1 /*- 2 * Copyright (c) 2016-2019 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 #ifdef NETFLIX_STATS 49 #include <sys/qmath.h> 50 #endif 51 #include <sys/socket.h> 52 #include <sys/socketvar.h> 53 #include <sys/sysctl.h> 54 #include <sys/systm.h> 55 #include <sys/tree.h> 56 #ifdef NETFLIX_STATS 57 #include <sys/stats.h> /* Must come after qmath.h and tree.h */ 58 #endif 59 #include <sys/refcount.h> 60 #include <sys/queue.h> 61 #include <sys/smp.h> 62 #include <sys/kthread.h> 63 #include <sys/kern_prefetch.h> 64 65 #include <vm/uma.h> 66 67 #include <net/route.h> 68 #include <net/vnet.h> 69 70 #define TCPSTATES /* for logging */ 71 72 #include <netinet/in.h> 73 #include <netinet/in_kdtrace.h> 74 #include <netinet/in_pcb.h> 75 #include <netinet/ip.h> 76 #include <netinet/ip_icmp.h> /* required for icmp_var.h */ 77 #include <netinet/icmp_var.h> /* for ICMP_BANDLIM */ 78 #include <netinet/ip_var.h> 79 #include <netinet/ip6.h> 80 #include <netinet6/in6_pcb.h> 81 #include <netinet6/ip6_var.h> 82 #define TCPOUTFLAGS 83 #include <netinet/tcp.h> 84 #include <netinet/tcp_fsm.h> 85 #include <netinet/tcp_log_buf.h> 86 #include <netinet/tcp_seq.h> 87 #include <netinet/tcp_timer.h> 88 #include <netinet/tcp_var.h> 89 #include <netinet/tcp_hpts.h> 90 #include <netinet/tcpip.h> 91 #include <netinet/cc/cc.h> 92 #include <netinet/tcp_fastopen.h> 93 #ifdef TCPDEBUG 94 #include <netinet/tcp_debug.h> 95 #endif /* TCPDEBUG */ 96 #ifdef TCP_OFFLOAD 97 #include <netinet/tcp_offload.h> 98 #endif 99 #ifdef INET6 100 #include <netinet6/tcp6_var.h> 101 #endif 102 103 #include <netipsec/ipsec_support.h> 104 105 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 106 #include <netipsec/ipsec.h> 107 #include <netipsec/ipsec6.h> 108 #endif /* IPSEC */ 109 110 #include <netinet/udp.h> 111 #include <netinet/udp_var.h> 112 #include <machine/in_cksum.h> 113 114 #ifdef MAC 115 #include <security/mac/mac_framework.h> 116 #endif 117 #include "sack_filter.h" 118 #include "tcp_rack.h" 119 #include "rack_bbr_common.h" 120 121 uma_zone_t rack_zone; 122 uma_zone_t rack_pcb_zone; 123 124 #ifndef TICKS2SBT 125 #define TICKS2SBT(__t) (tick_sbt * ((sbintime_t)(__t))) 126 #endif 127 128 struct sysctl_ctx_list rack_sysctl_ctx; 129 struct sysctl_oid *rack_sysctl_root; 130 131 #define CUM_ACKED 1 132 #define SACKED 2 133 134 /* 135 * The RACK module incorporates a number of 136 * TCP ideas that have been put out into the IETF 137 * over the last few years: 138 * - Matt Mathis's Rate Halving which slowly drops 139 * the congestion window so that the ack clock can 140 * be maintained during a recovery. 141 * - Yuchung Cheng's RACK TCP (for which its named) that 142 * will stop us using the number of dup acks and instead 143 * use time as the gage of when we retransmit. 144 * - Reorder Detection of RFC4737 and the Tail-Loss probe draft 145 * of Dukkipati et.al. 146 * RACK depends on SACK, so if an endpoint arrives that 147 * cannot do SACK the state machine below will shuttle the 148 * connection back to using the "default" TCP stack that is 149 * in FreeBSD. 150 * 151 * To implement RACK the original TCP stack was first decomposed 152 * into a functional state machine with individual states 153 * for each of the possible TCP connection states. The do_segement 154 * functions role in life is to mandate the connection supports SACK 155 * initially and then assure that the RACK state matches the conenction 156 * state before calling the states do_segment function. Each 157 * state is simplified due to the fact that the original do_segment 158 * has been decomposed and we *know* what state we are in (no 159 * switches on the state) and all tests for SACK are gone. This 160 * greatly simplifies what each state does. 161 * 162 * TCP output is also over-written with a new version since it 163 * must maintain the new rack scoreboard. 164 * 165 */ 166 static int32_t rack_precache = 1; 167 static int32_t rack_tlp_thresh = 1; 168 static int32_t rack_reorder_thresh = 2; 169 static int32_t rack_reorder_fade = 60000; /* 0 - never fade, def 60,000 170 * - 60 seconds */ 171 static int32_t rack_pkt_delay = 1; 172 static int32_t rack_inc_var = 0;/* For TLP */ 173 static int32_t rack_reduce_largest_on_idle = 0; 174 static int32_t rack_min_pace_time = 0; 175 static int32_t rack_min_pace_time_seg_req=6; 176 static int32_t rack_early_recovery = 1; 177 static int32_t rack_early_recovery_max_seg = 6; 178 static int32_t rack_send_a_lot_in_prr = 1; 179 static int32_t rack_min_to = 1; /* Number of ms minimum timeout */ 180 static int32_t rack_tlp_in_recovery = 1; /* Can we do TLP in recovery? */ 181 static int32_t rack_verbose_logging = 0; 182 static int32_t rack_ignore_data_after_close = 1; 183 static int32_t rack_map_entries_limit = 1024; 184 static int32_t rack_map_split_limit = 256; 185 186 /* 187 * Currently regular tcp has a rto_min of 30ms 188 * the backoff goes 12 times so that ends up 189 * being a total of 122.850 seconds before a 190 * connection is killed. 191 */ 192 static int32_t rack_tlp_min = 10; 193 static int32_t rack_rto_min = 30; /* 30ms same as main freebsd */ 194 static int32_t rack_rto_max = 30000; /* 30 seconds */ 195 static const int32_t rack_free_cache = 2; 196 static int32_t rack_hptsi_segments = 40; 197 static int32_t rack_rate_sample_method = USE_RTT_LOW; 198 static int32_t rack_pace_every_seg = 1; 199 static int32_t rack_delayed_ack_time = 200; /* 200ms */ 200 static int32_t rack_slot_reduction = 4; 201 static int32_t rack_lower_cwnd_at_tlp = 0; 202 static int32_t rack_use_proportional_reduce = 0; 203 static int32_t rack_proportional_rate = 10; 204 static int32_t rack_tlp_max_resend = 2; 205 static int32_t rack_limited_retran = 0; 206 static int32_t rack_always_send_oldest = 0; 207 static int32_t rack_sack_block_limit = 128; 208 static int32_t rack_use_sack_filter = 1; 209 static int32_t rack_tlp_threshold_use = TLP_USE_TWO_ONE; 210 211 /* Rack specific counters */ 212 counter_u64_t rack_badfr; 213 counter_u64_t rack_badfr_bytes; 214 counter_u64_t rack_rtm_prr_retran; 215 counter_u64_t rack_rtm_prr_newdata; 216 counter_u64_t rack_timestamp_mismatch; 217 counter_u64_t rack_reorder_seen; 218 counter_u64_t rack_paced_segments; 219 counter_u64_t rack_unpaced_segments; 220 counter_u64_t rack_saw_enobuf; 221 counter_u64_t rack_saw_enetunreach; 222 223 /* Tail loss probe counters */ 224 counter_u64_t rack_tlp_tot; 225 counter_u64_t rack_tlp_newdata; 226 counter_u64_t rack_tlp_retran; 227 counter_u64_t rack_tlp_retran_bytes; 228 counter_u64_t rack_tlp_retran_fail; 229 counter_u64_t rack_to_tot; 230 counter_u64_t rack_to_arm_rack; 231 counter_u64_t rack_to_arm_tlp; 232 counter_u64_t rack_to_alloc; 233 counter_u64_t rack_to_alloc_hard; 234 counter_u64_t rack_to_alloc_emerg; 235 counter_u64_t rack_to_alloc_limited; 236 counter_u64_t rack_alloc_limited_conns; 237 counter_u64_t rack_split_limited; 238 239 counter_u64_t rack_sack_proc_all; 240 counter_u64_t rack_sack_proc_short; 241 counter_u64_t rack_sack_proc_restart; 242 counter_u64_t rack_runt_sacks; 243 counter_u64_t rack_used_tlpmethod; 244 counter_u64_t rack_used_tlpmethod2; 245 counter_u64_t rack_enter_tlp_calc; 246 counter_u64_t rack_input_idle_reduces; 247 counter_u64_t rack_tlp_does_nada; 248 249 /* Temp CPU counters */ 250 counter_u64_t rack_find_high; 251 252 counter_u64_t rack_progress_drops; 253 counter_u64_t rack_out_size[TCP_MSS_ACCT_SIZE]; 254 counter_u64_t rack_opts_arry[RACK_OPTS_SIZE]; 255 256 /* 257 * This was originally defined in tcp_timer.c, but is now reproduced here given 258 * the unification of the SYN and non-SYN retransmit timer exponents combined 259 * with wanting to retain previous behaviour for previously deployed stack 260 * versions. 261 */ 262 int tcp_syn_backoff[TCP_MAXRXTSHIFT + 1] = 263 { 1, 1, 1, 1, 1, 2, 4, 8, 16, 32, 64, 64, 64 }; 264 265 static void 266 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick, int event, int line); 267 268 static int 269 rack_process_ack(struct mbuf *m, struct tcphdr *th, 270 struct socket *so, struct tcpcb *tp, struct tcpopt *to, 271 uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t * ret_val); 272 static int 273 rack_process_data(struct mbuf *m, struct tcphdr *th, 274 struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, 275 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); 276 static void 277 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, 278 struct tcphdr *th, uint16_t nsegs, uint16_t type, int32_t recovery); 279 static struct rack_sendmap *rack_alloc(struct tcp_rack *rack); 280 static struct rack_sendmap *rack_alloc_limit(struct tcp_rack *rack, 281 uint8_t limit_type); 282 static struct rack_sendmap * 283 rack_check_recovery_mode(struct tcpcb *tp, 284 uint32_t tsused); 285 static void 286 rack_cong_signal(struct tcpcb *tp, struct tcphdr *th, 287 uint32_t type); 288 static void rack_counter_destroy(void); 289 static int 290 rack_ctloutput(struct socket *so, struct sockopt *sopt, 291 struct inpcb *inp, struct tcpcb *tp); 292 static int32_t rack_ctor(void *mem, int32_t size, void *arg, int32_t how); 293 static void 294 rack_do_segment(struct mbuf *m, struct tcphdr *th, 295 struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, 296 uint8_t iptos); 297 static void rack_dtor(void *mem, int32_t size, void *arg); 298 static void 299 rack_earlier_retran(struct tcpcb *tp, struct rack_sendmap *rsm, 300 uint32_t t, uint32_t cts); 301 static struct rack_sendmap * 302 rack_find_high_nonack(struct tcp_rack *rack, 303 struct rack_sendmap *rsm); 304 static struct rack_sendmap *rack_find_lowest_rsm(struct tcp_rack *rack); 305 static void rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm); 306 static void rack_fini(struct tcpcb *tp, int32_t tcb_is_purged); 307 static int 308 rack_get_sockopt(struct socket *so, struct sockopt *sopt, 309 struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack); 310 static int32_t rack_handoff_ok(struct tcpcb *tp); 311 static int32_t rack_init(struct tcpcb *tp); 312 static void rack_init_sysctls(void); 313 static void 314 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, 315 struct tcphdr *th); 316 static void 317 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len, 318 uint32_t seq_out, uint8_t th_flags, int32_t err, uint32_t ts, 319 uint8_t pass, struct rack_sendmap *hintrsm); 320 static void 321 rack_log_sack_passed(struct tcpcb *tp, struct tcp_rack *rack, 322 struct rack_sendmap *rsm); 323 static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num); 324 static int32_t rack_output(struct tcpcb *tp); 325 static void 326 rack_hpts_do_segment(struct mbuf *m, struct tcphdr *th, 327 struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, 328 uint8_t iptos, int32_t nxt_pkt, struct timeval *tv); 329 330 static uint32_t 331 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, 332 struct sackblk *sack, struct tcpopt *to, struct rack_sendmap **prsm, 333 uint32_t cts); 334 static void rack_post_recovery(struct tcpcb *tp, struct tcphdr *th); 335 static void rack_remxt_tmr(struct tcpcb *tp); 336 static int 337 rack_set_sockopt(struct socket *so, struct sockopt *sopt, 338 struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack); 339 static void rack_set_state(struct tcpcb *tp, struct tcp_rack *rack); 340 static int32_t rack_stopall(struct tcpcb *tp); 341 static void 342 rack_timer_activate(struct tcpcb *tp, uint32_t timer_type, 343 uint32_t delta); 344 static int32_t rack_timer_active(struct tcpcb *tp, uint32_t timer_type); 345 static void rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line); 346 static void rack_timer_stop(struct tcpcb *tp, uint32_t timer_type); 347 static uint32_t 348 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack, 349 struct rack_sendmap *rsm, uint32_t ts, int32_t * lenp); 350 static void 351 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack, 352 struct rack_sendmap *rsm, uint32_t ts); 353 static int 354 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack, 355 struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type); 356 static int32_t tcp_addrack(module_t mod, int32_t type, void *data); 357 static void 358 rack_challenge_ack(struct mbuf *m, struct tcphdr *th, 359 struct tcpcb *tp, int32_t * ret_val); 360 static int 361 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, 362 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, 363 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); 364 static int 365 rack_do_closing(struct mbuf *m, struct tcphdr *th, 366 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, 367 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); 368 static void rack_do_drop(struct mbuf *m, struct tcpcb *tp); 369 static void 370 rack_do_dropafterack(struct mbuf *m, struct tcpcb *tp, 371 struct tcphdr *th, int32_t thflags, int32_t tlen, int32_t * ret_val); 372 static void 373 rack_do_dropwithreset(struct mbuf *m, struct tcpcb *tp, 374 struct tcphdr *th, int32_t rstreason, int32_t tlen); 375 static int 376 rack_do_established(struct mbuf *m, struct tcphdr *th, 377 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, 378 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); 379 static int 380 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, 381 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, 382 int32_t tlen, uint32_t tiwin, int32_t nxt_pkt); 383 static int 384 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, 385 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, 386 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); 387 static int 388 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, 389 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, 390 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); 391 static int 392 rack_do_lastack(struct mbuf *m, struct tcphdr *th, 393 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, 394 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); 395 static int 396 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, 397 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, 398 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); 399 static int 400 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, 401 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, 402 