefully been retired) that have bad VJ header 2978 * compression code which trashes TCP segments containing 2979 * unknown-to-them TCP options. 2980 */ 2981 if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) && 2982 (tp->t_rxtshift == 3)) 2983 tp->t_flags &= ~(TF_REQ_SCALE | TF_REQ_TSTMP | TF_SACK_PERMIT); 2984 /* 2985 * If we backed off this far, our srtt estimate is probably bogus. 2986 * Clobber it so we'll take the next rtt measurement as our srtt; 2987 * move the current srtt into rttvar to keep the current retransmit 2988 * times until then. 2989 */ 2990 if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) { 2991 #ifdef INET6 2992 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0) 2993 in6_losing(tp->t_inpcb); 2994 else 2995 #endif 2996 in_losing(tp->t_inpcb); 2997 tp->t_rttvar += (tp->t_srtt >> TCP_RTT_SHIFT); 2998 tp->t_srtt = 0; 2999 } 3000 if (rack_use_sack_filter) 3001 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una); 3002 tp->snd_recover = tp->snd_max; 3003 tp->t_flags |= TF_ACKNOW; 3004 tp->t_rtttime = 0; 3005 rack_cong_signal(tp, NULL, CC_RTO); 3006 out: 3007 return (retval); 3008 } 3009 3010 static int 3011 rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling) 3012 { 3013 int32_t ret = 0; 3014 int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK); 3015 3016 if (timers == 0) { 3017 return (0); 3018 } 3019 if (tp->t_state == TCPS_LISTEN) { 3020 /* no timers on listen sockets */ 3021 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) 3022 return (0); 3023 return (1); 3024 } 3025 if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) { 3026 uint32_t left; 3027 3028 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) { 3029 ret = -1; 3030 rack_log_to_processing(rack, cts, ret, 0); 3031 return (0); 3032 } 3033 if (hpts_calling == 0) { 3034 ret = -2; 3035 rack_log_to_processing(rack, cts, ret, 0); 3036 return (0); 3037 } 3038 /* 3039 * Ok our timer went off early and we are not paced false 3040 * alarm, go back to sleep. 3041 */ 3042 ret = -3; 3043 left = rack->r_ctl.rc_timer_exp - cts; 3044 tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(left)); 3045 rack_log_to_processing(rack, cts, ret, left); 3046 rack->rc_last_pto_set = 0; 3047 return (1); 3048 } 3049 rack->rc_tmr_stopped = 0; 3050 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK; 3051 if (timers & PACE_TMR_DELACK) { 3052 ret = rack_timeout_delack(tp, rack, cts); 3053 } else if (timers & PACE_TMR_RACK) { 3054 ret = rack_timeout_rack(tp, rack, cts); 3055 } else if (timers & PACE_TMR_TLP) { 3056 ret = rack_timeout_tlp(tp, rack, cts); 3057 } else if (timers & PACE_TMR_RXT) { 3058 ret = rack_timeout_rxt(tp, rack, cts); 3059 } else if (timers & PACE_TMR_PERSIT) { 3060 ret = rack_timeout_persist(tp, rack, cts); 3061 } else if (timers & PACE_TMR_KEEP) { 3062 ret = rack_timeout_keepalive(tp, rack, cts); 3063 } 3064 rack_log_to_processing(rack, cts, ret, timers); 3065 return (ret); 3066 } 3067 3068 static void 3069 rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line) 3070 { 3071 uint8_t hpts_removed = 0; 3072 3073 if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) && 3074 TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to)) { 3075 tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT); 3076 hpts_removed = 1; 3077 } 3078 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) { 3079 rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK; 3080 if (rack->rc_inp->inp_in_hpts && 3081 ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) { 3082 /* 3083 * Canceling timer's when we have no output being 3084 * paced. We also must remove ourselves from the 3085 * hpts. 3086 */ 3087 tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT); 3088 hpts_removed = 1; 3089 } 3090 rack_log_to_cancel(rack, hpts_removed, line); 3091 rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK); 3092 } 3093 } 3094 3095 static void 3096 rack_timer_stop(struct tcpcb *tp, uint32_t timer_type) 3097 { 3098 return; 3099 } 3100 3101 static int 3102 rack_stopall(struct tcpcb *tp) 3103 { 3104 struct tcp_rack *rack; 3105 rack = (struct tcp_rack *)tp->t_fb_ptr; 3106 rack->t_timers_stopped = 1; 3107 return (0); 3108 } 3109 3110 static void 3111 rack_timer_activate(struct tcpcb *tp, uint32_t timer_type, uint32_t delta) 3112 { 3113 return; 3114 } 3115 3116 static int 3117 rack_timer_active(struct tcpcb *tp, uint32_t timer_type) 3118 { 3119 return (0); 3120 } 3121 3122 static void 3123 rack_stop_all_timers(struct tcpcb *tp) 3124 { 3125 struct tcp_rack *rack; 3126 3127 /* 3128 * Assure no timers are running. 3129 */ 3130 if (tcp_timer_active(tp, TT_PERSIST)) { 3131 /* We enter in persists, set the flag appropriately */ 3132 rack = (struct tcp_rack *)tp->t_fb_ptr; 3133 rack->rc_in_persist = 1; 3134 } 3135 tcp_timer_suspend(tp, TT_PERSIST); 3136 tcp_timer_suspend(tp, TT_REXMT); 3137 tcp_timer_suspend(tp, TT_KEEP); 3138 tcp_timer_suspend(tp, TT_DELACK); 3139 } 3140 3141 static void 3142 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack, 3143 struct rack_sendmap *rsm, uint32_t ts) 3144 { 3145 int32_t idx; 3146 3147 rsm->r_rtr_cnt++; 3148 rsm->r_sndcnt++; 3149 if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) { 3150 rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS; 3151 rsm->r_flags |= RACK_OVERMAX; 3152 } 3153 if ((rsm->r_rtr_cnt > 1) && (rack->r_tlp_running == 0)) { 3154 rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start); 3155 rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start); 3156 } 3157 idx = rsm->r_rtr_cnt - 1; 3158 rsm->r_tim_lastsent[idx] = ts; 3159 if (rsm->r_flags & RACK_ACKED) { 3160 /* Problably MTU discovery messing with us */ 3161 rsm->r_flags &= ~RACK_ACKED; 3162 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); 3163 } 3164 if (rsm->r_in_tmap) { 3165 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext); 3166 } 3167 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext); 3168 rsm->r_in_tmap = 1; 3169 if (rsm->r_flags & RACK_SACK_PASSED) { 3170 /* We have retransmitted due to the SACK pass */ 3171 rsm->r_flags &= ~RACK_SACK_PASSED; 3172 rsm->r_flags |= RACK_WAS_SACKPASS; 3173 } 3174 /* Update memory for next rtr */ 3175 rack->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next); 3176 } 3177 3178 3179 static uint32_t 3180 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack, 3181 struct rack_sendmap *rsm, uint32_t ts, int32_t * lenp) 3182 { 3183 /* 3184 * We (re-)transmitted starting at rsm->r_start for some length 3185 * (possibly less than r_end. 3186 */ 3187 struct rack_sendmap *nrsm; 3188 uint32_t c_end; 3189 int32_t len; 3190 int32_t idx; 3191 3192 len = *lenp; 3193 c_end = rsm->r_start + len; 3194 if (SEQ_GEQ(c_end, rsm->r_end)) { 3195 /* 3196 * We retransmitted the whole piece or more than the whole 3197 * slopping into the next rsm. 3198 */ 3199 rack_update_rsm(tp, rack, rsm, ts); 3200 if (c_end == rsm->r_end) { 3201 *lenp = 0; 3202 return (0); 3203 } else { 3204 int32_t act_len; 3205 3206 /* Hangs over the end return whats left */ 3207 act_len = rsm->r_end - rsm->r_start; 3208 *lenp = (len - act_len); 3209 return (rsm->r_end); 3210 } 3211 /* We don't get out of this block. */ 3212 } 3213 /* 3214 * Here we retransmitted less than the whole thing which means we 3215 * have to split this into what was transmitted and what was not. 3216 */ 3217 nrsm = rack_alloc(rack); 3218 if (nrsm == NULL) { 3219 /* 3220 * We can't get memory, so lets not proceed. 3221 */ 3222 *lenp = 0; 3223 return (0); 3224 } 3225 /* 3226 * So here we are going to take the original rsm and make it what we 3227 * retransmitted. nrsm will be the tail portion we did not 3228 * retransmit. For example say the chunk was 1, 11 (10 bytes). And 3229 * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to 3230 * 1, 6 and the new piece will be 6, 11. 3231 */ 3232 nrsm->r_start = c_end; 3233 nrsm->r_end = rsm->r_end; 3234 nrsm->r_rtr_cnt = rsm->r_rtr_cnt; 3235 nrsm->r_flags = rsm->r_flags; 3236 nrsm->r_sndcnt = rsm->r_sndcnt; 3237 nrsm->r_rtr_bytes = 0; 3238 rsm->r_end = c_end; 3239 for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { 3240 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; 3241 } 3242 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_map, rsm, nrsm, r_next); 3243 if (rsm->r_in_tmap) { 3244 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 3245 nrsm->r_in_tmap = 1; 3246 } 3247 rsm->r_flags &= (~RACK_HAS_FIN); 3248 rack_update_rsm(tp, rack, rsm, ts); 3249 *lenp = 0; 3250 return (0); 3251 } 3252 3253 3254 static void 3255 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len, 3256 uint32_t seq_out, uint8_t th_flags, int32_t err, uint32_t ts, 3257 uint8_t pass, struct rack_sendmap *hintrsm) 3258 { 3259 struct tcp_rack *rack; 3260 struct rack_sendmap *rsm, *nrsm; 3261 register uint32_t snd_max, snd_una; 3262 int32_t idx; 3263 3264 /* 3265 * Add to the RACK log of packets in flight or retransmitted. If 3266 * there is a TS option we will use the TS echoed, if not we will 3267 * grab a TS. 3268 * 3269 * Retransmissions will increment the count and move the ts to its 3270 * proper place. Note that if options do not include TS's then we 3271 * won't be able to effectively use the ACK for an RTT on a retran. 3272 * 3273 * Notes about r_start and r_end. Lets consider a send starting at 3274 * sequence 1 for 10 bytes. In such an example the r_start would be 3275 * 1 (starting sequence) but the r_end would be r_start+len i.e. 11. 3276 * This means that r_end is actually the first sequence for the next 3277 * slot (11). 3278 * 3279 */ 3280 /* 3281 * If err is set what do we do XXXrrs? should we not add the thing? 3282 * -- i.e. return if err != 0 or should we pretend we sent it? -- 3283 * i.e. proceed with add ** do this for now. 3284 */ 3285 INP_WLOCK_ASSERT(tp->t_inpcb); 3286 if (err) 3287 /* 3288 * We don't log errors -- we could but snd_max does not 3289 * advance in this case either. 3290 */ 3291 return; 3292 3293 if (th_flags & TH_RST) { 3294 /* 3295 * We don't log resets and we return immediately from 3296 * sending 3297 */ 3298 return; 3299 } 3300 rack = (struct tcp_rack *)tp->t_fb_ptr; 3301 snd_una = tp->snd_una; 3302 if (SEQ_LEQ((seq_out + len), snd_una)) { 3303 /* Are sending an old segment to induce an ack (keep-alive)? */ 3304 return; 3305 } 3306 if (SEQ_LT(seq_out, snd_una)) { 3307 /* huh? should we panic? */ 3308 uint32_t end; 3309 3310 end = seq_out + len; 3311 seq_out = snd_una; 3312 len = end - seq_out; 3313 } 3314 snd_max = tp->snd_max; 3315 if (th_flags & (TH_SYN | TH_FIN)) { 3316 /* 3317 * The call to rack_log_output is made before bumping 3318 * snd_max. This means we can record one extra byte on a SYN 3319 * or FIN if seq_out is adding more on and a FIN is present 3320 * (and we are not resending). 3321 */ 3322 if (th_flags & TH_SYN) 3323 len++; 3324 if (th_flags & TH_FIN) 3325 len++; 3326 if (SEQ_LT(snd_max, tp->snd_nxt)) { 3327 /* 3328 * The add/update as not been done for the FIN/SYN 3329 * yet. 3330 */ 3331 snd_max = tp->snd_nxt; 3332 } 3333 } 3334 if (len == 0) { 3335 /* We don't log zero window probes */ 3336 return; 3337 } 3338 rack->r_ctl.rc_time_last_sent = ts; 3339 if (IN_RECOVERY(tp->t_flags)) { 3340 rack->r_ctl.rc_prr_out += len; 3341 } 3342 /* First question is it a retransmission? */ 3343 if (seq_out == snd_max) { 3344 again: 3345 rsm = rack_alloc(rack); 3346 if (rsm == NULL) { 3347 /* 3348 * Hmm out of memory and the tcb got destroyed while 3349 * we tried to wait. 3350 */ 3351 #ifdef INVARIANTS 3352 panic("Out of memory when we should not be rack:%p", rack); 3353 #endif 3354 return; 3355 } 3356 if (th_flags & TH_FIN) { 3357 rsm->r_flags = RACK_HAS_FIN; 3358 } else { 3359 rsm->r_flags = 0; 3360 } 3361 rsm->r_tim_lastsent[0] = ts; 3362 rsm->r_rtr_cnt = 1; 3363 rsm->r_rtr_bytes = 0; 3364 if (th_flags & TH_SYN) { 3365 /* The data space is one beyond snd_una */ 3366 rsm->r_start = seq_out + 1; 3367 rsm->r_end = rsm->r_start + (len - 1); 3368 } else { 3369 /* Normal case */ 3370 rsm->r_start = seq_out; 3371 rsm->r_end = rsm->r_start + len; 3372 } 3373 rsm->r_sndcnt = 0; 3374 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_map, rsm, r_next); 3375 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext); 3376 rsm->r_in_tmap = 1; 3377 return; 3378 } 3379 /* 3380 * If we reach here its a retransmission and we need to find it. 3381 */ 3382 more: 3383 if (hintrsm && (hintrsm->r_start == seq_out)) { 3384 rsm = hintrsm; 3385 hintrsm = NULL; 3386 } else if (rack->r_ctl.rc_next) { 3387 /* We have a hint from a previous run */ 3388 rsm = rack->r_ctl.rc_next; 3389 } else { 3390 /* No hints sorry */ 3391 rsm = NULL; 3392 } 3393 if ((rsm) && (rsm->r_start == seq_out)) { 3394 /* 3395 * We used rc_next or hintrsm to retransmit, hopefully the 3396 * likely case. 3397 */ 3398 seq_out = rack_update_entry(tp, rack, rsm, ts, &len); 3399 if (len == 0) { 3400 return; 3401 } else { 3402 goto more; 3403 } 3404 } 3405 /* Ok it was not the last pointer go through it the hard way. */ 3406 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_map, r_next) { 3407 if (rsm->r_start == seq_out) { 3408 seq_out = rack_update_entry(tp, rack, rsm, ts, &len); 3409 rack->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next); 3410 if (len == 0) { 3411 return; 3412 } else { 3413 continue; 3414 } 3415 } 3416 if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) { 3417 /* Transmitted within this piece */ 3418 /* 3419 * Ok we must split off the front and then let the 3420 * update do the rest 3421 */ 3422 nrsm = rack_alloc(rack); 3423 if (nrsm == NULL) { 3424 #ifdef INVARIANTS 3425 panic("Ran out of memory that was preallocated? rack:%p", rack); 3426 #endif 3427 rack_update_rsm(tp, rack, rsm, ts); 3428 return; 3429 } 3430 /* 3431 * copy rsm to nrsm and then trim the front of rsm 3432 * to not include this part. 3433 */ 3434 nrsm->r_start = seq_out; 3435 nrsm->r_end = rsm->r_end; 3436 nrsm->r_rtr_cnt = rsm->r_rtr_cnt; 3437 nrsm->r_flags = rsm->r_flags; 3438 nrsm->r_sndcnt = rsm->r_sndcnt; 3439 nrsm->r_rtr_bytes = 0; 3440 for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { 3441 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; 3442 } 3443 rsm->r_end = nrsm->r_start; 3444 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_map, rsm, nrsm, r_next); 3445 if (rsm->r_in_tmap) { 3446 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 3447 nrsm->r_in_tmap = 1; 3448 } 3449 rsm->r_flags &= (~RACK_HAS_FIN); 3450 seq_out = rack_update_entry(tp, rack, nrsm, ts, &len); 3451 if (len == 0) { 3452 return; 3453 } 3454 } 3455 } 3456 /* 3457 * Hmm not found in map did they retransmit both old and on into the 3458 * new? 3459 */ 3460 if (seq_out == tp->snd_max) { 3461 goto again; 3462 } else if (SEQ_LT(seq_out, tp->snd_max)) { 3463 #ifdef INVARIANTS 3464 printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n", 3465 seq_out, len, tp->snd_una, tp->snd_max); 3466 printf("Starting Dump of all rack entries\n"); 3467 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_map, r_next) { 3468 printf("rsm:%p start:%u end:%u\n", 3469 rsm, rsm->r_start, rsm->r_end); 3470 } 3471 printf("Dump complete\n"); 3472 panic("seq_out not found rack:%p tp:%p", 3473 rack, tp); 3474 #endif 3475 } else { 3476 #ifdef INVARIANTS 3477 /* 3478 * Hmm beyond sndmax? (only if we are using the new rtt-pack 3479 * flag) 3480 */ 3481 panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p", 3482 seq_out, len, tp->snd_max, tp); 3483 #endif 3484 } 3485 } 3486 3487 /* 3488 * Record one of the RTT updates from an ack into 3489 * our sample structure. 3490 */ 3491 static void 3492 tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt) 3493 { 3494 if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) || 3495 (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) { 3496 rack->r_ctl.rack_rs.rs_rtt_lowest = rtt; 3497 } 3498 if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) || 3499 (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) { 3500 rack->r_ctl.rack_rs.rs_rtt_highest = rtt; 3501 } 3502 rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID; 3503 rack->r_ctl.rack_rs.rs_rtt_tot += rtt; 3504 rack->r_ctl.rack_rs.rs_rtt_cnt++; 3505 } 3506 3507 /* 3508 * Collect new round-trip time estimate 3509 * and update averages and current timeout. 3510 */ 3511 static void 3512 tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp) 3513 { 3514 int32_t delta; 3515 uint32_t o_srtt, o_var; 3516 int32_t rtt; 3517 3518 if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) 3519 /* No valid sample */ 3520 return; 3521 if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) { 3522 /* We are to use the lowest RTT seen in a single ack */ 3523 rtt = rack->r_ctl.rack_rs.rs_rtt_lowest; 3524 } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) { 3525 /* We are to use the highest RTT seen in a single ack */ 3526 rtt = rack->r_ctl.rack_rs.rs_rtt_highest; 3527 } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) { 3528 /* We are to use the average RTT seen in a single ack */ 3529 rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot / 3530 (uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt); 3531 } else { 3532 #ifdef INVARIANTS 3533 panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method); 3534 #endif 3535 return; 3536 } 3537 if (rtt == 0) 3538 rtt = 1; 3539 rack_log_rtt_sample(rack, rtt); 3540 o_srtt = tp->t_srtt; 3541 o_var = tp->t_rttvar; 3542 rack = (struct tcp_rack *)tp->t_fb_ptr; 3543 if (tp->t_srtt != 0) { 3544 /* 3545 * srtt is stored as fixed point with 5 bits after the 3546 * binary point (i.e., scaled by 8). The following magic is 3547 * equivalent to the smoothing algorithm in rfc793 with an 3548 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point). 3549 * Adjust rtt to origin 0. 3550 */ 3551 delta = ((rtt - 1) << TCP_DELTA_SHIFT) 3552 - (tp->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT)); 3553 3554 tp->t_srtt += delta; 3555 if (tp->t_srtt <= 0) 3556 tp->t_srtt = 1; 3557 3558 /* 3559 * We accumulate a smoothed rtt variance (actually, a 3560 * smoothed mean difference), then set the retransmit timer 3561 * to smoothed rtt + 4 times the smoothed variance. rttvar 3562 * is stored as fixed point with 4 bits after the binary 3563 * point (scaled by 16). The following is equivalent to 3564 * rfc793 smoothing with an alpha of .75 (rttvar = 3565 * rttvar*3/4 + |delta| / 4). This replaces rfc793's 3566 * wired-in beta. 3567 */ 3568 if (delta < 0) 3569 delta = -delta; 3570 delta -= tp->t_rttvar >> (TCP_RTTVAR_SHIFT - TCP_DELTA_SHIFT); 3571 tp->t_rttvar += delta; 3572 if (tp->t_rttvar <= 0) 3573 tp->t_rttvar = 1; 3574 if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar) 3575 tp->t_rttbest = tp->t_srtt + tp->t_rttvar; 3576 } else { 3577 /* 3578 * No rtt measurement yet - use the unsmoothed rtt. Set the 3579 * variance to half the rtt (so our first retransmit happens 3580 * at 3*rtt). 3581 */ 3582 tp->t_srtt = rtt << TCP_RTT_SHIFT; 3583 tp->t_rttvar = rtt << (TCP_RTTVAR_SHIFT - 1); 3584 tp->t_rttbest = tp->t_srtt + tp->t_rttvar; 3585 } 3586 TCPSTAT_INC(tcps_rttupdated); 3587 rack_log_rtt_upd(tp, rack, rtt, o_srtt, o_var); 3588 tp->t_rttupdated++; 3589 #ifdef NETFLIX_STATS 3590 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt)); 3591 #endif 3592 tp->t_rxtshift = 0; 3593 3594 /* 3595 * the retransmit should happen at rtt + 4 * rttvar. Because of the 3596 * way we do the smoothing, srtt and rttvar will each average +1/2 3597 * tick of bias. When we compute the retransmit timer, we want 1/2 3598 * tick of rounding and 1 extra tick because of +-1/2 tick 3599 * uncertainty in the firing of the timer. The bias will give us 3600 * exactly the 1.5 tick we need. But, because the bias is 3601 * statistical, we have to test that we don't drop below the minimum 3602 * feasible timer (which is 2 ticks). 3603 */ 3604 TCPT_RANGESET(tp->t_rxtcur, TCP_REXMTVAL(tp), 3605 max(MSEC_2_TICKS(rack_rto_min), rtt + 2), MSEC_2_TICKS(rack_rto_max)); 3606 tp->t_softerror = 0; 3607 } 3608 3609 static void 3610 rack_earlier_retran(struct tcpcb *tp, struct rack_sendmap *rsm, 3611 uint32_t t, uint32_t cts) 3612 { 3613 /* 3614 * For this RSM, we acknowledged the data from a previous 3615 * transmission, not the last one we made. This means we did a false 3616 * retransmit. 3617 */ 3618 struct tcp_rack *rack; 3619 3620 if (rsm->r_flags & RACK_HAS_FIN) { 3621 /* 3622 * The sending of the FIN often is multiple sent when we 3623 * have everything outstanding ack'd. We ignore this case 3624 * since its over now. 3625 */ 3626 return; 3627 } 3628 if (rsm->r_flags & RACK_TLP) { 3629 /* 3630 * We expect TLP's to have this occur. 3631 */ 3632 return; 3633 } 3634 rack = (struct tcp_rack *)tp->t_fb_ptr; 3635 /* should we undo cc changes and exit recovery? */ 3636 if (IN_RECOVERY(tp->t_flags)) { 3637 if (rack->r_ctl.rc_rsm_start == rsm->r_start) { 3638 /* 3639 * Undo what we ratched down and exit recovery if 3640 * possible 3641 */ 3642 EXIT_RECOVERY(tp->t_flags); 3643 tp->snd_recover = tp->snd_una; 3644 if (rack->r_ctl.rc_cwnd_at > tp->snd_cwnd) 3645 tp->snd_cwnd = rack->r_ctl.rc_cwnd_at; 3646 if (rack->r_ctl.rc_ssthresh_at > tp->snd_ssthresh) 3647 tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at; 3648 } 3649 } 3650 if (rsm->r_flags & RACK_WAS_SACKPASS) { 3651 /* 3652 * We retransmitted based on a sack and the earlier 3653 * retransmission ack'd it - re-ordering is occuring. 3654 */ 3655 counter_u64_add(rack_reorder_seen, 1); 3656 rack->r_ctl.rc_reorder_ts = cts; 3657 } 3658 counter_u64_add(rack_badfr, 1); 3659 counter_u64_add(rack_badfr_bytes, (rsm->r_end - rsm->r_start)); 3660 } 3661 3662 3663 static int 3664 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack, 3665 struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type) 3666 { 3667 int32_t i; 3668 uint32_t t; 3669 3670 if (rsm->r_flags & RACK_ACKED) 3671 /* Already done */ 3672 return (0); 3673 3674 3675 if ((rsm->r_rtr_cnt == 1) || 3676 ((ack_type == CUM_ACKED) && 3677 (to->to_flags & TOF_TS) && 3678 (to->to_tsecr) && 3679 (rsm->r_tim_lastsent[rsm->r_rtr_cnt - 1] == to->to_tsecr)) 3680 ) { 3681 /* 3682 * We will only find a matching timestamp if its cum-acked. 3683 * But if its only one retransmission its for-sure matching 3684 * :-) 3685 */ 3686 t = cts - rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]; 3687 if ((int)t <= 0) 3688 t = 1; 3689 if (!tp->t_rttlow || tp->t_rttlow > t) 3690 tp->t_rttlow = t; 3691 if (!rack->r_ctl.rc_rack_min_rtt || 3692 SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) { 3693 rack->r_ctl.rc_rack_min_rtt = t; 3694 if (rack->r_ctl.rc_rack_min_rtt == 0) { 3695 rack->r_ctl.rc_rack_min_rtt = 1; 3696 } 3697 } 3698 tcp_rack_xmit_timer(rack, TCP_TS_TO_TICKS(t) + 1); 3699 if ((rsm->r_flags & RACK_TLP) && 3700 (!IN_RECOVERY(tp->t_flags))) { 3701 /* Segment was a TLP and our retrans matched */ 3702 if (rack->r_ctl.rc_tlp_cwnd_reduce) { 3703 rack->r_ctl.rc_rsm_start = tp->snd_max; 3704 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd; 3705 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh; 3706 rack_cong_signal(tp, NULL, CC_NDUPACK); 3707 /* 3708 * When we enter recovery we need to assure 3709 * we send one packet. 3710 */ 3711 rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; 3712 } else 3713 rack->r_ctl.rc_tlp_rtx_out = 0; 3714 } 3715 if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) { 3716 /* New more recent rack_tmit_time */ 3717 rack->r_ctl.rc_rack_tmit_time = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]; 3718 rack->rc_rack_rtt = t; 3719 } 3720 return (1); 3721 } 3722 /* 3723 * We clear the soft/rxtshift since we got an ack. 3724 * There is no assurance we will call the commit() function 3725 * so we need to clear these to avoid incorrect handling. 3726 */ 3727 tp->t_rxtshift = 0; 3728 tp->t_softerror = 0; 3729 if ((to->to_flags & TOF_TS) && 3730 (ack_type == CUM_ACKED) && 3731 (to->to_tsecr) && 3732 ((rsm->r_flags & (RACK_DEFERRED | RACK_OVERMAX)) == 0)) { 3733 /* 3734 * Now which timestamp does it match? In this block the ACK 3735 * must be coming from a previous transmission. 3736 */ 3737 for (i = 0; i < rsm->r_rtr_cnt; i++) { 3738 if (rsm->r_tim_lastsent[i] == to->to_tsecr) { 3739 t = cts - rsm->r_tim_lastsent[i]; 3740 if ((int)t <= 0) 3741 t = 1; 3742 if ((i + 1) < rsm->r_rtr_cnt) { 3743 /* Likely */ 3744 rack_earlier_retran(tp, rsm, t, cts); 3745 } 3746 if (!tp->t_rttlow || tp->t_rttlow > t) 3747 tp->t_rttlow = t; 3748 if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) { 3749 rack->r_ctl.rc_rack_min_rtt = t; 3750 if (rack->r_ctl.rc_rack_min_rtt == 0) { 3751 rack->r_ctl.rc_rack_min_rtt = 1; 3752 } 3753 } 3754 /* 3755 * Note the following calls to 3756 * tcp_rack_xmit_timer() are being commented 3757 * out for now. They give us no more accuracy 3758 * and often lead to a wrong choice. We have 3759 * enough samples that have not been 3760 * retransmitted. I leave the commented out 3761 * code in here in case in the future we 3762 * decide to add it back (though I can't forsee 3763 * doing that). That way we will easily see 3764 * where they need to be placed. 