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); 403 static int 404 rack_drop_checks(struct tcpopt *to, struct mbuf *m, 405 struct tcphdr *th, struct tcpcb *tp, int32_t * tlenp, int32_t * thf, 406 int32_t * drop_hdrlen, int32_t * ret_val); 407 static int 408 rack_process_rst(struct mbuf *m, struct tcphdr *th, 409 struct socket *so, struct tcpcb *tp); 410 struct rack_sendmap * 411 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, 412 uint32_t tsused); 413 static void tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt); 414 static void 415 tcp_rack_partialack(struct tcpcb *tp, struct tcphdr *th); 416 417 static int 418 rack_ts_check(struct mbuf *m, struct tcphdr *th, 419 struct tcpcb *tp, int32_t tlen, int32_t thflags, int32_t * ret_val); 420 421 int32_t rack_clear_counter=0; 422 423 424 static int 425 sysctl_rack_clear(SYSCTL_HANDLER_ARGS) 426 { 427 uint32_t stat; 428 int32_t error; 429 430 error = SYSCTL_OUT(req, &rack_clear_counter, sizeof(uint32_t)); 431 if (error || req->newptr == NULL) 432 return error; 433 434 error = SYSCTL_IN(req, &stat, sizeof(uint32_t)); 435 if (error) 436 return (error); 437 if (stat == 1) { 438 #ifdef INVARIANTS 439 printf("Clearing RACK counters\n"); 440 #endif 441 counter_u64_zero(rack_badfr); 442 counter_u64_zero(rack_badfr_bytes); 443 counter_u64_zero(rack_rtm_prr_retran); 444 counter_u64_zero(rack_rtm_prr_newdata); 445 counter_u64_zero(rack_timestamp_mismatch); 446 counter_u64_zero(rack_reorder_seen); 447 counter_u64_zero(rack_tlp_tot); 448 counter_u64_zero(rack_tlp_newdata); 449 counter_u64_zero(rack_tlp_retran); 450 counter_u64_zero(rack_tlp_retran_bytes); 451 counter_u64_zero(rack_tlp_retran_fail); 452 counter_u64_zero(rack_to_tot); 453 counter_u64_zero(rack_to_arm_rack); 454 counter_u64_zero(rack_to_arm_tlp); 455 counter_u64_zero(rack_paced_segments); 456 counter_u64_zero(rack_unpaced_segments); 457 counter_u64_zero(rack_saw_enobuf); 458 counter_u64_zero(rack_saw_enetunreach); 459 counter_u64_zero(rack_to_alloc_hard); 460 counter_u64_zero(rack_to_alloc_emerg); 461 counter_u64_zero(rack_sack_proc_all); 462 counter_u64_zero(rack_sack_proc_short); 463 counter_u64_zero(rack_sack_proc_restart); 464 counter_u64_zero(rack_to_alloc); 465 counter_u64_zero(rack_to_alloc_limited); 466 counter_u64_zero(rack_alloc_limited_conns); 467 counter_u64_zero(rack_split_limited); 468 counter_u64_zero(rack_find_high); 469 counter_u64_zero(rack_runt_sacks); 470 counter_u64_zero(rack_used_tlpmethod); 471 counter_u64_zero(rack_used_tlpmethod2); 472 counter_u64_zero(rack_enter_tlp_calc); 473 counter_u64_zero(rack_progress_drops); 474 counter_u64_zero(rack_tlp_does_nada); 475 } 476 rack_clear_counter = 0; 477 return (0); 478 } 479 480 481 482 static void 483 rack_init_sysctls() 484 { 485 SYSCTL_ADD_S32(&rack_sysctl_ctx, 486 SYSCTL_CHILDREN(rack_sysctl_root), 487 OID_AUTO, "map_limit", CTLFLAG_RW, 488 &rack_map_entries_limit , 1024, 489 "Is there a limit on how big the sendmap can grow? "); 490 491 SYSCTL_ADD_S32(&rack_sysctl_ctx, 492 SYSCTL_CHILDREN(rack_sysctl_root), 493 OID_AUTO, "map_splitlimit", CTLFLAG_RW, 494 &rack_map_split_limit , 256, 495 "Is there a limit on how much splitting a peer can do?"); 496 497 SYSCTL_ADD_S32(&rack_sysctl_ctx, 498 SYSCTL_CHILDREN(rack_sysctl_root), 499 OID_AUTO, "rate_sample_method", CTLFLAG_RW, 500 &rack_rate_sample_method , USE_RTT_LOW, 501 "What method should we use for rate sampling 0=high, 1=low "); 502 SYSCTL_ADD_S32(&rack_sysctl_ctx, 503 SYSCTL_CHILDREN(rack_sysctl_root), 504 OID_AUTO, "data_after_close", CTLFLAG_RW, 505 &rack_ignore_data_after_close, 0, 506 "Do we hold off sending a RST until all pending data is ack'd"); 507 SYSCTL_ADD_S32(&rack_sysctl_ctx, 508 SYSCTL_CHILDREN(rack_sysctl_root), 509 OID_AUTO, "tlpmethod", CTLFLAG_RW, 510 &rack_tlp_threshold_use, TLP_USE_TWO_ONE, 511 "What method do we do for TLP time calc 0=no-de-ack-comp, 1=ID, 2=2.1, 3=2.2"); 512 SYSCTL_ADD_S32(&rack_sysctl_ctx, 513 SYSCTL_CHILDREN(rack_sysctl_root), 514 OID_AUTO, "min_pace_time", CTLFLAG_RW, 515 &rack_min_pace_time, 0, 516 "Should we enforce a minimum pace time of 1ms"); 517 SYSCTL_ADD_S32(&rack_sysctl_ctx, 518 SYSCTL_CHILDREN(rack_sysctl_root), 519 OID_AUTO, "min_pace_segs", CTLFLAG_RW, 520 &rack_min_pace_time_seg_req, 6, 521 "How many segments have to be in the len to enforce min-pace-time"); 522 SYSCTL_ADD_S32(&rack_sysctl_ctx, 523 SYSCTL_CHILDREN(rack_sysctl_root), 524 OID_AUTO, "idle_reduce_high", CTLFLAG_RW, 525 &rack_reduce_largest_on_idle, 0, 526 "Should we reduce the largest cwnd seen to IW on idle reduction"); 527 SYSCTL_ADD_S32(&rack_sysctl_ctx, 528 SYSCTL_CHILDREN(rack_sysctl_root), 529 OID_AUTO, "bb_verbose", CTLFLAG_RW, 530 &rack_verbose_logging, 0, 531 "Should RACK black box logging be verbose"); 532 SYSCTL_ADD_S32(&rack_sysctl_ctx, 533 SYSCTL_CHILDREN(rack_sysctl_root), 534 OID_AUTO, "sackfiltering", CTLFLAG_RW, 535 &rack_use_sack_filter, 1, 536 "Do we use sack filtering?"); 537 SYSCTL_ADD_S32(&rack_sysctl_ctx, 538 SYSCTL_CHILDREN(rack_sysctl_root), 539 OID_AUTO, "delayed_ack", CTLFLAG_RW, 540 &rack_delayed_ack_time, 200, 541 "Delayed ack time (200ms)"); 542 SYSCTL_ADD_S32(&rack_sysctl_ctx, 543 SYSCTL_CHILDREN(rack_sysctl_root), 544 OID_AUTO, "tlpminto", CTLFLAG_RW, 545 &rack_tlp_min, 10, 546 "TLP minimum timeout per the specification (10ms)"); 547 SYSCTL_ADD_S32(&rack_sysctl_ctx, 548 SYSCTL_CHILDREN(rack_sysctl_root), 549 OID_AUTO, "precache", CTLFLAG_RW, 550 &rack_precache, 0, 551 "Where should we precache the mcopy (0 is not at all)"); 552 SYSCTL_ADD_S32(&rack_sysctl_ctx, 553 SYSCTL_CHILDREN(rack_sysctl_root), 554 OID_AUTO, "sblklimit", CTLFLAG_RW, 555 &rack_sack_block_limit, 128, 556 "When do we start paying attention to small sack blocks"); 557 SYSCTL_ADD_S32(&rack_sysctl_ctx, 558 SYSCTL_CHILDREN(rack_sysctl_root), 559 OID_AUTO, "send_oldest", CTLFLAG_RW, 560 &rack_always_send_oldest, 1, 561 "Should we always send the oldest TLP and RACK-TLP"); 562 SYSCTL_ADD_S32(&rack_sysctl_ctx, 563 SYSCTL_CHILDREN(rack_sysctl_root), 564 OID_AUTO, "rack_tlp_in_recovery", CTLFLAG_RW, 565 &rack_tlp_in_recovery, 1, 566 "Can we do a TLP during recovery?"); 567 SYSCTL_ADD_S32(&rack_sysctl_ctx, 568 SYSCTL_CHILDREN(rack_sysctl_root), 569 OID_AUTO, "rack_tlimit", CTLFLAG_RW, 570 &rack_limited_retran, 0, 571 "How many times can a rack timeout drive out sends"); 572 SYSCTL_ADD_S32(&rack_sysctl_ctx, 573 SYSCTL_CHILDREN(rack_sysctl_root), 574 OID_AUTO, "minrto", CTLFLAG_RW, 575 &rack_rto_min, 0, 576 "Minimum RTO in ms -- set with caution below 1000 due to TLP"); 577 SYSCTL_ADD_S32(&rack_sysctl_ctx, 578 SYSCTL_CHILDREN(rack_sysctl_root), 579 OID_AUTO, "maxrto", CTLFLAG_RW, 580 &rack_rto_max, 0, 581 "Maxiumum RTO in ms -- should be at least as large as min_rto"); 582 SYSCTL_ADD_S32(&rack_sysctl_ctx, 583 SYSCTL_CHILDREN(rack_sysctl_root), 584 OID_AUTO, "tlp_retry", CTLFLAG_RW, 585 &rack_tlp_max_resend, 2, 586 "How many times does TLP retry a single segment or multiple with no ACK"); 587 SYSCTL_ADD_S32(&rack_sysctl_ctx, 588 SYSCTL_CHILDREN(rack_sysctl_root), 589 OID_AUTO, "recovery_loss_prop", CTLFLAG_RW, 590 &rack_use_proportional_reduce, 0, 591 "Should we proportionaly reduce cwnd based on the number of losses "); 592 SYSCTL_ADD_S32(&rack_sysctl_ctx, 593 SYSCTL_CHILDREN(rack_sysctl_root), 594 OID_AUTO, "recovery_prop", CTLFLAG_RW, 595 &rack_proportional_rate, 10, 596 "What percent reduction per loss"); 597 SYSCTL_ADD_S32(&rack_sysctl_ctx, 598 SYSCTL_CHILDREN(rack_sysctl_root), 599 OID_AUTO, "tlp_cwnd_flag", CTLFLAG_RW, 600 &rack_lower_cwnd_at_tlp, 0, 601 "When a TLP completes a retran should we enter recovery?"); 602 SYSCTL_ADD_S32(&rack_sysctl_ctx, 603 SYSCTL_CHILDREN(rack_sysctl_root), 604 OID_AUTO, "hptsi_reduces", CTLFLAG_RW, 605 &rack_slot_reduction, 4, 606 "When setting a slot should we reduce by divisor"); 607 SYSCTL_ADD_S32(&rack_sysctl_ctx, 608 SYSCTL_CHILDREN(rack_sysctl_root), 609 OID_AUTO, "hptsi_every_seg", CTLFLAG_RW, 610 &rack_pace_every_seg, 1, 611 "Should we pace out every segment hptsi"); 612 SYSCTL_ADD_S32(&rack_sysctl_ctx, 613 SYSCTL_CHILDREN(rack_sysctl_root), 614 OID_AUTO, "hptsi_seg_max", CTLFLAG_RW, 615 &rack_hptsi_segments, 6, 616 "Should we pace out only a limited size of segments"); 617 SYSCTL_ADD_S32(&rack_sysctl_ctx, 618 SYSCTL_CHILDREN(rack_sysctl_root), 619 OID_AUTO, "prr_sendalot", CTLFLAG_RW, 620 &rack_send_a_lot_in_prr, 1, 621 "Send a lot in prr"); 622 SYSCTL_ADD_S32(&rack_sysctl_ctx, 623 SYSCTL_CHILDREN(rack_sysctl_root), 624 OID_AUTO, "minto", CTLFLAG_RW, 625 &rack_min_to, 1, 626 "Minimum rack timeout in milliseconds"); 627 SYSCTL_ADD_S32(&rack_sysctl_ctx, 628 SYSCTL_CHILDREN(rack_sysctl_root), 629 OID_AUTO, "earlyrecoveryseg", CTLFLAG_RW, 630 &rack_early_recovery_max_seg, 6, 631 "Max segments in early recovery"); 632 SYSCTL_ADD_S32(&rack_sysctl_ctx, 633 SYSCTL_CHILDREN(rack_sysctl_root), 634 OID_AUTO, "earlyrecovery", CTLFLAG_RW, 635 &rack_early_recovery, 1, 636 "Do we do early recovery with rack"); 637 SYSCTL_ADD_S32(&rack_sysctl_ctx, 638 SYSCTL_CHILDREN(rack_sysctl_root), 639 OID_AUTO, "reorder_thresh", CTLFLAG_RW, 640 &rack_reorder_thresh, 2, 641 "What factor for rack will be added when seeing reordering (shift right)"); 642 SYSCTL_ADD_S32(&rack_sysctl_ctx, 643 SYSCTL_CHILDREN(rack_sysctl_root), 644 OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW, 645 &rack_tlp_thresh, 1, 646 "what divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)"); 647 SYSCTL_ADD_S32(&rack_sysctl_ctx, 648 SYSCTL_CHILDREN(rack_sysctl_root), 649 OID_AUTO, "reorder_fade", CTLFLAG_RW, 650 &rack_reorder_fade, 0, 651 "Does reorder detection fade, if so how many ms (0 means never)"); 652 SYSCTL_ADD_S32(&rack_sysctl_ctx, 653 SYSCTL_CHILDREN(rack_sysctl_root), 654 OID_AUTO, "pktdelay", CTLFLAG_RW, 655 &rack_pkt_delay, 1, 656 "Extra RACK time (in ms) besides reordering thresh"); 657 SYSCTL_ADD_S32(&rack_sysctl_ctx, 658 SYSCTL_CHILDREN(rack_sysctl_root), 659 OID_AUTO, "inc_var", CTLFLAG_RW, 660 &rack_inc_var, 0, 661 "Should rack add to the TLP timer the variance in rtt calculation"); 662 rack_badfr = counter_u64_alloc(M_WAITOK); 663 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 664 SYSCTL_CHILDREN(rack_sysctl_root), 665 OID_AUTO, "badfr", CTLFLAG_RD, 666 &rack_badfr, "Total number of bad FRs"); 667 rack_badfr_bytes = counter_u64_alloc(M_WAITOK); 668 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 669 SYSCTL_CHILDREN(rack_sysctl_root), 670 OID_AUTO, "badfr_bytes", CTLFLAG_RD, 671 &rack_badfr_bytes, "Total number of bad FRs"); 672 rack_rtm_prr_retran = counter_u64_alloc(M_WAITOK); 673 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 674 SYSCTL_CHILDREN(rack_sysctl_root), 675 OID_AUTO, "prrsndret", CTLFLAG_RD, 676 &rack_rtm_prr_retran, 677 "Total number of prr based retransmits"); 678 rack_rtm_prr_newdata = counter_u64_alloc(M_WAITOK); 679 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 680 SYSCTL_CHILDREN(rack_sysctl_root), 681 OID_AUTO, "prrsndnew", CTLFLAG_RD, 682 &rack_rtm_prr_newdata, 683 "Total number of prr based new transmits"); 684 rack_timestamp_mismatch = counter_u64_alloc(M_WAITOK); 685 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 686 SYSCTL_CHILDREN(rack_sysctl_root), 687 OID_AUTO, "tsnf", CTLFLAG_RD, 688 &rack_timestamp_mismatch, 689 "Total number of timestamps that we could not find the reported ts"); 690 rack_find_high = counter_u64_alloc(M_WAITOK); 691 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 692 SYSCTL_CHILDREN(rack_sysctl_root), 693 OID_AUTO, "findhigh", CTLFLAG_RD, 694 &rack_find_high, 695 "Total number of FIN causing find-high"); 696 rack_reorder_seen = counter_u64_alloc(M_WAITOK); 697 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 698 SYSCTL_CHILDREN(rack_sysctl_root), 699 OID_AUTO, "reordering", CTLFLAG_RD, 700 &rack_reorder_seen, 701 "Total number of times we added delay due to reordering"); 702 rack_tlp_tot = counter_u64_alloc(M_WAITOK); 703 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 704 SYSCTL_CHILDREN(rack_sysctl_root), 705 OID_AUTO, "tlp_to_total", CTLFLAG_RD, 706 &rack_tlp_tot, 707 "Total number of tail loss probe expirations"); 708 rack_tlp_newdata = counter_u64_alloc(M_WAITOK); 709 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 710 SYSCTL_CHILDREN(rack_sysctl_root), 711 OID_AUTO, "tlp_new", CTLFLAG_RD, 712 &rack_tlp_newdata, 713 "Total number of tail loss probe sending new data"); 714 715 rack_tlp_retran = counter_u64_alloc(M_WAITOK); 716 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 717 SYSCTL_CHILDREN(rack_sysctl_root), 718 OID_AUTO, "tlp_retran", CTLFLAG_RD, 719 &rack_tlp_retran, 720 "Total number of tail loss probe sending retransmitted data"); 721 rack_tlp_retran_bytes = counter_u64_alloc(M_WAITOK); 722 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 723 SYSCTL_CHILDREN(rack_sysctl_root), 724 OID_AUTO, "tlp_retran_bytes", CTLFLAG_RD, 725 &rack_tlp_retran_bytes, 726 "Total bytes of tail loss probe sending retransmitted data"); 727 rack_tlp_retran_fail = counter_u64_alloc(M_WAITOK); 728 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 729 SYSCTL_CHILDREN(rack_sysctl_root), 730 OID_AUTO, "tlp_retran_fail", CTLFLAG_RD, 731 &rack_tlp_retran_fail, 732 "Total number of tail loss probe sending retransmitted data that failed (wait for t3)"); 733 rack_to_tot = counter_u64_alloc(M_WAITOK); 734 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 735 SYSCTL_CHILDREN(rack_sysctl_root), 736 OID_AUTO, "rack_to_tot", CTLFLAG_RD, 737 &rack_to_tot, 738 "Total number of times the rack to expired?"); 739 rack_to_arm_rack = counter_u64_alloc(M_WAITOK); 740 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 741 SYSCTL_CHILDREN(rack_sysctl_root), 742 OID_AUTO, "arm_rack", CTLFLAG_RD, 743 &rack_to_arm_rack, 744 "Total number of times the rack timer armed?"); 745 rack_to_arm_tlp = counter_u64_alloc(M_WAITOK); 746 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 747 SYSCTL_CHILDREN(rack_sysctl_root), 748 OID_AUTO, "arm_tlp", CTLFLAG_RD, 749 &rack_to_arm_tlp, 750 "Total number of times the tlp timer armed?"); 751 rack_paced_segments = counter_u64_alloc(M_WAITOK); 752 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 753 SYSCTL_CHILDREN(rack_sysctl_root), 754 OID_AUTO, "paced", CTLFLAG_RD, 755 &rack_paced_segments, 756 "Total number of times a segment send caused hptsi"); 757 rack_unpaced_segments = counter_u64_alloc(M_WAITOK); 758 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 759 SYSCTL_CHILDREN(rack_sysctl_root), 760 OID_AUTO, "unpaced", CTLFLAG_RD, 761 &rack_unpaced_segments, 762 "Total number of times a segment did not cause hptsi"); 763 rack_saw_enobuf = counter_u64_alloc(M_WAITOK); 764 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 765 SYSCTL_CHILDREN(rack_sysctl_root), 766 OID_AUTO, "saw_enobufs", CTLFLAG_RD, 767 &rack_saw_enobuf, 768 "Total number of times a segment did not cause hptsi"); 769 rack_saw_enetunreach = counter_u64_alloc(M_WAITOK); 770 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 771 SYSCTL_CHILDREN(rack_sysctl_root), 772 OID_AUTO, "saw_enetunreach", CTLFLAG_RD, 773 &rack_saw_enetunreach, 774 "Total number of times a segment did not cause hptsi"); 775 rack_to_alloc = counter_u64_alloc(M_WAITOK); 776 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 777 SYSCTL_CHILDREN(rack_sysctl_root), 778 OID_AUTO, "allocs", CTLFLAG_RD, 779 &rack_to_alloc, 780 "Total allocations of tracking structures"); 781 rack_to_alloc_hard = counter_u64_alloc(M_WAITOK); 782 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 783 SYSCTL_CHILDREN(rack_sysctl_root), 784 OID_AUTO, "allochard", CTLFLAG_RD, 785 &rack_to_alloc_hard, 786 "Total allocations done with sleeping the hard way"); 787 rack_to_alloc_emerg = counter_u64_alloc(M_WAITOK); 788 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 789 SYSCTL_CHILDREN(rack_sysctl_root), 790 OID_AUTO, "allocemerg", CTLFLAG_RD, 791 &rack_to_alloc_emerg, 792 "Total allocations done from emergency cache"); 793 rack_to_alloc_limited = counter_u64_alloc(M_WAITOK); 794 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 795 SYSCTL_CHILDREN(rack_sysctl_root), 796 OID_AUTO, "alloc_limited", CTLFLAG_RD, 797 &rack_to_alloc_limited, 798 "Total allocations dropped due to limit"); 799 rack_alloc_limited_conns = counter_u64_alloc(M_WAITOK); 800 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 801 SYSCTL_CHILDREN(rack_sysctl_root), 802 OID_AUTO, "alloc_limited_conns", CTLFLAG_RD, 803 &rack_alloc_limited_conns, 804 "Connections with allocations dropped due to limit"); 805 rack_split_limited = counter_u64_alloc(M_WAITOK); 806 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 807 SYSCTL_CHILDREN(rack_sysctl_root), 808 OID_AUTO, "split_limited", CTLFLAG_RD, 809 &rack_split_limited, 810 "Split allocations dropped due to limit"); 811 rack_sack_proc_all = counter_u64_alloc(M_WAITOK); 812 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 813 SYSCTL_CHILDREN(rack_sysctl_root), 814 OID_AUTO, "sack_long", CTLFLAG_RD, 815 &rack_sack_proc_all, 816 "Total times we had to walk whole list for sack processing"); 817 818 rack_sack_proc_restart = counter_u64_alloc(M_WAITOK); 819 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 820 SYSCTL_CHILDREN(rack_sysctl_root), 821 OID_AUTO, "sack_restart", CTLFLAG_RD, 822 &rack_sack_proc_restart, 823 "Total times we had to walk whole list due to a restart"); 824 rack_sack_proc_short = counter_u64_alloc(M_WAITOK); 825 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 826 SYSCTL_CHILDREN(rack_sysctl_root), 827 OID_AUTO, "sack_short", CTLFLAG_RD, 828 &rack_sack_proc_short, 829 "Total times we took shortcut for sack processing"); 830 rack_enter_tlp_calc = counter_u64_alloc(M_WAITOK); 831 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 832 SYSCTL_CHILDREN(rack_sysctl_root), 833 OID_AUTO, "tlp_calc_entered", CTLFLAG_RD, 834 &rack_enter_tlp_calc, 835 "Total times we called calc-tlp"); 836 rack_used_tlpmethod = counter_u64_alloc(M_WAITOK); 837 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 838 SYSCTL_CHILDREN(rack_sysctl_root), 839 OID_AUTO, "hit_tlp_method", CTLFLAG_RD, 840 &rack_used_tlpmethod, 841 "Total number of runt sacks"); 842 rack_used_tlpmethod2 = counter_u64_alloc(M_WAITOK); 843 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 844 SYSCTL_CHILDREN(rack_sysctl_root), 845 OID_AUTO, "hit_tlp_method2", CTLFLAG_RD, 846 &rack_used_tlpmethod2, 847 "Total number of runt sacks 2"); 848 rack_runt_sacks = counter_u64_alloc(M_WAITOK); 849 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 850 SYSCTL_CHILDREN(rack_sysctl_root), 851 OID_AUTO, "runtsacks", CTLFLAG_RD, 852 &rack_runt_sacks, 853 "Total number of runt sacks"); 854 rack_progress_drops = counter_u64_alloc(M_WAITOK); 855 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 856 SYSCTL_CHILDREN(rack_sysctl_root), 857 OID_AUTO, "prog_drops", CTLFLAG_RD, 858 &rack_progress_drops, 859 "Total number of progress drops"); 860 rack_input_idle_reduces = counter_u64_alloc(M_WAITOK); 861 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 862 SYSCTL_CHILDREN(rack_sysctl_root), 863 OID_AUTO, "idle_reduce_oninput", CTLFLAG_RD, 864 &rack_input_idle_reduces, 865 "Total number of idle reductions on input"); 866 rack_tlp_does_nada = counter_u64_alloc(M_WAITOK); 867 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 868 SYSCTL_CHILDREN(rack_sysctl_root), 869 OID_AUTO, "tlp_nada", CTLFLAG_RD, 870 &rack_tlp_does_nada, 871 "Total number of nada tlp calls"); 872 COUNTER_ARRAY_ALLOC(rack_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK); 873 SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), 874 OID_AUTO, "outsize", CTLFLAG_RD, 875 rack_out_size, TCP_MSS_ACCT_SIZE, "MSS send sizes"); 876 COUNTER_ARRAY_ALLOC(rack_opts_arry, RACK_OPTS_SIZE, M_WAITOK); 877 SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), 878 OID_AUTO, "opts", CTLFLAG_RD, 879 rack_opts_arry, RACK_OPTS_SIZE, "RACK Option Stats"); 880 SYSCTL_ADD_PROC(&rack_sysctl_ctx, 881 SYSCTL_CHILDREN(rack_sysctl_root), 882 OID_AUTO, "clear", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE, 883 &rack_clear_counter, 0, sysctl_rack_clear, "IU", "Clear counters"); 884 } 885 886 static inline int32_t 887 rack_progress_timeout_check(struct tcpcb *tp) 888 { 889 #ifdef NETFLIX_PROGRESS 890 if (tp->t_maxunacktime && tp->t_acktime && TSTMP_GT(ticks, tp->t_acktime)) { 891 if ((ticks - tp->t_acktime) >= tp->t_maxunacktime) { 892 /* 893 * There is an assumption that the caller 894 * will drop the connection so we will 895 * increment the counters here. 896 */ 897 struct tcp_rack *rack; 898 rack = (struct tcp_rack *)tp->t_fb_ptr; 899 counter_u64_add(rack_progress_drops, 1); 900 TCPSTAT_INC(tcps_progdrops); 901 rack_log_progress_event(rack, tp, ticks, PROGRESS_DROP, __LINE__); 902 return (1); 903 } 904 } 905 #endif 906 return (0); 907 } 908 909 910 static void 911 rack_log_to_start(struct tcp_rack *rack, uint32_t cts, uint32_t to, int32_t slot, uint8_t which) 912 { 913 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 914 union tcp_log_stackspecific log; 915 916 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 917 log.u_bbr.flex1 = TICKS_2_MSEC(rack->rc_tp->t_srtt >> TCP_RTT_SHIFT); 918 log.u_bbr.flex2 = to; 919 log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags; 920 log.u_bbr.flex4 = slot; 921 log.u_bbr.flex5 = rack->rc_inp->inp_hptsslot; 922 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur; 923 log.u_bbr.flex8 = which; 924 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; 925 log.u_bbr.ininput = rack->rc_inp->inp_in_input; 926 TCP_LOG_EVENT(rack->rc_tp, NULL, 927 &rack->rc_inp->inp_socket->so_rcv, 928 &rack->rc_inp->inp_socket->so_snd, 929 BBR_LOG_TIMERSTAR, 0, 930 0, &log, false); 931 } 932 } 933 934 static void 935 rack_log_to_event(struct tcp_rack *rack, int32_t to_num) 936 { 937 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 938 union tcp_log_stackspecific log; 939 940 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 941 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; 942 log.u_bbr.ininput = rack->rc_inp->inp_in_input; 943 log.u_bbr.flex8 = to_num; 944 log.u_bbr.flex1 = rack->r_ctl.rc_rack_min_rtt; 945 log.u_bbr.flex2 = rack->rc_rack_rtt; 946 TCP_LOG_EVENT(rack->rc_tp, NULL, 947 &rack->rc_inp->inp_socket->so_rcv, 948 &rack->rc_inp->inp_socket->so_snd, 949 BBR_LOG_RTO, 0, 950 0, &log, false); 951 } 952 } 953 954 static void 955 rack_log_rtt_upd(struct tcpcb *tp, struct tcp_rack *rack, int32_t t, 956 uint32_t o_srtt, uint32_t o_var) 957 { 958 if (tp->t_logstate != TCP_LOG_STATE_OFF) { 959 union tcp_log_stackspecific log; 960 961 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 962 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; 963 log.u_bbr.ininput = rack->rc_inp->inp_in_input; 964 log.u_bbr.flex1 = t; 965 log.u_bbr.flex2 = o_srtt; 966 log.u_bbr.flex3 = o_var; 967 log.u_bbr.flex4 = rack->r_ctl.rack_rs.rs_rtt_lowest; 968 log.u_bbr.flex5 = rack->r_ctl.rack_rs.rs_rtt_highest; 969 log.u_bbr.flex6 = rack->r_ctl.rack_rs.rs_rtt_cnt; 970 log.u_bbr.rttProp = rack->r_ctl.rack_rs.rs_rtt_tot; 971 log.u_bbr.flex8 = rack->r_ctl.rc_rate_sample_method; 972 TCP_LOG_EVENT(tp, NULL, 973 &rack->rc_inp->inp_socket->so_rcv, 974 &rack->rc_inp->inp_socket->so_snd, 975 BBR_LOG_BBRRTT, 0, 976 0, &log, false); 977 } 978 } 979 980 static void 981 rack_log_rtt_sample(struct tcp_rack *rack, uint32_t rtt) 982 { 983 /* 984 * Log the rtt sample we are 985 * applying to the srtt algorithm in 986 * useconds. 987 */ 988 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 989 union tcp_log_stackspecific log; 990 struct timeval tv; 991 992 memset(&log, 0, sizeof(log)); 993 /* Convert our ms to a microsecond */ 994 log.u_bbr.flex1 = rtt * 1000; 995 log.u_bbr.timeStamp = tcp_get_usecs(&tv); 996 TCP_LOG_EVENTP(rack->rc_tp, NULL, 997 &rack->rc_inp->inp_socket->so_rcv, 998 &rack->rc_inp->inp_socket->so_snd, 999 TCP_LOG_RTT, 0, 1000 0, &log, false, &tv); 1001 } 1002 } 1003 1004 1005 static inline void 1006 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick, int event, int line) 1007 { 1008 if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) { 1009 union tcp_log_stackspecific log; 1010 1011 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 1012 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; 1013 log.u_bbr.ininput = rack->rc_inp->inp_in_input; 1014 log.u_bbr.flex1 = line; 1015 log.u_bbr.flex2 = tick; 1016 log.u_bbr.flex3 = tp->t_maxunacktime; 1017 log.u_bbr.flex4 = tp->t_acktime; 1018 log.u_bbr.flex8 = event; 1019 TCP_LOG_EVENT(tp, NULL, 1020 &rack->rc_inp->inp_socket->so_rcv, 1021 &rack->rc_inp->inp_socket->so_snd, 1022 BBR_LOG_PROGRESS, 0, 1023 0, &log, false); 1024 } 1025 } 1026 1027 static void 1028 rack_log_type_bbrsnd(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint32_t cts) 1029 { 1030 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 1031 union tcp_log_stackspecific log; 1032 1033 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 1034 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; 1035 log.u_bbr.ininput = rack->rc_inp->inp_in_input; 1036 log.u_bbr.flex1 = slot; 1037 log.u_bbr.flex7 = (0x0000ffff & rack->r_ctl.rc_hpts_flags); 1038 log.u_bbr.flex8 = rack->rc_in_persist; 1039 TCP_LOG_EVENT(rack->rc_tp, NULL, 1040 &rack->rc_inp->inp_socket->so_rcv, 1041 &rack->rc_inp->inp_socket->so_snd, 1042 BBR_LOG_BBRSND, 0, 1043 0, &log, false); 1044 } 1045 } 1046 1047 static void 1048 rack_log_doseg_done(struct tcp_rack *rack, uint32_t cts, int32_t nxt_pkt, int32_t did_out, int way_out) 1049 { 1050 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 1051 union tcp_log_stackspecific log; 1052 1053 memset(&log, 0, sizeof(log)); 1054 log.u_bbr.flex1 = did_out; 1055 log.u_bbr.flex2 = nxt_pkt; 1056 log.u_bbr.flex3 = way_out; 1057 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags; 1058 log.u_bbr.flex7 = rack->r_wanted_output; 1059 log.u_bbr.flex8 = rack->rc_in_persist; 1060 TCP_LOG_EVENT(rack->rc_tp, NULL, 1061 &rack->rc_inp->inp_socket->so_rcv, 1062 &rack->rc_inp->inp_socket->so_snd, 1063 BBR_LOG_DOSEG_DONE, 0, 1064 0, &log, false); 1065 } 1066 } 1067 1068 1069 static void 1070 rack_log_type_just_return(struct tcp_rack *rack, uint32_t cts, uint32_t tlen, uint32_t slot, uint8_t hpts_calling) 1071 { 1072 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 1073 union tcp_log_stackspecific log; 1074 1075 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 1076 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; 1077 log.u_bbr.ininput = rack->rc_inp->inp_in_input; 1078 log.u_bbr.flex1 = slot; 1079 log.u_bbr.flex2 = rack->r_ctl.rc_hpts_flags; 1080 log.u_bbr.flex7 = hpts_calling; 1081 log.u_bbr.flex8 = rack->rc_in_persist; 1082 TCP_LOG_EVENT(rack->rc_tp, NULL, 1083 &rack->rc_inp->inp_socket->so_rcv, 1084 &rack->rc_inp->inp_socket->so_snd, 1085 BBR_LOG_JUSTRET, 0, 1086 tlen, &log, false); 1087 } 1088 } 1089 1090 static void 1091 rack_log_to_cancel(struct tcp_rack *rack, int32_t hpts_removed, int line) 1092 { 1093 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 1094 union tcp_log_stackspecific log; 1095 1096 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 1097 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; 1098 log.u_bbr.ininput = rack->rc_inp->inp_in_input; 1099 log.u_bbr.flex1 = line; 1100 log.u_bbr.flex2 = 0; 1101 log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags; 1102 log.u_bbr.flex4 = 0; 1103 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur; 1104 log.u_bbr.flex8 = hpts_removed; 1105 TCP_LOG_EVENT(rack->rc_tp, NULL, 1106 &rack->rc_inp->inp_socket->so_rcv, 1107 &rack->rc_inp->inp_socket->so_snd, 1108 BBR_LOG_TIMERCANC, 0, 1109 0, &log, false); 1110 } 1111 } 1112 1113 static void 1114 rack_log_to_processing(struct tcp_rack *rack, uint32_t cts, int32_t ret, int32_t timers) 1115 { 1116 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 1117 union tcp_log_stackspecific log; 1118 1119 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 1120 log.u_bbr.flex1 = timers; 1121 log.u_bbr.flex2 = ret; 1122 log.u_bbr.flex3 = rack->r_ctl.rc_timer_exp; 1123 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags; 1124 log.u_bbr.flex5 = cts; 1125 TCP_LOG_EVENT(rack->rc_tp, NULL, 1126 &rack->rc_inp->inp_socket->so_rcv, 1127 &rack->rc_inp->inp_socket->so_snd, 1128 BBR_LOG_TO_PROCESS, 0, 1129 0, &log, false); 1130 } 1131 } 1132 1133 static void 1134 rack_counter_destroy() 1135 { 1136 counter_u64_free(rack_badfr); 1137 counter_u64_free(rack_badfr_bytes); 1138 counter_u64_free(rack_rtm_prr_retran); 