3765 */ 3766 if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, 3767 rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) { 3768 /* New more recent rack_tmit_time */ 3769 rack->r_ctl.rc_rack_tmit_time = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]; 3770 rack->rc_rack_rtt = t; 3771 } 3772 return (1); 3773 } 3774 } 3775 goto ts_not_found; 3776 } else { 3777 /* 3778 * Ok its a SACK block that we retransmitted. or a windows 3779 * machine without timestamps. We can tell nothing from the 3780 * time-stamp since its not there or the time the peer last 3781 * recieved a segment that moved forward its cum-ack point. 3782 */ 3783 ts_not_found: 3784 i = rsm->r_rtr_cnt - 1; 3785 t = cts - rsm->r_tim_lastsent[i]; 3786 if ((int)t <= 0) 3787 t = 1; 3788 if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) { 3789 /* 3790 * We retransmitted and the ack came back in less 3791 * than the smallest rtt we have observed. We most 3792 * likey did an improper retransmit as outlined in 3793 * 4.2 Step 3 point 2 in the rack-draft. 3794 */ 3795 i = rsm->r_rtr_cnt - 2; 3796 t = cts - rsm->r_tim_lastsent[i]; 3797 rack_earlier_retran(tp, rsm, t, cts); 3798 } else if (rack->r_ctl.rc_rack_min_rtt) { 3799 /* 3800 * We retransmitted it and the retransmit did the 3801 * job. 3802 */ 3803 if (!rack->r_ctl.rc_rack_min_rtt || 3804 SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) { 3805 rack->r_ctl.rc_rack_min_rtt = t; 3806 if (rack->r_ctl.rc_rack_min_rtt == 0) { 3807 rack->r_ctl.rc_rack_min_rtt = 1; 3808 } 3809 } 3810 if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, rsm->r_tim_lastsent[i])) { 3811 /* New more recent rack_tmit_time */ 3812 rack->r_ctl.rc_rack_tmit_time = rsm->r_tim_lastsent[i]; 3813 rack->rc_rack_rtt = t; 3814 } 3815 return (1); 3816 } 3817 } 3818 return (0); 3819 } 3820 3821 /* 3822 * Mark the SACK_PASSED flag on all entries prior to rsm send wise. 3823 */ 3824 static void 3825 rack_log_sack_passed(struct tcpcb *tp, 3826 struct tcp_rack *rack, struct rack_sendmap *rsm) 3827 { 3828 struct rack_sendmap *nrsm; 3829 uint32_t ts; 3830 int32_t idx; 3831 3832 idx = rsm->r_rtr_cnt - 1; 3833 ts = rsm->r_tim_lastsent[idx]; 3834 nrsm = rsm; 3835 TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap, 3836 rack_head, r_tnext) { 3837 if (nrsm == rsm) { 3838 /* Skip orginal segment he is acked */ 3839 continue; 3840 } 3841 if (nrsm->r_flags & RACK_ACKED) { 3842 /* Skip ack'd segments */ 3843 continue; 3844 } 3845 idx = nrsm->r_rtr_cnt - 1; 3846 if (ts == nrsm->r_tim_lastsent[idx]) { 3847 /* 3848 * For this case lets use seq no, if we sent in a 3849 * big block (TSO) we would have a bunch of segments 3850 * sent at the same time. 3851 * 3852 * We would only get a report if its SEQ is earlier. 3853 * If we have done multiple retransmits the times 3854 * would not be equal. 3855 */ 3856 if (SEQ_LT(nrsm->r_start, rsm->r_start)) { 3857 nrsm->r_flags |= RACK_SACK_PASSED; 3858 nrsm->r_flags &= ~RACK_WAS_SACKPASS; 3859 } 3860 } else { 3861 /* 3862 * Here they were sent at different times, not a big 3863 * block. Since we transmitted this one later and 3864 * see it sack'd then this must also be missing (or 3865 * we would have gotten a sack block for it) 3866 */ 3867 nrsm->r_flags |= RACK_SACK_PASSED; 3868 nrsm->r_flags &= ~RACK_WAS_SACKPASS; 3869 } 3870 } 3871 } 3872 3873 static uint32_t 3874 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack, 3875 struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts) 3876 { 3877 int32_t idx; 3878 int32_t times = 0; 3879 uint32_t start, end, changed = 0; 3880 struct rack_sendmap *rsm, *nrsm; 3881 int32_t used_ref = 1; 3882 3883 start = sack->start; 3884 end = sack->end; 3885 rsm = *prsm; 3886 if (rsm && SEQ_LT(start, rsm->r_start)) { 3887 TAILQ_FOREACH_REVERSE_FROM(rsm, &rack->r_ctl.rc_map, rack_head, r_next) { 3888 if (SEQ_GEQ(start, rsm->r_start) && 3889 SEQ_LT(start, rsm->r_end)) { 3890 goto do_rest_ofb; 3891 } 3892 } 3893 } 3894 if (rsm == NULL) { 3895 start_at_beginning: 3896 rsm = NULL; 3897 used_ref = 0; 3898 } 3899 /* First lets locate the block where this guy is */ 3900 TAILQ_FOREACH_FROM(rsm, &rack->r_ctl.rc_map, r_next) { 3901 if (SEQ_GEQ(start, rsm->r_start) && 3902 SEQ_LT(start, rsm->r_end)) { 3903 break; 3904 } 3905 } 3906 do_rest_ofb: 3907 if (rsm == NULL) { 3908 /* 3909 * This happens when we get duplicate sack blocks with the 3910 * same end. For example SACK 4: 100 SACK 3: 100 The sort 3911 * will not change there location so we would just start at 3912 * the end of the first one and get lost. 3913 */ 3914 if (tp->t_flags & TF_SENTFIN) { 3915 /* 3916 * Check to see if we have not logged the FIN that 3917 * went out. 3918 */ 3919 nrsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_map, rack_sendmap, r_next); 3920 if (nrsm && (nrsm->r_end + 1) == tp->snd_max) { 3921 /* 3922 * Ok we did not get the FIN logged. 3923 */ 3924 nrsm->r_end++; 3925 rsm = nrsm; 3926 goto do_rest_ofb; 3927 } 3928 } 3929 if (times == 1) { 3930 #ifdef INVARIANTS 3931 panic("tp:%p rack:%p sack:%p to:%p prsm:%p", 3932 tp, rack, sack, to, prsm); 3933 #else 3934 goto out; 3935 #endif 3936 } 3937 times++; 3938 counter_u64_add(rack_sack_proc_restart, 1); 3939 goto start_at_beginning; 3940 } 3941 /* Ok we have an ACK for some piece of rsm */ 3942 if (rsm->r_start != start) { 3943 /* 3944 * Need to split this in two pieces the before and after. 3945 */ 3946 nrsm = rack_alloc(rack); 3947 if (nrsm == NULL) { 3948 /* 3949 * failed XXXrrs what can we do but loose the sack 3950 * info? 3951 */ 3952 goto out; 3953 } 3954 nrsm->r_start = start; 3955 nrsm->r_rtr_bytes = 0; 3956 nrsm->r_end = rsm->r_end; 3957 nrsm->r_rtr_cnt = rsm->r_rtr_cnt; 3958 nrsm->r_flags = rsm->r_flags; 3959 nrsm->r_sndcnt = rsm->r_sndcnt; 3960 for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { 3961 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; 3962 } 3963 rsm->r_end = nrsm->r_start; 3964 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_map, rsm, nrsm, r_next); 3965 if (rsm->r_in_tmap) { 3966 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 3967 nrsm->r_in_tmap = 1; 3968 } 3969 rsm->r_flags &= (~RACK_HAS_FIN); 3970 rsm = nrsm; 3971 } 3972 if (SEQ_GEQ(end, rsm->r_end)) { 3973 /* 3974 * The end of this block is either beyond this guy or right 3975 * at this guy. 3976 */ 3977 3978 if ((rsm->r_flags & RACK_ACKED) == 0) { 3979 rack_update_rtt(tp, rack, rsm, to, cts, SACKED); 3980 changed += (rsm->r_end - rsm->r_start); 3981 rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start); 3982 rack_log_sack_passed(tp, rack, rsm); 3983 /* Is Reordering occuring? */ 3984 if (rsm->r_flags & RACK_SACK_PASSED) { 3985 counter_u64_add(rack_reorder_seen, 1); 3986 rack->r_ctl.rc_reorder_ts = cts; 3987 } 3988 rsm->r_flags |= RACK_ACKED; 3989 rsm->r_flags &= ~RACK_TLP; 3990 if (rsm->r_in_tmap) { 3991 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext); 3992 rsm->r_in_tmap = 0; 3993 } 3994 } 3995 if (end == rsm->r_end) { 3996 /* This block only - done */ 3997 goto out; 3998 } 3999 /* There is more not coverend by this rsm move on */ 4000 start = rsm->r_end; 4001 nrsm = TAILQ_NEXT(rsm, r_next); 4002 rsm = nrsm; 4003 times = 0; 4004 goto do_rest_ofb; 4005 } 4006 /* Ok we need to split off this one at the tail */ 4007 nrsm = rack_alloc(rack); 4008 if (nrsm == NULL) { 4009 /* failed rrs what can we do but loose the sack info? */ 4010 goto out; 4011 } 4012 /* Clone it */ 4013 nrsm->r_start = end; 4014 nrsm->r_end = rsm->r_end; 4015 nrsm->r_rtr_bytes = 0; 4016 nrsm->r_rtr_cnt = rsm->r_rtr_cnt; 4017 nrsm->r_flags = rsm->r_flags; 4018 nrsm->r_sndcnt = rsm->r_sndcnt; 4019 for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { 4020 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; 4021 } 4022 /* The sack block does not cover this guy fully */ 4023 rsm->r_flags &= (~RACK_HAS_FIN); 4024 rsm->r_end = end; 4025 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_map, rsm, nrsm, r_next); 4026 if (rsm->r_in_tmap) { 4027 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 4028 nrsm->r_in_tmap = 1; 4029 } 4030 if (rsm->r_flags & RACK_ACKED) { 4031 /* Been here done that */ 4032 goto out; 4033 } 4034 rack_update_rtt(tp, rack, rsm, to, cts, SACKED); 4035 changed += (rsm->r_end - rsm->r_start); 4036 rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start); 4037 rack_log_sack_passed(tp, rack, rsm); 4038 /* Is Reordering occuring? */ 4039 if (rsm->r_flags & RACK_SACK_PASSED) { 4040 counter_u64_add(rack_reorder_seen, 1); 4041 rack->r_ctl.rc_reorder_ts = cts; 4042 } 4043 rsm->r_flags |= RACK_ACKED; 4044 rsm->r_flags &= ~RACK_TLP; 4045 if (rsm->r_in_tmap) { 4046 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext); 4047 rsm->r_in_tmap = 0; 4048 } 4049 out: 4050 if (used_ref == 0) { 4051 counter_u64_add(rack_sack_proc_all, 1); 4052 } else { 4053 counter_u64_add(rack_sack_proc_short, 1); 4054 } 4055 /* Save off where we last were */ 4056 if (rsm) 4057 rack->r_ctl.rc_sacklast = TAILQ_NEXT(rsm, r_next); 4058 else 4059 rack->r_ctl.rc_sacklast = NULL; 4060 *prsm = rsm; 4061 return (changed); 4062 } 4063 4064 static void inline 4065 rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack) 4066 { 4067 struct rack_sendmap *tmap; 4068 4069 tmap = NULL; 4070 while (rsm && (rsm->r_flags & RACK_ACKED)) { 4071 /* Its no longer sacked, mark it so */ 4072 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); 4073 #ifdef INVARIANTS 4074 if (rsm->r_in_tmap) { 4075 panic("rack:%p rsm:%p flags:0x%x in tmap?", 4076 rack, rsm, rsm->r_flags); 4077 } 4078 #endif 4079 rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS); 4080 /* Rebuild it into our tmap */ 4081 if (tmap == NULL) { 4082 TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext); 4083 tmap = rsm; 4084 } else { 4085 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext); 4086 tmap = rsm; 4087 } 4088 tmap->r_in_tmap = 1; 4089 rsm = TAILQ_NEXT(rsm, r_next); 4090 } 4091 /* 4092 * Now lets possibly clear the sack filter so we start 4093 * recognizing sacks that cover this area. 4094 */ 4095 if (rack_use_sack_filter) 4096 sack_filter_clear(&rack->r_ctl.rack_sf, th_ack); 4097 4098 } 4099 4100 static void 4101 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th) 4102 { 4103 uint32_t changed, last_seq, entered_recovery = 0; 4104 struct tcp_rack *rack; 4105 struct rack_sendmap *rsm; 4106 struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1]; 4107 register uint32_t th_ack; 4108 int32_t i, j, k, num_sack_blks = 0; 4109 uint32_t cts, acked, ack_point, sack_changed = 0; 4110 4111 INP_WLOCK_ASSERT(tp->t_inpcb); 4112 if (th->th_flags & TH_RST) { 4113 /* We don't log resets */ 4114 return; 4115 } 4116 rack = (struct tcp_rack *)tp->t_fb_ptr; 4117 cts = tcp_ts_getticks(); 4118 rsm = TAILQ_FIRST(&rack->r_ctl.rc_map); 4119 changed = 0; 4120 th_ack = th->th_ack; 4121 4122 if (SEQ_GT(th_ack, tp->snd_una)) { 4123 rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__); 4124 tp->t_acktime = ticks; 4125 } 4126 if (rsm && SEQ_GT(th_ack, rsm->r_start)) 4127 changed = th_ack - rsm->r_start; 4128 if (changed) { 4129 /* 4130 * The ACK point is advancing to th_ack, we must drop off 4131 * the packets in the rack log and calculate any eligble 4132 * RTT's. 4133 */ 4134 rack->r_wanted_output++; 4135 more: 4136 rsm = TAILQ_FIRST(&rack->r_ctl.rc_map); 4137 if (rsm == NULL) { 4138 if ((th_ack - 1) == tp->iss) { 4139 /* 4140 * For the SYN incoming case we will not 4141 * have called tcp_output for the sending of 4142 * the SYN, so there will be no map. All 4143 * other cases should probably be a panic. 4144 */ 4145 goto proc_sack; 4146 } 4147 if (tp->t_flags & TF_SENTFIN) { 4148 /* if we send a FIN we will not hav a map */ 4149 goto proc_sack; 4150 } 4151 #ifdef INVARIANTS 4152 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", 4153 tp, 4154 th, tp->t_state, rack, 4155 tp->snd_una, tp->snd_max, tp->snd_nxt, changed); 4156 #endif 4157 goto proc_sack; 4158 } 4159 if (SEQ_LT(th_ack, rsm->r_start)) { 4160 /* Huh map is missing this */ 4161 #ifdef INVARIANTS 4162 printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n", 4163 rsm->r_start, 4164 th_ack, tp->t_state, rack->r_state); 4165 #endif 4166 goto proc_sack; 4167 } 4168 rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED); 4169 /* Now do we consume the whole thing? */ 4170 if (SEQ_GEQ(th_ack, rsm->r_end)) { 4171 /* Its all consumed. */ 4172 uint32_t left; 4173 4174 rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; 4175 rsm->r_rtr_bytes = 0; 4176 TAILQ_REMOVE(&rack->r_ctl.rc_map, rsm, r_next); 4177 if (rsm->r_in_tmap) { 4178 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext); 4179 rsm->r_in_tmap = 0; 4180 } 4181 if (rack->r_ctl.rc_next == rsm) { 4182 /* scoot along the marker */ 4183 rack->r_ctl.rc_next = TAILQ_FIRST(&rack->r_ctl.rc_map); 4184 } 4185 if (rsm->r_flags & RACK_ACKED) { 4186 /* 4187 * It was acked on the scoreboard -- remove 4188 * it from total 4189 */ 4190 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); 4191 } else if (rsm->r_flags & RACK_SACK_PASSED) { 4192 /* 4193 * There are acked segments ACKED on the 4194 * scoreboard further up. We are seeing 4195 * reordering. 4196 */ 4197 counter_u64_add(rack_reorder_seen, 1); 4198 rsm->r_flags |= RACK_ACKED; 4199 rack->r_ctl.rc_reorder_ts = cts; 4200 } 4201 left = th_ack - rsm->r_end; 4202 if (rsm->r_rtr_cnt > 1) { 4203 /* 4204 * Technically we should make r_rtr_cnt be 4205 * monotonicly increasing and just mod it to 4206 * the timestamp it is replacing.. that way 4207 * we would have the last 3 retransmits. Now 4208 * rc_loss_count will be wrong if we 4209 * retransmit something more than 2 times in 4210 * recovery :( 4211 */ 4212 rack->r_ctl.rc_loss_count += (rsm->r_rtr_cnt - 1); 4213 } 4214 /* Free back to zone */ 4215 rack_free(rack, rsm); 4216 if (left) { 4217 goto more; 4218 } 4219 goto proc_sack; 4220 } 4221 if (rsm->r_flags & RACK_ACKED) { 4222 /* 4223 * It was acked on the scoreboard -- remove it from 4224 * total for the part being cum-acked. 4225 */ 4226 rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start); 4227 } 4228 rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; 4229 rsm->r_rtr_bytes = 0; 4230 rsm->r_start = th_ack; 4231 } 4232 proc_sack: 4233 /* Check for reneging */ 4234 rsm = TAILQ_FIRST(&rack->r_ctl.rc_map); 4235 if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) { 4236 /* 4237 * The peer has moved snd_una up to 4238 * the edge of this send, i.e. one 4239 * that it had previously acked. The only 4240 * way that can be true if the peer threw 4241 * away data (space issues) that it had 4242 * previously sacked (else it would have 4243 * given us snd_una up to (rsm->r_end). 4244 * We need to undo the acked markings here. 4245 * 4246 * Note we have to look to make sure th_ack is 4247 * our rsm->r_start in case we get an old ack 4248 * where th_ack is behind snd_una. 4249 */ 4250 rack_peer_reneges(rack, rsm, th->th_ack); 4251 } 4252 if ((to->to_flags & TOF_SACK) == 0) { 4253 /* We are done nothing left to log */ 4254 goto out; 4255 } 4256 rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_map, rack_sendmap, r_next); 4257 if (rsm) { 4258 last_seq = rsm->r_end; 4259 } else { 4260 last_seq = tp->snd_max; 4261 } 4262 /* Sack block processing */ 4263 if (SEQ_GT(th_ack, tp->snd_una)) 4264 ack_point = th_ack; 4265 else 4266 ack_point = tp->snd_una; 4267 for (i = 0; i < to->to_nsacks; i++) { 4268 bcopy((to->to_sacks + i * TCPOLEN_SACK), 4269 &sack, sizeof(sack)); 4270 sack.start = ntohl(sack.start); 4271 sack.end = ntohl(sack.end); 4272 if (SEQ_GT(sack.end, sack.start) && 4273 SEQ_GT(sack.start, ack_point) && 4274 SEQ_LT(sack.start, tp->snd_max) && 4275 SEQ_GT(sack.end, ack_point) && 4276 SEQ_LEQ(sack.end, tp->snd_max)) { 4277 if ((rack->r_ctl.rc_num_maps_alloced > rack_sack_block_limit) && 4278 (SEQ_LT(sack.end, last_seq)) && 4279 ((sack.end - sack.start) < (tp->t_maxseg / 8))) { 4280 /* 4281 * Not the last piece and its smaller than 4282 * 1/8th of a MSS. We ignore this. 4283 */ 4284 counter_u64_add(rack_runt_sacks, 1); 4285 continue; 4286 } 4287 sack_blocks[num_sack_blks] = sack; 4288 num_sack_blks++; 4289 #ifdef NETFLIX_STATS 4290 } else if (SEQ_LEQ(sack.start, th_ack) && 4291 SEQ_LEQ(sack.end, th_ack)) { 4292 /* 4293 * Its a D-SACK block. 4294 */ 4295 tcp_record_dsack(sack.start, sack.end); 4296 #endif 4297 } 4298 4299 } 4300 if (num_sack_blks == 0) 4301 goto out; 4302 /* 4303 * Sort the SACK blocks so we can update the rack scoreboard with 4304 * just one pass. 4305 */ 4306 if (rack_use_sack_filter) { 4307 num_sack_blks = sack_filter_blks(&rack->r_ctl.rack_sf, sack_blocks, num_sack_blks, th->th_ack); 4308 } 4309 if (num_sack_blks < 2) { 4310 goto do_sack_work; 4311 } 4312 /* Sort the sacks */ 4313 for (i = 0; i < num_sack_blks; i++) { 4314 for (j = i + 1; j < num_sack_blks; j++) { 4315 if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) { 4316 sack = sack_blocks[i]; 4317 sack_blocks[i] = sack_blocks[j]; 4318 sack_blocks[j] = sack; 4319 } 4320 } 4321 } 4322 /* 4323 * Now are any of the sack block ends the same (yes some 4324 * implememtations send these)? 4325 */ 4326 again: 4327 if (num_sack_blks > 1) { 4328 for (i = 0; i < num_sack_blks; i++) { 4329 for (j = i + 1; j < num_sack_blks; j++) { 4330 if (sack_blocks[i].end == sack_blocks[j].end) { 4331 /* 4332 * Ok these two have the same end we 4333 * want the smallest end and then 4334 * throw away the larger and start 4335 * again. 4336 */ 4337 if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) { 4338 /* 4339 * The second block covers 4340 * more area use that 4341 */ 4342 sack_blocks[i].start = sack_blocks[j].start; 4343 } 4344 /* 4345 * Now collapse out the dup-sack and 4346 * lower the count 4347 */ 4348 for (k = (j + 1); k < num_sack_blks; k++) { 4349 sack_blocks[j].start = sack_blocks[k].start; 4350 sack_blocks[j].end = sack_blocks[k].end; 4351 j++; 4352 } 4353 num_sack_blks--; 4354 goto again; 4355 } 4356 } 4357 } 4358 } 4359 do_sack_work: 4360 rsm = rack->r_ctl.rc_sacklast; 4361 for (i = 0; i < num_sack_blks; i++) { 4362 acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts); 4363 if (acked) { 4364 rack->r_wanted_output++; 4365 changed += acked; 4366 sack_changed += acked; 4367 } 4368 } 4369 out: 4370 if (changed) { 4371 /* Something changed cancel the rack timer */ 4372 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 4373 } 4374 if ((sack_changed) && (!IN_RECOVERY(tp->t_flags))) { 4375 /* 4376 * Ok we have a high probability that we need to go in to 4377 * recovery since we have data sack'd 4378 */ 4379 struct rack_sendmap *rsm; 4380 uint32_t tsused; 4381 4382 tsused = tcp_ts_getticks(); 4383 rsm = tcp_rack_output(tp, rack, tsused); 4384 if (rsm) { 4385 /* Enter recovery */ 4386 rack->r_ctl.rc_rsm_start = rsm->r_start; 4387 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd; 4388 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh; 4389 entered_recovery = 1; 4390 rack_cong_signal(tp, NULL, CC_NDUPACK); 4391 /* 4392 * When we enter recovery we need to assure we send 4393 * one packet. 4394 */ 4395 rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; 4396 rack->r_timer_override = 1; 4397 } 4398 } 4399 if (IN_RECOVERY(tp->t_flags) && (entered_recovery == 0)) { 4400 /* Deal with changed an PRR here (in recovery only) */ 4401 uint32_t pipe, snd_una; 4402 4403 rack->r_ctl.rc_prr_delivered += changed; 4404 /* Compute prr_sndcnt */ 4405 if (SEQ_GT(tp->snd_una, th_ack)) { 4406 snd_una = tp->snd_una; 4407 } else { 4408 snd_una = th_ack; 4409 } 4410 pipe = ((tp->snd_max - snd_una) - rack->r_ctl.rc_sacked) + rack->r_ctl.rc_holes_rxt; 4411 if (pipe > tp->snd_ssthresh) { 4412 long sndcnt; 4413 4414 sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh; 4415 if (rack->r_ctl.rc_prr_recovery_fs > 0) 4416 sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs; 4417 else { 4418 rack->r_ctl.rc_prr_sndcnt = 0; 4419 sndcnt = 0; 4420 } 4421 sndcnt++; 4422 if (sndcnt > (long)rack->r_ctl.rc_prr_out) 4423 sndcnt -= rack->r_ctl.rc_prr_out; 4424 else 4425 sndcnt = 0; 4426 rack->r_ctl.rc_prr_sndcnt = sndcnt; 4427 } else { 4428 uint32_t limit; 4429 4430 if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out) 4431 limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out); 4432 else 4433 limit = 0; 4434 if (changed > limit) 4435 limit = changed; 4436 limit += tp->t_maxseg; 4437 if (tp->snd_ssthresh > pipe) { 4438 rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit); 4439 } else { 4440 rack->r_ctl.rc_prr_sndcnt = min(0, limit); 4441 } 4442 } 4443 if (rack->r_ctl.rc_prr_sndcnt >= tp->t_maxseg) { 4444 rack->r_timer_override = 1; 4445 } 4446 } 4447 } 4448 4449 /* 4450 * Return value of 1, we do not need to call rack_process_data(). 4451 * return value of 0, rack_process_data can be called. 4452 * For ret_val if its 0 the TCP is locked, if its non-zero 4453 * its unlocked and probably unsafe to touch the TCB. 4454 */ 4455 static int 4456 rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so, 4457 struct tcpcb *tp, struct tcpopt *to, 4458 uint32_t tiwin, int32_t tlen, 4459 int32_t * ofia, int32_t thflags, int32_t * ret_val) 4460 { 4461 int32_t ourfinisacked = 0; 4462 int32_t nsegs, acked_amount; 4463 int32_t acked; 4464 struct mbuf *mfree; 4465 struct tcp_rack *rack; 4466 int32_t recovery = 0; 4467 4468 rack = (struct tcp_rack *)tp->t_fb_ptr; 4469 if (SEQ_GT(th->th_ack, tp->snd_max)) { 4470 rack_do_dropafterack(m, tp, th, thflags, tlen, ret_val); 4471 return (1); 4472 } 4473 if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) { 4474 rack_log_ack(tp, to, th); 4475 } 4476 if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) { 4477 /* 4478 * Old ack, behind (or duplicate to) the last one rcv'd 4479 * Note: Should mark reordering is occuring! We should also 4480 * look for sack blocks arriving e.g. ack 1, 4-4 then ack 1, 4481 * 3-3, 4-4 would be reording. As well as ack 1, 3-3 <no 4482 * retran and> ack 3 4483 */ 4484 return (0); 4485 } 4486 /* 4487 * If we reach this point, ACK is not a duplicate, i.e., it ACKs 4488 * something we sent. 4489 */ 4490 if (tp->t_flags & TF_NEEDSYN) { 4491 /* 4492 * T/TCP: Connection was half-synchronized, and our SYN has 4493 * been ACK'd (so connection is now fully synchronized). Go 4494 * to non-starred state, increment snd_una for ACK of SYN, 4495 * and check if we can do window scaling. 4496 */ 4497 tp->t_flags &= ~TF_NEEDSYN; 4498 tp->snd_una++; 4499 /* Do window scaling? */ 4500 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == 4501 (TF_RCVD_SCALE | TF_REQ_SCALE)) { 4502 tp->rcv_scale = tp->request_r_scale; 4503 /* Send window already scaled. */ 4504 } 4505 } 4506 nsegs = max(1, m->m_pkthdr.lro_nsegs); 4507 INP_WLOCK_ASSERT(tp->t_inpcb); 4508 4509 acked = BYTES_THIS_ACK(tp, th); 4510 TCPSTAT_ADD(tcps_rcvackpack, nsegs); 4511 TCPSTAT_ADD(tcps_rcvackbyte, acked); 4512 4513 /* 4514 * If we just performed our first retransmit, and the ACK arrives 4515 * within our recovery window, then it was a mistake to do the 4516 * retransmit in the first place. Recover our original cwnd and 4517 * ssthresh, and proceed to transmit where we left off. 4518 */ 4519 if (tp->t_flags & TF_PREVVALID) { 4520 tp->t_flags &= ~TF_PREVVALID; 4521 if (tp->t_rxtshift == 1 && 4522 (int)(ticks - tp->t_badrxtwin) < 0) 4523 rack_cong_signal(tp, th, CC_RTO_ERR); 4524 } 4525 /* 4526 * If we have a timestamp reply, update smoothed round trip time. If 4527 * no timestamp is present but transmit timer is running and timed 4528 * sequence number was acked, update smoothed round trip time. Since 4529 * we now have an rtt measurement, cancel the timer backoff (cf., 4530 * Phil Karn's retransmit alg.). Recompute the initial retransmit 4531 * timer. 4532 * 4533 * Some boxes send broken timestamp replies during the SYN+ACK 4534 * phase, ignore timestamps of 0 or we could calculate a huge RTT 4535 * and blow up the retransmit timer. 4536 */ 4537 /* 4538 * If all outstanding data is acked, stop retransmit timer and 4539 * remember to restart (more output or persist). If there is more 4540 * data to be acked, restart retransmit timer, using current 4541 * (possibly backed-off) value. 4542 */ 4543 if (th->th_ack == tp->snd_max) { 4544 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 4545 rack->r_wanted_output++; 4546 } 4547 /* 4548 * If no data (only SYN) was ACK'd, skip rest of ACK processing. 