1139 counter_u64_free(rack_rtm_prr_newdata); 1140 counter_u64_free(rack_timestamp_mismatch); 1141 counter_u64_free(rack_reorder_seen); 1142 counter_u64_free(rack_tlp_tot); 1143 counter_u64_free(rack_tlp_newdata); 1144 counter_u64_free(rack_tlp_retran); 1145 counter_u64_free(rack_tlp_retran_bytes); 1146 counter_u64_free(rack_tlp_retran_fail); 1147 counter_u64_free(rack_to_tot); 1148 counter_u64_free(rack_to_arm_rack); 1149 counter_u64_free(rack_to_arm_tlp); 1150 counter_u64_free(rack_paced_segments); 1151 counter_u64_free(rack_unpaced_segments); 1152 counter_u64_free(rack_saw_enobuf); 1153 counter_u64_free(rack_saw_enetunreach); 1154 counter_u64_free(rack_to_alloc_hard); 1155 counter_u64_free(rack_to_alloc_emerg); 1156 counter_u64_free(rack_sack_proc_all); 1157 counter_u64_free(rack_sack_proc_short); 1158 counter_u64_free(rack_sack_proc_restart); 1159 counter_u64_free(rack_to_alloc); 1160 counter_u64_free(rack_to_alloc_limited); 1161 counter_u64_free(rack_split_limited); 1162 counter_u64_free(rack_find_high); 1163 counter_u64_free(rack_runt_sacks); 1164 counter_u64_free(rack_enter_tlp_calc); 1165 counter_u64_free(rack_used_tlpmethod); 1166 counter_u64_free(rack_used_tlpmethod2); 1167 counter_u64_free(rack_progress_drops); 1168 counter_u64_free(rack_input_idle_reduces); 1169 counter_u64_free(rack_tlp_does_nada); 1170 COUNTER_ARRAY_FREE(rack_out_size, TCP_MSS_ACCT_SIZE); 1171 COUNTER_ARRAY_FREE(rack_opts_arry, RACK_OPTS_SIZE); 1172 } 1173 1174 static struct rack_sendmap * 1175 rack_alloc(struct tcp_rack *rack) 1176 { 1177 struct rack_sendmap *rsm; 1178 1179 rsm = uma_zalloc(rack_zone, M_NOWAIT); 1180 if (rsm) { 1181 rack->r_ctl.rc_num_maps_alloced++; 1182 counter_u64_add(rack_to_alloc, 1); 1183 return (rsm); 1184 } 1185 if (rack->rc_free_cnt) { 1186 counter_u64_add(rack_to_alloc_emerg, 1); 1187 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free); 1188 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_next); 1189 rack->rc_free_cnt--; 1190 return (rsm); 1191 } 1192 return (NULL); 1193 } 1194 1195 static struct rack_sendmap * 1196 rack_alloc_full_limit(struct tcp_rack *rack) 1197 { 1198 if ((rack_map_entries_limit > 0) && 1199 (rack->r_ctl.rc_num_maps_alloced >= rack_map_entries_limit)) { 1200 counter_u64_add(rack_to_alloc_limited, 1); 1201 if (!rack->alloc_limit_reported) { 1202 rack->alloc_limit_reported = 1; 1203 counter_u64_add(rack_alloc_limited_conns, 1); 1204 } 1205 return (NULL); 1206 } 1207 return (rack_alloc(rack)); 1208 } 1209 1210 /* wrapper to allocate a sendmap entry, subject to a specific limit */ 1211 static struct rack_sendmap * 1212 rack_alloc_limit(struct tcp_rack *rack, uint8_t limit_type) 1213 { 1214 struct rack_sendmap *rsm; 1215 1216 if (limit_type) { 1217 /* currently there is only one limit type */ 1218 if (rack_map_split_limit > 0 && 1219 rack->r_ctl.rc_num_split_allocs >= rack_map_split_limit) { 1220 counter_u64_add(rack_split_limited, 1); 1221 if (!rack->alloc_limit_reported) { 1222 rack->alloc_limit_reported = 1; 1223 counter_u64_add(rack_alloc_limited_conns, 1); 1224 } 1225 return (NULL); 1226 } 1227 } 1228 1229 /* allocate and mark in the limit type, if set */ 1230 rsm = rack_alloc(rack); 1231 if (rsm != NULL && limit_type) { 1232 rsm->r_limit_type = limit_type; 1233 rack->r_ctl.rc_num_split_allocs++; 1234 } 1235 return (rsm); 1236 } 1237 1238 static void 1239 rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm) 1240 { 1241 if (rsm->r_limit_type) { 1242 /* currently there is only one limit type */ 1243 rack->r_ctl.rc_num_split_allocs--; 1244 } 1245 if (rack->r_ctl.rc_tlpsend == rsm) 1246 rack->r_ctl.rc_tlpsend = NULL; 1247 if (rack->r_ctl.rc_next == rsm) 1248 rack->r_ctl.rc_next = NULL; 1249 if (rack->r_ctl.rc_sacklast == rsm) 1250 rack->r_ctl.rc_sacklast = NULL; 1251 if (rack->rc_free_cnt < rack_free_cache) { 1252 memset(rsm, 0, sizeof(struct rack_sendmap)); 1253 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_next); 1254 rsm->r_limit_type = 0; 1255 rack->rc_free_cnt++; 1256 return; 1257 } 1258 rack->r_ctl.rc_num_maps_alloced--; 1259 uma_zfree(rack_zone, rsm); 1260 } 1261 1262 /* 1263 * CC wrapper hook functions 1264 */ 1265 static void 1266 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, struct tcphdr *th, uint16_t nsegs, 1267 uint16_t type, int32_t recovery) 1268 { 1269 #ifdef NETFLIX_STATS 1270 int32_t gput; 1271 #endif 1272 1273 INP_WLOCK_ASSERT(tp->t_inpcb); 1274 1275 tp->ccv->nsegs = nsegs; 1276 tp->ccv->bytes_this_ack = BYTES_THIS_ACK(tp, th); 1277 if ((recovery) && (rack->r_ctl.rc_early_recovery_segs)) { 1278 uint32_t max; 1279 1280 max = rack->r_ctl.rc_early_recovery_segs * tp->t_maxseg; 1281 if (tp->ccv->bytes_this_ack > max) { 1282 tp->ccv->bytes_this_ack = max; 1283 } 1284 } 1285 if (tp->snd_cwnd <= tp->snd_wnd) 1286 tp->ccv->flags |= CCF_CWND_LIMITED; 1287 else 1288 tp->ccv->flags &= ~CCF_CWND_LIMITED; 1289 1290 if (type == CC_ACK) { 1291 #ifdef NETFLIX_STATS 1292 stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF, 1293 ((int32_t) tp->snd_cwnd) - tp->snd_wnd); 1294 if ((tp->t_flags & TF_GPUTINPROG) && 1295 SEQ_GEQ(th->th_ack, tp->gput_ack)) { 1296 gput = (((int64_t) (th->th_ack - tp->gput_seq)) << 3) / 1297 max(1, tcp_ts_getticks() - tp->gput_ts); 1298 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT, 1299 gput); 1300 /* 1301 * XXXLAS: This is a temporary hack, and should be 1302 * chained off VOI_TCP_GPUT when stats(9) grows an 1303 * API to deal with chained VOIs. 1304 */ 1305 if (tp->t_stats_gput_prev > 0) 1306 stats_voi_update_abs_s32(tp->t_stats, 1307 VOI_TCP_GPUT_ND, 1308 ((gput - tp->t_stats_gput_prev) * 100) / 1309 tp->t_stats_gput_prev); 1310 tp->t_flags &= ~TF_GPUTINPROG; 1311 tp->t_stats_gput_prev = gput; 1312 if (tp->t_maxpeakrate) { 1313 /* 1314 * We update t_peakrate_thr. This gives us roughly 1315 * one update per round trip time. 1316 */ 1317 tcp_update_peakrate_thr(tp); 1318 } 1319 } 1320 #endif 1321 if (tp->snd_cwnd > tp->snd_ssthresh) { 1322 tp->t_bytes_acked += min(tp->ccv->bytes_this_ack, 1323 nsegs * V_tcp_abc_l_var * tp->t_maxseg); 1324 if (tp->t_bytes_acked >= tp->snd_cwnd) { 1325 tp->t_bytes_acked -= tp->snd_cwnd; 1326 tp->ccv->flags |= CCF_ABC_SENTAWND; 1327 } 1328 } else { 1329 tp->ccv->flags &= ~CCF_ABC_SENTAWND; 1330 tp->t_bytes_acked = 0; 1331 } 1332 } 1333 if (CC_ALGO(tp)->ack_received != NULL) { 1334 /* XXXLAS: Find a way to live without this */ 1335 tp->ccv->curack = th->th_ack; 1336 CC_ALGO(tp)->ack_received(tp->ccv, type); 1337 } 1338 #ifdef NETFLIX_STATS 1339 stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, tp->snd_cwnd); 1340 #endif 1341 if (rack->r_ctl.rc_rack_largest_cwnd < tp->snd_cwnd) { 1342 rack->r_ctl.rc_rack_largest_cwnd = tp->snd_cwnd; 1343 } 1344 /* we enforce max peak rate if it is set. */ 1345 if (tp->t_peakrate_thr && tp->snd_cwnd > tp->t_peakrate_thr) { 1346 tp->snd_cwnd = tp->t_peakrate_thr; 1347 } 1348 } 1349 1350 static void 1351 tcp_rack_partialack(struct tcpcb *tp, struct tcphdr *th) 1352 { 1353 struct tcp_rack *rack; 1354 1355 rack = (struct tcp_rack *)tp->t_fb_ptr; 1356 INP_WLOCK_ASSERT(tp->t_inpcb); 1357 if (rack->r_ctl.rc_prr_sndcnt > 0) 1358 rack->r_wanted_output++; 1359 } 1360 1361 static void 1362 rack_post_recovery(struct tcpcb *tp, struct tcphdr *th) 1363 { 1364 struct tcp_rack *rack; 1365 1366 INP_WLOCK_ASSERT(tp->t_inpcb); 1367 rack = (struct tcp_rack *)tp->t_fb_ptr; 1368 if (CC_ALGO(tp)->post_recovery != NULL) { 1369 tp->ccv->curack = th->th_ack; 1370 CC_ALGO(tp)->post_recovery(tp->ccv); 1371 } 1372 /* 1373 * Here we can in theory adjust cwnd to be based on the number of 1374 * losses in the window (rack->r_ctl.rc_loss_count). This is done 1375 * based on the rack_use_proportional flag. 1376 */ 1377 if (rack->r_ctl.rc_prop_reduce && rack->r_ctl.rc_prop_rate) { 1378 int32_t reduce; 1379 1380 reduce = (rack->r_ctl.rc_loss_count * rack->r_ctl.rc_prop_rate); 1381 if (reduce > 50) { 1382 reduce = 50; 1383 } 1384 tp->snd_cwnd -= ((reduce * tp->snd_cwnd) / 100); 1385 } else { 1386 if (tp->snd_cwnd > tp->snd_ssthresh) { 1387 /* Drop us down to the ssthresh (1/2 cwnd at loss) */ 1388 tp->snd_cwnd = tp->snd_ssthresh; 1389 } 1390 } 1391 if (rack->r_ctl.rc_prr_sndcnt > 0) { 1392 /* Suck the next prr cnt back into cwnd */ 1393 tp->snd_cwnd += rack->r_ctl.rc_prr_sndcnt; 1394 rack->r_ctl.rc_prr_sndcnt = 0; 1395 } 1396 tp->snd_recover = tp->snd_una; 1397 EXIT_RECOVERY(tp->t_flags); 1398 } 1399 1400 static void 1401 rack_cong_signal(struct tcpcb *tp, struct tcphdr *th, uint32_t type) 1402 { 1403 struct tcp_rack *rack; 1404 1405 INP_WLOCK_ASSERT(tp->t_inpcb); 1406 1407 rack = (struct tcp_rack *)tp->t_fb_ptr; 1408 switch (type) { 1409 case CC_NDUPACK: 1410 /* rack->r_ctl.rc_ssthresh_set = 1;*/ 1411 if (!IN_FASTRECOVERY(tp->t_flags)) { 1412 rack->r_ctl.rc_tlp_rtx_out = 0; 1413 rack->r_ctl.rc_prr_delivered = 0; 1414 rack->r_ctl.rc_prr_out = 0; 1415 rack->r_ctl.rc_loss_count = 0; 1416 rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; 1417 rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una; 1418 tp->snd_recover = tp->snd_max; 1419 if (tp->t_flags & TF_ECN_PERMIT) 1420 tp->t_flags |= TF_ECN_SND_CWR; 1421 } 1422 break; 1423 case CC_ECN: 1424 if (!IN_CONGRECOVERY(tp->t_flags)) { 1425 TCPSTAT_INC(tcps_ecn_rcwnd); 1426 tp->snd_recover = tp->snd_max; 1427 if (tp->t_flags & TF_ECN_PERMIT) 1428 tp->t_flags |= TF_ECN_SND_CWR; 1429 } 1430 break; 1431 case CC_RTO: 1432 tp->t_dupacks = 0; 1433 tp->t_bytes_acked = 0; 1434 EXIT_RECOVERY(tp->t_flags); 1435 tp->snd_ssthresh = max(2, min(tp->snd_wnd, tp->snd_cwnd) / 2 / 1436 tp->t_maxseg) * tp->t_maxseg; 1437 tp->snd_cwnd = tp->t_maxseg; 1438 break; 1439 case CC_RTO_ERR: 1440 TCPSTAT_INC(tcps_sndrexmitbad); 1441 /* RTO was unnecessary, so reset everything. */ 1442 tp->snd_cwnd = tp->snd_cwnd_prev; 1443 tp->snd_ssthresh = tp->snd_ssthresh_prev; 1444 tp->snd_recover = tp->snd_recover_prev; 1445 if (tp->t_flags & TF_WASFRECOVERY) 1446 ENTER_FASTRECOVERY(tp->t_flags); 1447 if (tp->t_flags & TF_WASCRECOVERY) 1448 ENTER_CONGRECOVERY(tp->t_flags); 1449 tp->snd_nxt = tp->snd_max; 1450 tp->t_badrxtwin = 0; 1451 break; 1452 } 1453 1454 if (CC_ALGO(tp)->cong_signal != NULL) { 1455 if (th != NULL) 1456 tp->ccv->curack = th->th_ack; 1457 CC_ALGO(tp)->cong_signal(tp->ccv, type); 1458 } 1459 } 1460 1461 1462 1463 static inline void 1464 rack_cc_after_idle(struct tcpcb *tp, int reduce_largest) 1465 { 1466 uint32_t i_cwnd; 1467 1468 INP_WLOCK_ASSERT(tp->t_inpcb); 1469 1470 #ifdef NETFLIX_STATS 1471 TCPSTAT_INC(tcps_idle_restarts); 1472 if (tp->t_state == TCPS_ESTABLISHED) 1473 TCPSTAT_INC(tcps_idle_estrestarts); 1474 #endif 1475 if (CC_ALGO(tp)->after_idle != NULL) 1476 CC_ALGO(tp)->after_idle(tp->ccv); 1477 1478 if (V_tcp_initcwnd_segments) 1479 i_cwnd = min((V_tcp_initcwnd_segments * tp->t_maxseg), 1480 max(2 * tp->t_maxseg, 14600)); 1481 else if (V_tcp_do_rfc3390) 1482 i_cwnd = min(4 * tp->t_maxseg, 1483 max(2 * tp->t_maxseg, 4380)); 1484 else { 1485 /* Per RFC5681 Section 3.1 */ 1486 if (tp->t_maxseg > 2190) 1487 i_cwnd = 2 * tp->t_maxseg; 1488 else if (tp->t_maxseg > 1095) 1489 i_cwnd = 3 * tp->t_maxseg; 1490 else 1491 i_cwnd = 4 * tp->t_maxseg; 1492 } 1493 if (reduce_largest) { 1494 /* 1495 * Do we reduce the largest cwnd to make 1496 * rack play nice on restart hptsi wise? 1497 */ 1498 if (((struct tcp_rack *)tp->t_fb_ptr)->r_ctl.rc_rack_largest_cwnd > i_cwnd) 1499 ((struct tcp_rack *)tp->t_fb_ptr)->r_ctl.rc_rack_largest_cwnd = i_cwnd; 1500 } 1501 /* 1502 * Being idle is no differnt than the initial window. If the cc 1503 * clamps it down below the initial window raise it to the initial 1504 * window. 1505 */ 1506 if (tp->snd_cwnd < i_cwnd) { 1507 tp->snd_cwnd = i_cwnd; 1508 } 1509 } 1510 1511 1512 /* 1513 * Indicate whether this ack should be delayed. We can delay the ack if 1514 * following conditions are met: 1515 * - There is no delayed ack timer in progress. 1516 * - Our last ack wasn't a 0-sized window. We never want to delay 1517 * the ack that opens up a 0-sized window. 1518 * - LRO wasn't used for this segment. We make sure by checking that the 1519 * segment size is not larger than the MSS. 1520 * - Delayed acks are enabled or this is a half-synchronized T/TCP 1521 * connection. 1522 */ 1523 #define DELAY_ACK(tp, tlen) \ 1524 (((tp->t_flags & TF_RXWIN0SENT) == 0) && \ 1525 ((tp->t_flags & TF_DELACK) == 0) && \ 1526 (tlen <= tp->t_maxseg) && \ 1527 (tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN))) 1528 1529 static inline void 1530 rack_calc_rwin(struct socket *so, struct tcpcb *tp) 1531 { 1532 int32_t win; 1533 1534 /* 1535 * Calculate amount of space in receive window, and then do TCP 1536 * input processing. Receive window is amount of space in rcv queue, 1537 * but not less than advertised window. 1538 */ 1539 win = sbspace(&so->so_rcv); 1540 if (win < 0) 1541 win = 0; 1542 tp->rcv_wnd = imax(win, (int)(tp->rcv_adv - tp->rcv_nxt)); 1543 } 1544 1545 static void 1546 rack_do_drop(struct mbuf *m, struct tcpcb *tp) 1547 { 1548 /* 1549 * Drop space held by incoming segment and return. 1550 */ 1551 if (tp != NULL) 1552 INP_WUNLOCK(tp->t_inpcb); 1553 if (m) 1554 m_freem(m); 1555 } 1556 1557 static void 1558 rack_do_dropwithreset(struct mbuf *m, struct tcpcb *tp, struct tcphdr *th, int32_t rstreason, int32_t tlen) 1559 { 1560 if (tp != NULL) { 1561 tcp_dropwithreset(m, th, tp, tlen, rstreason); 1562 INP_WUNLOCK(tp->t_inpcb); 1563 } else 1564 tcp_dropwithreset(m, th, NULL, tlen, rstreason); 1565 } 1566 1567 /* 1568 * The value in ret_val informs the caller 1569 * if we dropped the tcb (and lock) or not. 1570 * 1 = we dropped it, 0 = the TCB is still locked 1571 * and valid. 1572 */ 1573 static void 1574 rack_do_dropafterack(struct mbuf *m, struct tcpcb *tp, struct tcphdr *th, int32_t thflags, int32_t tlen, int32_t * ret_val) 1575 { 1576 /* 1577 * Generate an ACK dropping incoming segment if it occupies sequence 1578 * space, where the ACK reflects our state. 1579 * 1580 * We can now skip the test for the RST flag since all paths to this 1581 * code happen after packets containing RST have been dropped. 