4549 */ 4550 if (acked == 0) { 4551 if (ofia) 4552 *ofia = ourfinisacked; 4553 return (0); 4554 } 4555 if (rack->r_ctl.rc_early_recovery) { 4556 if (IN_FASTRECOVERY(tp->t_flags)) { 4557 if (SEQ_LT(th->th_ack, tp->snd_recover)) { 4558 tcp_rack_partialack(tp, th); 4559 } else { 4560 rack_post_recovery(tp, th); 4561 recovery = 1; 4562 } 4563 } 4564 } 4565 /* 4566 * Let the congestion control algorithm update congestion control 4567 * related information. This typically means increasing the 4568 * congestion window. 4569 */ 4570 rack_ack_received(tp, rack, th, nsegs, CC_ACK, recovery); 4571 SOCKBUF_LOCK(&so->so_snd); 4572 acked_amount = min(acked, (int)sbavail(&so->so_snd)); 4573 tp->snd_wnd -= acked_amount; 4574 mfree = sbcut_locked(&so->so_snd, acked_amount); 4575 if ((sbused(&so->so_snd) == 0) && 4576 (acked > acked_amount) && 4577 (tp->t_state >= TCPS_FIN_WAIT_1)) { 4578 ourfinisacked = 1; 4579 } 4580 /* NB: sowwakeup_locked() does an implicit unlock. */ 4581 sowwakeup_locked(so); 4582 m_freem(mfree); 4583 if (rack->r_ctl.rc_early_recovery == 0) { 4584 if (IN_FASTRECOVERY(tp->t_flags)) { 4585 if (SEQ_LT(th->th_ack, tp->snd_recover)) { 4586 tcp_rack_partialack(tp, th); 4587 } else { 4588 rack_post_recovery(tp, th); 4589 } 4590 } 4591 } 4592 tp->snd_una = th->th_ack; 4593 if (SEQ_GT(tp->snd_una, tp->snd_recover)) 4594 tp->snd_recover = tp->snd_una; 4595 4596 if (SEQ_LT(tp->snd_nxt, tp->snd_una)) { 4597 tp->snd_nxt = tp->snd_una; 4598 } 4599 if (tp->snd_una == tp->snd_max) { 4600 /* Nothing left outstanding */ 4601 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__); 4602 tp->t_acktime = 0; 4603 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 4604 /* Set need output so persist might get set */ 4605 rack->r_wanted_output++; 4606 if (rack_use_sack_filter) 4607 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una); 4608 if ((tp->t_state >= TCPS_FIN_WAIT_1) && 4609 (sbavail(&so->so_snd) == 0) && 4610 (tp->t_flags2 & TF2_DROP_AF_DATA)) { 4611 /* 4612 * The socket was gone and the 4613 * peer sent data, time to 4614 * reset him. 4615 */ 4616 *ret_val = 1; 4617 tp = tcp_close(tp); 4618 rack_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen); 4619 return (1); 4620 } 4621 } 4622 if (ofia) 4623 *ofia = ourfinisacked; 4624 return (0); 4625 } 4626 4627 4628 /* 4629 * Return value of 1, the TCB is unlocked and most 4630 * likely gone, return value of 0, the TCP is still 4631 * locked. 4632 */ 4633 static int 4634 rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so, 4635 struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, 4636 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 4637 { 4638 /* 4639 * Update window information. Don't look at window if no ACK: TAC's 4640 * send garbage on first SYN. 4641 */ 4642 int32_t nsegs; 4643 int32_t tfo_syn; 4644 struct tcp_rack *rack; 4645 4646 rack = (struct tcp_rack *)tp->t_fb_ptr; 4647 INP_WLOCK_ASSERT(tp->t_inpcb); 4648 nsegs = max(1, m->m_pkthdr.lro_nsegs); 4649 if ((thflags & TH_ACK) && 4650 (SEQ_LT(tp->snd_wl1, th->th_seq) || 4651 (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) || 4652 (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) { 4653 /* keep track of pure window updates */ 4654 if (tlen == 0 && 4655 tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd) 4656 TCPSTAT_INC(tcps_rcvwinupd); 4657 tp->snd_wnd = tiwin; 4658 tp->snd_wl1 = th->th_seq; 4659 tp->snd_wl2 = th->th_ack; 4660 if (tp->snd_wnd > tp->max_sndwnd) 4661 tp->max_sndwnd = tp->snd_wnd; 4662 rack->r_wanted_output++; 4663 } else if (thflags & TH_ACK) { 4664 if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) { 4665 tp->snd_wnd = tiwin; 4666 tp->snd_wl1 = th->th_seq; 4667 tp->snd_wl2 = th->th_ack; 4668 } 4669 } 4670 /* Was persist timer active and now we have window space? */ 4671 if ((rack->rc_in_persist != 0) && tp->snd_wnd) { 4672 rack_exit_persist(tp, rack); 4673 tp->snd_nxt = tp->snd_max; 4674 /* Make sure we output to start the timer */ 4675 rack->r_wanted_output++; 4676 } 4677 if (tp->t_flags2 & TF2_DROP_AF_DATA) { 4678 m_freem(m); 4679 return (0); 4680 } 4681 /* 4682 * Process segments with URG. 4683 */ 4684 if ((thflags & TH_URG) && th->th_urp && 4685 TCPS_HAVERCVDFIN(tp->t_state) == 0) { 4686 /* 4687 * This is a kludge, but if we receive and accept random 4688 * urgent pointers, we'll crash in soreceive. It's hard to 4689 * imagine someone actually wanting to send this much urgent 4690 * data. 4691 */ 4692 SOCKBUF_LOCK(&so->so_rcv); 4693 if (th->th_urp + sbavail(&so->so_rcv) > sb_max) { 4694 th->th_urp = 0; /* XXX */ 4695 thflags &= ~TH_URG; /* XXX */ 4696 SOCKBUF_UNLOCK(&so->so_rcv); /* XXX */ 4697 goto dodata; /* XXX */ 4698 } 4699 /* 4700 * If this segment advances the known urgent pointer, then 4701 * mark the data stream. This should not happen in 4702 * CLOSE_WAIT, CLOSING, LAST_ACK or TIME_WAIT STATES since a 4703 * FIN has been received from the remote side. In these 4704 * states we ignore the URG. 4705 * 4706 * According to RFC961 (Assigned Protocols), the urgent 4707 * pointer points to the last octet of urgent data. We 4708 * continue, however, to consider it to indicate the first 4709 * octet of data past the urgent section as the original 4710 * spec states (in one of two places). 4711 */ 4712 if (SEQ_GT(th->th_seq + th->th_urp, tp->rcv_up)) { 4713 tp->rcv_up = th->th_seq + th->th_urp; 4714 so->so_oobmark = sbavail(&so->so_rcv) + 4715 (tp->rcv_up - tp->rcv_nxt) - 1; 4716 if (so->so_oobmark == 0) 4717 so->so_rcv.sb_state |= SBS_RCVATMARK; 4718 sohasoutofband(so); 4719 tp->t_oobflags &= ~(TCPOOB_HAVEDATA | TCPOOB_HADDATA); 4720 } 4721 SOCKBUF_UNLOCK(&so->so_rcv); 4722 /* 4723 * Remove out of band data so doesn't get presented to user. 4724 * This can happen independent of advancing the URG pointer, 4725 * but if two URG's are pending at once, some out-of-band 4726 * data may creep in... ick. 4727 */ 4728 if (th->th_urp <= (uint32_t) tlen && 4729 !(so->so_options & SO_OOBINLINE)) { 4730 /* hdr drop is delayed */ 4731 tcp_pulloutofband(so, th, m, drop_hdrlen); 4732 } 4733 } else { 4734 /* 4735 * If no out of band data is expected, pull receive urgent 4736 * pointer along with the receive window. 4737 */ 4738 if (SEQ_GT(tp->rcv_nxt, tp->rcv_up)) 4739 tp->rcv_up = tp->rcv_nxt; 4740 } 4741 dodata: /* XXX */ 4742 INP_WLOCK_ASSERT(tp->t_inpcb); 4743 4744 /* 4745 * Process the segment text, merging it into the TCP sequencing 4746 * queue, and arranging for acknowledgment of receipt if necessary. 4747 * This process logically involves adjusting tp->rcv_wnd as data is 4748 * presented to the user (this happens in tcp_usrreq.c, case 4749 * PRU_RCVD). If a FIN has already been received on this connection 4750 * then we just ignore the text. 4751 */ 4752 tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) && 4753 IS_FASTOPEN(tp->t_flags)); 4754 if ((tlen || (thflags & TH_FIN) || tfo_syn) && 4755 TCPS_HAVERCVDFIN(tp->t_state) == 0) { 4756 tcp_seq save_start = th->th_seq; 4757 4758 m_adj(m, drop_hdrlen); /* delayed header drop */ 4759 /* 4760 * Insert segment which includes th into TCP reassembly 4761 * queue with control block tp. Set thflags to whether 4762 * reassembly now includes a segment with FIN. This handles 4763 * the common case inline (segment is the next to be 4764 * received on an established connection, and the queue is 4765 * empty), avoiding linkage into and removal from the queue 4766 * and repetition of various conversions. Set DELACK for 4767 * segments received in order, but ack immediately when 4768 * segments are out of order (so fast retransmit can work). 4769 */ 4770 if (th->th_seq == tp->rcv_nxt && 4771 SEGQ_EMPTY(tp) && 4772 (TCPS_HAVEESTABLISHED(tp->t_state) || 4773 tfo_syn)) { 4774 if (DELAY_ACK(tp, tlen) || tfo_syn) { 4775 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 4776 tp->t_flags |= TF_DELACK; 4777 } else { 4778 rack->r_wanted_output++; 4779 tp->t_flags |= TF_ACKNOW; 4780 } 4781 tp->rcv_nxt += tlen; 4782 thflags = th->th_flags & TH_FIN; 4783 TCPSTAT_ADD(tcps_rcvpack, nsegs); 4784 TCPSTAT_ADD(tcps_rcvbyte, tlen); 4785 SOCKBUF_LOCK(&so->so_rcv); 4786 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) 4787 m_freem(m); 4788 else 4789 sbappendstream_locked(&so->so_rcv, m, 0); 4790 /* NB: sorwakeup_locked() does an implicit unlock. */ 4791 sorwakeup_locked(so); 4792 } else { 4793 /* 4794 * XXX: Due to the header drop above "th" is 4795 * theoretically invalid by now. Fortunately 4796 * m_adj() doesn't actually frees any mbufs when 4797 * trimming from the head. 4798 */ 4799 thflags = tcp_reass(tp, th, &save_start, &tlen, m); 4800 tp->t_flags |= TF_ACKNOW; 4801 } 4802 if (tlen > 0) 4803 tcp_update_sack_list(tp, save_start, save_start + tlen); 4804 } else { 4805 m_freem(m); 4806 thflags &= ~TH_FIN; 4807 } 4808 4809 /* 4810 * If FIN is received ACK the FIN and let the user know that the 4811 * connection is closing. 4812 */ 4813 if (thflags & TH_FIN) { 4814 if (TCPS_HAVERCVDFIN(tp->t_state) == 0) { 4815 socantrcvmore(so); 4816 /* 4817 * If connection is half-synchronized (ie NEEDSYN 4818 * flag on) then delay ACK, so it may be piggybacked 4819 * when SYN is sent. Otherwise, since we received a 4820 * FIN then no more input can be expected, send ACK 4821 * now. 4822 */ 4823 if (tp->t_flags & TF_NEEDSYN) { 4824 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 4825 tp->t_flags |= TF_DELACK; 4826 } else { 4827 tp->t_flags |= TF_ACKNOW; 4828 } 4829 tp->rcv_nxt++; 4830 } 4831 switch (tp->t_state) { 4832 4833 /* 4834 * In SYN_RECEIVED and ESTABLISHED STATES enter the 4835 * CLOSE_WAIT state. 4836 */ 4837 case TCPS_SYN_RECEIVED: 4838 tp->t_starttime = ticks; 4839 /* FALLTHROUGH */ 4840 case TCPS_ESTABLISHED: 4841 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 4842 tcp_state_change(tp, TCPS_CLOSE_WAIT); 4843 break; 4844 4845 /* 4846 * If still in FIN_WAIT_1 STATE FIN has not been 4847 * acked so enter the CLOSING state. 4848 */ 4849 case TCPS_FIN_WAIT_1: 4850 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 4851 tcp_state_change(tp, TCPS_CLOSING); 4852 break; 4853 4854 /* 4855 * In FIN_WAIT_2 state enter the TIME_WAIT state, 4856 * starting the time-wait timer, turning off the 4857 * other standard timers. 4858 */ 4859 case TCPS_FIN_WAIT_2: 4860 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 4861 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 4862 tcp_twstart(tp); 4863 return (1); 4864 } 4865 } 4866 /* 4867 * Return any desired output. 4868 */ 4869 if ((tp->t_flags & TF_ACKNOW) || (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) { 4870 rack->r_wanted_output++; 4871 } 4872 INP_WLOCK_ASSERT(tp->t_inpcb); 4873 return (0); 4874 } 4875 4876 /* 4877 * Here nothing is really faster, its just that we 4878 * have broken out the fast-data path also just like 4879 * the fast-ack. 4880 */ 4881 static int 4882 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so, 4883 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 4884 uint32_t tiwin, int32_t nxt_pkt) 4885 { 4886 int32_t nsegs; 4887 int32_t newsize = 0; /* automatic sockbuf scaling */ 4888 struct tcp_rack *rack; 4889 #ifdef TCPDEBUG 4890 /* 4891 * The size of tcp_saveipgen must be the size of the max ip header, 4892 * now IPv6. 4893 */ 4894 u_char tcp_saveipgen[IP6_HDR_LEN]; 4895 struct tcphdr tcp_savetcp; 4896 short ostate = 0; 4897 4898 #endif 4899 /* 4900 * If last ACK falls within this segment's sequence numbers, record 4901 * the timestamp. NOTE that the test is modified according to the 4902 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26). 4903 */ 4904 if (__predict_false(th->th_seq != tp->rcv_nxt)) { 4905 return (0); 4906 } 4907 if (__predict_false(tp->snd_nxt != tp->snd_max)) { 4908 return (0); 4909 } 4910 if (tiwin && tiwin != tp->snd_wnd) { 4911 return (0); 4912 } 4913 if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) { 4914 return (0); 4915 } 4916 if (__predict_false((to->to_flags & TOF_TS) && 4917 (TSTMP_LT(to->to_tsval, tp->ts_recent)))) { 4918 return (0); 4919 } 4920 if (__predict_false((th->th_ack != tp->snd_una))) { 4921 return (0); 4922 } 4923 if (__predict_false(tlen > sbspace(&so->so_rcv))) { 4924 return (0); 4925 } 4926 if ((to->to_flags & TOF_TS) != 0 && 4927 SEQ_LEQ(th->th_seq, tp->last_ack_sent)) { 4928 tp->ts_recent_age = tcp_ts_getticks(); 4929 tp->ts_recent = to->to_tsval; 4930 } 4931 rack = (struct tcp_rack *)tp->t_fb_ptr; 4932 /* 4933 * This is a pure, in-sequence data packet with nothing on the 4934 * reassembly queue and we have enough buffer space to take it. 4935 */ 4936 nsegs = max(1, m->m_pkthdr.lro_nsegs); 4937 4938 4939 /* Clean receiver SACK report if present */ 4940 if (tp->rcv_numsacks) 4941 tcp_clean_sackreport(tp); 4942 TCPSTAT_INC(tcps_preddat); 4943 tp->rcv_nxt += tlen; 4944 /* 4945 * Pull snd_wl1 up to prevent seq wrap relative to th_seq. 4946 */ 4947 tp->snd_wl1 = th->th_seq; 4948 /* 4949 * Pull rcv_up up to prevent seq wrap relative to rcv_nxt. 4950 */ 4951 tp->rcv_up = tp->rcv_nxt; 4952 TCPSTAT_ADD(tcps_rcvpack, nsegs); 4953 TCPSTAT_ADD(tcps_rcvbyte, tlen); 4954 #ifdef TCPDEBUG 4955 if (so->so_options & SO_DEBUG) 4956 tcp_trace(TA_INPUT, ostate, tp, 4957 (void *)tcp_saveipgen, &tcp_savetcp, 0); 4958 #endif 4959 newsize = tcp_autorcvbuf(m, th, so, tp, tlen); 4960 4961 /* Add data to socket buffer. */ 4962 SOCKBUF_LOCK(&so->so_rcv); 4963 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 4964 m_freem(m); 4965 } else { 4966 /* 4967 * Set new socket buffer size. Give up when limit is 4968 * reached. 4969 */ 4970 if (newsize) 4971 if (!sbreserve_locked(&so->so_rcv, 4972 newsize, so, NULL)) 4973 so->so_rcv.sb_flags &= ~SB_AUTOSIZE; 4974 m_adj(m, drop_hdrlen); /* delayed header drop */ 4975 sbappendstream_locked(&so->so_rcv, m, 0); 4976 rack_calc_rwin(so, tp); 4977 } 4978 /* NB: sorwakeup_locked() does an implicit unlock. */ 4979 sorwakeup_locked(so); 4980 if (DELAY_ACK(tp, tlen)) { 4981 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 4982 tp->t_flags |= TF_DELACK; 4983 } else { 4984 tp->t_flags |= TF_ACKNOW; 4985 rack->r_wanted_output++; 4986 } 4987 if ((tp->snd_una == tp->snd_max) && rack_use_sack_filter) 4988 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una); 4989 return (1); 4990 } 4991 4992 /* 4993 * This subfunction is used to try to highly optimize the 4994 * fast path. We again allow window updates that are 4995 * in sequence to remain in the fast-path. We also add 4996 * in the __predict's to attempt to help the compiler. 4997 * Note that if we return a 0, then we can *not* process 4998 * it and the caller should push the packet into the 4999 * slow-path. 5000 */ 5001 static int 5002 rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so, 5003 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 5004 uint32_t tiwin, int32_t nxt_pkt, uint32_t cts) 5005 { 5006 int32_t acked; 5007 int32_t nsegs; 5008 5009 #ifdef TCPDEBUG 5010 /* 5011 * The size of tcp_saveipgen must be the size of the max ip header, 5012 * now IPv6. 5013 */ 5014 u_char tcp_saveipgen[IP6_HDR_LEN]; 5015 struct tcphdr tcp_savetcp; 5016 short ostate = 0; 5017 5018 #endif 5019 struct tcp_rack *rack; 5020 5021 if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) { 5022 /* Old ack, behind (or duplicate to) the last one rcv'd */ 5023 return (0); 5024 } 5025 if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) { 5026 /* Above what we have sent? */ 5027 return (0); 5028 } 5029 if (__predict_false(tp->snd_nxt != tp->snd_max)) { 5030 /* We are retransmitting */ 5031 return (0); 5032 } 5033 if (__predict_false(tiwin == 0)) { 5034 /* zero window */ 5035 return (0); 5036 } 5037 if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) { 5038 /* We need a SYN or a FIN, unlikely.. */ 5039 return (0); 5040 } 5041 if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) { 5042 /* Timestamp is behind .. old ack with seq wrap? */ 5043 return (0); 5044 } 5045 if (__predict_false(IN_RECOVERY(tp->t_flags))) { 5046 /* Still recovering */ 5047 return (0); 5048 } 5049 rack = (struct tcp_rack *)tp->t_fb_ptr; 5050 if (rack->r_ctl.rc_sacked) { 5051 /* We have sack holes on our scoreboard */ 5052 return (0); 5053 } 5054 /* Ok if we reach here, we can process a fast-ack */ 5055 nsegs = max(1, m->m_pkthdr.lro_nsegs); 5056 rack_log_ack(tp, to, th); 5057 /* Did the window get updated? */ 5058 if (tiwin != tp->snd_wnd) { 5059 tp->snd_wnd = tiwin; 5060 tp->snd_wl1 = th->th_seq; 5061 if (tp->snd_wnd > tp->max_sndwnd) 5062 tp->max_sndwnd = tp->snd_wnd; 5063 } 5064 if ((rack->rc_in_persist != 0) && (tp->snd_wnd >= tp->t_maxseg)) { 5065 rack_exit_persist(tp, rack); 5066 } 5067 /* 5068 * If last ACK falls within this segment's sequence numbers, record 5069 * the timestamp. NOTE that the test is modified according to the 5070 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26). 5071 */ 5072 if ((to->to_flags & TOF_TS) != 0 && 5073 SEQ_LEQ(th->th_seq, tp->last_ack_sent)) { 5074 tp->ts_recent_age = tcp_ts_getticks(); 5075 tp->ts_recent = to->to_tsval; 5076 } 5077 /* 5078 * This is a pure ack for outstanding data. 5079 */ 5080 TCPSTAT_INC(tcps_predack); 5081 5082 /* 5083 * "bad retransmit" recovery. 5084 */ 5085 if (tp->t_flags & TF_PREVVALID) { 5086 tp->t_flags &= ~TF_PREVVALID; 5087 if (tp->t_rxtshift == 1 && 5088 (int)(ticks - tp->t_badrxtwin) < 0) 5089 rack_cong_signal(tp, th, CC_RTO_ERR); 5090 } 5091 /* 5092 * Recalculate the transmit timer / rtt. 5093 * 5094 * Some boxes send broken timestamp replies during the SYN+ACK 5095 * phase, ignore timestamps of 0 or we could calculate a huge RTT 5096 * and blow up the retransmit timer. 5097 */ 5098 acked = BYTES_THIS_ACK(tp, th); 5099 5100 #ifdef TCP_HHOOK 5101 /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */ 5102 hhook_run_tcp_est_in(tp, th, to); 5103 #endif 5104 5105 TCPSTAT_ADD(tcps_rcvackpack, nsegs); 5106 TCPSTAT_ADD(tcps_rcvackbyte, acked); 5107 sbdrop(&so->so_snd, acked); 5108 /* 5109 * Let the congestion control algorithm update congestion control 5110 * related information. This typically means increasing the 5111 * congestion window. 5112 */ 5113 rack_ack_received(tp, rack, th, nsegs, CC_ACK, 0); 5114 5115 tp->snd_una = th->th_ack; 5116 /* 5117 * Pull snd_wl2 up to prevent seq wrap relative to th_ack. 5118 */ 5119 tp->snd_wl2 = th->th_ack; 5120 tp->t_dupacks = 0; 5121 m_freem(m); 5122 /* ND6_HINT(tp); *//* Some progress has been made. */ 5123 5124 /* 5125 * If all outstanding data are acked, stop retransmit timer, 5126 * otherwise restart timer using current (possibly backed-off) 5127 * value. If process is waiting for space, wakeup/selwakeup/signal. 5128 * If data are ready to send, let tcp_output decide between more 5129 * output or persist. 5130 */ 5131 #ifdef TCPDEBUG 5132 if (so->so_options & SO_DEBUG) 5133 tcp_trace(TA_INPUT, ostate, tp, 5134 (void *)tcp_saveipgen, 5135 &tcp_savetcp, 0); 5136 #endif 5137 if (tp->snd_una == tp->snd_max) { 5138 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__); 5139 tp->t_acktime = 0; 5140 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 5141 } 5142 /* Wake up the socket if we have room to write more */ 5143 sowwakeup(so); 5144 if (sbavail(&so->so_snd)) { 5145 rack->r_wanted_output++; 5146 } 5147 return (1); 5148 } 5149 5150 /* 5151 * Return value of 1, the TCB is unlocked and most 5152 * likely gone, return value of 0, the TCP is still 5153 * locked. 5154 */ 5155 static int 5156 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so, 5157 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 5158 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 5159 { 5160 int32_t ret_val = 0; 5161 int32_t todrop; 5162 int32_t ourfinisacked = 0; 5163 5164 rack_calc_rwin(so, tp); 5165 /* 5166 * If the state is SYN_SENT: if seg contains an ACK, but not for our 5167 * SYN, drop the input. if seg contains a RST, then drop the 5168 * connection. if seg does not contain SYN, then drop it. Otherwise 5169 * this is an acceptable SYN segment initialize tp->rcv_nxt and 5170 * tp->irs if seg contains ack then advance tp->snd_una if seg 5171 * contains an ECE and ECN support is enabled, the stream is ECN 5172 * capable. if SYN has been acked change to ESTABLISHED else 5173 * SYN_RCVD state arrange for segment to be acked (eventually) 5174 * continue processing rest of data/controls, beginning with URG 5175 */ 5176 if ((thflags & TH_ACK) && 5177 (SEQ_LEQ(th->th_ack, tp->iss) || 5178 SEQ_GT(th->th_ack, tp->snd_max))) { 5179 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 5180 return (1); 5181 } 5182 if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) { 5183 TCP_PROBE5(connect__refused, NULL, tp, 5184 mtod(m, const char *), tp, th); 5185 tp = tcp_drop(tp, ECONNREFUSED); 5186 rack_do_drop(m, tp); 5187 return (1); 5188 } 5189 if (thflags & TH_RST) { 5190 rack_do_drop(m, tp); 5191 return (1); 5192 } 5193 if (!(thflags & TH_SYN)) { 5194 rack_do_drop(m, tp); 5195 return (1); 5196 } 5197 tp->irs = th->th_seq; 5198 tcp_rcvseqinit(tp); 5199 if (thflags & TH_ACK) { 5200 int tfo_partial = 0; 5201 5202 TCPSTAT_INC(tcps_connects); 5203 soisconnected(so); 5204 #ifdef MAC 5205 mac_socketpeer_set_from_mbuf(m, so); 5206 #endif 5207 /* Do window scaling on this connection? */ 5208 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == 5209 (TF_RCVD_SCALE | TF_REQ_SCALE)) { 5210 tp->rcv_scale = tp->request_r_scale; 5211 } 5212 tp->rcv_adv += min(tp->rcv_wnd, 5213 TCP_MAXWIN << tp->rcv_scale); 5214 /* 5215 * If not all the data that was sent in the TFO SYN 5216 * has been acked, resend the remainder right away. 5217 */ 5218 if (IS_FASTOPEN(tp->t_flags) && 5219 (tp->snd_una != tp->snd_max)) { 5220 tp->snd_nxt = th->th_ack; 5221 tfo_partial = 1; 5222 } 5223 /* 5224 * If there's data, delay ACK; if there's also a FIN ACKNOW 5225 * will be turned on later. 5226 */ 5227 if (DELAY_ACK(tp, tlen) && tlen != 0 && (tfo_partial == 0)) { 5228 rack_timer_cancel(tp, (struct tcp_rack *)tp->t_fb_ptr, 5229 ((struct tcp_rack *)tp->t_fb_ptr)->r_ctl.rc_rcvtime, __LINE__); 5230 tp->t_flags |= TF_DELACK; 5231 } else { 5232 ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output++; 5233 tp->t_flags |= TF_ACKNOW; 5234 } 5235 5236 if ((thflags & TH_ECE) && V_tcp_do_ecn) { 5237 tp->t_flags |= TF_ECN_PERMIT; 5238 TCPSTAT_INC(tcps_ecn_shs); 5239 } 5240 if (SEQ_GT(th->th_ack, tp->snd_una)) { 5241 /* 5242 * We advance snd_una for the 5243 * fast open case. If th_ack is 5244 * acknowledging data beyond 5245 * snd_una we can't just call 5246 * ack-processing since the 5247 * data stream in our send-map 5248 * will start at snd_una + 1 (one 5249 * beyond the SYN). If its just 5250 * equal we don't need to do that 5251 * and there is no send_map. 5252 */ 5253 tp->snd_una++; 5254 } 5255 /* 5256 * Received <SYN,ACK> in SYN_SENT[*] state. Transitions: 5257 * SYN_SENT --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1 5258 */ 5259 tp->t_starttime = ticks; 5260 if (tp->t_flags & TF_NEEDFIN) { 5261 tcp_state_change(tp, TCPS_FIN_WAIT_1); 5262 tp->t_flags &= ~TF_NEEDFIN; 5263 thflags &= ~TH_SYN; 5264 } else { 5265 tcp_state_change(tp, TCPS_ESTABLISHED); 5266 TCP_PROBE5(connect__established, NULL, tp, 5267 mtod(m, const char *), tp, th); 5268 cc_conn_init(tp); 5269 } 5270 } else { 5271 /* 5272 * Received initial SYN in SYN-SENT[*] state => simultaneous 5273 * open. If segment contains CC option and there is a 5274 * cached CC, apply TAO test. If it succeeds, connection is * 5275 * half-synchronized. Otherwise, do 3-way handshake: 5276 * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If 5277 * there was no CC option, clear cached CC value. 5278 */ 5279 tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN); 5280 tcp_state_change(tp, TCPS_SYN_RECEIVED); 5281 } 5282 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 5283 INP_WLOCK_ASSERT(tp->t_inpcb); 5284 /* 5285 * Advance th->th_seq to correspond to first data byte. If data, 5286 * trim to stay within window, dropping FIN if necessary. 5287 */ 5288 th->th_seq++; 5289 if (tlen > tp->rcv_wnd) { 5290 todrop = tlen - tp->rcv_wnd; 5291 m_adj(m, -todrop); 5292 tlen = tp->rcv_wnd; 5293 thflags &= ~TH_FIN; 5294 TCPSTAT_INC(tcps_rcvpackafterwin); 5295 TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop); 5296 } 5297 tp->snd_wl1 = th->th_seq - 1; 5298 tp->rcv_up = th->th_seq; 5299 /* 5300 * Client side of transaction: already sent SYN and data. If the 5301 * remote host used T/TCP to validate the SYN, our data will be 5302 * ACK'd; if so, enter normal data segment processing in the middle 5303 * of step 5, ack processing. Otherwise, goto step 6. 5304 */ 5305 if (thflags & TH_ACK) { 5306 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) 5307 return (ret_val); 5308 /* We may have changed to FIN_WAIT_1 above */ 5309 if (tp->t_state == TCPS_FIN_WAIT_1) { 5310 /* 5311 * In FIN_WAIT_1 STATE in addition to the processing 5312 * for the ESTABLISHED state if our FIN is now 5313 * acknowledged then enter FIN_WAIT_2. 5314 */ 5315 if (ourfinisacked) { 5316 /* 5317 * If we can't receive any more data, then 5318 * closing user can proceed. Starting the 5319 * timer is contrary to the specification, 5320 * but if we don't get a FIN we'll hang 5321 * forever. 5322 * 5323 * XXXjl: we should release the tp also, and 5324 * use a compressed state. 5325 */ 5326 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 5327 soisdisconnected(so); 5328 tcp_timer_activate(tp, TT_2MSL, 5329 (tcp_fast_finwait2_recycle ? 5330 tcp_finwait2_timeout : 5331 TP_MAXIDLE(tp))); 5332 } 5333 tcp_state_change(tp, TCPS_FIN_WAIT_2); 5334 } 5335 } 5336 } 5337 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5338 tiwin, thflags, nxt_pkt)); 5339 } 5340 5341 /* 5342 * Return value of 1, the TCB is unlocked and most 5343 * likely gone, return value of 0, the TCP is still 5344 * locked. 5345 */ 5346 static int 5347 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so, 5348 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 5349 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 5350 { 5351 int32_t ret_val = 0; 5352 int32_t ourfinisacked = 0; 5353 5354 rack_calc_rwin(so, tp); 5355 5356 if ((thflags & TH_ACK) && 5357 (SEQ_LEQ(th->th_ack, tp->snd_una) || 5358 SEQ_GT(th->th_ack, tp->snd_max))) { 5359 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 5360 return (1); 5361 } 5362 if (IS_FASTOPEN(tp->t_flags)) { 5363 /* 5364 * When a TFO connection is in SYN_RECEIVED, the 5365 * only valid packets are the initial SYN, a 5366 * retransmit/copy of the initial SYN (possibly with 5367 * a subset of the original data), a valid ACK, a 5368 * FIN, or a RST. 5369 */ 5370 if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) { 5371 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 5372 return (1); 5373 } else if (thflags & TH_SYN) { 5374 /* non-initial SYN is ignored */ 5375 struct tcp_rack *rack; 5376 5377 rack = (struct tcp_rack *)tp->t_fb_ptr; 5378 if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) || 5379 (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) || 5380 (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) { 5381 rack_do_drop(m, NULL); 5382 return (0); 5383 } 5384 } else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) { 5385 rack_do_drop(m, NULL); 5386 return (0); 5387 } 5388 } 5389 if (thflags & TH_RST) 5390 return (rack_process_rst(m, th, so, tp)); 5391 /* 5392 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 5393 * it's less than ts_recent, drop it. 5394 */ 5395 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 5396 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 5397 if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val)) 5398 return (ret_val); 5399 } 5400 /* 5401 * In the SYN-RECEIVED state, validate that the packet belongs to 5402 * this connection before trimming the data to fit the receive 5403 * window. Check the sequence number versus IRS since we know the 5404 * sequence numbers haven't wrapped. This is a partial fix for the 5405 * "LAND" DoS attack. 5406 */ 5407 if (SEQ_LT(th->th_seq, tp->irs)) { 5408 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 5409 return (1); 5410 } 5411 if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 5412 return (ret_val); 5413 } 5414 /* 5415 * If last ACK falls within this segment's sequence numbers, record 5416 * its timestamp. NOTE: 1) That the test incorporates suggestions 5417 * from the latest proposal of the tcplw@cray.com list (Braden 5418 * 1993/04/26). 2) That updating only on newer timestamps interferes 5419 * with our earlier PAWS tests, so this check should be solely 5420 * predicated on the sequence space of this segment. 3) That we 5421 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 5422 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 5423 * SEG.Len, This modified check allows us to overcome RFC1323's 5424 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 5425 * p.869. In such cases, we can still calculate the RTT correctly 5426 * when RCV.NXT == Last.ACK.Sent. 5427 */ 5428 if ((to->to_flags & TOF_TS) != 0 && 5429 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 5430 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 5431 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 5432 tp->ts_recent_age = tcp_ts_getticks(); 5433 tp->ts_recent = to->to_tsval; 5434 } 5435 /* 5436 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 5437 * is on (half-synchronized state), then queue data for later 5438 * processing; else drop segment and return. 5439 */ 5440 if ((thflags & TH_ACK) == 0) { 5441 if (IS_FASTOPEN(tp->t_flags)) { 5442 tp->snd_wnd = tiwin; 5443 cc_conn_init(tp); 5444 } 5445 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5446 tiwin, thflags, nxt_pkt)); 5447 } 5448 TCPSTAT_INC(tcps_connects); 5449 soisconnected(so); 5450 /* Do window scaling? */ 5451 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == 5452 (TF_RCVD_SCALE | TF_REQ_SCALE)) { 5453 tp->rcv_scale = tp->request_r_scale; 5454 tp->snd_wnd = tiwin; 5455 } 5456 /* 5457 * Make transitions: SYN-RECEIVED -> ESTABLISHED SYN-RECEIVED* -> 5458 * FIN-WAIT-1 5459 */ 5460 tp->t_starttime = ticks; 5461 if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) { 5462 tcp_fastopen_decrement_counter(tp->t_tfo_pending); 5463 tp->t_tfo_pending = NULL; 5464 5465 /* 5466 * Account for the ACK of our SYN prior to 5467 * regular ACK processing below. 5468 */ 5469 tp->snd_una++; 5470 } 5471 if (tp->t_flags & TF_NEEDFIN) { 5472 tcp_state_change(tp, TCPS_FIN_WAIT_1); 5473 tp->t_flags &= ~TF_NEEDFIN; 5474 } else { 5475 tcp_state_change(tp, TCPS_ESTABLISHED); 5476 TCP_PROBE5(accept__established, NULL, tp, 5477 mtod(m, const char *), tp, th); 5478 /* 5479 * TFO connections call cc_conn_init() during SYN 5480 * processing. Calling it again here for such connections 5481 * is not harmless as it would undo the snd_cwnd reduction 5482 * that occurs when a TFO SYN|ACK is retransmitted. 5483 */ 5484 if (!IS_FASTOPEN(tp->t_flags)) 5485 cc_conn_init(tp); 5486 } 5487 /* 5488 * If segment contains data or ACK, will call tcp_reass() later; if 5489 * not, do so now to pass queued data to user. 5490 */ 5491 if (tlen == 0 && (thflags & TH_FIN) == 0) 5492 (void) tcp_reass(tp, (struct tcphdr *)0, NULL, 0, 5493 (struct mbuf *)0); 5494 tp->snd_wl1 = th->th_seq - 1; 5495 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 5496 return (ret_val); 5497 } 5498 if (tp->t_state == TCPS_FIN_WAIT_1) { 5499 /* We could have went to FIN_WAIT_1 (or EST) above */ 5500 /* 5501 * In FIN_WAIT_1 STATE in addition to the processing for the 5502 * ESTABLISHED state if our FIN is now acknowledged then 5503 * enter FIN_WAIT_2. 5504 */ 5505 if (ourfinisacked) { 5506 /* 5507 * If we can't receive any more data, then closing 5508 * user can proceed. Starting the timer is contrary 5509 * to the specification, but if we don't get a FIN 5510 * we'll hang forever. 5511 * 5512 * XXXjl: we should release the tp also, and use a 5513 * compressed state. 5514 */ 5515 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 5516 soisdisconnected(so); 5517 tcp_timer_activate(tp, TT_2MSL, 5518 (tcp_fast_finwait2_recycle ? 5519 tcp_finwait2_timeout : 5520 TP_MAXIDLE(tp))); 5521 } 5522 tcp_state_change(tp, TCPS_FIN_WAIT_2); 5523 } 5524 } 5525 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5526 tiwin, thflags, nxt_pkt)); 5527 } 5528 5529 /* 5530 * Return value of 1, the TCB is unlocked and most 5531 * likely gone, return value of 0, the TCP is still 5532 * locked. 5533 */ 5534 static int 5535 rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so, 5536 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 5537 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 5538 { 5539 int32_t ret_val = 0; 5540 5541 /* 5542 * Header prediction: check for the two common cases of a 5543 * uni-directional data xfer. If the packet has no control flags, 5544 * is in-sequence, the window didn't change and we're not 5545 * retransmitting, it's a candidate. If the length is zero and the 5546 * ack moved forward, we're the sender side of the xfer. Just free 5547 * the data acked & wake any higher level process that was blocked 5548 * waiting for space. If the length is non-zero and the ack didn't 5549 * move, we're the receiver side. If we're getting packets in-order 5550 * (the reassembly queue is empty), add the data toc The socket 5551 * buffer and note that we need a delayed ack. Make sure that the 5552 * hidden state-flags are also off. Since we check for 5553 * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN. 5554 */ 5555 if (__predict_true(((to->to_flags & TOF_SACK) == 0)) && 5556 __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_URG | TH_ACK)) == TH_ACK) && 5557 __predict_true(SEGQ_EMPTY(tp)) && 5558 __predict_true(th->th_seq == tp->rcv_nxt)) { 5559 struct tcp_rack *rack; 5560 5561 rack = (struct tcp_rack *)tp->t_fb_ptr; 5562 if (tlen == 0) { 5563 if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen, 5564 tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) { 5565 return (0); 5566 } 5567 } else { 5568 if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen, 5569 tiwin, nxt_pkt)) { 5570 return (0); 5571 } 5572 } 5573 } 5574 rack_calc_rwin(so, tp); 5575 5576 if (thflags & TH_RST) 5577 return (rack_process_rst(m, th, so, tp)); 5578 5579 /* 5580 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 5581 * synchronized state. 5582 */ 5583 if (thflags & TH_SYN) { 5584 rack_challenge_ack(m, th, tp, &ret_val); 5585 return (ret_val); 5586 } 5587 /* 5588 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 5589 * it's less than ts_recent, drop it. 5590 */ 5591 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 5592 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 5593 if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val)) 5594 return (ret_val); 5595 } 5596 if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 5597 return (ret_val); 5598 } 5599 /* 5600 * If last ACK falls within this segment's sequence numbers, record 5601 * its timestamp. NOTE: 1) That the test incorporates suggestions 5602 * from the latest proposal of the tcplw@cray.com list (Braden 5603 * 1993/04/26). 2) That updating only on newer timestamps interferes 5604 * with our earlier PAWS tests, so this check should be solely 5605 * predicated on the sequence space of this segment. 3) That we 5606 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 5607 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 5608 * SEG.Len, This modified check allows us to overcome RFC1323's 5609 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 5610 * p.869. In such cases, we can still calculate the RTT correctly 5611 * when RCV.NXT == Last.ACK.Sent. 5612 */ 5613 if ((to->to_flags & TOF_TS) != 0 && 5614 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 5615 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 5616 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 5617 tp->ts_recent_age = tcp_ts_getticks(); 5618 tp->ts_recent = to->to_tsval; 5619 } 5620 /* 5621 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 5622 * is on (half-synchronized state), then queue data for later 5623 * processing; else drop segment and return. 5624 */ 5625 if ((thflags & TH_ACK) == 0) { 5626 if (tp->t_flags & TF_NEEDSYN) { 5627 5628 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5629 tiwin, thflags, nxt_pkt)); 5630 5631 } else if (tp->t_flags & TF_ACKNOW) { 5632 rack_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 5633 return (ret_val); 5634 } else { 5635 rack_do_drop(m, NULL); 5636 return (0); 5637 } 5638 } 5639 /* 5640 * Ack processing. 5641 */ 5642 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) { 5643 return (ret_val); 5644 } 5645 if (sbavail(&so->so_snd)) { 5646 if (rack_progress_timeout_check(tp)) { 5647 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); 5648 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 5649 return (1); 5650 } 5651 } 5652 /* State changes only happen in rack_process_data() */ 5653 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5654 tiwin, thflags, nxt_pkt)); 5655 } 5656 5657 /* 5658 * Return value of 1, the TCB is unlocked and most 5659 * likely gone, return value of 0, the TCP is still 5660 * locked. 5661 */ 5662 static int 5663 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so, 5664 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 5665 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 5666 { 5667 int32_t ret_val = 0; 5668 5669 rack_calc_rwin(so, tp); 5670 if (thflags & TH_RST) 5671 return (rack_process_rst(m, th, so, tp)); 5672 /* 5673 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 5674 * synchronized state. 5675 */ 5676 if (thflags & TH_SYN) { 5677 rack_challenge_ack(m, th, tp, &ret_val); 5678 return (ret_val); 5679 } 5680 /* 5681 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 5682 * it's less than ts_recent, drop it. 5683 */ 5684 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 5685 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 5686 if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val)) 5687 return (ret_val); 5688 } 5689 if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 5690 return (ret_val); 5691 } 5692 /* 5693 * If last ACK falls within this segment's sequence numbers, record 5694 * its timestamp. NOTE: 1) That the test incorporates suggestions 5695 * from the latest proposal of the tcplw@cray.com list (Braden 5696 * 1993/04/26). 2) That updating only on newer timestamps interferes 5697 * with our earlier PAWS tests, so this check should be solely 5698 * predicated on the sequence space of this segment. 3) That we 5699 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 5700 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 5701 * SEG.Len, This modified check allows us to overcome RFC1323's 5702 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 5703 * p.869. In such cases, we can still calculate the RTT correctly 5704 * when RCV.NXT == Last.ACK.Sent. 5705 */ 5706 if ((to->to_flags & TOF_TS) != 0 && 5707 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 5708 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 5709 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 5710 tp->ts_recent_age = tcp_ts_getticks(); 5711 tp->ts_recent = to->to_tsval; 5712 } 5713 /* 5714 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 5715 * is on (half-synchronized state), then queue data for later 5716 * processing; else drop segment and return. 5717 */ 5718 if ((thflags & TH_ACK) == 0) { 5719 if (tp->t_flags & TF_NEEDSYN) { 5720 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5721 tiwin, thflags, nxt_pkt)); 5722 5723 } else if (tp->t_flags & TF_ACKNOW) { 5724 rack_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 5725 return (ret_val); 5726 } else { 5727 rack_do_drop(m, NULL); 5728 return (0); 5729 } 5730 } 5731 /* 5732 * Ack processing. 5733 */ 5734 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) { 5735 return (ret_val); 5736 } 5737 if (sbavail(&so->so_snd)) { 5738 if (rack_progress_timeout_check(tp)) { 5739 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); 5740 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 5741 return (1); 5742 } 5743 } 5744 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5745 tiwin, thflags, nxt_pkt)); 5746 } 5747 5748 static int 5749 rack_check_data_after_close(struct mbuf *m, 5750 struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so) 5751 { 5752 struct tcp_rack *rack; 5753 5754 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 5755 rack = (struct tcp_rack *)tp->t_fb_ptr; 5756 if (rack->rc_allow_data_af_clo == 0) { 5757 close_now: 5758 tp = tcp_close(tp); 5759 TCPSTAT_INC(tcps_rcvafterclose); 5760 rack_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen)); 5761 return (1); 5762 } 5763 if (sbavail(&so->so_snd) == 0) 5764 goto close_now; 5765 /* Ok we allow data that is ignored and a followup reset */ 5766 tp->rcv_nxt = th->th_seq + *tlen; 5767 tp->t_flags2 |= TF2_DROP_AF_DATA; 5768 rack->r_wanted_output = 1; 5769 *tlen = 0; 5770 return (0); 5771 } 5772 5773 /* 5774 * Return value of 1, the TCB is unlocked and most 5775 * likely gone, return value of 0, the TCP is still 5776 * locked. 5777 */ 5778 static int 5779 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so, 5780 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 5781 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 5782 { 5783 int32_t ret_val = 0; 5784 int32_t ourfinisacked = 0; 5785 5786 rack_calc_rwin(so, tp); 5787 5788 if (thflags & TH_RST) 5789 return (rack_process_rst(m, th, so, tp)); 5790 /* 5791 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 5792 * synchronized state. 5793 */ 5794 if (thflags & TH_SYN) { 5795 rack_challenge_ack(m, th, tp, &ret_val); 5796 return (ret_val); 5797 } 5798 /* 5799 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 5800 * it's less than ts_recent, drop it. 5801 */ 5802 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 5803 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 5804 if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val)) 5805 return (ret_val); 5806 } 5807 if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 5808 return (ret_val); 5809 } 5810 /* 5811 * If new data are received on a connection after the user processes 5812 * are gone, then RST the other end. 5813 */ 5814 if ((so->so_state & SS_NOFDREF) && tlen) { 5815 if (rack_check_data_after_close(m, tp, &tlen, th, so)) 5816 return (1); 5817 } 5818 /* 5819 * If last ACK falls within this segment's sequence numbers, record 5820 * its timestamp. NOTE: 1) That the test incorporates suggestions 5821 * from the latest proposal of the tcplw@cray.com list (Braden 5822 * 1993/04/26). 2) That updating only on newer timestamps interferes 5823 * with our earlier PAWS tests, so this check should be solely 5824 * predicated on the sequence space of this segment. 3) That we 5825 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 5826 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 5827 * SEG.Len, This modified check allows us to overcome RFC1323's 5828 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 5829 * p.869. In such cases, we can still calculate the RTT correctly 5830 * when RCV.NXT == Last.ACK.Sent. 5831 */ 5832 if ((to->to_flags & TOF_TS) != 0 && 5833 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 5834 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 5835 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 5836 tp->ts_recent_age = tcp_ts_getticks(); 5837 tp->ts_recent = to->to_tsval; 5838 } 5839 /* 5840 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 5841 * is on (half-synchronized state), then queue data for later 5842 * processing; else drop segment and return. 5843 */ 5844 if ((thflags & TH_ACK) == 0) { 5845 if (tp->t_flags & TF_NEEDSYN) { 5846 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5847 tiwin, thflags, nxt_pkt)); 5848 } else if (tp->t_flags & TF_ACKNOW) { 5849 rack_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 5850 return (ret_val); 5851 } else { 5852 rack_do_drop(m, NULL); 5853 return (0); 5854 } 5855 } 5856 /* 5857 * Ack processing. 5858 */ 5859 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 5860 return (ret_val); 5861 } 5862 if (ourfinisacked) { 5863 /* 5864 * If we can't receive any more data, then closing user can 5865 * proceed. Starting the timer is contrary to the 5866 * specification, but if we don't get a FIN we'll hang 5867 * forever. 5868 * 5869 * XXXjl: we should release the tp also, and use a 5870 * compressed state. 5871 */ 5872 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 5873 soisdisconnected(so); 5874 tcp_timer_activate(tp, TT_2MSL, 5875 (tcp_fast_finwait2_recycle ? 5876 tcp_finwait2_timeout : 5877 TP_MAXIDLE(tp))); 5878 } 5879 tcp_state_change(tp, TCPS_FIN_WAIT_2); 5880 } 5881 if (sbavail(&so->so_snd)) { 5882 if (rack_progress_timeout_check(tp)) { 5883 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); 5884 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 5885 return (1); 5886 } 5887 } 5888 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5889 tiwin, thflags, nxt_pkt)); 5890 } 5891 5892 /* 5893 * Return value of 1, the TCB is unlocked and most 5894 * likely gone, return value of 0, the TCP is still 5895 * locked. 5896 */ 5897 static int 5898 rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so, 5899 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 5900 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 5901 { 5902 int32_t ret_val = 0; 5903 int32_t ourfinisacked = 0; 5904 5905 rack_calc_rwin(so, tp); 5906 5907 if (thflags & TH_RST) 5908 return (rack_process_rst(m, th, so, tp)); 5909 /* 5910 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 5911 * synchronized state. 5912 */ 5913 if (thflags & TH_SYN) { 5914 rack_challenge_ack(m, th, tp, &ret_val); 5915 return (ret_val); 5916 } 5917 /* 5918 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 5919 * it's less than ts_recent, drop it. 5920 */ 5921 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 5922 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 5923 if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val)) 5924 return (ret_val); 5925 } 5926 if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 5927 return (ret_val); 5928 } 5929 /* 5930 * If new data are received on a connection after the user processes 5931 * are gone, then RST the other end. 5932 */ 5933 if ((so->so_state & SS_NOFDREF) && tlen) { 5934 if (rack_check_data_after_close(m, tp, &tlen, th, so)) 5935 return (1); 5936 } 5937 /* 5938 * If last ACK falls within this segment's sequence numbers, record 5939 * its timestamp. NOTE: 1) That the test incorporates suggestions 5940 * from the latest proposal of the tcplw@cray.com list (Braden 5941 * 1993/04/26). 2) That updating only on newer timestamps interferes 5942 * with our earlier PAWS tests, so this check should be solely 5943 * predicated on the sequence space of this segment. 3) That we 5944 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 5945 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 5946 * SEG.Len, This modified check allows us to overcome RFC1323's 5947 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 5948 * p.869. In such cases, we can still calculate the RTT correctly 5949 * when RCV.NXT == Last.ACK.Sent. 5950 */ 5951 if ((to->to_flags & TOF_TS) != 0 && 5952 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 5953 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 5954 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 5955 tp->ts_recent_age = tcp_ts_getticks(); 5956 tp->ts_recent = to->to_tsval; 5957 } 5958 /* 5959 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 5960 * is on (half-synchronized state), then queue data for later 5961 * processing; else drop segment and return. 5962 */ 5963 if ((thflags & TH_ACK) == 0) { 5964 if (tp->t_flags & TF_NEEDSYN) { 5965 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5966 tiwin, thflags, nxt_pkt)); 5967 } else if (tp->t_flags & TF_ACKNOW) { 5968 rack_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 5969 return (ret_val); 5970 } else { 5971 rack_do_drop(m, NULL); 5972 return (0); 5973 } 5974 } 5975 /* 5976 * Ack processing. 5977 */ 5978 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 5979 return (ret_val); 5980 } 5981 if (ourfinisacked) { 5982 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 5983 tcp_twstart(tp); 5984 m_freem(m); 5985 return (1); 5986 } 5987 if (sbavail(&so->so_snd)) { 5988 if (rack_progress_timeout_check(tp)) { 5989 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); 5990 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 5991 return (1); 5992 } 5993 } 5994 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5995 tiwin, thflags, nxt_pkt)); 5996 } 5997 5998 /* 5999 * Return value of 1, the TCB is unlocked and most 6000 * likely gone, return value of 0, the TCP is still 6001 * locked. 6002 */ 6003 static int 6004 rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so, 6005 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 6006 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 6007 { 6008 int32_t ret_val = 0; 6009 int32_t ourfinisacked = 0; 6010 6011 rack_calc_rwin(so, tp); 6012 6013 if (thflags & TH_RST) 6014 return (rack_process_rst(m, th, so, tp)); 6015 /* 6016 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 6017 * synchronized state. 