1582 * 1583 * In the SYN-RECEIVED state, don't send an ACK unless the segment 1584 * we received passes the SYN-RECEIVED ACK test. If it fails send a 1585 * RST. This breaks the loop in the "LAND" DoS attack, and also 1586 * prevents an ACK storm between two listening ports that have been 1587 * sent forged SYN segments, each with the source address of the 1588 * other. 1589 */ 1590 struct tcp_rack *rack; 1591 1592 if (tp->t_state == TCPS_SYN_RECEIVED && (thflags & TH_ACK) && 1593 (SEQ_GT(tp->snd_una, th->th_ack) || 1594 SEQ_GT(th->th_ack, tp->snd_max))) { 1595 *ret_val = 1; 1596 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 1597 return; 1598 } else 1599 *ret_val = 0; 1600 rack = (struct tcp_rack *)tp->t_fb_ptr; 1601 rack->r_wanted_output++; 1602 tp->t_flags |= TF_ACKNOW; 1603 if (m) 1604 m_freem(m); 1605 } 1606 1607 1608 static int 1609 rack_process_rst(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp) 1610 { 1611 /* 1612 * RFC5961 Section 3.2 1613 * 1614 * - RST drops connection only if SEG.SEQ == RCV.NXT. - If RST is in 1615 * window, we send challenge ACK. 1616 * 1617 * Note: to take into account delayed ACKs, we should test against 1618 * last_ack_sent instead of rcv_nxt. Note 2: we handle special case 1619 * of closed window, not covered by the RFC. 1620 */ 1621 int dropped = 0; 1622 1623 if ((SEQ_GEQ(th->th_seq, (tp->last_ack_sent - 1)) && 1624 SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) || 1625 (tp->rcv_wnd == 0 && tp->last_ack_sent == th->th_seq)) { 1626 1627 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 1628 KASSERT(tp->t_state != TCPS_SYN_SENT, 1629 ("%s: TH_RST for TCPS_SYN_SENT th %p tp %p", 1630 __func__, th, tp)); 1631 1632 if (V_tcp_insecure_rst || 1633 (tp->last_ack_sent == th->th_seq) || 1634 (tp->rcv_nxt == th->th_seq) || 1635 ((tp->last_ack_sent - 1) == th->th_seq)) { 1636 TCPSTAT_INC(tcps_drops); 1637 /* Drop the connection. */ 1638 switch (tp->t_state) { 1639 case TCPS_SYN_RECEIVED: 1640 so->so_error = ECONNREFUSED; 1641 goto close; 1642 case TCPS_ESTABLISHED: 1643 case TCPS_FIN_WAIT_1: 1644 case TCPS_FIN_WAIT_2: 1645 case TCPS_CLOSE_WAIT: 1646 case TCPS_CLOSING: 1647 case TCPS_LAST_ACK: 1648 so->so_error = ECONNRESET; 1649 close: 1650 tcp_state_change(tp, TCPS_CLOSED); 1651 /* FALLTHROUGH */ 1652 default: 1653 tp = tcp_close(tp); 1654 } 1655 dropped = 1; 1656 rack_do_drop(m, tp); 1657 } else { 1658 TCPSTAT_INC(tcps_badrst); 1659 /* Send challenge ACK. */ 1660 tcp_respond(tp, mtod(m, void *), th, m, 1661 tp->rcv_nxt, tp->snd_nxt, TH_ACK); 1662 tp->last_ack_sent = tp->rcv_nxt; 1663 } 1664 } else { 1665 m_freem(m); 1666 } 1667 return (dropped); 1668 } 1669 1670 /* 1671 * The value in ret_val informs the caller 1672 * if we dropped the tcb (and lock) or not. 1673 * 1 = we dropped it, 0 = the TCB is still locked 1674 * and valid. 1675 */ 1676 static void 1677 rack_challenge_ack(struct mbuf *m, struct tcphdr *th, struct tcpcb *tp, int32_t * ret_val) 1678 { 1679 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 1680 1681 TCPSTAT_INC(tcps_badsyn); 1682 if (V_tcp_insecure_syn && 1683 SEQ_GEQ(th->th_seq, tp->last_ack_sent) && 1684 SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) { 1685 tp = tcp_drop(tp, ECONNRESET); 1686 *ret_val = 1; 1687 rack_do_drop(m, tp); 1688 } else { 1689 /* Send challenge ACK. */ 1690 tcp_respond(tp, mtod(m, void *), th, m, tp->rcv_nxt, 1691 tp->snd_nxt, TH_ACK); 1692 tp->last_ack_sent = tp->rcv_nxt; 1693 m = NULL; 1694 *ret_val = 0; 1695 rack_do_drop(m, NULL); 1696 } 1697 } 1698 1699 /* 1700 * rack_ts_check returns 1 for you should not proceed. It places 1701 * in ret_val what should be returned 1/0 by the caller. The 1 indicates 1702 * that the TCB is unlocked and probably dropped. The 0 indicates the 1703 * TCB is still valid and locked. 1704 */ 1705 static int 1706 rack_ts_check(struct mbuf *m, struct tcphdr *th, struct tcpcb *tp, int32_t tlen, int32_t thflags, int32_t * ret_val) 1707 { 1708 1709 /* Check to see if ts_recent is over 24 days old. */ 1710 if (tcp_ts_getticks() - tp->ts_recent_age > TCP_PAWS_IDLE) { 1711 /* 1712 * Invalidate ts_recent. If this segment updates ts_recent, 1713 * the age will be reset later and ts_recent will get a 1714 * valid value. If it does not, setting ts_recent to zero 1715 * will at least satisfy the requirement that zero be placed 1716 * in the timestamp echo reply when ts_recent isn't valid. 1717 * The age isn't reset until we get a valid ts_recent 1718 * because we don't want out-of-order segments to be dropped 1719 * when ts_recent is old. 1720 */ 1721 tp->ts_recent = 0; 1722 } else { 1723 TCPSTAT_INC(tcps_rcvduppack); 1724 TCPSTAT_ADD(tcps_rcvdupbyte, tlen); 1725 TCPSTAT_INC(tcps_pawsdrop); 1726 *ret_val = 0; 1727 if (tlen) { 1728 rack_do_dropafterack(m, tp, th, thflags, tlen, ret_val); 1729 } else { 1730 rack_do_drop(m, NULL); 1731 } 1732 return (1); 1733 } 1734 return (0); 1735 } 1736 1737 /* 1738 * rack_drop_checks returns 1 for you should not proceed. It places 1739 * in ret_val what should be returned 1/0 by the caller. The 1 indicates 1740 * that the TCB is unlocked and probably dropped. The 0 indicates the 1741 * TCB is still valid and locked. 1742 */ 1743 static int 1744 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) 1745 { 1746 int32_t todrop; 1747 int32_t thflags; 1748 int32_t tlen; 1749 1750 thflags = *thf; 1751 tlen = *tlenp; 1752 todrop = tp->rcv_nxt - th->th_seq; 1753 if (todrop > 0) { 1754 if (thflags & TH_SYN) { 1755 thflags &= ~TH_SYN; 1756 th->th_seq++; 1757 if (th->th_urp > 1) 1758 th->th_urp--; 1759 else 1760 thflags &= ~TH_URG; 1761 todrop--; 1762 } 1763 /* 1764 * Following if statement from Stevens, vol. 2, p. 960. 1765 */ 1766 if (todrop > tlen 1767 || (todrop == tlen && (thflags & TH_FIN) == 0)) { 1768 /* 1769 * Any valid FIN must be to the left of the window. 1770 * At this point the FIN must be a duplicate or out 1771 * of sequence; drop it. 1772 */ 1773 thflags &= ~TH_FIN; 1774 /* 1775 * Send an ACK to resynchronize and drop any data. 1776 * But keep on processing for RST or ACK. 1777 */ 1778 tp->t_flags |= TF_ACKNOW; 1779 todrop = tlen; 1780 TCPSTAT_INC(tcps_rcvduppack); 1781 TCPSTAT_ADD(tcps_rcvdupbyte, todrop); 1782 } else { 1783 TCPSTAT_INC(tcps_rcvpartduppack); 1784 TCPSTAT_ADD(tcps_rcvpartdupbyte, todrop); 1785 } 1786 if (tp->t_flags & TF_SACK_PERMIT) { 1787 /* 1788 * record the left, to-be-dropped edge of data 1789 * here, for use as dsack block further down 1790 */ 1791 tcp_update_sack_list(tp, th->th_seq, 1792 th->th_seq + todrop); 1793 } 1794 *drop_hdrlen += todrop; /* drop from the top afterwards */ 1795 th->th_seq += todrop; 1796 tlen -= todrop; 1797 if (th->th_urp > todrop) 1798 th->th_urp -= todrop; 1799 else { 1800 thflags &= ~TH_URG; 1801 th->th_urp = 0; 1802 } 1803 } 1804 /* 1805 * If segment ends after window, drop trailing data (and PUSH and 1806 * FIN); if nothing left, just ACK. 1807 */ 1808 todrop = (th->th_seq + tlen) - (tp->rcv_nxt + tp->rcv_wnd); 1809 if (todrop > 0) { 1810 TCPSTAT_INC(tcps_rcvpackafterwin); 1811 if (todrop >= tlen) { 1812 TCPSTAT_ADD(tcps_rcvbyteafterwin, tlen); 1813 /* 1814 * If window is closed can only take segments at 1815 * window edge, and have to drop data and PUSH from 1816 * incoming segments. Continue processing, but 1817 * remember to ack. Otherwise, drop segment and 1818 * ack. 1819 */ 1820 if (tp->rcv_wnd == 0 && th->th_seq == tp->rcv_nxt) { 1821 tp->t_flags |= TF_ACKNOW; 1822 TCPSTAT_INC(tcps_rcvwinprobe); 1823 } else { 1824 rack_do_dropafterack(m, tp, th, thflags, tlen, ret_val); 1825 return (1); 1826 } 1827 } else 1828 TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop); 1829 m_adj(m, -todrop); 1830 tlen -= todrop; 1831 thflags &= ~(TH_PUSH | TH_FIN); 1832 } 1833 *thf = thflags; 1834 *tlenp = tlen; 1835 return (0); 1836 } 1837 1838 static struct rack_sendmap * 1839 rack_find_lowest_rsm(struct tcp_rack *rack) 1840 { 1841 struct rack_sendmap *rsm; 1842 1843 /* 1844 * Walk the time-order transmitted list looking for an rsm that is 1845 * not acked. This will be the one that was sent the longest time 1846 * ago that is still outstanding. 1847 */ 1848 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) { 1849 if (rsm->r_flags & RACK_ACKED) { 1850 continue; 1851 } 1852 goto finish; 1853 } 1854 finish: 1855 return (rsm); 1856 } 1857 1858 static struct rack_sendmap * 1859 rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm) 1860 { 1861 struct rack_sendmap *prsm; 1862 1863 /* 1864 * Walk the sequence order list backward until we hit and arrive at 1865 * the highest seq not acked. In theory when this is called it 1866 * should be the last segment (which it was not). 1867 */ 1868 counter_u64_add(rack_find_high, 1); 1869 prsm = rsm; 1870 TAILQ_FOREACH_REVERSE_FROM(prsm, &rack->r_ctl.rc_map, rack_head, r_next) { 1871 if (prsm->r_flags & (RACK_ACKED | RACK_HAS_FIN)) { 1872 continue; 1873 } 1874 return (prsm); 1875 } 1876 return (NULL); 1877 } 1878 1879 1880 static uint32_t 1881 rack_calc_thresh_rack(struct tcp_rack *rack, uint32_t srtt, uint32_t cts) 1882 { 1883 int32_t lro; 1884 uint32_t thresh; 1885 1886 /* 1887 * lro is the flag we use to determine if we have seen reordering. 1888 * If it gets set we have seen reordering. The reorder logic either 1889 * works in one of two ways: 1890 * 1891 * If reorder-fade is configured, then we track the last time we saw 1892 * re-ordering occur. If we reach the point where enough time as 1893 * passed we no longer consider reordering has occuring. 1894 * 1895 * Or if reorder-face is 0, then once we see reordering we consider 1896 * the connection to alway be subject to reordering and just set lro 1897 * to 1. 1898 * 1899 * In the end if lro is non-zero we add the extra time for 1900 * reordering in. 1901 */ 1902 if (srtt == 0) 1903 srtt = 1; 1904 if (rack->r_ctl.rc_reorder_ts) { 1905 if (rack->r_ctl.rc_reorder_fade) { 1906 if (SEQ_GEQ(cts, rack->r_ctl.rc_reorder_ts)) { 1907 lro = cts - rack->r_ctl.rc_reorder_ts; 1908 if (lro == 0) { 1909 /* 1910 * No time as passed since the last 1911 * reorder, mark it as reordering. 1912 */ 1913 lro = 1; 1914 } 1915 } else { 1916 /* Negative time? */ 1917 lro = 0; 1918 } 1919 if (lro > rack->r_ctl.rc_reorder_fade) { 1920 /* Turn off reordering seen too */ 1921 rack->r_ctl.rc_reorder_ts = 0; 1922 lro = 0; 1923 } 1924 } else { 1925 /* Reodering does not fade */ 1926 lro = 1; 1927 } 1928 } else { 1929 lro = 0; 1930 } 1931 thresh = srtt + rack->r_ctl.rc_pkt_delay; 1932 if (lro) { 1933 /* It must be set, if not you get 1/4 rtt */ 1934 if (rack->r_ctl.rc_reorder_shift) 1935 thresh += (srtt >> rack->r_ctl.rc_reorder_shift); 1936 else 1937 thresh += (srtt >> 2); 1938 } else { 1939 thresh += 1; 1940 } 1941 /* We don't let the rack timeout be above a RTO */ 1942 1943 if (thresh > TICKS_2_MSEC(rack->rc_tp->t_rxtcur)) { 1944 thresh = TICKS_2_MSEC(rack->rc_tp->t_rxtcur); 1945 } 1946 /* And we don't want it above the RTO max either */ 1947 if (thresh > rack_rto_max) { 1948 thresh = rack_rto_max; 1949 } 1950 return (thresh); 1951 } 1952 1953 static uint32_t 1954 rack_calc_thresh_tlp(struct tcpcb *tp, struct tcp_rack *rack, 1955 struct rack_sendmap *rsm, uint32_t srtt) 1956 { 1957 struct rack_sendmap *prsm; 1958 uint32_t thresh, len; 1959 int maxseg; 1960 1961 if (srtt == 0) 1962 srtt = 1; 1963 if (rack->r_ctl.rc_tlp_threshold) 1964 thresh = srtt + (srtt / rack->r_ctl.rc_tlp_threshold); 1965 else 1966 thresh = (srtt * 2); 1967 1968 /* Get the previous sent packet, if any */ 1969 maxseg = tcp_maxseg(tp); 1970 counter_u64_add(rack_enter_tlp_calc, 1); 1971 len = rsm->r_end - rsm->r_start; 1972 if (rack->rack_tlp_threshold_use == TLP_USE_ID) { 1973 /* Exactly like the ID */ 1974 if (((tp->snd_max - tp->snd_una) - rack->r_ctl.rc_sacked + rack->r_ctl.rc_holes_rxt) <= maxseg) { 1975 uint32_t alt_thresh; 1976 /* 1977 * Compensate for delayed-ack with the d-ack time. 1978 */ 1979 counter_u64_add(rack_used_tlpmethod, 1); 1980 alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time; 1981 if (alt_thresh > thresh) 1982 thresh = alt_thresh; 1983 } 1984 } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_ONE) { 1985 /* 2.1 behavior */ 1986 prsm = TAILQ_PREV(rsm, rack_head, r_tnext); 1987 if (prsm && (len <= maxseg)) { 1988 /* 1989 * Two packets outstanding, thresh should be (2*srtt) + 1990 * possible inter-packet delay (if any). 