6018 */ 6019 if (thflags & TH_SYN) { 6020 rack_challenge_ack(m, th, tp, &ret_val); 6021 return (ret_val); 6022 } 6023 /* 6024 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 6025 * it's less than ts_recent, drop it. 6026 */ 6027 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 6028 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 6029 if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val)) 6030 return (ret_val); 6031 } 6032 if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 6033 return (ret_val); 6034 } 6035 /* 6036 * If new data are received on a connection after the user processes 6037 * are gone, then RST the other end. 6038 */ 6039 if ((so->so_state & SS_NOFDREF) && tlen) { 6040 if (rack_check_data_after_close(m, tp, &tlen, th, so)) 6041 return (1); 6042 } 6043 /* 6044 * If last ACK falls within this segment's sequence numbers, record 6045 * its timestamp. NOTE: 1) That the test incorporates suggestions 6046 * from the latest proposal of the tcplw@cray.com list (Braden 6047 * 1993/04/26). 2) That updating only on newer timestamps interferes 6048 * with our earlier PAWS tests, so this check should be solely 6049 * predicated on the sequence space of this segment. 3) That we 6050 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 6051 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 6052 * SEG.Len, This modified check allows us to overcome RFC1323's 6053 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 6054 * p.869. In such cases, we can still calculate the RTT correctly 6055 * when RCV.NXT == Last.ACK.Sent. 6056 */ 6057 if ((to->to_flags & TOF_TS) != 0 && 6058 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 6059 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 6060 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 6061 tp->ts_recent_age = tcp_ts_getticks(); 6062 tp->ts_recent = to->to_tsval; 6063 } 6064 /* 6065 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 6066 * is on (half-synchronized state), then queue data for later 6067 * processing; else drop segment and return. 6068 */ 6069 if ((thflags & TH_ACK) == 0) { 6070 if (tp->t_flags & TF_NEEDSYN) { 6071 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 6072 tiwin, thflags, nxt_pkt)); 6073 } else if (tp->t_flags & TF_ACKNOW) { 6074 rack_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 6075 return (ret_val); 6076 } else { 6077 rack_do_drop(m, NULL); 6078 return (0); 6079 } 6080 } 6081 /* 6082 * case TCPS_LAST_ACK: Ack processing. 6083 */ 6084 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 6085 return (ret_val); 6086 } 6087 if (ourfinisacked) { 6088 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 6089 tp = tcp_close(tp); 6090 rack_do_drop(m, tp); 6091 return (1); 6092 } 6093 if (sbavail(&so->so_snd)) { 6094 if (rack_progress_timeout_check(tp)) { 6095 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); 6096 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 6097 return (1); 6098 } 6099 } 6100 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 6101 tiwin, thflags, nxt_pkt)); 6102 } 6103 6104 6105 /* 6106 * Return value of 1, the TCB is unlocked and most 6107 * likely gone, return value of 0, the TCP is still 6108 * locked. 6109 */ 6110 static int 6111 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so, 6112 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 6113 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 6114 { 6115 int32_t ret_val = 0; 6116 int32_t ourfinisacked = 0; 6117 6118 rack_calc_rwin(so, tp); 6119 6120 /* Reset receive buffer auto scaling when not in bulk receive mode. */ 6121 if (thflags & TH_RST) 6122 return (rack_process_rst(m, th, so, tp)); 6123 /* 6124 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 6125 * synchronized state. 6126 */ 6127 if (thflags & TH_SYN) { 6128 rack_challenge_ack(m, th, tp, &ret_val); 6129 return (ret_val); 6130 } 6131 /* 6132 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 6133 * it's less than ts_recent, drop it. 6134 */ 6135 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 6136 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 6137 if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val)) 6138 return (ret_val); 6139 } 6140 if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 6141 return (ret_val); 6142 } 6143 /* 6144 * If new data are received on a connection after the user processes 6145 * are gone, then RST the other end. 6146 */ 6147 if ((so->so_state & SS_NOFDREF) && 6148 tlen) { 6149 if (rack_check_data_after_close(m, tp, &tlen, th, so)) 6150 return (1); 6151 } 6152 /* 6153 * If last ACK falls within this segment's sequence numbers, record 6154 * its timestamp. NOTE: 1) That the test incorporates suggestions 6155 * from the latest proposal of the tcplw@cray.com list (Braden 6156 * 1993/04/26). 2) That updating only on newer timestamps interferes 6157 * with our earlier PAWS tests, so this check should be solely 6158 * predicated on the sequence space of this segment. 3) That we 6159 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 6160 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 6161 * SEG.Len, This modified check allows us to overcome RFC1323's 6162 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 6163 * p.869. In such cases, we can still calculate the RTT correctly 6164 * when RCV.NXT == Last.ACK.Sent. 6165 */ 6166 if ((to->to_flags & TOF_TS) != 0 && 6167 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 6168 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 6169 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 6170 tp->ts_recent_age = tcp_ts_getticks(); 6171 tp->ts_recent = to->to_tsval; 6172 } 6173 /* 6174 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 6175 * is on (half-synchronized state), then queue data for later 6176 * processing; else drop segment and return. 6177 */ 6178 if ((thflags & TH_ACK) == 0) { 6179 if (tp->t_flags & TF_NEEDSYN) { 6180 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 6181 tiwin, thflags, nxt_pkt)); 6182 } else if (tp->t_flags & TF_ACKNOW) { 6183 rack_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 6184 return (ret_val); 6185 } else { 6186 rack_do_drop(m, NULL); 6187 return (0); 6188 } 6189 } 6190 /* 6191 * Ack processing. 6192 */ 6193 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 6194 return (ret_val); 6195 } 6196 if (sbavail(&so->so_snd)) { 6197 if (rack_progress_timeout_check(tp)) { 6198 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); 6199 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 6200 return (1); 6201 } 6202 } 6203 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 6204 tiwin, thflags, nxt_pkt)); 6205 } 6206 6207 6208 static void inline 6209 rack_clear_rate_sample(struct tcp_rack *rack) 6210 { 6211 rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY; 6212 rack->r_ctl.rack_rs.rs_rtt_cnt = 0; 6213 rack->r_ctl.rack_rs.rs_rtt_tot = 0; 6214 } 6215 6216 static int 6217 rack_init(struct tcpcb *tp) 6218 { 6219 struct tcp_rack *rack = NULL; 6220 6221 tp->t_fb_ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT); 6222 if (tp->t_fb_ptr == NULL) { 6223 /* 6224 * We need to allocate memory but cant. The INP and INP_INFO 6225 * locks and they are recusive (happens during setup. So a 6226 * scheme to drop the locks fails :( 6227 * 6228 */ 6229 return (ENOMEM); 6230 } 6231 memset(tp->t_fb_ptr, 0, sizeof(struct tcp_rack)); 6232 6233 rack = (struct tcp_rack *)tp->t_fb_ptr; 6234 TAILQ_INIT(&rack->r_ctl.rc_map); 6235 TAILQ_INIT(&rack->r_ctl.rc_free); 6236 TAILQ_INIT(&rack->r_ctl.rc_tmap); 6237 rack->rc_tp = tp; 6238 if (tp->t_inpcb) { 6239 rack->rc_inp = tp->t_inpcb; 6240 } 6241 /* Probably not needed but lets be sure */ 6242 rack_clear_rate_sample(rack); 6243 rack->r_cpu = 0; 6244 rack->r_ctl.rc_reorder_fade = rack_reorder_fade; 6245 rack->rc_allow_data_af_clo = rack_ignore_data_after_close; 6246 rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh; 6247 rack->rc_pace_reduce = rack_slot_reduction; 6248 if (V_tcp_delack_enabled) 6249 tp->t_delayed_ack = 1; 6250 else 6251 tp->t_delayed_ack = 0; 6252 rack->rc_pace_max_segs = rack_hptsi_segments; 6253 rack->r_ctl.rc_early_recovery_segs = rack_early_recovery_max_seg; 6254 rack->r_ctl.rc_reorder_shift = rack_reorder_thresh; 6255 rack->r_ctl.rc_pkt_delay = rack_pkt_delay; 6256 rack->r_ctl.rc_prop_reduce = rack_use_proportional_reduce; 6257 rack->r_idle_reduce_largest = rack_reduce_largest_on_idle; 6258 rack->r_enforce_min_pace = rack_min_pace_time; 6259 rack->r_min_pace_seg_thresh = rack_min_pace_time_seg_req; 6260 rack->r_ctl.rc_prop_rate = rack_proportional_rate; 6261 rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp; 6262 rack->r_ctl.rc_early_recovery = rack_early_recovery; 6263 rack->rc_always_pace = rack_pace_every_seg; 6264 rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method; 6265 rack->rack_tlp_threshold_use = rack_tlp_threshold_use; 6266 rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr; 6267 rack->r_ctl.rc_min_to = rack_min_to; 6268 rack->r_ctl.rc_prr_inc_var = rack_inc_var; 6269 rack_start_hpts_timer(rack, tp, tcp_ts_getticks(), __LINE__, 0, 0, 0); 6270 if (tp->snd_una != tp->snd_max) { 6271 /* Create a send map for the current outstanding data */ 6272 struct rack_sendmap *rsm; 6273 6274 rsm = rack_alloc(rack); 6275 if (rsm == NULL) { 6276 uma_zfree(rack_pcb_zone, tp->t_fb_ptr); 6277 tp->t_fb_ptr = NULL; 6278 return (ENOMEM); 6279 } 6280 rsm->r_flags = RACK_OVERMAX; 6281 rsm->r_tim_lastsent[0] = tcp_ts_getticks(); 6282 rsm->r_rtr_cnt = 1; 6283 rsm->r_rtr_bytes = 0; 6284 rsm->r_start = tp->snd_una; 6285 rsm->r_end = tp->snd_max; 6286 rsm->r_sndcnt = 0; 6287 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_map, rsm, r_next); 6288 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext); 6289 rsm->r_in_tmap = 1; 6290 } 6291 return (0); 6292 } 6293 6294 static int 6295 rack_handoff_ok(struct tcpcb *tp) 6296 { 6297 if ((tp->t_state == TCPS_CLOSED) || 6298 (tp->t_state == TCPS_LISTEN)) { 6299 /* Sure no problem though it may not stick */ 6300 return (0); 6301 } 6302 if ((tp->t_state == TCPS_SYN_SENT) || 6303 (tp->t_state == TCPS_SYN_RECEIVED)) { 6304 /* 6305 * We really don't know you have to get to ESTAB or beyond 6306 * to tell. 6307 */ 6308 return (EAGAIN); 6309 } 6310 if (tp->t_flags & TF_SACK_PERMIT) { 6311 return (0); 6312 } 6313 /* 6314 * If we reach here we don't do SACK on this connection so we can 6315 * never do rack. 6316 */ 6317 return (EINVAL); 6318 } 6319 6320 static void 6321 rack_fini(struct tcpcb *tp, int32_t tcb_is_purged) 6322 { 6323 if (tp->t_fb_ptr) { 6324 struct tcp_rack *rack; 6325 struct rack_sendmap *rsm; 6326 6327 rack = (struct tcp_rack *)tp->t_fb_ptr; 6328 #ifdef TCP_BLACKBOX 6329 tcp_log_flowend(tp); 6330 #endif 6331 rsm = TAILQ_FIRST(&rack->r_ctl.rc_map); 6332 while (rsm) { 6333 TAILQ_REMOVE(&rack->r_ctl.rc_map, rsm, r_next); 6334 uma_zfree(rack_zone, rsm); 6335 rsm = TAILQ_FIRST(&rack->r_ctl.rc_map); 6336 } 6337 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free); 6338 while (rsm) { 6339 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_next); 6340 uma_zfree(rack_zone, rsm); 6341 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free); 6342 } 6343 rack->rc_free_cnt = 0; 6344 uma_zfree(rack_pcb_zone, tp->t_fb_ptr); 6345 tp->t_fb_ptr = NULL; 6346 } 6347 } 6348 6349 static void 6350 rack_set_state(struct tcpcb *tp, struct tcp_rack *rack) 6351 { 6352 switch (tp->t_state) { 6353 case TCPS_SYN_SENT: 6354 rack->r_state = TCPS_SYN_SENT; 6355 rack->r_substate = rack_do_syn_sent; 6356 break; 6357 case TCPS_SYN_RECEIVED: 6358 rack->r_state = TCPS_SYN_RECEIVED; 6359 rack->r_substate = rack_do_syn_recv; 6360 break; 6361 case TCPS_ESTABLISHED: 6362 rack->r_state = TCPS_ESTABLISHED; 6363 rack->r_substate = rack_do_established; 6364 break; 6365 case TCPS_CLOSE_WAIT: 6366 rack->r_state = TCPS_CLOSE_WAIT; 6367 rack->r_substate = rack_do_close_wait; 6368 break; 6369 case TCPS_FIN_WAIT_1: 6370 rack->r_state = TCPS_FIN_WAIT_1; 6371 rack->r_substate = rack_do_fin_wait_1; 6372 break; 6373 case TCPS_CLOSING: 6374 rack->r_state = TCPS_CLOSING; 6375 rack->r_substate = rack_do_closing; 6376 break; 6377 case TCPS_LAST_ACK: 6378 rack->r_state = TCPS_LAST_ACK; 6379 rack->r_substate = rack_do_lastack; 6380 break; 6381 case TCPS_FIN_WAIT_2: 6382 rack->r_state = TCPS_FIN_WAIT_2; 6383 rack->r_substate = rack_do_fin_wait_2; 6384 break; 6385 case TCPS_LISTEN: 6386 case TCPS_CLOSED: 6387 case TCPS_TIME_WAIT: 6388 default: 6389 #ifdef INVARIANTS 6390 panic("tcp tp:%p state:%d sees impossible state?", tp, tp->t_state); 6391 #endif 6392 break; 6393 }; 6394 } 6395 6396 6397 static void 6398 rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb) 6399 { 6400 /* 6401 * We received an ack, and then did not 6402 * call send or were bounced out due to the 6403 * hpts was running. Now a timer is up as well, is 6404 * it the right timer? 6405 */ 6406 struct rack_sendmap *rsm; 6407 int tmr_up; 6408 6409 tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK; 6410 if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT)) 6411 return; 6412 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); 6413 if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) && 6414 (tmr_up == PACE_TMR_RXT)) { 6415 /* Should be an RXT */ 6416 return; 6417 } 6418 if (rsm == NULL) { 6419 /* Nothing outstanding? */ 6420 if (tp->t_flags & TF_DELACK) { 6421 if (tmr_up == PACE_TMR_DELACK) 6422 /* We are supposed to have delayed ack up and we do */ 6423 return; 6424 } else if (sbavail(&tp->t_inpcb->inp_socket->so_snd) && (tmr_up == PACE_TMR_RXT)) { 6425 /* 6426 * if we hit enobufs then we would expect the possiblity 6427 * of nothing outstanding and the RXT up (and the hptsi timer). 6428 */ 6429 return; 6430 } else if (((tcp_always_keepalive || 6431 rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) && 6432 (tp->t_state <= TCPS_CLOSING)) && 6433 (tmr_up == PACE_TMR_KEEP) && 6434 (tp->snd_max == tp->snd_una)) { 6435 /* We should have keep alive up and we do */ 6436 return; 6437 } 6438 } 6439 if (rsm && (rsm->r_flags & RACK_SACK_PASSED)) { 6440 if ((tp->t_flags & TF_SENTFIN) && 6441 ((tp->snd_max - tp->snd_una) == 1) && 6442 (rsm->r_flags & RACK_HAS_FIN)) { 6443 /* needs to be a RXT */ 6444 if (tmr_up == PACE_TMR_RXT) 6445 return; 6446 } else if (tmr_up == PACE_TMR_RACK) 6447 return; 6448 } else if (SEQ_GT(tp->snd_max,tp->snd_una) && 6449 ((tmr_up == PACE_TMR_TLP) || 6450 (tmr_up == PACE_TMR_RXT))) { 6451 /* 6452 * Either a TLP or RXT is fine if no sack-passed 6453 * is in place and data is outstanding. 6454 */ 6455 return; 6456 } else if (tmr_up == PACE_TMR_DELACK) { 6457 /* 6458 * If the delayed ack was going to go off 6459 * before the rtx/tlp/rack timer were going to 6460 * expire, then that would be the timer in control. 6461 * Note we don't check the time here trusting the 6462 * code is correct. 6463 */ 6464 return; 6465 } 6466 /* 6467 * Ok the timer originally started is not what we want now. 6468 * We will force the hpts to be stopped if any, and restart 6469 * with the slot set to what was in the saved slot. 6470 */ 6471 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 6472 rack_start_hpts_timer(rack, tp, tcp_ts_getticks(), __LINE__, 0, 0, 0); 6473 } 6474 6475 static void 6476 rack_hpts_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so, 6477 struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos, 6478 int32_t nxt_pkt, struct timeval *tv) 6479 { 6480 int32_t thflags, retval, did_out = 0; 6481 int32_t way_out = 0; 6482 uint32_t cts; 6483 uint32_t tiwin; 6484 struct tcpopt to; 6485 struct tcp_rack *rack; 6486 struct rack_sendmap *rsm; 6487 int32_t prev_state = 0; 6488 6489 cts = tcp_tv_to_mssectick(tv); 6490 rack = (struct tcp_rack *)tp->t_fb_ptr; 6491 6492 kern_prefetch(rack, &prev_state); 6493 prev_state = 0; 6494 thflags = th->th_flags; 6495 /* 6496 * If this is either a state-changing packet or current state isn't 6497 * established, we require a read lock on tcbinfo. Otherwise, we 6498 * allow the tcbinfo to be in either locked or unlocked, as the 6499 * caller may have unnecessarily acquired a lock due to a race. 6500 */ 6501 if ((thflags & (TH_SYN | TH_FIN | TH_RST)) != 0 || 6502 tp->t_state != TCPS_ESTABLISHED) { 6503 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 6504 } 6505 INP_WLOCK_ASSERT(tp->t_inpcb); 6506 KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN", 6507 __func__)); 6508 KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT", 6509 __func__)); 6510 { 6511 union tcp_log_stackspecific log; 6512 6513 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 6514 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; 6515 log.u_bbr.ininput = rack->rc_inp->inp_in_input; 6516 TCP_LOG_EVENT(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0, 6517 tlen, &log, true); 6518 } 6519 if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) { 6520 way_out = 4; 6521 goto done_with_input; 6522 } 6523 /* 6524 * If a segment with the ACK-bit set arrives in the SYN-SENT state 6525 * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9. 6526 */ 6527 if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) && 6528 (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) { 6529 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 6530 return; 6531 } 6532 /* 6533 * Segment received on connection. Reset idle time and keep-alive 6534 * timer. XXX: This should be done after segment validation to 6535 * ignore broken/spoofed segs. 6536 */ 6537 if (tp->t_idle_reduce && (tp->snd_max == tp->snd_una)) { 6538 #ifdef NETFLIX_CWV 6539 if ((tp->cwv_enabled) && 6540 ((tp->cwv_cwnd_valid == 0) && 6541 TCPS_HAVEESTABLISHED(tp->t_state) && 6542 (tp->snd_cwnd > tp->snd_cwv.init_cwnd))) { 6543 tcp_newcwv_nvp_closedown(tp); 6544 } else 6545 #endif 6546 if ((ticks - tp->t_rcvtime) >= tp->t_rxtcur) { 6547 counter_u64_add(rack_input_idle_reduces, 1); 6548 rack_cc_after_idle(tp, 6549 (rack->r_idle_reduce_largest ? 1 :0)); 6550 } 6551 } 6552 rack->r_ctl.rc_rcvtime = cts; 6553 tp->t_rcvtime = ticks; 6554 6555 #ifdef NETFLIX_CWV 6556 if (tp->cwv_enabled) { 6557 if ((tp->cwv_cwnd_valid == 0) && 6558 TCPS_HAVEESTABLISHED(tp->t_state) && 6559 (tp->snd_cwnd > tp->snd_cwv.init_cwnd)) 6560 tcp_newcwv_nvp_closedown(tp); 6561 } 6562 #endif 6563 /* 6564 * Unscale the window into a 32-bit value. For the SYN_SENT state 6565 * the scale is zero. 6566 */ 6567 tiwin = th->th_win << tp->snd_scale; 6568 #ifdef NETFLIX_STATS 6569 stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin); 6570 #endif 6571 /* 6572 * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move 6573 * this to occur after we've validated the segment. 6574 */ 6575 if (tp->t_flags & TF_ECN_PERMIT) { 6576 if (thflags & TH_CWR) 6577 tp->t_flags &= ~TF_ECN_SND_ECE; 6578 switch (iptos & IPTOS_ECN_MASK) { 6579 case IPTOS_ECN_CE: 6580 tp->t_flags |= TF_ECN_SND_ECE; 6581 TCPSTAT_INC(tcps_ecn_ce); 6582 break; 6583 case IPTOS_ECN_ECT0: 6584 TCPSTAT_INC(tcps_ecn_ect0); 6585 break; 6586 case IPTOS_ECN_ECT1: 6587 TCPSTAT_INC(tcps_ecn_ect1); 6588 break; 6589 } 6590 /* Congestion experienced. */ 6591 if (thflags & TH_ECE) { 6592 rack_cong_signal(tp, th, CC_ECN); 6593 } 6594 } 6595 /* 6596 * Parse options on any incoming segment. 6597 */ 6598 tcp_dooptions(&to, (u_char *)(th + 1), 6599 (th->th_off << 2) - sizeof(struct tcphdr), 6600 (thflags & TH_SYN) ? TO_SYN : 0); 6601 6602 /* 6603 * If echoed timestamp is later than the current time, fall back to 6604 * non RFC1323 RTT calculation. Normalize timestamp if syncookies 6605 * were used when this connection was established. 6606 */ 6607 if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) { 6608 to.to_tsecr -= tp->ts_offset; 6609 if (TSTMP_GT(to.to_tsecr, cts)) 6610 to.to_tsecr = 0; 6611 } 6612 /* 6613 * If its the first time in we need to take care of options and 6614 * verify we can do SACK for rack! 6615 */ 6616 if (rack->r_state == 0) { 6617 /* Should be init'd by rack_init() */ 6618 KASSERT(rack->rc_inp != NULL, 6619 ("%s: rack->rc_inp unexpectedly NULL", __func__)); 6620 if (rack->rc_inp == NULL) { 6621 rack->rc_inp = tp->t_inpcb; 6622 } 6623 6624 /* 6625 * Process options only when we get SYN/ACK back. The SYN 6626 * case for incoming connections is handled in tcp_syncache. 6627 * According to RFC1323 the window field in a SYN (i.e., a 6628 * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX 6629 * this is traditional behavior, may need to be cleaned up. 6630 */ 6631 rack->r_cpu = inp_to_cpuid(tp->t_inpcb); 6632 if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) { 6633 if ((to.to_flags & TOF_SCALE) && 6634 (tp->t_flags & TF_REQ_SCALE)) { 6635 tp->t_flags |= TF_RCVD_SCALE; 6636 tp->snd_scale = to.to_wscale; 6637 } 6638 /* 6639 * Initial send window. It will be updated with the 6640 * next incoming segment to the scaled value. 6641 */ 6642 tp->snd_wnd = th->th_win; 6643 if (to.to_flags & TOF_TS) { 6644 tp->t_flags |= TF_RCVD_TSTMP; 6645 tp->ts_recent = to.to_tsval; 6646 tp->ts_recent_age = cts; 6647 } 6648 if (to.to_flags & TOF_MSS) 6649 tcp_mss(tp, to.to_mss); 6650 if ((tp->t_flags & TF_SACK_PERMIT) && 6651 (to.to_flags & TOF_SACKPERM) == 0) 6652 tp->t_flags &= ~TF_SACK_PERMIT; 6653 if (IS_FASTOPEN(tp->t_flags)) { 6654 if (to.to_flags & TOF_FASTOPEN) { 6655 uint16_t mss; 6656 6657 if (to.to_flags & TOF_MSS) 6658 mss = to.to_mss; 6659 else 6660 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0) 6661 mss = TCP6_MSS; 6662 else 6663 mss = TCP_MSS; 6664 tcp_fastopen_update_cache(tp, mss, 6665 to.to_tfo_len, to.to_tfo_cookie); 6666 } else 6667 tcp_fastopen_disable_path(tp); 6668 } 6669 } 6670 /* 6671 * At this point we are at the initial call. Here we decide 6672 * if we are doing RACK or not. We do this by seeing if 6673 * TF_SACK_PERMIT is set, if not rack is *not* possible and 6674 * we switch to the default code. 6675 */ 6676 if ((tp->t_flags & TF_SACK_PERMIT) == 0) { 6677 tcp_switch_back_to_default(tp); 6678 (*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen, 6679 tlen, iptos); 6680 return; 6681 } 6682 /* Set the flag */ 6683 rack->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0; 6684 tcp_set_hpts(tp->t_inpcb); 6685 rack_stop_all_timers(tp); 6686 sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack); 6687 } 6688 /* 6689 * This is the one exception case where we set the rack state 6690 * always. All other times (timers etc) we must have a rack-state 6691 * set (so we assure we have done the checks above for SACK). 