1991 */ 1992 uint32_t inter_gap = 0; 1993 int idx, nidx; 1994 1995 counter_u64_add(rack_used_tlpmethod, 1); 1996 idx = rsm->r_rtr_cnt - 1; 1997 nidx = prsm->r_rtr_cnt - 1; 1998 if (TSTMP_GEQ(rsm->r_tim_lastsent[nidx], prsm->r_tim_lastsent[idx])) { 1999 /* Yes it was sent later (or at the same time) */ 2000 inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx]; 2001 } 2002 thresh += inter_gap; 2003 } else if (len <= maxseg) { 2004 /* 2005 * Possibly compensate for delayed-ack. 2006 */ 2007 uint32_t alt_thresh; 2008 2009 counter_u64_add(rack_used_tlpmethod2, 1); 2010 alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time; 2011 if (alt_thresh > thresh) 2012 thresh = alt_thresh; 2013 } 2014 } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_TWO) { 2015 /* 2.2 behavior */ 2016 if (len <= maxseg) { 2017 uint32_t alt_thresh; 2018 /* 2019 * Compensate for delayed-ack with the d-ack time. 2020 */ 2021 counter_u64_add(rack_used_tlpmethod, 1); 2022 alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time; 2023 if (alt_thresh > thresh) 2024 thresh = alt_thresh; 2025 } 2026 } 2027 /* Not above an RTO */ 2028 if (thresh > TICKS_2_MSEC(tp->t_rxtcur)) { 2029 thresh = TICKS_2_MSEC(tp->t_rxtcur); 2030 } 2031 /* Not above a RTO max */ 2032 if (thresh > rack_rto_max) { 2033 thresh = rack_rto_max; 2034 } 2035 /* Apply user supplied min TLP */ 2036 if (thresh < rack_tlp_min) { 2037 thresh = rack_tlp_min; 2038 } 2039 return (thresh); 2040 } 2041 2042 static struct rack_sendmap * 2043 rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused) 2044 { 2045 /* 2046 * Check to see that we don't need to fall into recovery. We will 2047 * need to do so if our oldest transmit is past the time we should 2048 * have had an ack. 2049 */ 2050 struct tcp_rack *rack; 2051 struct rack_sendmap *rsm; 2052 int32_t idx; 2053 uint32_t srtt_cur, srtt, thresh; 2054 2055 rack = (struct tcp_rack *)tp->t_fb_ptr; 2056 if (TAILQ_EMPTY(&rack->r_ctl.rc_map)) { 2057 return (NULL); 2058 } 2059 srtt_cur = tp->t_srtt >> TCP_RTT_SHIFT; 2060 srtt = TICKS_2_MSEC(srtt_cur); 2061 if (rack->rc_rack_rtt && (srtt > rack->rc_rack_rtt)) 2062 srtt = rack->rc_rack_rtt; 2063 2064 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); 2065 if (rsm == NULL) 2066 return (NULL); 2067 2068 if (rsm->r_flags & RACK_ACKED) { 2069 rsm = rack_find_lowest_rsm(rack); 2070 if (rsm == NULL) 2071 return (NULL); 2072 } 2073 idx = rsm->r_rtr_cnt - 1; 2074 thresh = rack_calc_thresh_rack(rack, srtt, tsused); 2075 if (tsused < rsm->r_tim_lastsent[idx]) { 2076 return (NULL); 2077 } 2078 if ((tsused - rsm->r_tim_lastsent[idx]) < thresh) { 2079 return (NULL); 2080 } 2081 /* Ok if we reach here we are over-due */ 2082 rack->r_ctl.rc_rsm_start = rsm->r_start; 2083 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd; 2084 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh; 2085 rack_cong_signal(tp, NULL, CC_NDUPACK); 2086 return (rsm); 2087 } 2088 2089 static uint32_t 2090 rack_get_persists_timer_val(struct tcpcb *tp, struct tcp_rack *rack) 2091 { 2092 int32_t t; 2093 int32_t tt; 2094 uint32_t ret_val; 2095 2096 t = TICKS_2_MSEC((tp->t_srtt >> TCP_RTT_SHIFT) + ((tp->t_rttvar * 4) >> TCP_RTT_SHIFT)); 2097 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], 2098 tcp_persmin, tcp_persmax); 2099 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 2100 tp->t_rxtshift++; 2101 rack->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT; 2102 ret_val = (uint32_t)tt; 2103 return (ret_val); 2104 } 2105 2106 static uint32_t 2107 rack_timer_start(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) 2108 { 2109 /* 2110 * Start the FR timer, we do this based on getting the first one in 2111 * the rc_tmap. Note that if its NULL we must stop the timer. in all 2112 * events we need to stop the running timer (if its running) before 2113 * starting the new one. 2114 */ 2115 uint32_t thresh, exp, to, srtt, time_since_sent; 2116 uint32_t srtt_cur; 2117 int32_t idx; 2118 int32_t is_tlp_timer = 0; 2119 struct rack_sendmap *rsm; 2120 2121 if (rack->t_timers_stopped) { 2122 /* All timers have been stopped none are to run */ 2123 return (0); 2124 } 2125 if (rack->rc_in_persist) { 2126 /* We can't start any timer in persists */ 2127 return (rack_get_persists_timer_val(tp, rack)); 2128 } 2129 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); 2130 if (rsm == NULL) { 2131 /* Nothing on the send map */ 2132 activate_rxt: 2133 if (SEQ_LT(tp->snd_una, tp->snd_max) || sbavail(&(tp->t_inpcb->inp_socket->so_snd))) { 2134 rack->r_ctl.rc_hpts_flags |= PACE_TMR_RXT; 2135 to = TICKS_2_MSEC(tp->t_rxtcur); 2136 if (to == 0) 2137 to = 1; 2138 return (to); 2139 } 2140 return (0); 2141 } 2142 if (rsm->r_flags & RACK_ACKED) { 2143 rsm = rack_find_lowest_rsm(rack); 2144 if (rsm == NULL) { 2145 /* No lowest? */ 2146 goto activate_rxt; 2147 } 2148 } 2149 /* Convert from ms to usecs */ 2150 if (rsm->r_flags & RACK_SACK_PASSED) { 2151 if ((tp->t_flags & TF_SENTFIN) && 2152 ((tp->snd_max - tp->snd_una) == 1) && 2153 (rsm->r_flags & RACK_HAS_FIN)) { 2154 /* 2155 * We don't start a rack timer if all we have is a 2156 * FIN outstanding. 2157 */ 2158 goto activate_rxt; 2159 } 2160 if (tp->t_srtt) { 2161 srtt_cur = (tp->t_srtt >> TCP_RTT_SHIFT); 2162 srtt = TICKS_2_MSEC(srtt_cur); 2163 } else 2164 srtt = RACK_INITIAL_RTO; 2165 2166 thresh = rack_calc_thresh_rack(rack, srtt, cts); 2167 idx = rsm->r_rtr_cnt - 1; 2168 exp = rsm->r_tim_lastsent[idx] + thresh; 2169 if (SEQ_GEQ(exp, cts)) { 2170 to = exp - cts; 2171 if (to < rack->r_ctl.rc_min_to) { 2172 to = rack->r_ctl.rc_min_to; 2173 } 2174 } else { 2175 to = rack->r_ctl.rc_min_to; 2176 } 2177 } else { 2178 /* Ok we need to do a TLP not RACK */ 2179 if ((rack->rc_tlp_in_progress != 0) || 2180 (rack->r_ctl.rc_tlp_rtx_out != 0)) { 2181 /* 2182 * The previous send was a TLP or a tlp_rtx is in 2183 * process. 2184 */ 2185 goto activate_rxt; 2186 } 2187 if ((tp->snd_max - tp->snd_una) > tp->snd_wnd) { 2188 /* 2189 * Peer collapsed rwnd, don't do TLP. 2190 */ 2191 goto activate_rxt; 2192 } 2193 rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext); 2194 if (rsm == NULL) { 2195 /* We found no rsm to TLP with. */ 2196 goto activate_rxt; 2197 } 2198 if (rsm->r_flags & RACK_HAS_FIN) { 2199 /* If its a FIN we dont do TLP */ 2200 rsm = NULL; 2201 goto activate_rxt; 2202 } 2203 idx = rsm->r_rtr_cnt - 1; 2204 if (TSTMP_GT(cts, rsm->r_tim_lastsent[idx])) 2205 time_since_sent = cts - rsm->r_tim_lastsent[idx]; 2206 else 2207 time_since_sent = 0; 2208 is_tlp_timer = 1; 2209 if (tp->t_srtt) { 2210 srtt_cur = (tp->t_srtt >> TCP_RTT_SHIFT); 2211 srtt = TICKS_2_MSEC(srtt_cur); 2212 } else 2213 srtt = RACK_INITIAL_RTO; 2214 thresh = rack_calc_thresh_tlp(tp, rack, rsm, srtt); 2215 if (thresh > time_since_sent) 2216 to = thresh - time_since_sent; 2217 else 2218 to = rack->r_ctl.rc_min_to; 2219 if (to > TCPTV_REXMTMAX) { 2220 /* 2221 * If the TLP time works out to larger than the max 2222 * RTO lets not do TLP.. just RTO. 2223 */ 2224 goto activate_rxt; 2225 } 2226 if (rsm->r_start != rack->r_ctl.rc_last_tlp_seq) { 2227 /* 2228 * The tail is no longer the last one I did a probe 2229 * on 2230 */ 2231 rack->r_ctl.rc_tlp_seg_send_cnt = 0; 2232 rack->r_ctl.rc_last_tlp_seq = rsm->r_start; 2233 } 2234 } 2235 if (is_tlp_timer == 0) { 2236 rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK; 2237 } else { 2238 if ((rack->r_ctl.rc_tlp_send_cnt > rack_tlp_max_resend) || 2239 (rack->r_ctl.rc_tlp_seg_send_cnt > rack_tlp_max_resend)) { 2240 /* 2241 * We have exceeded how many times we can retran the 2242 * current TLP timer, switch to the RTO timer. 2243 */ 2244 goto activate_rxt; 2245 } else { 2246 rack->r_ctl.rc_hpts_flags |= PACE_TMR_TLP; 2247 } 2248 } 2249 if (to == 0) 2250 to = 1; 2251 return (to); 2252 } 2253 2254 static void 2255 rack_enter_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) 2256 { 2257 if (rack->rc_in_persist == 0) { 2258 if (((tp->t_flags & TF_SENTFIN) == 0) && 2259 (tp->snd_max - tp->snd_una) >= sbavail(&rack->rc_inp->inp_socket->so_snd)) 2260 /* Must need to send more data to enter persist */ 2261 return; 2262 rack->r_ctl.rc_went_idle_time = cts; 2263 rack_timer_cancel(tp, rack, cts, __LINE__); 2264 tp->t_rxtshift = 0; 2265 rack->rc_in_persist = 1; 2266 } 2267 } 2268 2269 static void 2270 rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack) 2271 { 2272 if (rack->rc_inp->inp_in_hpts) { 2273 tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT); 2274 rack->r_ctl.rc_hpts_flags = 0; 2275 } 2276 rack->rc_in_persist = 0; 2277 rack->r_ctl.rc_went_idle_time = 0; 2278 tp->t_flags &= ~TF_FORCEDATA; 2279 tp->t_rxtshift = 0; 2280 } 2281 2282 static void 2283 rack_start_hpts_timer(struct tcp_rack *rack, struct tcpcb *tp, uint32_t cts, int32_t line, 2284 int32_t slot, uint32_t tot_len_this_send, int32_t frm_out_sbavail) 2285 { 2286 struct inpcb *inp; 2287 uint32_t delayed_ack = 0; 2288 uint32_t hpts_timeout; 2289 uint8_t stopped; 2290 uint32_t left = 0; 2291 2292 inp = tp->t_inpcb; 2293 if (inp->inp_in_hpts) { 2294 /* A previous call is already set up */ 2295 return; 2296 } 2297 2298 if ((tp->t_state == TCPS_CLOSED) || 2299 (tp->t_state == TCPS_LISTEN)) { 2300 return; 2301 } 2302 stopped = rack->rc_tmr_stopped; 2303 if (stopped && TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) { 2304 left = rack->r_ctl.rc_timer_exp - cts; 2305 } 2306 rack->r_ctl.rc_timer_exp = 0; 2307 if (rack->rc_inp->inp_in_hpts == 0) { 2308 rack->r_ctl.rc_hpts_flags = 0; 2309 } 2310 if (slot) { 2311 /* We are hptsi too */ 2312 rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT; 2313 } else if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) { 2314 /* 2315 * We are still left on the hpts when the to goes 2316 * it will be for output. 2317 */ 2318 if (TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) 2319 slot = rack->r_ctl.rc_last_output_to - cts; 2320 else 2321 slot = 1; 2322 } 2323 if ((tp->snd_wnd == 0) && TCPS_HAVEESTABLISHED(tp->t_state)) { 2324 /* No send window.. we must enter persist */ 2325 rack_enter_persist(tp, rack, cts); 2326 } else if ((frm_out_sbavail && 2327 (frm_out_sbavail > (tp->snd_max - tp->snd_una)) && 2328 (tp->snd_wnd < tp->t_maxseg)) && 2329 TCPS_HAVEESTABLISHED(tp->t_state)) { 2330 /* 2331 * If we have no window or we can't send a segment (and have 2332 * data to send.. we cheat here and frm_out_sbavail is 2333 * passed in with the sbavail(sb) only from bbr_output) and 2334 * we are established, then we must enter persits (if not 2335 * already in persits). 2336 */ 2337 rack_enter_persist(tp, rack, cts); 2338 } 2339 hpts_timeout = rack_timer_start(tp, rack, cts); 2340 if (tp->t_flags & TF_DELACK) { 2341 delayed_ack = tcp_delacktime; 2342 rack->r_ctl.rc_hpts_flags |= PACE_TMR_DELACK; 2343 } 2344 if (delayed_ack && ((hpts_timeout == 0) || 2345 (delayed_ack < hpts_timeout))) 2346 hpts_timeout = delayed_ack; 2347 else 2348 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK; 2349 /* 2350 * If no timers are going to run and we will fall off the hptsi 2351 * wheel, we resort to a keep-alive timer if its configured. 2352 */ 2353 if ((hpts_timeout == 0) && 2354 (slot == 0)) { 2355 if ((tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && 2356 (tp->t_state <= TCPS_CLOSING)) { 2357 /* 2358 * Ok we have no timer (persists, rack, tlp, rxt or 2359 * del-ack), we don't have segments being paced. So 2360 * all that is left is the keepalive timer. 2361 */ 2362 if (TCPS_HAVEESTABLISHED(tp->t_state)) { 2363 /* Get the established keep-alive time */ 2364 hpts_timeout = TP_KEEPIDLE(tp); 2365 } else { 2366 /* Get the initial setup keep-alive time */ 2367 hpts_timeout = TP_KEEPINIT(tp); 2368 } 2369 rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP; 2370 } 2371 } 2372 if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) == 2373 (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) { 2374 /* 2375 * RACK, TLP, persists and RXT timers all are restartable 2376 * based on actions input .. i.e we received a packet (ack 2377 * or sack) and that changes things (rw, or snd_una etc). 2378 * Thus we can restart them with a new value. For 2379 * keep-alive, delayed_ack we keep track of what was left 2380 * and restart the timer with a smaller value. 2381 */ 2382 if (left < hpts_timeout) 2383 hpts_timeout = left; 2384 } 2385 if (hpts_timeout) { 2386 /* 2387 * Hack alert for now we can't time-out over 2,147,483 2388 * seconds (a bit more than 596 hours), which is probably ok 2389 * :). 2390 */ 2391 if (hpts_timeout > 0x7ffffffe) 2392 hpts_timeout = 0x7ffffffe; 2393 rack->r_ctl.rc_timer_exp = cts + hpts_timeout; 2394 } 2395 if (slot) { 2396 rack->r_ctl.rc_last_output_to = cts + slot; 2397 if ((hpts_timeout == 0) || (hpts_timeout > slot)) { 2398 if (rack->rc_inp->inp_in_hpts == 0) 2399 tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(slot)); 2400 rack_log_to_start(rack, cts, hpts_timeout, slot, 1); 2401 } else { 2402 /* 2403 * Arrange for the hpts to kick back in after the 2404 * t-o if the t-o does not cause a send. 2405 */ 2406 if (rack->rc_inp->inp_in_hpts == 0) 2407 tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(hpts_timeout)); 2408 rack_log_to_start(rack, cts, hpts_timeout, slot, 0); 2409 } 2410 } else if (hpts_timeout) { 2411 if (rack->rc_inp->inp_in_hpts == 0) 2412 tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(hpts_timeout)); 2413 rack_log_to_start(rack, cts, hpts_timeout, slot, 0); 2414 } else { 2415 /* No timer starting */ 2416 #ifdef INVARIANTS 2417 if (SEQ_GT(tp->snd_max, tp->snd_una)) { 2418 panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?", 2419 tp, rack, tot_len_this_send, cts, slot, hpts_timeout); 2420 } 2421 #endif 2422 } 2423 rack->rc_tmr_stopped = 0; 2424 if (slot) 2425 rack_log_type_bbrsnd(rack, tot_len_this_send, slot, cts); 2426 } 2427 2428 /* 2429 * RACK Timer, here we simply do logging and house keeping. 2430 * the normal rack_output() function will call the 2431 * appropriate thing to check if we need to do a RACK retransmit. 2432 * We return 1, saying don't proceed with rack_output only 2433 * when all timers have been stopped (destroyed PCB?). 2434 */ 2435 static int 2436 rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) 2437 { 2438 /* 2439 * This timer simply provides an internal trigger to send out data. 2440 * The check_recovery_mode call will see if there are needed 2441 * retransmissions, if so we will enter fast-recovery. The output 2442 * call may or may not do the same thing depending on sysctl 2443 * settings. 2444 */ 2445 struct rack_sendmap *rsm; 2446 int32_t recovery; 2447 2448 if (tp->t_timers->tt_flags & TT_STOPPED) { 2449 return (1); 2450 } 2451 if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) { 2452 /* Its not time yet */ 2453 return (0); 2454 } 2455 rack_log_to_event(rack, RACK_TO_FRM_RACK); 2456 recovery = IN_RECOVERY(tp->t_flags); 2457 counter_u64_add(rack_to_tot, 1); 2458 if (rack->r_state && (rack->r_state != tp->t_state)) 2459 rack_set_state(tp, rack); 2460 rsm = rack_check_recovery_mode(tp, cts); 2461 if (rsm) { 2462 uint32_t rtt; 2463 2464 rtt = rack->rc_rack_rtt; 2465 if (rtt == 0) 2466 rtt = 1; 2467 if ((recovery == 0) && 2468 (rack->r_ctl.rc_prr_sndcnt < tp->t_maxseg)) { 2469 /* 2470 * The rack-timeout that enter's us into recovery 2471 * will force out one MSS and set us up so that we 2472 * can do one more send in 2*rtt (transitioning the 2473 * rack timeout into a rack-tlp). 2474 */ 2475 rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; 2476 } else if ((rack->r_ctl.rc_prr_sndcnt < tp->t_maxseg) && 2477 ((rsm->r_end - rsm->r_start) > rack->r_ctl.rc_prr_sndcnt)) { 2478 /* 2479 * When a rack timer goes, we have to send at 2480 * least one segment. They will be paced a min of 1ms 2481 * apart via the next rack timer (or further 2482 * if the rack timer dictates it). 2483 */ 2484 rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; 2485 } 2486 } else { 2487 /* This is a case that should happen rarely if ever */ 2488 counter_u64_add(rack_tlp_does_nada, 1); 2489 #ifdef TCP_BLACKBOX 2490 tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true); 2491 #endif 2492 rack->r_ctl.rc_resend = TAILQ_FIRST(&rack->r_ctl.rc_tmap); 2493 } 2494 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK; 2495 return (0); 2496 } 2497 2498 static struct rack_sendmap * 2499 rack_merge_rsm(struct tcp_rack *rack, 2500 struct rack_sendmap *l_rsm, 2501 struct rack_sendmap *r_rsm) 2502 { 2503 /* 2504 * We are merging two ack'd RSM's, 2505 * the l_rsm is on the left (lower seq 2506 * values) and the r_rsm is on the right 2507 * (higher seq value). The simplest way 2508 * to merge these is to move the right 2509 * one into the left. I don't think there 2510 * is any reason we need to try to find 2511 * the oldest (or last oldest retransmitted). 2512 */ 2513 l_rsm->r_end = r_rsm->r_end; 2514 if (r_rsm->r_rtr_bytes) 2515 l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes; 2516 if (r_rsm->r_in_tmap) { 2517 /* This really should not happen */ 2518 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, r_rsm, r_tnext); 2519 } 2520 /* Now the flags */ 2521 if (r_rsm->r_flags & RACK_HAS_FIN) 2522 l_rsm->r_flags |= RACK_HAS_FIN; 2523 if (r_rsm->r_flags & RACK_TLP) 2524 l_rsm->r_flags |= RACK_TLP; 2525 TAILQ_REMOVE(&rack->r_ctl.rc_map, r_rsm, r_next); 2526 if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) { 2527 /* Transfer the split limit to the map we free */ 2528 r_rsm->r_limit_type = l_rsm->r_limit_type; 2529 l_rsm->r_limit_type = 0; 2530 } 2531 rack_free(rack, r_rsm); 2532 return(l_rsm); 2533 } 2534 2535 /* 2536 * TLP Timer, here we simply setup what segment we want to 2537 * have the TLP expire on, the normal rack_output() will then 2538 * send it out. 2539 * 2540 * We return 1, saying don't proceed with rack_output only 2541 * when all timers have been stopped (destroyed PCB?). 2542 */ 2543 static int 2544 rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) 2545 { 2546 /* 2547 * Tail Loss Probe. 2548 */ 2549 struct rack_sendmap *rsm = NULL; 2550 struct socket *so; 2551 uint32_t amm, old_prr_snd = 0; 2552 uint32_t out, avail; 2553 2554 if (tp->t_timers->tt_flags & TT_STOPPED) { 2555 return (1); 2556 } 2557 if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) { 2558 /* Its not time yet */ 2559 return (0); 2560 } 2561 if (rack_progress_timeout_check(tp)) { 2562 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); 2563 return (1); 2564 } 2565 /* 2566 * A TLP timer has expired. We have been idle for 2 rtts. So we now 2567 * need to figure out how to force a full MSS segment out. 2568 */ 2569 rack_log_to_event(rack, RACK_TO_FRM_TLP); 2570 counter_u64_add(rack_tlp_tot, 1); 2571 if (rack->r_state && (rack->r_state != tp->t_state)) 2572 rack_set_state(tp, rack); 2573 so = tp->t_inpcb->inp_socket; 2574 avail = sbavail(&so->so_snd); 2575 out = tp->snd_max - tp->snd_una; 2576 rack->rc_timer_up = 1; 2577 /* 2578 * If we are in recovery we can jazz out a segment if new data is 2579 * present simply by setting rc_prr_sndcnt to a segment. 2580 */ 2581 if ((avail > out) && 2582 ((rack_always_send_oldest == 0) || (TAILQ_EMPTY(&rack->r_ctl.rc_tmap)))) { 2583 /* New data is available */ 2584 amm = avail - out; 2585 if (amm > tp->t_maxseg) { 2586 amm = tp->t_maxseg; 2587 } else if ((amm < tp->t_maxseg) && ((tp->t_flags & TF_NODELAY) == 0)) { 2588 /* not enough to fill a MTU and no-delay is off */ 2589 goto need_retran; 2590 } 2591 if (IN_RECOVERY(tp->t_flags)) { 2592 /* Unlikely */ 2593 old_prr_snd = rack->r_ctl.rc_prr_sndcnt; 2594 if (out + amm <= tp->snd_wnd) 2595 rack->r_ctl.rc_prr_sndcnt = amm; 2596 else 2597 goto need_retran; 2598 } else { 2599 /* Set the send-new override */ 2600 if (out + amm <= tp->snd_wnd) 2601 rack->r_ctl.rc_tlp_new_data = amm; 2602 else 2603 goto need_retran; 2604 } 2605 rack->r_ctl.rc_tlp_seg_send_cnt = 0; 2606 rack->r_ctl.rc_last_tlp_seq = tp->snd_max; 2607 rack->r_ctl.rc_tlpsend = NULL; 2608 counter_u64_add(rack_tlp_newdata, 1); 2609 goto send; 2610 } 2611 need_retran: 2612 /* 2613 * Ok we need to arrange the last un-acked segment to be re-sent, or 2614 * optionally the first un-acked segment. 2615 */ 2616 if (rack_always_send_oldest) 2617 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); 2618 else { 2619 rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_map, rack_sendmap, r_next); 2620 if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) { 2621 rsm = rack_find_high_nonack(rack, rsm); 2622 } 2623 } 2624 if (rsm == NULL) { 2625 counter_u64_add(rack_tlp_does_nada, 1); 2626 #ifdef TCP_BLACKBOX 2627 tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true); 2628 #endif 2629 goto out; 2630 } 2631 if ((rsm->r_end - rsm->r_start) > tp->t_maxseg) { 2632 /* 2633 * We need to split this the last segment in two. 2634 */ 2635 int32_t idx; 2636 struct rack_sendmap *nrsm; 2637 2638 nrsm = rack_alloc_full_limit(rack); 2639 if (nrsm == NULL) { 2640 /* 2641 * No memory to split, we will just exit and punt 2642 * off to the RXT timer. 2643 */ 2644 counter_u64_add(rack_tlp_does_nada, 1); 2645 goto out; 2646 } 2647 nrsm->r_start = (rsm->r_end - tp->t_maxseg); 2648 nrsm->r_end = rsm->r_end; 2649 nrsm->r_rtr_cnt = rsm->r_rtr_cnt; 2650 nrsm->r_flags = rsm->r_flags; 2651 nrsm->r_sndcnt = rsm->r_sndcnt; 2652 nrsm->r_rtr_bytes = 0; 2653 rsm->r_end = nrsm->r_start; 2654 for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { 2655 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; 2656 } 2657 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_map, rsm, nrsm, r_next); 2658 if (rsm->r_in_tmap) { 2659 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 2660 nrsm->r_in_tmap = 1; 2661 } 2662 rsm->r_flags &= (~RACK_HAS_FIN); 2663 rsm = nrsm; 2664 } 2665 rack->r_ctl.rc_tlpsend = rsm; 2666 rack->r_ctl.rc_tlp_rtx_out = 1; 2667 if (rsm->r_start == rack->r_ctl.rc_last_tlp_seq) { 2668 rack->r_ctl.rc_tlp_seg_send_cnt++; 2669 tp->t_rxtshift++; 2670 } else { 2671 rack->r_ctl.rc_last_tlp_seq = rsm->r_start; 2672 rack->r_ctl.rc_tlp_seg_send_cnt = 1; 2673 } 2674 send: 2675 rack->r_ctl.rc_tlp_send_cnt++; 2676 if (rack->r_ctl.rc_tlp_send_cnt > rack_tlp_max_resend) { 2677 /* 2678 * Can't [re]/transmit a segment we have not heard from the 2679 * peer in max times. We need the retransmit timer to take 2680 * over. 2681 */ 2682 restore: 2683 rack->r_ctl.rc_tlpsend = NULL; 2684 if (rsm) 2685 rsm->r_flags &= ~RACK_TLP; 2686 rack->r_ctl.rc_prr_sndcnt = old_prr_snd; 2687 counter_u64_add(rack_tlp_retran_fail, 1); 2688 goto out; 2689 } else if (rsm) { 2690 rsm->r_flags |= RACK_TLP; 2691 } 2692 if (rsm && (rsm->r_start == rack->r_ctl.rc_last_tlp_seq) && 2693 (rack->r_ctl.rc_tlp_seg_send_cnt > rack_tlp_max_resend)) { 2694 /* 2695 * We don't want to send a single segment more than the max 2696 * either. 2697 */ 2698 goto restore; 2699 } 2700 rack->r_timer_override = 1; 2701 rack->r_tlp_running = 1; 2702 rack->rc_tlp_in_progress = 1; 2703 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP; 2704 return (0); 2705 out: 2706 rack->rc_timer_up = 0; 2707 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP; 2708 return (0); 2709 } 2710 2711 /* 2712 * Delayed ack Timer, here we simply need to setup the 2713 * ACK_NOW flag and remove the DELACK flag. From there 2714 * the output routine will send the ack out. 2715 * 2716 * We only return 1, saying don't proceed, if all timers 2717 * are stopped (destroyed PCB?). 2718 */ 2719 static int 2720 rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) 2721 { 2722 if (tp->t_timers->tt_flags & TT_STOPPED) { 2723 return (1); 2724 } 2725 rack_log_to_event(rack, RACK_TO_FRM_DELACK); 2726 tp->t_flags &= ~TF_DELACK; 2727 tp->t_flags |= TF_ACKNOW; 2728 TCPSTAT_INC(tcps_delack); 2729 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK; 2730 return (0); 2731 } 2732 2733 /* 2734 * Persists timer, here we simply need to setup the 2735 * FORCE-DATA flag the output routine will send 2736 * the one byte send. 2737 * 2738 * We only return 1, saying don't proceed, if all timers 2739 * are stopped (destroyed PCB?). 2740 */ 2741 static int 2742 rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) 2743 { 2744 struct inpcb *inp; 2745 int32_t retval = 0; 2746 2747 inp = tp->t_inpcb; 2748 2749 if (tp->t_timers->tt_flags & TT_STOPPED) { 2750 return (1); 2751 } 2752 if (rack->rc_in_persist == 0) 2753 return (0); 2754 if (rack_progress_timeout_check(tp)) { 2755 tcp_set_inp_to_drop(inp, ETIMEDOUT); 2756 return (1); 2757 } 2758 KASSERT(inp != NULL, ("%s: tp %p tp->t_inpcb == NULL", __func__, tp)); 2759 /* 2760 * Persistence timer into zero window. Force a byte to be output, if 2761 * possible. 2762 */ 2763 TCPSTAT_INC(tcps_persisttimeo); 2764 /* 2765 * Hack: if the peer is dead/unreachable, we do not time out if the 2766 * window is closed. After a full backoff, drop the connection if 2767 * the idle time (no responses to probes) reaches the maximum 2768 * backoff that we would use if retransmitting. 2769 */ 2770 if (tp->t_rxtshift == TCP_MAXRXTSHIFT && 2771 (ticks - tp->t_rcvtime >= tcp_maxpersistidle || 2772 ticks - tp->t_rcvtime >= TCP_REXMTVAL(tp) * tcp_totbackoff)) { 2773 TCPSTAT_INC(tcps_persistdrop); 2774 retval = 1; 2775 tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT); 2776 goto out; 2777 } 2778 if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) && 2779 tp->snd_una == tp->snd_max) 2780 rack_exit_persist(tp, rack); 2781 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT; 2782 /* 2783 * If the user has closed the socket then drop a persisting 2784 * connection after a much reduced timeout. 2785 */ 2786 if (tp->t_state > TCPS_CLOSE_WAIT && 2787 (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) { 2788 retval = 1; 2789 TCPSTAT_INC(tcps_persistdrop); 2790 tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT); 2791 goto out; 2792 } 2793 tp->t_flags |= TF_FORCEDATA; 2794 out: 2795 rack_log_to_event(rack, RACK_TO_FRM_PERSIST); 2796 return (retval); 2797 } 2798 2799 /* 2800 * If a keepalive goes off, we had no other timers 2801 * happening. We always return 1 here since this 2802 * routine either drops the connection or sends 2803 * out a segment with respond. 2804 */ 2805 static int 2806 rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) 2807 { 2808 struct tcptemp *t_template; 2809 struct inpcb *inp; 2810 2811 if (tp->t_timers->tt_flags & TT_STOPPED) { 2812 return (1); 2813 } 2814 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP; 2815 inp = tp->t_inpcb; 2816 rack_log_to_event(rack, RACK_TO_FRM_KEEP); 2817 /* 2818 * Keep-alive timer went off; send something or drop connection if 2819 * idle for too long. 2820 */ 2821 TCPSTAT_INC(tcps_keeptimeo); 2822 if (tp->t_state < TCPS_ESTABLISHED) 2823 goto dropit; 2824 if ((tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && 2825 tp->t_state <= TCPS_CLOSING) { 2826 if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp)) 2827 goto dropit; 2828 /* 2829 * Send a packet designed to force a response if the peer is 2830 * up and reachable: either an ACK if the connection is 2831 * still alive, or an RST if the peer has closed the 2832 * connection due to timeout or reboot. Using sequence 2833 * number tp->snd_una-1 causes the transmitted zero-length 2834 * segment to lie outside the receive window; by the 2835 * protocol spec, this requires the correspondent TCP to 2836 * respond. 2837 */ 2838 TCPSTAT_INC(tcps_keepprobe); 2839 t_template = tcpip_maketemplate(inp); 2840 if (t_template) { 2841 tcp_respond(tp, t_template->tt_ipgen, 2842 &t_template->tt_t, (struct mbuf *)NULL, 2843 tp->rcv_nxt, tp->snd_una - 1, 0); 2844 free(t_template, M_TEMP); 2845 } 2846 } 2847 rack_start_hpts_timer(rack, tp, cts, __LINE__, 0, 0, 0); 2848 return (1); 2849 dropit: 2850 TCPSTAT_INC(tcps_keepdrops); 2851 tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT); 2852 return (1); 2853 } 2854 2855 /* 2856 * Retransmit helper function, clear up all the ack 2857 * flags and take care of important book keeping. 2858 */ 2859 static void 2860 rack_remxt_tmr(struct tcpcb *tp) 2861 { 2862 /* 2863 * The retransmit timer went off, all sack'd blocks must be 2864 * un-acked. 2865 */ 2866 struct rack_sendmap *rsm, *trsm = NULL; 2867 struct tcp_rack *rack; 2868 int32_t cnt = 0; 2869 2870 rack = (struct tcp_rack *)tp->t_fb_ptr; 2871 rack_timer_cancel(tp, rack, tcp_ts_getticks(), __LINE__); 2872 rack_log_to_event(rack, RACK_TO_FRM_TMR); 2873 if (rack->r_state && (rack->r_state != tp->t_state)) 2874 rack_set_state(tp, rack); 2875 /* 2876 * Ideally we would like to be able to 2877 * mark SACK-PASS on anything not acked here. 2878 * However, if we do that we would burst out 2879 * all that data 1ms apart. This would be unwise, 2880 * so for now we will just let the normal rxt timer 2881 * and tlp timer take care of it. 2882 */ 2883 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_map, r_next) { 2884 if (rsm->r_flags & RACK_ACKED) { 2885 cnt++; 2886 rsm->r_sndcnt = 0; 2887 if (rsm->r_in_tmap == 0) { 2888 /* We must re-add it back to the tlist */ 2889 if (trsm == NULL) { 2890 TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext); 2891 } else { 2892 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext); 2893 } 2894 rsm->r_in_tmap = 1; 2895 trsm = rsm; 2896 } 2897 } 2898 rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS); 2899 } 2900 /* Clear the count (we just un-acked them) */ 2901 rack->r_ctl.rc_sacked = 0; 2902 /* Clear the tlp rtx mark */ 2903 rack->r_ctl.rc_tlp_rtx_out = 0; 2904 rack->r_ctl.rc_tlp_seg_send_cnt = 0; 2905 rack->r_ctl.rc_resend = TAILQ_FIRST(&rack->r_ctl.rc_map); 2906 /* Setup so we send one segment */ 2907 if (rack->r_ctl.rc_prr_sndcnt < tp->t_maxseg) 2908 rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; 2909 rack->r_timer_override = 1; 2910 } 2911 2912 /* 2913 * Re-transmit timeout! If we drop the PCB we will return 1, otherwise 2914 * we will setup to retransmit the lowest seq number outstanding. 2915 */ 2916 static int 2917 rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) 2918 { 2919 int32_t rexmt; 2920 struct inpcb *inp; 2921 int32_t retval = 0; 2922 2923 inp = tp->t_inpcb; 2924 if (tp->t_timers->tt_flags & TT_STOPPED) { 2925 return (1); 2926 } 2927 if (rack_progress_timeout_check(tp)) { 2928 tcp_set_inp_to_drop(inp, ETIMEDOUT); 2929 return (1); 2930 } 2931 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT; 2932 if (TCPS_HAVEESTABLISHED(tp->t_state) && 2933 (tp->snd_una == tp->snd_max)) { 2934 /* Nothing outstanding .. nothing to do */ 2935 return (0); 2936 } 2937 /* 2938 * Retransmission timer went off. Message has not been acked within 2939 * retransmit interval. Back off to a longer retransmit interval 2940 * and retransmit one segment. 