6692 */ 6693 if (rack->r_state != tp->t_state) 6694 rack_set_state(tp, rack); 6695 if (SEQ_GT(th->th_ack, tp->snd_una) && (rsm = TAILQ_FIRST(&rack->r_ctl.rc_map)) != NULL) 6696 kern_prefetch(rsm, &prev_state); 6697 prev_state = rack->r_state; 6698 rack->r_ctl.rc_tlp_send_cnt = 0; 6699 rack_clear_rate_sample(rack); 6700 retval = (*rack->r_substate) (m, th, so, 6701 tp, &to, drop_hdrlen, 6702 tlen, tiwin, thflags, nxt_pkt); 6703 #ifdef INVARIANTS 6704 if ((retval == 0) && 6705 (tp->t_inpcb == NULL)) { 6706 panic("retval:%d tp:%p t_inpcb:NULL state:%d", 6707 retval, tp, prev_state); 6708 } 6709 #endif 6710 if (retval == 0) { 6711 /* 6712 * If retval is 1 the tcb is unlocked and most likely the tp 6713 * is gone. 6714 */ 6715 INP_WLOCK_ASSERT(tp->t_inpcb); 6716 tcp_rack_xmit_timer_commit(rack, tp); 6717 if (((tp->snd_max - tp->snd_una) > tp->snd_wnd) && 6718 (rack->rc_in_persist == 0)){ 6719 /* 6720 * The peer shrunk its window on us to the point 6721 * where we have sent too much. The only thing 6722 * we can do here is stop any timers and 6723 * enter persist. We most likely lost the last 6724 * bytes we sent but oh well, we will have to 6725 * retransmit them after the peer is caught up. 6726 */ 6727 if (rack->rc_inp->inp_in_hpts) 6728 tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT); 6729 rack_timer_cancel(tp, rack, cts, __LINE__); 6730 rack_enter_persist(tp, rack, cts); 6731 rack_start_hpts_timer(rack, tp, tcp_ts_getticks(), __LINE__, 0, 0, 0); 6732 way_out = 3; 6733 goto done_with_input; 6734 } 6735 if (nxt_pkt == 0) { 6736 if (rack->r_wanted_output != 0) { 6737 did_out = 1; 6738 (void)tp->t_fb->tfb_tcp_output(tp); 6739 } 6740 rack_start_hpts_timer(rack, tp, cts, __LINE__, 0, 0, 0); 6741 } 6742 if (((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) && 6743 (SEQ_GT(tp->snd_max, tp->snd_una) || 6744 (tp->t_flags & TF_DELACK) || 6745 ((tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) && 6746 (tp->t_state <= TCPS_CLOSING)))) { 6747 /* We could not send (probably in the hpts but stopped the timer earlier)? */ 6748 if ((tp->snd_max == tp->snd_una) && 6749 ((tp->t_flags & TF_DELACK) == 0) && 6750 (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) { 6751 /* keep alive not needed if we are hptsi output yet */ 6752 ; 6753 } else { 6754 if (rack->rc_inp->inp_in_hpts) 6755 tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT); 6756 rack_start_hpts_timer(rack, tp, tcp_ts_getticks(), __LINE__, 0, 0, 0); 6757 } 6758 way_out = 1; 6759 } else { 6760 /* Do we have the correct timer running? */ 6761 rack_timer_audit(tp, rack, &so->so_snd); 6762 way_out = 2; 6763 } 6764 done_with_input: 6765 rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out); 6766 if (did_out) 6767 rack->r_wanted_output = 0; 6768 #ifdef INVARIANTS 6769 if (tp->t_inpcb == NULL) { 6770 panic("OP:%d retval:%d tp:%p t_inpcb:NULL state:%d", 6771 did_out, 6772 retval, tp, prev_state); 6773 } 6774 #endif 6775 INP_WUNLOCK(tp->t_inpcb); 6776 } 6777 } 6778 6779 void 6780 rack_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so, 6781 struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos) 6782 { 6783 struct timeval tv; 6784 #ifdef RSS 6785 struct tcp_function_block *tfb; 6786 struct tcp_rack *rack; 6787 struct epoch_tracker et; 6788 6789 rack = (struct tcp_rack *)tp->t_fb_ptr; 6790 if (rack->r_state == 0) { 6791 /* 6792 * Initial input (ACK to SYN-ACK etc)lets go ahead and get 6793 * it processed 6794 */ 6795 INP_INFO_RLOCK_ET(&V_tcbinfo, et); 6796 tcp_get_usecs(&tv); 6797 rack_hpts_do_segment(m, th, so, tp, drop_hdrlen, 6798 tlen, iptos, 0, &tv); 6799 INP_INFO_RUNLOCK_ET(&V_tcbinfo, et); 6800 return; 6801 } 6802 tcp_queue_to_input(tp, m, th, tlen, drop_hdrlen, iptos); 6803 INP_WUNLOCK(tp->t_inpcb); 6804 #else 6805 tcp_get_usecs(&tv); 6806 rack_hpts_do_segment(m, th, so, tp, drop_hdrlen, 6807 tlen, iptos, 0, &tv); 6808 #endif 6809 } 6810 6811 struct rack_sendmap * 6812 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused) 6813 { 6814 struct rack_sendmap *rsm = NULL; 6815 int32_t idx; 6816 uint32_t srtt_cur, srtt = 0, thresh = 0, ts_low = 0; 6817 6818 /* Return the next guy to be re-transmitted */ 6819 if (TAILQ_EMPTY(&rack->r_ctl.rc_map)) { 6820 return (NULL); 6821 } 6822 if (tp->t_flags & TF_SENTFIN) { 6823 /* retran the end FIN? */ 6824 return (NULL); 6825 } 6826 /* ok lets look at this one */ 6827 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); 6828 if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) { 6829 goto check_it; 6830 } 6831 rsm = rack_find_lowest_rsm(rack); 6832 if (rsm == NULL) { 6833 return (NULL); 6834 } 6835 check_it: 6836 srtt_cur = tp->t_srtt >> TCP_RTT_SHIFT; 6837 srtt = TICKS_2_MSEC(srtt_cur); 6838 if (rack->rc_rack_rtt && (srtt > rack->rc_rack_rtt)) 6839 srtt = rack->rc_rack_rtt; 6840 if (rsm->r_flags & RACK_ACKED) { 6841 return (NULL); 6842 } 6843 if ((rsm->r_flags & RACK_SACK_PASSED) == 0) { 6844 /* Its not yet ready */ 6845 return (NULL); 6846 } 6847 idx = rsm->r_rtr_cnt - 1; 6848 ts_low = rsm->r_tim_lastsent[idx]; 6849 thresh = rack_calc_thresh_rack(rack, srtt, tsused); 6850 if (tsused <= ts_low) { 6851 return (NULL); 6852 } 6853 if ((tsused - ts_low) >= thresh) { 6854 return (rsm); 6855 } 6856 return (NULL); 6857 } 6858 6859 static int 6860 rack_output(struct tcpcb *tp) 6861 { 6862 struct socket *so; 6863 uint32_t recwin, sendwin; 6864 uint32_t sb_offset; 6865 int32_t len, flags, error = 0; 6866 struct mbuf *m; 6867 struct mbuf *mb; 6868 uint32_t if_hw_tsomaxsegcount = 0; 6869 uint32_t if_hw_tsomaxsegsize; 6870 long tot_len_this_send = 0; 6871 struct ip *ip = NULL; 6872 #ifdef TCPDEBUG 6873 struct ipovly *ipov = NULL; 6874 #endif 6875 struct udphdr *udp = NULL; 6876 struct tcp_rack *rack; 6877 struct tcphdr *th; 6878 uint8_t pass = 0; 6879 uint8_t wanted_cookie = 0; 6880 u_char opt[TCP_MAXOLEN]; 6881 unsigned ipoptlen, optlen, hdrlen, ulen=0; 6882 uint32_t rack_seq; 6883 6884 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 6885 unsigned ipsec_optlen = 0; 6886 6887 #endif 6888 int32_t idle, sendalot; 6889 int32_t sub_from_prr = 0; 6890 volatile int32_t sack_rxmit; 6891 struct rack_sendmap *rsm = NULL; 6892 int32_t tso, mtu, would_have_fin = 0; 6893 struct tcpopt to; 6894 int32_t slot = 0; 6895 uint32_t cts; 6896 uint8_t hpts_calling, doing_tlp = 0; 6897 int32_t do_a_prefetch; 6898 int32_t prefetch_rsm = 0; 6899 int32_t prefetch_so_done = 0; 6900 struct tcp_log_buffer *lgb = NULL; 6901 struct inpcb *inp; 6902 struct sockbuf *sb; 6903 #ifdef INET6 6904 struct ip6_hdr *ip6 = NULL; 6905 int32_t isipv6; 6906 #endif 6907 /* setup and take the cache hits here */ 6908 rack = (struct tcp_rack *)tp->t_fb_ptr; 6909 inp = rack->rc_inp; 6910 so = inp->inp_socket; 6911 sb = &so->so_snd; 6912 kern_prefetch(sb, &do_a_prefetch); 6913 do_a_prefetch = 1; 6914 6915 INP_WLOCK_ASSERT(inp); 6916 #ifdef TCP_OFFLOAD 6917 if (tp->t_flags & TF_TOE) 6918 return (tcp_offload_output(tp)); 6919 #endif 6920 #ifdef INET6 6921 if (rack->r_state) { 6922 /* Use the cache line loaded if possible */ 6923 isipv6 = rack->r_is_v6; 6924 } else { 6925 isipv6 = (inp->inp_vflag & INP_IPV6) != 0; 6926 } 6927 #endif 6928 cts = tcp_ts_getticks(); 6929 if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) && 6930 inp->inp_in_hpts) { 6931 /* 6932 * We are on the hpts for some timer but not hptsi output. 6933 * Remove from the hpts unconditionally. 6934 */ 6935 rack_timer_cancel(tp, rack, cts, __LINE__); 6936 } 6937 /* Mark that we have called rack_output(). */ 6938 if ((rack->r_timer_override) || 6939 (tp->t_flags & TF_FORCEDATA) || 6940 (tp->t_state < TCPS_ESTABLISHED)) { 6941 if (tp->t_inpcb->inp_in_hpts) 6942 tcp_hpts_remove(tp->t_inpcb, HPTS_REMOVE_OUTPUT); 6943 } else if (tp->t_inpcb->inp_in_hpts) { 6944 /* 6945 * On the hpts you can't pass even if ACKNOW is on, we will 6946 * when the hpts fires. 6947 */ 6948 counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1); 6949 return (0); 6950 } 6951 hpts_calling = inp->inp_hpts_calls; 6952 inp->inp_hpts_calls = 0; 6953 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) { 6954 if (rack_process_timers(tp, rack, cts, hpts_calling)) { 6955 counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1); 6956 return (0); 6957 } 6958 } 6959 rack->r_wanted_output = 0; 6960 rack->r_timer_override = 0; 6961 /* 6962 * For TFO connections in SYN_SENT or SYN_RECEIVED, 6963 * only allow the initial SYN or SYN|ACK and those sent 6964 * by the retransmit timer. 6965 */ 6966 if (IS_FASTOPEN(tp->t_flags) && 6967 ((tp->t_state == TCPS_SYN_RECEIVED) || 6968 (tp->t_state == TCPS_SYN_SENT)) && 6969 SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */ 6970 (tp->t_rxtshift == 0)) /* not a retransmit */ 6971 return (0); 6972 /* 6973 * Determine length of data that should be transmitted, and flags 6974 * that will be used. If there is some data or critical controls 6975 * (SYN, RST) to send, then transmit; otherwise, investigate 6976 * further. 6977 */ 6978 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una); 6979 #ifdef NETFLIX_CWV 6980 if (tp->cwv_enabled) { 6981 if ((tp->cwv_cwnd_valid == 0) && 6982 TCPS_HAVEESTABLISHED(tp->t_state) && 6983 (tp->snd_cwnd > tp->snd_cwv.init_cwnd)) 6984 tcp_newcwv_nvp_closedown(tp); 6985 } else 6986 #endif 6987 if (tp->t_idle_reduce) { 6988 if (idle && ((ticks - tp->t_rcvtime) >= tp->t_rxtcur)) 6989 rack_cc_after_idle(tp, 6990 (rack->r_idle_reduce_largest ? 1 :0)); 6991 } 6992 tp->t_flags &= ~TF_LASTIDLE; 6993 if (idle) { 6994 if (tp->t_flags & TF_MORETOCOME) { 6995 tp->t_flags |= TF_LASTIDLE; 6996 idle = 0; 6997 } 6998 } 6999 again: 7000 /* 7001 * If we've recently taken a timeout, snd_max will be greater than 7002 * snd_nxt. There may be SACK information that allows us to avoid 7003 * resending already delivered data. Adjust snd_nxt accordingly. 7004 */ 7005 sendalot = 0; 7006 cts = tcp_ts_getticks(); 7007 tso = 0; 7008 mtu = 0; 7009 sb_offset = tp->snd_max - tp->snd_una; 7010 sendwin = min(tp->snd_wnd, tp->snd_cwnd); 7011 7012 flags = tcp_outflags[tp->t_state]; 7013 /* 7014 * Send any SACK-generated retransmissions. If we're explicitly 7015 * trying to send out new data (when sendalot is 1), bypass this 7016 * function. If we retransmit in fast recovery mode, decrement 7017 * snd_cwnd, since we're replacing a (future) new transmission with 7018 * a retransmission now, and we previously incremented snd_cwnd in 7019 * tcp_input(). 7020 */ 7021 /* 7022 * Still in sack recovery , reset rxmit flag to zero. 7023 */ 7024 while (rack->rc_free_cnt < rack_free_cache) { 7025 rsm = rack_alloc(rack); 7026 if (rsm == NULL) { 7027 if (inp->inp_hpts_calls) 7028 /* Retry in a ms */ 7029 slot = 1; 7030 goto just_return_nolock; 7031 } 7032 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_next); 7033 rack->rc_free_cnt++; 7034 rsm = NULL; 7035 } 7036 if (inp->inp_hpts_calls) 7037 inp->inp_hpts_calls = 0; 7038 sack_rxmit = 0; 7039 len = 0; 7040 rsm = NULL; 7041 if (flags & TH_RST) { 7042 SOCKBUF_LOCK(sb); 7043 goto send; 7044 } 7045 if (rack->r_ctl.rc_tlpsend) { 7046 /* Tail loss probe */ 7047 long cwin; 7048 long tlen; 7049 7050 doing_tlp = 1; 7051 rsm = rack->r_ctl.rc_tlpsend; 7052 rack->r_ctl.rc_tlpsend = NULL; 7053 sack_rxmit = 1; 7054 tlen = rsm->r_end - rsm->r_start; 7055 if (tlen > tp->t_maxseg) 7056 tlen = tp->t_maxseg; 7057 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start), 7058 ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p", 7059 __func__, __LINE__, 7060 rsm->r_start, tp->snd_una, tp, rack, rsm)); 7061 sb_offset = rsm->r_start - tp->snd_una; 7062 cwin = min(tp->snd_wnd, tlen); 7063 len = cwin; 7064 } else if (rack->r_ctl.rc_resend) { 7065 /* Retransmit timer */ 7066 rsm = rack->r_ctl.rc_resend; 7067 rack->r_ctl.rc_resend = NULL; 7068 len = rsm->r_end - rsm->r_start; 7069 sack_rxmit = 1; 7070 sendalot = 0; 7071 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start), 7072 ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p", 7073 __func__, __LINE__, 7074 rsm->r_start, tp->snd_una, tp, rack, rsm)); 7075 sb_offset = rsm->r_start - tp->snd_una; 7076 if (len >= tp->t_maxseg) { 7077 len = tp->t_maxseg; 7078 } 7079 } else if ((rack->rc_in_persist == 0) && 7080 ((rsm = tcp_rack_output(tp, rack, cts)) != NULL)) { 7081 long tlen; 7082 7083 if ((!IN_RECOVERY(tp->t_flags)) && 7084 ((tp->t_flags & (TF_WASFRECOVERY | TF_WASCRECOVERY)) == 0)) { 7085 /* Enter recovery if not induced by a time-out */ 7086 rack->r_ctl.rc_rsm_start = rsm->r_start; 7087 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd; 7088 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh; 7089 rack_cong_signal(tp, NULL, CC_NDUPACK); 7090 /* 7091 * When we enter recovery we need to assure we send 7092 * one packet. 7093 */ 7094 rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; 7095 } 7096 #ifdef INVARIANTS 7097 if (SEQ_LT(rsm->r_start, tp->snd_una)) { 7098 panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n", 7099 tp, rack, rsm, rsm->r_start, tp->snd_una); 7100 } 7101 #endif 7102 tlen = rsm->r_end - rsm->r_start; 7103 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start), 7104 ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p", 7105 __func__, __LINE__, 7106 rsm->r_start, tp->snd_una, tp, rack, rsm)); 7107 sb_offset = rsm->r_start - tp->snd_una; 7108 if (tlen > rack->r_ctl.rc_prr_sndcnt) { 7109 len = rack->r_ctl.rc_prr_sndcnt; 7110 } else { 7111 len = tlen; 7112 } 7113 if (len >= tp->t_maxseg) { 7114 sendalot = 1; 7115 len = tp->t_maxseg; 7116 } else { 7117 sendalot = 0; 7118 if ((rack->rc_timer_up == 0) && 7119 (len < tlen)) { 7120 /* 7121 * If its not a timer don't send a partial 7122 * segment. 7123 */ 7124 len = 0; 7125 goto just_return_nolock; 7126 } 7127 } 7128 if (len > 0) { 7129 sub_from_prr = 1; 7130 sack_rxmit = 1; 7131 TCPSTAT_INC(tcps_sack_rexmits); 7132 TCPSTAT_ADD(tcps_sack_rexmit_bytes, 7133 min(len, tp->t_maxseg)); 7134 counter_u64_add(rack_rtm_prr_retran, 1); 7135 } 7136 } 7137 if (rsm && (rsm->r_flags & RACK_HAS_FIN)) { 7138 /* we are retransmitting the fin */ 7139 len--; 7140 if (len) { 7141 /* 7142 * When retransmitting data do *not* include the 7143 * FIN. This could happen from a TLP probe. 7144 */ 7145 flags &= ~TH_FIN; 7146 } 7147 } 7148 #ifdef INVARIANTS 7149 /* For debugging */ 7150 rack->r_ctl.rc_rsm_at_retran = rsm; 7151 #endif 7152 /* 7153 * Get standard flags, and add SYN or FIN if requested by 'hidden' 7154 * state flags. 7155 */ 7156 if (tp->t_flags & TF_NEEDFIN) 7157 flags |= TH_FIN; 7158 if (tp->t_flags & TF_NEEDSYN) 7159 flags |= TH_SYN; 7160 if ((sack_rxmit == 0) && (prefetch_rsm == 0)) { 7161 void *end_rsm; 7162 end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext); 7163 if (end_rsm) 7164 kern_prefetch(end_rsm, &prefetch_rsm); 7165 prefetch_rsm = 1; 7166 } 7167 SOCKBUF_LOCK(sb); 7168 /* 7169 * If in persist timeout with window of 0, send 1 byte. Otherwise, 7170 * if window is small but nonzero and time TF_SENTFIN expired, we 7171 * will send what we can and go to transmit state. 7172 */ 7173 if (tp->t_flags & TF_FORCEDATA) { 7174 if (sendwin == 0) { 7175 /* 7176 * If we still have some data to send, then clear 7177 * the FIN bit. Usually this would happen below 7178 * when it realizes that we aren't sending all the 7179 * data. However, if we have exactly 1 byte of 7180 * unsent data, then it won't clear the FIN bit 7181 * below, and if we are in persist state, we wind up 7182 * sending the packet without recording that we sent 7183 * the FIN bit. 7184 * 7185 * We can't just blindly clear the FIN bit, because 7186 * if we don't have any more data to send then the 7187 * probe will be the FIN itself. 7188 */ 7189 if (sb_offset < sbused(sb)) 7190 flags &= ~TH_FIN; 7191 sendwin = 1; 7192 } else { 7193 if (rack->rc_in_persist) 7194 rack_exit_persist(tp, rack); 7195 /* 7196 * If we are dropping persist mode then we need to 7197 * correct snd_nxt/snd_max and off. 7198 */ 7199 tp->snd_nxt = tp->snd_max; 7200 sb_offset = tp->snd_nxt - tp->snd_una; 7201 } 7202 } 7203 /* 7204 * If snd_nxt == snd_max and we have transmitted a FIN, the 7205 * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a 7206 * negative length. This can also occur when TCP opens up its 7207 * congestion window while receiving additional duplicate acks after 7208 * fast-retransmit because TCP will reset snd_nxt to snd_max after 7209 * the fast-retransmit. 7210 * 7211 * In the normal retransmit-FIN-only case, however, snd_nxt will be 7212 * set to snd_una, the sb_offset will be 0, and the length may wind 7213 * up 0. 7214 * 7215 * If sack_rxmit is true we are retransmitting from the scoreboard 7216 * in which case len is already set. 7217 */ 7218 if (sack_rxmit == 0) { 7219 uint32_t avail; 7220 7221 avail = sbavail(sb); 7222 if (SEQ_GT(tp->snd_nxt, tp->snd_una) && avail) 7223 sb_offset = tp->snd_nxt - tp->snd_una; 7224 else 7225 sb_offset = 0; 7226 if (IN_RECOVERY(tp->t_flags) == 0) { 7227 if (rack->r_ctl.rc_tlp_new_data) { 7228 /* TLP is forcing out new data */ 7229 if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) { 7230 rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset); 7231 } 7232 if (rack->r_ctl.rc_tlp_new_data > tp->snd_wnd) 7233 len = tp->snd_wnd; 7234 else 7235 len = rack->r_ctl.rc_tlp_new_data; 7236 rack->r_ctl.rc_tlp_new_data = 0; 7237 doing_tlp = 1; 7238 } else { 7239 if (sendwin > avail) { 7240 /* use the available */ 7241 if (avail > sb_offset) { 7242 len = (int32_t)(avail - sb_offset); 7243 } else { 7244 len = 0; 7245 } 7246 } else { 7247 if (sendwin > sb_offset) { 7248 len = (int32_t)(sendwin - sb_offset); 7249 } else { 7250 len = 0; 7251 } 7252 } 7253 } 7254 } else { 7255 uint32_t outstanding; 7256 7257 /* 7258 * We are inside of a SACK recovery episode and are 7259 * sending new data, having retransmitted all the 7260 * data possible so far in the scoreboard. 7261 */ 7262 outstanding = tp->snd_max - tp->snd_una; 7263 if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) 7264 len = 0; 7265 else if (avail > sb_offset) 7266 len = avail - sb_offset; 7267 else 7268 len = 0; 7269 if (len > 0) { 7270 if (len > rack->r_ctl.rc_prr_sndcnt) 7271 len = rack->r_ctl.rc_prr_sndcnt; 7272 7273 if (len > 0) { 7274 sub_from_prr = 1; 7275 counter_u64_add(rack_rtm_prr_newdata, 1); 7276 } 7277 } 7278 if (len > tp->t_maxseg) { 7279 /* 7280 * We should never send more than a MSS when 7281 * retransmitting or sending new data in prr 7282 * mode unless the override flag is on. Most 7283 * likely the PRR algorithm is not going to 7284 * let us send a lot as well :-) 7285 */ 7286 if (rack->r_ctl.rc_prr_sendalot == 0) 7287 len = tp->t_maxseg; 7288 } else if (len < tp->t_maxseg) { 7289 /* 7290 * Do we send any? The idea here is if the 7291 * send empty's the socket buffer we want to 7292 * do it. However if not then lets just wait 7293 * for our prr_sndcnt to get bigger. 7294 */ 7295 long leftinsb; 7296 7297 leftinsb = sbavail(sb) - sb_offset; 7298 if (leftinsb > len) { 7299 /* This send does not empty the sb */ 7300 len = 0; 7301 } 7302 } 7303 } 7304 } 7305 if (prefetch_so_done == 0) { 7306 kern_prefetch(so, &prefetch_so_done); 7307 prefetch_so_done = 1; 7308 } 7309 /* 7310 * Lop off SYN bit if it has already been sent. However, if this is 7311 * SYN-SENT state and if segment contains data and if we don't know 7312 * that foreign host supports TAO, suppress sending segment. 7313 */ 7314 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una) && 7315 ((sack_rxmit == 0) && (tp->t_rxtshift == 0))) { 7316 if (tp->t_state != TCPS_SYN_RECEIVED) 7317 flags &= ~TH_SYN; 7318 /* 7319 * When sending additional segments following a TFO SYN|ACK, 7320 * do not include the SYN bit. 7321 */ 7322 if (IS_FASTOPEN(tp->t_flags) && 7323 (tp->t_state == TCPS_SYN_RECEIVED)) 7324 flags &= ~TH_SYN; 7325 sb_offset--, len++; 7326 } 7327 /* 7328 * Be careful not to send data and/or FIN on SYN segments. This 7329 * measure is needed to prevent interoperability problems with not 7330 * fully conformant TCP implementations. 7331 */ 7332 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) { 7333 len = 0; 7334 flags &= ~TH_FIN; 7335 } 7336 /* 7337 * On TFO sockets, ensure no data is sent in the following cases: 7338 * 7339 * - When retransmitting SYN|ACK on a passively-created socket 7340 * 7341 * - When retransmitting SYN on an actively created socket 7342 * 7343 * - When sending a zero-length cookie (cookie request) on an 7344 * actively created socket 7345 * 7346 * - When the socket is in the CLOSED state (RST is being sent) 7347 */ 7348 if (IS_FASTOPEN(tp->t_flags) && 7349 (((flags & TH_SYN) && (tp->t_rxtshift > 0)) || 7350 ((tp->t_state == TCPS_SYN_SENT) && 7351 (tp->t_tfo_client_cookie_len == 0)) || 7352 (flags & TH_RST))) { 7353 sack_rxmit = 0; 7354 len = 0; 7355 } 7356 /* Without fast-open there should never be data sent on a SYN */ 7357 if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags))) 7358 len = 0; 7359 if (len <= 0) { 7360 /* 7361 * If FIN has been sent but not acked, but we haven't been 7362 * called to retransmit, len will be < 0. Otherwise, window 7363 * shrank after we sent into it. If window shrank to 0, 7364 * cancel pending retransmit, pull snd_nxt back to (closed) 7365 * window, and set the persist timer if it isn't already 7366 * going. If the window didn't close completely, just wait 7367 * for an ACK. 7368 * 7369 * We also do a general check here to ensure that we will 7370 * set the persist timer when we have data to send, but a 7371 * 0-byte window. This makes sure the persist timer is set 7372 * even if the packet hits one of the "goto send" lines 7373 * below. 7374 */ 7375 len = 0; 7376 if ((tp->snd_wnd == 0) && 7377 (TCPS_HAVEESTABLISHED(tp->t_state)) && 7378 (sb_offset < (int)sbavail(sb))) { 7379 tp->snd_nxt = tp->snd_una; 7380 rack_enter_persist(tp, rack, cts); 7381 } 7382 } 7383 /* len will be >= 0 after this point. */ 7384 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 7385 tcp_sndbuf_autoscale(tp, so, sendwin); 7386 /* 7387 * Decide if we can use TCP Segmentation Offloading (if supported by 7388 * hardware). 7389 * 7390 * TSO may only be used if we are in a pure bulk sending state. The 7391 * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP 7392 * options prevent using TSO. With TSO the TCP header is the same 7393 * (except for the sequence number) for all generated packets. This 7394 * makes it impossible to transmit any options which vary per 7395 * generated segment or packet. 7396 * 7397 * IPv4 handling has a clear separation of ip options and ip header 7398 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does 7399 * the right thing below to provide length of just ip options and thus 7400 * checking for ipoptlen is enough to decide if ip options are present. 7401 */ 7402 7403 #ifdef INET6 7404 if (isipv6) 7405 ipoptlen = ip6_optlen(tp->t_inpcb); 7406 else 7407 #endif 7408 if (tp->t_inpcb->inp_options) 7409 ipoptlen = tp->t_inpcb->inp_options->m_len - 7410 offsetof(struct ipoption, ipopt_list); 7411 else 7412 ipoptlen = 0; 7413 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 7414 /* 7415 * Pre-calculate here as we save another lookup into the darknesses 7416 * of IPsec that way and can actually decide if TSO is ok. 7417 */ 7418 #ifdef INET6 7419 if (isipv6 && IPSEC_ENABLED(ipv6)) 7420 ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb); 7421 #ifdef INET 7422 else 7423 #endif 7424 #endif /* INET6 */ 7425 #ifdef INET 7426 if (IPSEC_ENABLED(ipv4)) 7427 ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb); 7428 #endif /* INET */ 7429 #endif 7430 7431 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 7432 ipoptlen += ipsec_optlen; 7433 #endif 7434 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg && 7435 (tp->t_port == 0) && 7436 ((tp->t_flags & TF_SIGNATURE) == 0) && 7437 tp->rcv_numsacks == 0 && sack_rxmit == 0 && 7438 ipoptlen == 0) 7439 tso = 1; 7440 { 7441 uint32_t outstanding; 7442 7443 outstanding = tp->snd_max - tp->snd_una; 7444 if (tp->t_flags & TF_SENTFIN) { 7445 /* 7446 * If we sent a fin, snd_max is 1 higher than 7447 * snd_una 7448 */ 7449 outstanding--; 7450 } 7451 if (outstanding > 0) { 7452 /* 7453 * This is sub-optimal. We only send a stand alone 7454 * FIN on its own segment. 7455 */ 7456 if (flags & TH_FIN) { 7457 flags &= ~TH_FIN; 7458 would_have_fin = 1; 7459 } 7460 } else if (sack_rxmit) { 7461 if ((rsm->r_flags & RACK_HAS_FIN) == 0) 7462 flags &= ~TH_FIN; 7463 } else { 7464 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + 7465 sbused(sb))) 7466 flags &= ~TH_FIN; 7467 } 7468 } 7469 recwin = sbspace(&so->so_rcv); 7470 7471 /* 7472 * Sender silly window avoidance. We transmit under the following 7473 * conditions when len is non-zero: 7474 * 7475 * - We have a full segment (or more with TSO) - This is the last 7476 * buffer in a write()/send() and we are either idle or running 7477 * NODELAY - we've timed out (e.g. persist timer) - we have more 7478 * then 1/2 the maximum send window's worth of data (receiver may be 7479 * limited the window size) - we need to retransmit 7480 */ 7481 if (len) { 7482 if (len >= tp->t_maxseg) { 7483 pass = 1; 7484 goto send; 7485 } 7486 /* 7487 * NOTE! on localhost connections an 'ack' from the remote 7488 * end may occur synchronously with the output and cause us 7489 * to flush a buffer queued with moretocome. XXX 7490 * 7491 */ 7492 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */ 7493 (idle || (tp->t_flags & TF_NODELAY)) && 7494 ((uint32_t)len + (uint32_t)sb_offset >= sbavail(&so->so_snd)) && 7495 (tp->t_flags & TF_NOPUSH) == 0) { 7496 pass = 2; 7497 goto send; 7498 } 7499 if (tp->t_flags & TF_FORCEDATA) { /* typ. timeout case */ 7500 pass = 3; 7501 goto send; 7502 } 7503 if ((tp->snd_una == tp->snd_max) && len) { /* Nothing outstanding */ 7504 goto send; 7505 } 7506 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) { 7507 pass = 4; 7508 goto send; 7509 } 7510 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { /* retransmit case */ 7511 pass = 5; 7512 goto send; 7513 } 7514 if (sack_rxmit) { 7515 pass = 6; 7516 goto send; 7517 } 7518 } 7519 /* 7520 * Sending of standalone window updates. 7521 * 7522 * Window updates are important when we close our window due to a 7523 * full socket buffer and are opening it again after the application 7524 * reads data from it. Once the window has opened again and the 7525 * remote end starts to send again the ACK clock takes over and 7526 * provides the most current window information. 7527 * 7528 * We must avoid the silly window syndrome whereas every read from 7529 * the receive buffer, no matter how small, causes a window update 7530 * to be sent. We also should avoid sending a flurry of window 7531 * updates when the socket buffer had queued a lot of data and the 7532 * application is doing small reads. 7533 * 7534 * Prevent a flurry of pointless window updates by only sending an 7535 * update when we can increase the advertized window by more than 7536 * 1/4th of the socket buffer capacity. When the buffer is getting 7537 * full or is very small be more aggressive and send an update 7538 * whenever we can increase by two mss sized segments. In all other 7539 * situations the ACK's to new incoming data will carry further 7540 * window increases. 7541 * 7542 * Don't send an independent window update if a delayed ACK is 7543 * pending (it will get piggy-backed on it) or the remote side 7544 * already has done a half-close and won't send more data. Skip 7545 * this if the connection is in T/TCP half-open state. 7546 */ 7547 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) && 7548 !(tp->t_flags & TF_DELACK) && 7549 !TCPS_HAVERCVDFIN(tp->t_state)) { 7550 /* 7551 * "adv" is the amount we could increase the window, taking 7552 * into account that we are limited by TCP_MAXWIN << 7553 * tp->rcv_scale. 7554 */ 7555 int32_t adv; 7556 int oldwin; 7557 7558 adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale); 7559 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) { 7560 oldwin = (tp->rcv_adv - tp->rcv_nxt); 7561 adv -= oldwin; 7562 } else 7563 oldwin = 0; 7564 7565 /* 7566 * If the new window size ends up being the same as the old 7567 * size when it is scaled, then don't force a window update. 7568 */ 7569 if (oldwin >> tp->rcv_scale == (adv + oldwin) >> tp->rcv_scale) 7570 goto dontupdate; 7571 7572 if (adv >= (int32_t)(2 * tp->t_maxseg) && 7573 (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) || 7574 recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) || 7575 so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg)) { 7576 pass = 7; 7577 goto send; 7578 } 7579 if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) 7580 goto send; 7581 } 7582 dontupdate: 7583 7584 /* 7585 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW 7586 * is also a catch-all for the retransmit timer timeout case. 7587 */ 7588 if (tp->t_flags & TF_ACKNOW) { 7589 pass = 8; 7590 goto send; 7591 } 7592 if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) { 7593 pass = 9; 7594 goto send; 7595 } 7596 if (SEQ_GT(tp->snd_up, tp->snd_una)) { 7597 pass = 10; 7598 goto send; 7599 } 7600 /* 7601 * If our state indicates that FIN should be sent and we have not 7602 * yet done so, then we need to send. 7603 */ 7604 if ((flags & TH_FIN) && 7605 (tp->snd_nxt == tp->snd_una)) { 7606 pass = 11; 7607 goto send; 7608 } 7609 /* 7610 * No reason to send a segment, just return. 7611 */ 7612 just_return: 7613 SOCKBUF_UNLOCK(sb); 7614 just_return_nolock: 7615 if (tot_len_this_send == 0) 7616 counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1); 7617 rack_start_hpts_timer(rack, tp, cts, __LINE__, slot, tot_len_this_send, 1); 7618 rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling); 7619 tp->t_flags &= ~TF_FORCEDATA; 7620 return (0); 7621 7622 send: 7623 if (doing_tlp == 0) { 7624 /* 7625 * Data not a TLP, and its not the rxt firing. If it is the 7626 * rxt firing, we want to leave the tlp_in_progress flag on 7627 * so we don't send another TLP. It has to be a rack timer 7628 * or normal send (response to acked data) to clear the tlp 7629 * in progress flag. 7630 */ 7631 rack->rc_tlp_in_progress = 0; 7632 } 7633 SOCKBUF_LOCK_ASSERT(sb); 7634 if (len > 0) { 7635 if (len >= tp->t_maxseg) 7636 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT; 7637 else 7638 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT; 7639 } 7640 /* 7641 * Before ESTABLISHED, force sending of initial options unless TCP 7642 * set not to do any options. NOTE: we assume that the IP/TCP header 7643 * plus TCP options always fit in a single mbuf, leaving room for a 7644 * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr) 7645 * + optlen <= MCLBYTES 7646 */ 7647 optlen = 0; 7648 #ifdef INET6 7649 if (isipv6) 7650 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 7651 else 7652 #endif 7653 hdrlen = sizeof(struct tcpiphdr); 7654 7655 /* 7656 * Compute options for segment. We only have to care about SYN and 7657 * established connection segments. Options for SYN-ACK segments 7658 * are handled in TCP syncache. 7659 */ 7660 to.to_flags = 0; 7661 if ((tp->t_flags & TF_NOOPT) == 0) { 7662 /* Maximum segment size. */ 7663 if (flags & TH_SYN) { 7664 tp->snd_nxt = tp->iss; 7665 to.to_mss = tcp_mssopt(&inp->inp_inc); 7666 #ifdef NETFLIX_TCPOUDP 7667 if (tp->t_port) 7668 to.to_mss -= V_tcp_udp_tunneling_overhead; 7669 #endif 7670 to.to_flags |= TOF_MSS; 7671 7672 /* 7673 * On SYN or SYN|ACK transmits on TFO connections, 7674 * only include the TFO option if it is not a 7675 * retransmit, as the presence of the TFO option may 7676 * have caused the original SYN or SYN|ACK to have 7677 * been dropped by a middlebox. 7678 */ 7679 if (IS_FASTOPEN(tp->t_flags) && 7680 (tp->t_rxtshift == 0)) { 7681 if (tp->t_state == TCPS_SYN_RECEIVED) { 7682 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; 7683 to.to_tfo_cookie = 7684 (u_int8_t *)&tp->t_tfo_cookie.server; 7685 to.to_flags |= TOF_FASTOPEN; 7686 wanted_cookie = 1; 7687 } else if (tp->t_state == TCPS_SYN_SENT) { 7688 to.to_tfo_len = 7689 tp->t_tfo_client_cookie_len; 7690 to.to_tfo_cookie = 7691 tp->t_tfo_cookie.client; 7692 to.to_flags |= TOF_FASTOPEN; 7693 wanted_cookie = 1; 7694 /* 7695 * If we wind up having more data to 7696 * send with the SYN than can fit in 7697 * one segment, don't send any more 7698 * until the SYN|ACK comes back from 7699 * the other end. 7700 */ 7701 sendalot = 0; 7702 } 7703 } 7704 } 7705 /* Window scaling. */ 7706 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) { 7707 to.to_wscale = tp->request_r_scale; 7708 to.to_flags |= TOF_SCALE; 7709 } 7710 /* Timestamps. */ 7711 if ((tp->t_flags & TF_RCVD_TSTMP) || 7712 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) { 7713 to.to_tsval = cts + tp->ts_offset; 7714 to.to_tsecr = tp->ts_recent; 7715 to.to_flags |= TOF_TS; 7716 } 7717 /* Set receive buffer autosizing timestamp. */ 7718 if (tp->rfbuf_ts == 0 && 7719 (so->so_rcv.sb_flags & SB_AUTOSIZE)) 7720 tp->rfbuf_ts = tcp_ts_getticks(); 7721 /* Selective ACK's. */ 7722 if (flags & TH_SYN) 7723 to.to_flags |= TOF_SACKPERM; 7724 else if (TCPS_HAVEESTABLISHED(tp->t_state) && 7725 tp->rcv_numsacks > 0) { 7726 to.to_flags |= TOF_SACK; 7727 to.to_nsacks = tp->rcv_numsacks; 7728 to.to_sacks = (u_char *)tp->sackblks; 7729 } 7730 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 7731 /* TCP-MD5 (RFC2385). */ 7732 if (tp->t_flags & TF_SIGNATURE) 7733 to.to_flags |= TOF_SIGNATURE; 7734 #endif /* TCP_SIGNATURE */ 7735 7736 /* Processing the options. */ 7737 hdrlen += optlen = tcp_addoptions(&to, opt); 7738 /* 7739 * If we wanted a TFO option to be added, but it was unable 7740 * to fit, ensure no data is sent. 7741 */ 7742 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie && 7743 !(to.to_flags & TOF_FASTOPEN)) 7744 len = 0; 7745 } 7746 #ifdef NETFLIX_TCPOUDP 7747 if (tp->t_port) { 7748 if (V_tcp_udp_tunneling_port == 0) { 7749 /* The port was removed?? */ 7750 SOCKBUF_UNLOCK(&so->so_snd); 7751 return (EHOSTUNREACH); 7752 } 7753 hdrlen += sizeof(struct udphdr); 7754 } 7755 #endif 7756 ipoptlen = 0; 7757 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 7758 ipoptlen += ipsec_optlen; 7759 #endif 7760 7761 /* 7762 * Adjust data length if insertion of options will bump the packet 7763 * length beyond the t_maxseg length. Clear the FIN bit because we 7764 * cut off the tail of the segment. 7765 */ 7766 if (len + optlen + ipoptlen > tp->t_maxseg) { 7767 if (flags & TH_FIN) { 7768 would_have_fin = 1; 7769 flags &= ~TH_FIN; 7770 } 7771 if (tso) { 7772 uint32_t if_hw_tsomax; 7773 uint32_t moff; 7774 int32_t max_len; 7775 7776 /* extract TSO information */ 7777 if_hw_tsomax = tp->t_tsomax; 7778 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount; 7779 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize; 7780 KASSERT(ipoptlen == 0, 7781 ("%s: TSO can't do IP options", __func__)); 7782 7783 /* 7784 * Check if we should limit by maximum payload 7785 * length: 7786 */ 7787 if (if_hw_tsomax != 0) { 7788 /* compute maximum TSO length */ 7789 max_len = (if_hw_tsomax - hdrlen - 7790 max_linkhdr); 7791 if (max_len <= 0) { 7792 len = 0; 7793 } else if (len > max_len) { 7794 sendalot = 1; 7795 len = max_len; 7796 } 7797 } 7798 /* 7799 * Prevent the last segment from being fractional 7800 * unless the send sockbuf can be emptied: 7801 */ 7802 max_len = (tp->t_maxseg - optlen); 7803 if ((sb_offset + len) < sbavail(sb)) { 7804 moff = len % (u_int)max_len; 7805 if (moff != 0) { 7806 len -= moff; 7807 sendalot = 1; 7808 } 7809 } 7810 /* 7811 * In case there are too many small fragments don't 7812 * use TSO: 7813 */ 7814 if (len <= max_len) { 7815 len = max_len; 7816 sendalot = 1; 7817 tso = 0; 7818 } 7819 /* 7820 * Send the FIN in a separate segment after the bulk 7821 * sending is done. We don't trust the TSO 7822 * implementations to clear the FIN flag on all but 7823 * the last segment. 7824 */ 7825 if (tp->t_flags & TF_NEEDFIN) 7826 sendalot = 1; 7827 7828 } else { 7829 len = tp->t_maxseg - optlen - ipoptlen; 7830 sendalot = 1; 7831 } 7832 } else 7833 tso = 0; 7834 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET, 7835 ("%s: len > IP_MAXPACKET", __func__)); 7836 #ifdef DIAGNOSTIC 7837 #ifdef INET6 7838 if (max_linkhdr + hdrlen > MCLBYTES) 7839 #else 7840 if (max_linkhdr + hdrlen > MHLEN) 7841 #endif 7842 panic("tcphdr too big"); 7843 #endif 7844 7845 /* 7846 * This KASSERT is here to catch edge cases at a well defined place. 7847 * Before, those had triggered (random) panic conditions further 7848 * down. 7849 */ 7850 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 7851 if ((len == 0) && 7852 (flags & TH_FIN) && 7853 (sbused(sb))) { 7854 /* 7855 * We have outstanding data, don't send a fin by itself!. 7856 */ 7857 goto just_return; 7858 } 7859 /* 7860 * Grab a header mbuf, attaching a copy of data to be transmitted, 7861 * and initialize the header from the template for sends on this 7862 * connection. 7863 */ 7864 if (len) { 7865 uint32_t max_val; 7866 uint32_t moff; 7867 7868 if (rack->rc_pace_max_segs) 7869 max_val = rack->rc_pace_max_segs * tp->t_maxseg; 7870 else 7871 max_val = len; 7872 /* 7873 * We allow a limit on sending with hptsi. 7874 */ 7875 if (len > max_val) { 7876 len = max_val; 7877 } 7878 #ifdef INET6 7879 if (MHLEN < hdrlen + max_linkhdr) 7880 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); 7881 else 7882 #endif 7883 m = m_gethdr(M_NOWAIT, MT_DATA); 7884 7885 if (m == NULL) { 7886 SOCKBUF_UNLOCK(sb); 7887 error = ENOBUFS; 7888 sack_rxmit = 0; 7889 goto out; 7890 } 7891 m->m_data += max_linkhdr; 7892 m->m_len = hdrlen; 7893 7894 /* 7895 * Start the m_copy functions from the closest mbuf to the 7896 * sb_offset in the socket buffer chain. 7897 */ 7898 mb = sbsndptr_noadv(sb, sb_offset, &moff); 7899 if (len <= MHLEN - hdrlen - max_linkhdr) { 7900 m_copydata(mb, moff, (int)len, 7901 mtod(m, caddr_t)+hdrlen); 7902 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 7903 sbsndptr_adv(sb, mb, len); 7904 m->m_len += len; 7905 } else { 7906 struct sockbuf *msb; 7907 7908 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 7909 msb = NULL; 7910 else 7911 msb = sb; 7912 m->m_next = tcp_m_copym(mb, moff, &len, 7913 if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb); 7914 if (len <= (tp->t_maxseg - optlen)) { 7915 /* 7916 * Must have ran out of mbufs for the copy 7917 * shorten it to no longer need tso. Lets 7918 * not put on sendalot since we are low on 7919 * mbufs. 7920 */ 7921 tso = 0; 7922 } 7923 if (m->m_next == NULL) { 7924 SOCKBUF_UNLOCK(sb); 7925 (void)m_free(m); 7926 error = ENOBUFS; 7927 sack_rxmit = 0; 7928 goto out; 7929 } 7930 } 7931 if ((tp->t_flags & TF_FORCEDATA) && len == 1) { 7932 TCPSTAT_INC(tcps_sndprobe); 7933 #ifdef NETFLIX_STATS 7934 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 7935 stats_voi_update_abs_u32(tp->t_stats, 7936 VOI_TCP_RETXPB, len); 7937 else 7938 stats_voi_update_abs_u64(tp->t_stats, 7939 VOI_TCP_TXPB, len); 7940 #endif 7941 } else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) { 7942 if (rsm && (rsm->r_flags & RACK_TLP)) { 7943 /* 7944 * TLP should not count in retran count, but 7945 * in its own bin 7946 */ 7947 counter_u64_add(rack_tlp_retran, 1); 7948 counter_u64_add(rack_tlp_retran_bytes, len); 7949 } else { 7950 tp->t_sndrexmitpack++; 7951 TCPSTAT_INC(tcps_sndrexmitpack); 7952 TCPSTAT_ADD(tcps_sndrexmitbyte, len); 7953 } 7954 #ifdef NETFLIX_STATS 7955 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB, 7956 len); 7957 #endif 7958 } else { 7959 TCPSTAT_INC(tcps_sndpack); 7960 TCPSTAT_ADD(tcps_sndbyte, len); 7961 #ifdef NETFLIX_STATS 7962 stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB, 7963 len); 7964 #endif 7965 } 7966 /* 7967 * If we're sending everything we've got, set PUSH. (This 7968 * will keep happy those implementations which only give 7969 * data to the user when a buffer fills or a PUSH comes in.) 7970 */ 7971 if (sb_offset + len == sbused(sb) && 7972 sbused(sb) && 7973 !(flags & TH_SYN)) 7974 flags |= TH_PUSH; 7975 7976 /* 7977 * Are we doing hptsi, if so we must calculate the slot. We 7978 * only do hptsi in ESTABLISHED and with no RESET being 7979 * sent where we have data to send. 7980 */ 7981 if (((tp->t_state == TCPS_ESTABLISHED) || 7982 (tp->t_state == TCPS_CLOSE_WAIT) || 7983 ((tp->t_state == TCPS_FIN_WAIT_1) && 7984 ((tp->t_flags & TF_SENTFIN) == 0) && 7985 ((flags & TH_FIN) == 0))) && 7986 ((flags & TH_RST) == 0) && 7987 (rack->rc_always_pace)) { 7988 /* 7989 * We use the most optimistic possible cwnd/srtt for 7990 * sending calculations. This will make our 7991 * calculation anticipate getting more through 7992 * quicker then possible. But thats ok we don't want 7993 * the peer to have a gap in data sending. 7994 */ 7995 uint32_t srtt, cwnd, tr_perms = 0; 7996 7997 if (rack->r_ctl.rc_rack_min_rtt) 7998 srtt = rack->r_ctl.rc_rack_min_rtt; 7999 else 8000 srtt = TICKS_2_MSEC((tp->t_srtt >> TCP_RTT_SHIFT)); 8001 if (rack->r_ctl.rc_rack_largest_cwnd) 8002 cwnd = rack->r_ctl.rc_rack_largest_cwnd; 8003 else 8004 cwnd = tp->snd_cwnd; 8005 tr_perms = cwnd / srtt; 8006 if (tr_perms == 0) { 8007 tr_perms = tp->t_maxseg; 8008 } 8009 tot_len_this_send += len; 8010 /* 8011 * Calculate how long this will take to drain, if 8012 * the calculation comes out to zero, thats ok we 8013 * will use send_a_lot to possibly spin around for 8014 * more increasing tot_len_this_send to the point 8015 * that its going to require a pace, or we hit the 8016 * cwnd. Which in that case we are just waiting for 8017 * a ACK. 8018 */ 8019 slot = tot_len_this_send / tr_perms; 8020 /* Now do we reduce the time so we don't run dry? */ 8021 if (slot && rack->rc_pace_reduce) { 8022 int32_t reduce; 8023 8024 reduce = (slot / rack->rc_pace_reduce); 8025 if (reduce < slot) { 8026 slot -= reduce; 8027 } else 8028 slot = 0; 8029 } 8030 if (rack->r_enforce_min_pace && 8031 (slot == 0) && 8032 (tot_len_this_send >= (rack->r_min_pace_seg_thresh * tp->t_maxseg))) { 8033 /* We are enforcing a minimum pace time of 1ms */ 8034 slot = rack->r_enforce_min_pace; 8035 } 8036 } 8037 SOCKBUF_UNLOCK(sb); 8038 } else { 8039 SOCKBUF_UNLOCK(sb); 8040 if (tp->t_flags & TF_ACKNOW) 8041 TCPSTAT_INC(tcps_sndacks); 8042 else if (flags & (TH_SYN | TH_FIN | TH_RST)) 8043 TCPSTAT_INC(tcps_sndctrl); 8044 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 8045 TCPSTAT_INC(tcps_sndurg); 8046 else 8047 TCPSTAT_INC(tcps_sndwinup); 8048 8049 m = m_gethdr(M_NOWAIT, MT_DATA); 8050 if (m == NULL) { 8051 error = ENOBUFS; 8052 sack_rxmit = 0; 8053 goto out; 8054 } 8055 #ifdef INET6 8056 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && 8057 MHLEN >= hdrlen) { 8058 M_ALIGN(m, hdrlen); 8059 } else 8060 #endif 8061 m->m_data += max_linkhdr; 8062 m->m_len = hdrlen; 8063 } 8064 SOCKBUF_UNLOCK_ASSERT(sb); 8065 m->m_pkthdr.rcvif = (struct ifnet *)0; 8066 #ifdef MAC 8067 mac_inpcb_create_mbuf(inp, m); 8068 #endif 8069 #ifdef INET6 8070 if (isipv6) { 8071 ip6 = mtod(m, struct ip6_hdr *); 8072 #ifdef NETFLIX_TCPOUDP 8073 if (tp->t_port) { 8074 udp = (struct udphdr *)((caddr_t)ip6 + ipoptlen + sizeof(struct ip6_hdr)); 8075 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 8076 udp->uh_dport = tp->t_port; 8077 ulen = hdrlen + len - sizeof(struct ip6_hdr); 8078 udp->uh_ulen = htons(ulen); 8079 th = (struct tcphdr *)(udp + 1); 8080 } else 8081 #endif 8082 th = (struct tcphdr *)(ip6 + 1); 8083 tcpip_fillheaders(inp, ip6, th); 8084 } else 8085 #endif /* INET6 */ 8086 { 8087 ip = mtod(m, struct ip *); 8088 #ifdef TCPDEBUG 8089 ipov = (struct ipovly *)ip; 8090 #endif 8091 #ifdef NETFLIX_TCPOUDP 8092 if (tp->t_port) { 8093 udp = (struct udphdr *)((caddr_t)ip + ipoptlen + sizeof(struct ip)); 8094 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 8095 udp->uh_dport = tp->t_port; 8096 ulen = hdrlen + len - sizeof(struct ip); 8097 udp->uh_ulen = htons(ulen); 8098 th = (struct tcphdr *)(udp + 1); 8099 } else 8100 #endif 8101 th = (struct tcphdr *)(ip + 1); 8102 tcpip_fillheaders(inp, ip, th); 8103 } 8104 /* 8105 * Fill in fields, remembering maximum advertised window for use in 8106 * delaying messages about window sizes. If resending a FIN, be sure 8107 * not to use a new sequence number. 8108 */ 8109 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && 8110 tp->snd_nxt == tp->snd_max) 8111 tp->snd_nxt--; 8112 /* 8113 * If we are starting a connection, send ECN setup SYN packet. If we 8114 * are on a retransmit, we may resend those bits a number of times 8115 * as per RFC 3168. 8116 */ 8117 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) { 8118 if (tp->t_rxtshift >= 1) { 8119 if (tp->t_rxtshift <= V_tcp_ecn_maxretries) 8120 flags |= TH_ECE | TH_CWR; 8121 } else 8122 flags |= TH_ECE | TH_CWR; 8123 } 8124 if (tp->t_state == TCPS_ESTABLISHED && 8125 (tp->t_flags & TF_ECN_PERMIT)) { 8126 /* 8127 * If the peer has ECN, mark data packets with ECN capable 8128 * transmission (ECT). Ignore pure ack packets, 8129 * retransmissions and window probes. 8130 */ 8131 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) && 8132 !((tp->t_flags & TF_FORCEDATA) && len == 1)) { 8133 #ifdef INET6 8134 if (isipv6) 8135 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20); 8136 else 8137 #endif 8138 ip->ip_tos |= IPTOS_ECN_ECT0; 8139 TCPSTAT_INC(tcps_ecn_ect0); 8140 } 8141 /* 8142 * Reply with proper ECN notifications. 8143 */ 8144 if (tp->t_flags & TF_ECN_SND_CWR) { 8145 flags |= TH_CWR; 8146 tp->t_flags &= ~TF_ECN_SND_CWR; 8147 } 8148 if (tp->t_flags & TF_ECN_SND_ECE) 8149 flags |= TH_ECE; 8150 } 8151 /* 8152 * If we are doing retransmissions, then snd_nxt will not reflect 8153 * the first unsent octet. For ACK only packets, we do not want the 8154 * sequence number of the retransmitted packet, we want the sequence 8155 * number of the next unsent octet. So, if there is no data (and no 8156 * SYN or FIN), use snd_max instead of snd_nxt when filling in 8157 * ti_seq. But if we are in persist state, snd_max might reflect 8158 * one byte beyond the right edge of the window, so use snd_nxt in 8159 * that case, since we know we aren't doing a retransmission. 8160 * (retransmit and persist are mutually exclusive...) 8161 */ 8162 if (sack_rxmit == 0) { 8163 if (len || (flags & (TH_SYN | TH_FIN)) || 8164 rack->rc_in_persist) { 8165 th->th_seq = htonl(tp->snd_nxt); 8166 rack_seq = tp->snd_nxt; 8167 } else if (flags & TH_RST) { 8168 /* 8169 * For a Reset send the last cum ack in sequence 8170 * (this like any other choice may still generate a 8171 * challenge ack, if a ack-update packet is in 8172 * flight). 8173 */ 8174 th->th_seq = htonl(tp->snd_una); 8175 rack_seq = tp->snd_una; 8176 } else { 8177 th->th_seq = htonl(tp->snd_max); 8178 rack_seq = tp->snd_max; 8179 } 8180 } else { 8181 th->th_seq = htonl(rsm->r_start); 8182 rack_seq = rsm->r_start; 8183 } 8184 th->th_ack = htonl(tp->rcv_nxt); 8185 if (optlen) { 8186 bcopy(opt, th + 1, optlen); 8187 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 8188 } 8189 th->th_flags = flags; 8190 /* 8191 * Calculate receive window. Don't shrink window, but avoid silly 8192 * window syndrome. 8193 * If a RST segment is sent, advertise a window of zero. 8194 */ 8195 if (flags & TH_RST) { 8196 recwin = 0; 8197 } else { 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 /* 8209 * According to RFC1323 the window field in a SYN (i.e., a <SYN> or 8210 * <SYN,ACK>) segment itself is never scaled. The <SYN,ACK> case is 8211 * handled in syncache. 8212 */ 8213 if (flags & TH_SYN) 8214 th->th_win = htons((u_short) 8215 (min(sbspace(&so->so_rcv), TCP_MAXWIN))); 8216 else 8217 th->th_win = htons((u_short)(recwin >> tp->rcv_scale)); 8218 /* 8219 * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0 8220 * window. This may cause the remote transmitter to stall. This 8221 * flag tells soreceive() to disable delayed acknowledgements when 8222 * draining the buffer. This can occur if the receiver is 8223 * attempting to read more data than can be buffered prior to 8224 * transmitting on the connection. 8225 */ 8226 if (th->th_win == 0) { 8227 tp->t_sndzerowin++; 8228 tp->t_flags |= TF_RXWIN0SENT; 8229 } else 8230 tp->t_flags &= ~TF_RXWIN0SENT; 8231 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 8232 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); 8233 th->th_flags |= TH_URG; 8234 } else 8235 /* 8236 * If no urgent pointer to send, then we pull the urgent 8237 * pointer to the left edge of the send window so that it 8238 * doesn't drift into the send window on sequence number 8239 * wraparound. 8240 */ 8241 tp->snd_up = tp->snd_una; /* drag it along */ 8242 8243 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 8244 if (to.to_flags & TOF_SIGNATURE) { 8245 /* 8246 * Calculate MD5 signature and put it into the place 8247 * determined before. 8248 * NOTE: since TCP options buffer doesn't point into 8249 * mbuf's data, calculate offset and use it. 8250 */ 8251 if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th, 8252 (u_char *)(th + 1) + (to.to_signature - opt)) != 0) { 8253 /* 8254 * Do not send segment if the calculation of MD5 8255 * digest has failed. 8256 */ 8257 goto out; 8258 } 8259 } 8260 #endif 8261 8262 /* 8263 * Put TCP length in extended header, and then checksum extended 8264 * header and data. 8265 */ 8266 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 8267 #ifdef INET6 8268 if (isipv6) { 8269 /* 8270 * ip6_plen is not need to be filled now, and will be filled 8271 * in ip6_output. 8272 */ 8273 if (tp->t_port) { 8274 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6; 8275 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 8276 udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0); 8277 th->th_sum = htons(0); 8278 } else { 8279 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 8280 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 8281 th->th_sum = in6_cksum_pseudo(ip6, 8282 sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP, 8283 0); 8284 } 8285 } 8286 #endif 8287 #if defined(INET6) && defined(INET) 8288 else 8289 #endif 8290 #ifdef INET 8291 { 8292 if (tp->t_port) { 8293 m->m_pkthdr.csum_flags = CSUM_UDP; 8294 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 8295 udp->uh_sum = in_pseudo(ip->ip_src.s_addr, 8296 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP)); 8297 th->th_sum = htons(0); 8298 } else { 8299 m->m_pkthdr.csum_flags = CSUM_TCP; 8300 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 8301 th->th_sum = in_pseudo(ip->ip_src.s_addr, 8302 ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) + 8303 IPPROTO_TCP + len + optlen)); 8304 } 8305 /* IP version must be set here for ipv4/ipv6 checking later */ 8306 KASSERT(ip->ip_v == IPVERSION, 8307 ("%s: IP version incorrect: %d", __func__, ip->ip_v)); 8308 } 8309 #endif 8310 8311 /* 8312 * Enable TSO and specify the size of the segments. The TCP pseudo 8313 * header checksum is always provided. XXX: Fixme: This is currently 8314 * not the case for IPv6. 8315 */ 8316 if (tso) { 8317 KASSERT(len > tp->t_maxseg - optlen, 8318 ("%s: len <= tso_segsz", __func__)); 8319 m->m_pkthdr.csum_flags |= CSUM_TSO; 8320 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen; 8321 } 8322 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 8323 KASSERT(len + hdrlen + ipoptlen - ipsec_optlen == m_length(m, NULL), 8324 ("%s: mbuf chain shorter than expected: %d + %u + %u - %u != %u", 8325 __func__, len, hdrlen, ipoptlen, ipsec_optlen, m_length(m, NULL))); 8326 #else 8327 KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL), 8328 ("%s: mbuf chain shorter than expected: %d + %u + %u != %u", 8329 __func__, len, hdrlen, ipoptlen, m_length(m, NULL))); 8330 #endif 8331 8332 #ifdef TCP_HHOOK 8333 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */ 8334 hhook_run_tcp_est_out(tp, th, &to, len, tso); 8335 #endif 8336 8337 #ifdef TCPDEBUG 8338 /* 8339 * Trace. 8340 */ 8341 if (so->so_options & SO_DEBUG) { 8342 u_short save = 0; 8343 8344 #ifdef INET6 8345 if (!isipv6) 8346 #endif 8347 { 8348 save = ipov->ih_len; 8349 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + 8350 * (th->th_off << 2) */ ); 8351 } 8352 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0); 8353 #ifdef INET6 8354 if (!isipv6) 8355 #endif 8356 ipov->ih_len = save; 8357 } 8358 #endif /* TCPDEBUG */ 8359 8360 /* We're getting ready to send; log now. */ 8361 if (tp->t_logstate != TCP_LOG_STATE_OFF) { 8362 union tcp_log_stackspecific log; 8363 8364 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 8365 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; 8366 log.u_bbr.ininput = rack->rc_inp->inp_in_input; 8367 log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt; 8368 if (rsm || sack_rxmit) { 8369 log.u_bbr.flex8 = 1; 8370 } else { 8371 log.u_bbr.flex8 = 0; 8372 } 8373 lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK, 8374 len, &log, false, NULL, NULL, 0, NULL); 8375 } else 8376 lgb = NULL; 8377 8378 /* 8379 * Fill in IP length and desired time to live and send to IP level. 8380 * There should be a better way to handle ttl and tos; we could keep 8381 * them in the template, but need a way to checksum without them. 8382 */ 8383 /* 8384 * m->m_pkthdr.len should have been set before cksum calcuration, 8385 * because in6_cksum() need it. 8386 */ 8387 #ifdef INET6 8388 if (isipv6) { 8389 /* 8390 * we separately set hoplimit for every segment, since the 8391 * user might want to change the value via setsockopt. Also, 8392 * desired default hop limit might be changed via Neighbor 8393 * Discovery. 8394 */ 8395 ip6->ip6_hlim = in6_selecthlim(inp, NULL); 8396 8397 /* 8398 * Set the packet size here for the benefit of DTrace 8399 * probes. ip6_output() will set it properly; it's supposed 8400 * to include the option header lengths as well. 8401 */ 8402 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6)); 8403 8404 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) 8405 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 8406 else 8407 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 8408 8409 if (tp->t_state == TCPS_SYN_SENT) 8410 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th); 8411 8412 TCP_PROBE5(send, NULL, tp, ip6, tp, th); 8413 /* TODO: IPv6 IP6TOS_ECT bit on */ 8414 error = ip6_output(m, tp->t_inpcb->in6p_outputopts, 8415 &inp->inp_route6, 8416 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 8417 NULL, NULL, inp); 8418 8419 if (error == EMSGSIZE && inp->inp_route6.ro_rt != NULL) 8420 mtu = inp->inp_route6.ro_rt->rt_mtu; 8421 } 8422 #endif /* INET6 */ 8423 #if defined(INET) && defined(INET6) 8424 else 8425 #endif 8426 #ifdef INET 8427 { 8428 ip->ip_len = htons(m->m_pkthdr.len); 8429 #ifdef INET6 8430 if (inp->inp_vflag & INP_IPV6PROTO) 8431 ip->ip_ttl = in6_selecthlim(inp, NULL); 8432 #endif /* INET6 */ 8433 /* 8434 * If we do path MTU discovery, then we set DF on every 8435 * packet. This might not be the best thing to do according 8436 * to RFC3390 Section 2. However the tcp hostcache migitates 8437 * the problem so it affects only the first tcp connection 8438 * with a host. 8439 * 8440 * NB: Don't set DF on small MTU/MSS to have a safe 8441 * fallback. 8442 */ 8443 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) { 8444 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 8445 if (tp->t_port == 0 || len < V_tcp_minmss) { 8446 ip->ip_off |= htons(IP_DF); 8447 } 8448 } else { 8449 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 8450 } 8451 8452 if (tp->t_state == TCPS_SYN_SENT) 8453 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th); 8454 8455 TCP_PROBE5(send, NULL, tp, ip, tp, th); 8456 8457 error = ip_output(m, tp->t_inpcb->inp_options, &inp->inp_route, 8458 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, 8459 inp); 8460 if (error == EMSGSIZE && inp->inp_route.ro_rt != NULL) 8461 mtu = inp->inp_route.ro_rt->rt_mtu; 8462 } 8463 #endif /* INET */ 8464 8465 out: 8466 if (lgb) { 8467 lgb->tlb_errno = error; 8468 lgb = NULL; 8469 } 8470 /* 8471 * In transmit state, time the transmission and arrange for the 8472 * retransmit. In persist state, just set snd_max. 8473 */ 8474 if (error == 0) { 8475 if (len == 0) 8476 counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1); 8477 else if (len == 1) { 8478 counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1); 8479 } else if (len > 1) { 8480 int idx; 8481 8482 idx = (len / tp->t_maxseg) + 3; 8483 if (idx >= TCP_MSS_ACCT_ATIMER) 8484 counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1); 8485 else 8486 counter_u64_add(rack_out_size[idx], 1); 8487 } 8488 } 8489 if (sub_from_prr && (error == 0)) { 8490 rack->r_ctl.rc_prr_sndcnt -= len; 8491 } 8492 sub_from_prr = 0; 8493 rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error, cts, 8494 pass, rsm); 8495 if ((tp->t_flags & TF_FORCEDATA) == 0 || 8496 (rack->rc_in_persist == 0)) { 8497 tcp_seq startseq = tp->snd_nxt; 8498 8499 /* 8500 * Advance snd_nxt over sequence space of this segment. 8501 */ 8502 if (error) 8503 /* We don't log or do anything with errors */ 8504 goto timer; 8505 8506 if (flags & (TH_SYN | TH_FIN)) { 8507 if (flags & TH_SYN) 8508 tp->snd_nxt++; 8509 if (flags & TH_FIN) { 8510 tp->snd_nxt++; 8511 tp->t_flags |= TF_SENTFIN; 8512 } 8513 } 8514 /* In the ENOBUFS case we do *not* update snd_max */ 8515 if (sack_rxmit) 8516 goto timer; 8517 8518 tp->snd_nxt += len; 8519 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 8520 if (tp->snd_una == tp->snd_max) { 8521 /* 8522 * Update the time we just added data since 8523 * none was outstanding. 8524 */ 8525 rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__); 8526 tp->t_acktime = ticks; 8527 } 8528 tp->snd_max = tp->snd_nxt; 8529 /* 8530 * Time this transmission if not a retransmission and 8531 * not currently timing anything. 8532 * This is only relevant in case of switching back to 8533 * the base stack. 8534 */ 8535 if (tp->t_rtttime == 0) { 8536 tp->t_rtttime = ticks; 8537 tp->t_rtseq = startseq; 8538 TCPSTAT_INC(tcps_segstimed); 8539 } 8540 #ifdef NETFLIX_STATS 8541 if (!(tp->t_flags & TF_GPUTINPROG) && len) { 8542 tp->t_flags |= TF_GPUTINPROG; 8543 tp->gput_seq = startseq; 8544 tp->gput_ack = startseq + 8545 ulmin(sbavail(sb) - sb_offset, sendwin); 8546 tp->gput_ts = tcp_ts_getticks(); 8547 } 8548 #endif 8549 } 8550 /* 8551 * Set retransmit timer if not currently set, and not doing 8552 * a pure ack or a keep-alive probe. Initial value for 8553 * retransmit timer is smoothed round-trip time + 2 * 8554 * round-trip time variance. Initialize shift counter which 8555 * is used for backoff of retransmit time. 8556 */ 8557 timer: 8558 if ((tp->snd_wnd == 0) && 8559 TCPS_HAVEESTABLISHED(tp->t_state)) { 8560 /* 8561 * If the persists timer was set above (right before 8562 * the goto send), and still needs to be on. Lets 8563 * make sure all is canceled. If the persist timer 8564 * is not running, we want to get it up. 8565 */ 8566 if (rack->rc_in_persist == 0) { 8567 rack_enter_persist(tp, rack, cts); 8568 } 8569 } 8570 } else { 8571 /* 8572 * Persist case, update snd_max but since we are in persist 8573 * mode (no window) we do not update snd_nxt. 8574 */ 8575 int32_t xlen = len; 8576 8577 if (error) 8578 goto nomore; 8579 8580 if (flags & TH_SYN) 8581 ++xlen; 8582 if (flags & TH_FIN) { 8583 ++xlen; 8584 tp->t_flags |= TF_SENTFIN; 8585 } 8586 /* In the ENOBUFS case we do *not* update snd_max */ 8587 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) { 8588 if (tp->snd_una == tp->snd_max) { 8589 /* 8590 * Update the time we just added data since 8591 * none was outstanding. 8592 */ 8593 rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__); 8594 tp->t_acktime = ticks; 8595 } 8596 tp->snd_max = tp->snd_nxt + len; 8597 } 8598 } 8599 nomore: 8600 if (error) { 8601 SOCKBUF_UNLOCK_ASSERT(sb); /* Check gotos. */ 8602 /* 8603 * Failures do not advance the seq counter above. For the 8604 * case of ENOBUFS we will fall out and retry in 1ms with 8605 * the hpts. Everything else will just have to retransmit 8606 * with the timer. 8607 * 8608 * In any case, we do not want to loop around for another 8609 * send without a good reason. 8610 */ 8611 sendalot = 0; 8612 switch (error) { 8613 case EPERM: 8614 tp->t_flags &= ~TF_FORCEDATA; 8615 tp->t_softerror = error; 8616 return (error); 8617 case ENOBUFS: 8618 if (slot == 0) { 8619 /* 8620 * Pace us right away to retry in a some 8621 * time 8622 */ 8623 slot = 1 + rack->rc_enobuf; 8624 if (rack->rc_enobuf < 255) 8625 rack->rc_enobuf++; 8626 if (slot > (rack->rc_rack_rtt / 2)) { 8627 slot = rack->rc_rack_rtt / 2; 8628 } 8629 if (slot < 10) 8630 slot = 10; 8631 } 8632 counter_u64_add(rack_saw_enobuf, 1); 8633 error = 0; 8634 goto enobufs; 8635 case EMSGSIZE: 8636 /* 8637 * For some reason the interface we used initially 8638 * to send segments changed to another or lowered 8639 * its MTU. If TSO was active we either got an 8640 * interface without TSO capabilits or TSO was 8641 * turned off. If we obtained mtu from ip_output() 8642 * then update it and try again. 8643 */ 8644 if (tso) 8645 tp->t_flags &= ~TF_TSO; 8646 if (mtu != 0) { 8647 tcp_mss_update(tp, -1, mtu, NULL, NULL); 8648 goto again; 8649 } 8650 slot = 10; 8651 rack_start_hpts_timer(rack, tp, cts, __LINE__, slot, 0, 1); 8652 tp->t_flags &= ~TF_FORCEDATA; 8653 return (error); 8654 case ENETUNREACH: 8655 counter_u64_add(rack_saw_enetunreach, 1); 8656 case EHOSTDOWN: 8657 case EHOSTUNREACH: 8658 case ENETDOWN: 8659 if (TCPS_HAVERCVDSYN(tp->t_state)) { 8660 tp->t_softerror = error; 8661 } 8662 /* FALLTHROUGH */ 8663 default: 8664 slot = 10; 8665 rack_start_hpts_timer(rack, tp, cts, __LINE__, slot, 0, 1); 8666 tp->t_flags &= ~TF_FORCEDATA; 8667 return (error); 8668 } 8669 } else { 8670 rack->rc_enobuf = 0; 8671 } 8672 TCPSTAT_INC(tcps_sndtotal); 8673 8674 /* 8675 * Data sent (as far as we can tell). If this advertises a larger 8676 * window than any other segment, then remember the size of the 8677 * advertised window. Any pending ACK has now been sent. 8678 */ 8679 if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) 8680 tp->rcv_adv = tp->rcv_nxt + recwin; 8681 tp->last_ack_sent = tp->rcv_nxt; 8682 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); 8683 enobufs: 8684 rack->r_tlp_running = 0; 8685 if ((flags & TH_RST) || (would_have_fin == 1)) { 8686 /* 8687 * We don't send again after a RST. We also do *not* send 8688 * again if we would have had a find, but now have 8689 * outstanding data. 8690 */ 8691 slot = 0; 8692 sendalot = 0; 8693 } 8694 if (slot) { 8695 /* set the rack tcb into the slot N */ 8696 counter_u64_add(rack_paced_segments, 1); 8697 } else if (sendalot) { 8698 if (len) 8699 counter_u64_add(rack_unpaced_segments, 1); 8700 sack_rxmit = 0; 8701 tp->t_flags &= ~TF_FORCEDATA; 8702 goto again; 8703 } else if (len) { 8704 counter_u64_add(rack_unpaced_segments, 1); 8705 } 8706 tp->t_flags &= ~TF_FORCEDATA; 8707 rack_start_hpts_timer(rack, tp, cts, __LINE__, slot, tot_len_this_send, 1); 8708 return (error); 8709 } 8710 8711 /* 8712 * rack_ctloutput() must drop the inpcb lock before performing copyin on 8713 * socket option arguments. When it re-acquires the lock after the copy, it 8714 * has to revalidate that the connection is still valid for the socket 8715 * option. 8716 */ 8717 static int 8718 rack_set_sockopt(struct socket *so, struct sockopt *sopt, 8719 struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack) 8720 { 8721 int32_t error = 0, optval; 8722 8723 switch (sopt->sopt_name) { 8724 case TCP_RACK_PROP_RATE: 8725 case TCP_RACK_PROP: 8726 case TCP_RACK_TLP_REDUCE: 8727 case TCP_RACK_EARLY_RECOV: 8728 case TCP_RACK_PACE_ALWAYS: 8729 case TCP_DELACK: 8730 case TCP_RACK_PACE_REDUCE: 8731 case TCP_RACK_PACE_MAX_SEG: 8732 case TCP_RACK_PRR_SENDALOT: 8733 case TCP_RACK_MIN_TO: 8734 case TCP_RACK_EARLY_SEG: 8735 case TCP_RACK_REORD_THRESH: 8736 case TCP_RACK_REORD_FADE: 8737 case TCP_RACK_TLP_THRESH: 8738 case TCP_RACK_PKT_DELAY: 8739 case TCP_RACK_TLP_USE: 8740 case TCP_RACK_TLP_INC_VAR: 8741 case TCP_RACK_IDLE_REDUCE_HIGH: 8742 case TCP_RACK_MIN_PACE: 8743 case TCP_RACK_MIN_PACE_SEG: 8744 case TCP_BBR_RACK_RTT_USE: 8745 case TCP_DATA_AFTER_CLOSE: 8746 break; 8747 default: 8748 return (tcp_default_ctloutput(so, sopt, inp, tp)); 8749 break; 8750 } 8751 INP_WUNLOCK(inp); 8752 error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval)); 8753 if (error) 8754 return (error); 8755 INP_WLOCK(inp); 8756 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { 8757 INP_WUNLOCK(inp); 8758 return (ECONNRESET); 8759 } 8760 tp = intotcpcb(inp); 8761 rack = (struct tcp_rack *)tp->t_fb_ptr; 8762 switch (sopt->sopt_name) { 8763 case TCP_RACK_PROP_RATE: 8764 if ((optval <= 0) || (optval >= 100)) { 8765 error = EINVAL; 8766 break; 8767 } 8768 RACK_OPTS_INC(tcp_rack_prop_rate); 8769 rack->r_ctl.rc_prop_rate = optval; 8770 break; 8771 case TCP_RACK_TLP_USE: 8772 if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) { 8773 error = EINVAL; 8774 break; 8775 } 8776 RACK_OPTS_INC(tcp_tlp_use); 8777 rack->rack_tlp_threshold_use = optval; 8778 break; 8779 case TCP_RACK_PROP: 8780 /* RACK proportional rate reduction (bool) */ 8781 RACK_OPTS_INC(tcp_rack_prop); 8782 rack->r_ctl.rc_prop_reduce = optval; 8783 break; 8784 case TCP_RACK_TLP_REDUCE: 8785 /* RACK TLP cwnd reduction (bool) */ 8786 RACK_OPTS_INC(tcp_rack_tlp_reduce); 8787 rack->r_ctl.rc_tlp_cwnd_reduce = optval; 8788 break; 8789 case TCP_RACK_EARLY_RECOV: 8790 /* Should recovery happen early (bool) */ 8791 RACK_OPTS_INC(tcp_rack_early_recov); 8792 rack->r_ctl.rc_early_recovery = optval; 8793 break; 8794 case TCP_RACK_PACE_ALWAYS: 8795 /* Use the always pace method (bool) */ 8796 RACK_OPTS_INC(tcp_rack_pace_always); 8797 if (optval > 0) 8798 rack->rc_always_pace = 1; 8799 else 8800 rack->rc_always_pace = 0; 8801 break; 8802 case TCP_RACK_PACE_REDUCE: 8803 /* RACK Hptsi reduction factor (divisor) */ 8804 RACK_OPTS_INC(tcp_rack_pace_reduce); 8805 if (optval) 8806 /* Must be non-zero */ 8807 rack->rc_pace_reduce = optval; 8808 else 8809 error = EINVAL; 8810 break; 8811 case TCP_RACK_PACE_MAX_SEG: 8812 /* Max segments in a pace */ 8813 RACK_OPTS_INC(tcp_rack_max_seg); 8814 rack->rc_pace_max_segs = optval; 8815 break; 8816 case TCP_RACK_PRR_SENDALOT: 8817 /* Allow PRR to send more than one seg */ 8818 RACK_OPTS_INC(tcp_rack_prr_sendalot); 8819 rack->r_ctl.rc_prr_sendalot = optval; 8820 break; 8821 case TCP_RACK_MIN_TO: 8822 /* Minimum time between rack t-o's in ms */ 8823 RACK_OPTS_INC(tcp_rack_min_to); 8824 rack->r_ctl.rc_min_to = optval; 8825 break; 8826 case TCP_RACK_EARLY_SEG: 8827 /* If early recovery max segments */ 8828 RACK_OPTS_INC(tcp_rack_early_seg); 8829 rack->r_ctl.rc_early_recovery_segs = optval; 8830 break; 8831 case TCP_RACK_REORD_THRESH: 8832 /* RACK reorder threshold (shift amount) */ 8833 RACK_OPTS_INC(tcp_rack_reord_thresh); 8834 if ((optval > 0) && (optval < 31)) 8835 rack->r_ctl.rc_reorder_shift = optval; 8836 else 8837 error = EINVAL; 8838 break; 8839 case TCP_RACK_REORD_FADE: 8840 /* Does reordering fade after ms time */ 8841 RACK_OPTS_INC(tcp_rack_reord_fade); 8842 rack->r_ctl.rc_reorder_fade = optval; 8843 break; 8844 case TCP_RACK_TLP_THRESH: 8845 /* RACK TLP theshold i.e. srtt+(srtt/N) */ 8846 RACK_OPTS_INC(tcp_rack_tlp_thresh); 8847 if (optval) 8848 rack->r_ctl.rc_tlp_threshold = optval; 8849 else 8850 error = EINVAL; 8851 break; 8852 case TCP_RACK_PKT_DELAY: 8853 /* RACK added ms i.e. rack-rtt + reord + N */ 8854 RACK_OPTS_INC(tcp_rack_pkt_delay); 8855 rack->r_ctl.rc_pkt_delay = optval; 8856 break; 8857 case TCP_RACK_TLP_INC_VAR: 8858 /* Does TLP include rtt variance in t-o */ 8859 RACK_OPTS_INC(tcp_rack_tlp_inc_var); 8860 rack->r_ctl.rc_prr_inc_var = optval; 8861 break; 8862 case TCP_RACK_IDLE_REDUCE_HIGH: 8863 RACK_OPTS_INC(tcp_rack_idle_reduce_high); 8864 if (optval) 8865 rack->r_idle_reduce_largest = 1; 8866 else 8867 rack->r_idle_reduce_largest = 0; 8868 break; 8869 case TCP_DELACK: 8870 if (optval == 0) 8871 tp->t_delayed_ack = 0; 8872 else 8873 tp->t_delayed_ack = 1; 8874 if (tp->t_flags & TF_DELACK) { 8875 tp->t_flags &= ~TF_DELACK; 8876 tp->t_flags |= TF_ACKNOW; 8877 rack_output(tp); 8878 } 8879 break; 8880 case TCP_RACK_MIN_PACE: 8881 RACK_OPTS_INC(tcp_rack_min_pace); 8882 if (optval > 3) 8883 rack->r_enforce_min_pace = 3; 8884 else 8885 rack->r_enforce_min_pace = optval; 8886 break; 8887 case TCP_RACK_MIN_PACE_SEG: 8888 RACK_OPTS_INC(tcp_rack_min_pace_seg); 8889 if (optval >= 16) 8890 rack->r_min_pace_seg_thresh = 15; 8891 else 8892 rack->r_min_pace_seg_thresh = optval; 8893 break; 8894 case TCP_BBR_RACK_RTT_USE: 8895 if ((optval != USE_RTT_HIGH) && 8896 (optval != USE_RTT_LOW) && 8897 (optval != USE_RTT_AVG)) 8898 error = EINVAL; 8899 else 8900 rack->r_ctl.rc_rate_sample_method = optval; 8901 break; 8902 case TCP_DATA_AFTER_CLOSE: 8903 if (optval) 8904 rack->rc_allow_data_af_clo = 1; 8905 else 8906 rack->rc_allow_data_af_clo = 0; 8907 break; 8908 default: 8909 return (tcp_default_ctloutput(so, sopt, inp, tp)); 8910 break; 8911 } 8912 #ifdef NETFLIX_STATS 8913 tcp_log_socket_option(tp, sopt->sopt_name, optval, error); 8914 #endif 8915 INP_WUNLOCK(inp); 8916 return (error); 8917 } 8918 8919 static int 8920 rack_get_sockopt(struct socket *so, struct sockopt *sopt, 8921 struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack) 8922 { 8923 int32_t error, optval; 8924 8925 /* 8926 * Because all our options are either boolean or an int, we can just 8927 * pull everything into optval and then unlock and copy. If we ever 8928 * add a option that is not a int, then this will have quite an 8929 * impact to this routine. 8930 */ 8931 switch (sopt->sopt_name) { 8932 case TCP_RACK_PROP_RATE: 8933 optval = rack->r_ctl.rc_prop_rate; 8934 break; 8935 case TCP_RACK_PROP: 8936 /* RACK proportional rate reduction (bool) */ 8937 optval = rack->r_ctl.rc_prop_reduce; 8938 break; 8939 case TCP_RACK_TLP_REDUCE: 8940 /* RACK TLP cwnd reduction (bool) */ 8941 optval = rack->r_ctl.rc_tlp_cwnd_reduce; 8942 break; 8943 case TCP_RACK_EARLY_RECOV: 8944 /* Should recovery happen early (bool) */ 8945 optval = rack->r_ctl.rc_early_recovery; 8946 break; 8947 case TCP_RACK_PACE_REDUCE: 8948 /* RACK Hptsi reduction factor (divisor) */ 8949 optval = rack->rc_pace_reduce; 8950 break; 8951 case TCP_RACK_PACE_MAX_SEG: 8952 /* Max segments in a pace */ 8953 optval = rack->rc_pace_max_segs; 8954 break; 8955 case TCP_RACK_PACE_ALWAYS: 8956 /* Use the always pace method */ 8957 optval = rack->rc_always_pace; 8958 break; 8959 case TCP_RACK_PRR_SENDALOT: 8960 /* Allow PRR to send more than one seg */ 8961 optval = rack->r_ctl.rc_prr_sendalot; 8962 break; 8963 case TCP_RACK_MIN_TO: 8964 /* Minimum time between rack t-o's in ms */ 8965 optval = rack->r_ctl.rc_min_to; 8966 break; 8967 case TCP_RACK_EARLY_SEG: 8968 /* If early recovery max segments */ 8969 optval = rack->r_ctl.rc_early_recovery_segs; 8970 break; 8971 case TCP_RACK_REORD_THRESH: 8972 /* RACK reorder threshold (shift amount) */ 8973 optval = rack->r_ctl.rc_reorder_shift; 8974 break; 8975 case TCP_RACK_REORD_FADE: 8976 /* Does reordering fade after ms time */ 8977 optval = rack->r_ctl.rc_reorder_fade; 8978 break; 8979 case TCP_RACK_TLP_THRESH: 8980 /* RACK TLP theshold i.e. srtt+(srtt/N) */ 8981 optval = rack->r_ctl.rc_tlp_threshold; 8982 break; 8983 case TCP_RACK_PKT_DELAY: 8984 /* RACK added ms i.e. rack-rtt + reord + N */ 8985 optval = rack->r_ctl.rc_pkt_delay; 8986 break; 8987 case TCP_RACK_TLP_USE: 8988 optval = rack->rack_tlp_threshold_use; 8989 break; 8990 case TCP_RACK_TLP_INC_VAR: 8991 /* Does TLP include rtt variance in t-o */ 8992 optval = rack->r_ctl.rc_prr_inc_var; 8993 break; 8994 case TCP_RACK_IDLE_REDUCE_HIGH: 8995 optval = rack->r_idle_reduce_largest; 8996 break; 8997 case TCP_RACK_MIN_PACE: 8998 optval = rack->r_enforce_min_pace; 8999 break; 9000 case TCP_RACK_MIN_PACE_SEG: 9001 optval = rack->r_min_pace_seg_thresh; 9002 break; 9003 case TCP_BBR_RACK_RTT_USE: 9004 optval = rack->r_ctl.rc_rate_sample_method; 9005 break; 9006 case TCP_DELACK: 9007 optval = tp->t_delayed_ack; 9008 break; 9009 case TCP_DATA_AFTER_CLOSE: 9010 optval = rack->rc_allow_data_af_clo; 9011 break; 9012 default: 9013 return (tcp_default_ctloutput(so, sopt, inp, tp)); 9014 break; 9015 } 9016 INP_WUNLOCK(inp); 9017 error = sooptcopyout(sopt, &optval, sizeof optval); 9018 return (error); 9019 } 9020 9021 static int 9022 rack_ctloutput(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp) 9023 { 9024 int32_t error = EINVAL; 9025 struct tcp_rack *rack; 9026 9027 rack = (struct tcp_rack *)tp->t_fb_ptr; 9028 if (rack == NULL) { 9029 /* Huh? */ 9030 goto out; 9031 } 9032 if (sopt->sopt_dir == SOPT_SET) { 9033 return (rack_set_sockopt(so, sopt, inp, tp, rack)); 9034 } else if (sopt->sopt_dir == SOPT_GET) { 9035 return (rack_get_sockopt(so, sopt, inp, tp, rack)); 9036 } 9037 out: 9038 INP_WUNLOCK(inp); 9039 return (error); 9040 } 9041 9042 9043 struct tcp_function_block __tcp_rack = { 9044 .tfb_tcp_block_name = __XSTRING(STACKNAME), 9045 .tfb_tcp_output = rack_output, 9046 .tfb_tcp_do_segment = rack_do_segment, 9047 .tfb_tcp_hpts_do_segment = rack_hpts_do_segment, 9048 .tfb_tcp_ctloutput = rack_ctloutput, 9049 .tfb_tcp_fb_init = rack_init, 9050 .tfb_tcp_fb_fini = rack_fini, 9051 .tfb_tcp_timer_stop_all = rack_stopall, 9052 .tfb_tcp_timer_activate = rack_timer_activate, 9053 .tfb_tcp_timer_active = rack_timer_active, 9054 .tfb_tcp_timer_stop = rack_timer_stop, 9055 .tfb_tcp_rexmit_tmr = rack_remxt_tmr, 9056 .tfb_tcp_handoff_ok = rack_handoff_ok 9057 }; 9058 9059 static const char *rack_stack_names[] = { 9060 __XSTRING(STACKNAME), 9061 #ifdef STACKALIAS 9062 __XSTRING(STACKALIAS), 9063 #endif 9064 }; 9065 9066 static int 9067 rack_ctor(void *mem, int32_t size, void *arg, int32_t how) 9068 { 9069 memset(mem, 0, size); 9070 return (0); 9071 } 9072 9073 static void 9074 rack_dtor(void *mem, int32_t size, void *arg) 9075 { 9076 9077 } 9078 9079 static bool rack_mod_inited = false; 9080 9081 static int 9082 tcp_addrack(module_t mod, int32_t type, void *data) 9083 { 9084 int32_t err = 0; 9085 int num_stacks; 9086 9087 switch (type) { 9088 case MOD_LOAD: 9089 rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map", 9090 sizeof(struct rack_sendmap), 9091 rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0); 9092 9093 rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb", 9094 sizeof(struct tcp_rack), 9095 rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0); 9096 9097 sysctl_ctx_init(&rack_sysctl_ctx); 9098 rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx, 9099 SYSCTL_STATIC_CHILDREN(_net_inet_tcp), 9100 OID_AUTO, 9101 __XSTRING(STACKNAME), 9102 CTLFLAG_RW, 0, 9103 ""); 9104 if (rack_sysctl_root == NULL) { 9105 printf("Failed to add sysctl node\n"); 9106 err = EFAULT; 9107 goto free_uma; 9108 } 9109 rack_init_sysctls(); 9110 num_stacks = nitems(rack_stack_names); 9111 err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK, 9112 rack_stack_names, &num_stacks); 9113 if (err) { 9114 printf("Failed to register %s stack name for " 9115 "%s module\n", rack_stack_names[num_stacks], 9116 __XSTRING(MODNAME)); 9117 sysctl_ctx_free(&rack_sysctl_ctx); 9118 free_uma: 9119 uma_zdestroy(rack_zone); 9120 uma_zdestroy(rack_pcb_zone); 9121 rack_counter_destroy(); 9122 printf("Failed to register rack module -- err:%d\n", err); 9123 return (err); 9124 } 9125 rack_mod_inited = true; 9126 break; 9127 case MOD_QUIESCE: 9128 err = deregister_tcp_functions(&__tcp_rack, true, false); 9129 break; 9130 case MOD_UNLOAD: 9131 err = deregister_tcp_functions(&__tcp_rack, false, true); 9132 if (err == EBUSY) 9133 break; 9134 if (rack_mod_inited) { 9135 uma_zdestroy(rack_zone); 9136 uma_zdestroy(rack_pcb_zone); 9137 sysctl_ctx_free(&rack_sysctl_ctx); 9138 rack_counter_destroy(); 9139 rack_mod_inited = false; 9140 } 9141 err = 0; 9142 break; 9143 default: 9144 return (EOPNOTSUPP); 9145 } 9146 return (err); 9147 } 9148 9149 static moduledata_t tcp_rack = { 9150 .name = __XSTRING(MODNAME), 9151 .evhand = tcp_addrack, 9152 .priv = 0 9153 }; 9154 9155 MODULE_VERSION(MODNAME, 1); 9156 DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY); 9157 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1); 9158