2941 */ 2942 if (++tp->t_rxtshift > TCP_MAXRXTSHIFT) { 2943 tp->t_rxtshift = TCP_MAXRXTSHIFT; 2944 TCPSTAT_INC(tcps_timeoutdrop); 2945 retval = 1; 2946 tcp_set_inp_to_drop(rack->rc_inp, 2947 (tp->t_softerror ? (uint16_t) tp->t_softerror : ETIMEDOUT)); 2948 goto out; 2949 } 2950 rack_remxt_tmr(tp); 2951 if (tp->t_state == TCPS_SYN_SENT) { 2952 /* 2953 * If the SYN was retransmitted, indicate CWND to be limited 2954 * to 1 segment in cc_conn_init(). 2955 */ 2956 tp->snd_cwnd = 1; 2957 } else if (tp->t_rxtshift == 1) { 2958 /* 2959 * first retransmit; record ssthresh and cwnd so they can be 2960 * recovered if this turns out to be a "bad" retransmit. A 2961 * retransmit is considered "bad" if an ACK for this segment 2962 * is received within RTT/2 interval; the assumption here is 2963 * that the ACK was already in flight. See "On Estimating 2964 * End-to-End Network Path Properties" by Allman and Paxson 2965 * for more details. 2966 */ 2967 tp->snd_cwnd_prev = tp->snd_cwnd; 2968 tp->snd_ssthresh_prev = tp->snd_ssthresh; 2969 tp->snd_recover_prev = tp->snd_recover; 2970 if (IN_FASTRECOVERY(tp->t_flags)) 2971 tp->t_flags |= TF_WASFRECOVERY; 2972 else 2973 tp->t_flags &= ~TF_WASFRECOVERY; 2974 if (IN_CONGRECOVERY(tp->t_flags)) 2975 tp->t_flags |= TF_WASCRECOVERY; 2976 else 2977 tp->t_flags &= ~TF_WASCRECOVERY; 2978 tp->t_badrxtwin = ticks + (tp->t_srtt >> (TCP_RTT_SHIFT + 1)); 2979 tp->t_flags |= TF_PREVVALID; 2980 } else 2981 tp->t_flags &= ~TF_PREVVALID; 2982 TCPSTAT_INC(tcps_rexmttimeo); 2983 if ((tp->t_state == TCPS_SYN_SENT) || 2984 (tp->t_state == TCPS_SYN_RECEIVED)) 2985 rexmt = MSEC_2_TICKS(RACK_INITIAL_RTO * tcp_syn_backoff[tp->t_rxtshift]); 2986 else 2987 rexmt = TCP_REXMTVAL(tp) * tcp_backoff[tp->t_rxtshift]; 2988 TCPT_RANGESET(tp->t_rxtcur, rexmt, 2989 max(MSEC_2_TICKS(rack_rto_min), rexmt), 2990 MSEC_2_TICKS(rack_rto_max)); 2991 /* 2992 * We enter the path for PLMTUD if connection is established or, if 2993 * connection is FIN_WAIT_1 status, reason for the last is that if 2994 * amount of data we send is very small, we could send it in couple 2995 * of packets and process straight to FIN. In that case we won't 2996 * catch ESTABLISHED state. 2997 */ 2998 if (V_tcp_pmtud_blackhole_detect && (((tp->t_state == TCPS_ESTABLISHED)) 2999 || (tp->t_state == TCPS_FIN_WAIT_1))) { 3000 #ifdef INET6 3001 int32_t isipv6; 3002 #endif 3003 3004 /* 3005 * Idea here is that at each stage of mtu probe (usually, 3006 * 1448 -> 1188 -> 524) should be given 2 chances to recover 3007 * before further clamping down. 'tp->t_rxtshift % 2 == 0' 3008 * should take care of that. 3009 */ 3010 if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) == 3011 (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) && 3012 (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 && 3013 tp->t_rxtshift % 2 == 0)) { 3014 /* 3015 * Enter Path MTU Black-hole Detection mechanism: - 3016 * Disable Path MTU Discovery (IP "DF" bit). - 3017 * Reduce MTU to lower value than what we negotiated 3018 * with peer. 3019 */ 3020 if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) { 3021 /* Record that we may have found a black hole. */ 3022 tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE; 3023 /* Keep track of previous MSS. */ 3024 tp->t_pmtud_saved_maxseg = tp->t_maxseg; 3025 } 3026 3027 /* 3028 * Reduce the MSS to blackhole value or to the 3029 * default in an attempt to retransmit. 3030 */ 3031 #ifdef INET6 3032 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) ? 1 : 0; 3033 if (isipv6 && 3034 tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) { 3035 /* Use the sysctl tuneable blackhole MSS. */ 3036 tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss; 3037 TCPSTAT_INC(tcps_pmtud_blackhole_activated); 3038 } else if (isipv6) { 3039 /* Use the default MSS. */ 3040 tp->t_maxseg = V_tcp_v6mssdflt; 3041 /* 3042 * Disable Path MTU Discovery when we switch 3043 * to minmss. 3044 */ 3045 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 3046 TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss); 3047 } 3048 #endif 3049 #if defined(INET6) && defined(INET) 3050 else 3051 #endif 3052 #ifdef INET 3053 if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) { 3054 /* Use the sysctl tuneable blackhole MSS. */ 3055 tp->t_maxseg = V_tcp_pmtud_blackhole_mss; 3056 TCPSTAT_INC(tcps_pmtud_blackhole_activated); 3057 } else { 3058 /* Use the default MSS. */ 3059 tp->t_maxseg = V_tcp_mssdflt; 3060 /* 3061 * Disable Path MTU Discovery when we switch 3062 * to minmss. 3063 */ 3064 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 3065 TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss); 3066 } 3067 #endif 3068 } else { 3069 /* 3070 * If further retransmissions are still unsuccessful 3071 * with a lowered MTU, maybe this isn't a blackhole 3072 * and we restore the previous MSS and blackhole 3073 * detection flags. The limit '6' is determined by 3074 * giving each probe stage (1448, 1188, 524) 2 3075 * chances to recover. 3076 */ 3077 if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) && 3078 (tp->t_rxtshift >= 6)) { 3079 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 3080 tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE; 3081 tp->t_maxseg = tp->t_pmtud_saved_maxseg; 3082 TCPSTAT_INC(tcps_pmtud_blackhole_failed); 3083 } 3084 } 3085 } 3086 /* 3087 * Disable RFC1323 and SACK if we haven't got any response to our 3088 * third SYN to work-around some broken terminal servers (most of 3089 * which have hopefully been retired) that have bad VJ header 3090 * compression code which trashes TCP segments containing 3091 * unknown-to-them TCP options. 3092 */ 3093 if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) && 3094 (tp->t_rxtshift == 3)) 3095 tp->t_flags &= ~(TF_REQ_SCALE | TF_REQ_TSTMP | TF_SACK_PERMIT); 3096 /* 3097 * If we backed off this far, our srtt estimate is probably bogus. 3098 * Clobber it so we'll take the next rtt measurement as our srtt; 3099 * move the current srtt into rttvar to keep the current retransmit 3100 * times until then. 3101 */ 3102 if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) { 3103 #ifdef INET6 3104 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0) 3105 in6_losing(tp->t_inpcb); 3106 else 3107 #endif 3108 in_losing(tp->t_inpcb); 3109 tp->t_rttvar += (tp->t_srtt >> TCP_RTT_SHIFT); 3110 tp->t_srtt = 0; 3111 } 3112 if (rack_use_sack_filter) 3113 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una); 3114 tp->snd_recover = tp->snd_max; 3115 tp->t_flags |= TF_ACKNOW; 3116 tp->t_rtttime = 0; 3117 rack_cong_signal(tp, NULL, CC_RTO); 3118 out: 3119 return (retval); 3120 } 3121 3122 static int 3123 rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling) 3124 { 3125 int32_t ret = 0; 3126 int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK); 3127 3128 if (timers == 0) { 3129 return (0); 3130 } 3131 if (tp->t_state == TCPS_LISTEN) { 3132 /* no timers on listen sockets */ 3133 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) 3134 return (0); 3135 return (1); 3136 } 3137 if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) { 3138 uint32_t left; 3139 3140 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) { 3141 ret = -1; 3142 rack_log_to_processing(rack, cts, ret, 0); 3143 return (0); 3144 } 3145 if (hpts_calling == 0) { 3146 ret = -2; 3147 rack_log_to_processing(rack, cts, ret, 0); 3148 return (0); 3149 } 3150 /* 3151 * Ok our timer went off early and we are not paced false 3152 * alarm, go back to sleep. 3153 */ 3154 ret = -3; 3155 left = rack->r_ctl.rc_timer_exp - cts; 3156 tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(left)); 3157 rack_log_to_processing(rack, cts, ret, left); 3158 rack->rc_last_pto_set = 0; 3159 return (1); 3160 } 3161 rack->rc_tmr_stopped = 0; 3162 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK; 3163 if (timers & PACE_TMR_DELACK) { 3164 ret = rack_timeout_delack(tp, rack, cts); 3165 } else if (timers & PACE_TMR_RACK) { 3166 ret = rack_timeout_rack(tp, rack, cts); 3167 } else if (timers & PACE_TMR_TLP) { 3168 ret = rack_timeout_tlp(tp, rack, cts); 3169 } else if (timers & PACE_TMR_RXT) { 3170 ret = rack_timeout_rxt(tp, rack, cts); 3171 } else if (timers & PACE_TMR_PERSIT) { 3172 ret = rack_timeout_persist(tp, rack, cts); 3173 } else if (timers & PACE_TMR_KEEP) { 3174 ret = rack_timeout_keepalive(tp, rack, cts); 3175 } 3176 rack_log_to_processing(rack, cts, ret, timers); 3177 return (ret); 3178 } 3179 3180 static void 3181 rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line) 3182 { 3183 uint8_t hpts_removed = 0; 3184 3185 if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) && 3186 TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to)) { 3187 tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT); 3188 hpts_removed = 1; 3189 } 3190 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) { 3191 rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK; 3192 if (rack->rc_inp->inp_in_hpts && 3193 ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) { 3194 /* 3195 * Canceling timer's when we have no output being 3196 * paced. We also must remove ourselves from the 3197 * hpts. 3198 */ 3199 tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT); 3200 hpts_removed = 1; 3201 } 3202 rack_log_to_cancel(rack, hpts_removed, line); 3203 rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK); 3204 } 3205 } 3206 3207 static void 3208 rack_timer_stop(struct tcpcb *tp, uint32_t timer_type) 3209 { 3210 return; 3211 } 3212 3213 static int 3214 rack_stopall(struct tcpcb *tp) 3215 { 3216 struct tcp_rack *rack; 3217 rack = (struct tcp_rack *)tp->t_fb_ptr; 3218 rack->t_timers_stopped = 1; 3219 return (0); 3220 } 3221 3222 static void 3223 rack_timer_activate(struct tcpcb *tp, uint32_t timer_type, uint32_t delta) 3224 { 3225 return; 3226 } 3227 3228 static int 3229 rack_timer_active(struct tcpcb *tp, uint32_t timer_type) 3230 { 3231 return (0); 3232 } 3233 3234 static void 3235 rack_stop_all_timers(struct tcpcb *tp) 3236 { 3237 struct tcp_rack *rack; 3238 3239 /* 3240 * Assure no timers are running. 3241 */ 3242 if (tcp_timer_active(tp, TT_PERSIST)) { 3243 /* We enter in persists, set the flag appropriately */ 3244 rack = (struct tcp_rack *)tp->t_fb_ptr; 3245 rack->rc_in_persist = 1; 3246 } 3247 tcp_timer_suspend(tp, TT_PERSIST); 3248 tcp_timer_suspend(tp, TT_REXMT); 3249 tcp_timer_suspend(tp, TT_KEEP); 3250 tcp_timer_suspend(tp, TT_DELACK); 3251 } 3252 3253 static void 3254 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack, 3255 struct rack_sendmap *rsm, uint32_t ts) 3256 { 3257 int32_t idx; 3258 3259 rsm->r_rtr_cnt++; 3260 rsm->r_sndcnt++; 3261 if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) { 3262 rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS; 3263 rsm->r_flags |= RACK_OVERMAX; 3264 } 3265 if ((rsm->r_rtr_cnt > 1) && (rack->r_tlp_running == 0)) { 3266 rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start); 3267 rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start); 3268 } 3269 idx = rsm->r_rtr_cnt - 1; 3270 rsm->r_tim_lastsent[idx] = ts; 3271 if (rsm->r_flags & RACK_ACKED) { 3272 /* Problably MTU discovery messing with us */ 3273 rsm->r_flags &= ~RACK_ACKED; 3274 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); 3275 } 3276 if (rsm->r_in_tmap) { 3277 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext); 3278 } 3279 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext); 3280 rsm->r_in_tmap = 1; 3281 if (rsm->r_flags & RACK_SACK_PASSED) { 3282 /* We have retransmitted due to the SACK pass */ 3283 rsm->r_flags &= ~RACK_SACK_PASSED; 3284 rsm->r_flags |= RACK_WAS_SACKPASS; 3285 } 3286 /* Update memory for next rtr */ 3287 rack->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next); 3288 } 3289 3290 3291 static uint32_t 3292 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack, 3293 struct rack_sendmap *rsm, uint32_t ts, int32_t * lenp) 3294 { 3295 /* 3296 * We (re-)transmitted starting at rsm->r_start for some length 3297 * (possibly less than r_end. 3298 */ 3299 struct rack_sendmap *nrsm; 3300 uint32_t c_end; 3301 int32_t len; 3302 int32_t idx; 3303 3304 len = *lenp; 3305 c_end = rsm->r_start + len; 3306 if (SEQ_GEQ(c_end, rsm->r_end)) { 3307 /* 3308 * We retransmitted the whole piece or more than the whole 3309 * slopping into the next rsm. 3310 */ 3311 rack_update_rsm(tp, rack, rsm, ts); 3312 if (c_end == rsm->r_end) { 3313 *lenp = 0; 3314 return (0); 3315 } else { 3316 int32_t act_len; 3317 3318 /* Hangs over the end return whats left */ 3319 act_len = rsm->r_end - rsm->r_start; 3320 *lenp = (len - act_len); 3321 return (rsm->r_end); 3322 } 3323 /* We don't get out of this block. */ 3324 } 3325 /* 3326 * Here we retransmitted less than the whole thing which means we 3327 * have to split this into what was transmitted and what was not. 3328 */ 3329 nrsm = rack_alloc_full_limit(rack); 3330 if (nrsm == NULL) { 3331 /* 3332 * We can't get memory, so lets not proceed. 3333 */ 3334 *lenp = 0; 3335 return (0); 3336 } 3337 /* 3338 * So here we are going to take the original rsm and make it what we 3339 * retransmitted. nrsm will be the tail portion we did not 3340 * retransmit. For example say the chunk was 1, 11 (10 bytes). And 3341 * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to 3342 * 1, 6 and the new piece will be 6, 11. 3343 */ 3344 nrsm->r_start = c_end; 3345 nrsm->r_end = rsm->r_end; 3346 nrsm->r_rtr_cnt = rsm->r_rtr_cnt; 3347 nrsm->r_flags = rsm->r_flags; 3348 nrsm->r_sndcnt = rsm->r_sndcnt; 3349 nrsm->r_rtr_bytes = 0; 3350 rsm->r_end = c_end; 3351 for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { 3352 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; 3353 } 3354 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_map, rsm, nrsm, r_next); 3355 if (rsm->r_in_tmap) { 3356 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 3357 nrsm->r_in_tmap = 1; 3358 } 3359 rsm->r_flags &= (~RACK_HAS_FIN); 3360 rack_update_rsm(tp, rack, rsm, ts); 3361 *lenp = 0; 3362 return (0); 3363 } 3364 3365 3366 static void 3367 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len, 3368 uint32_t seq_out, uint8_t th_flags, int32_t err, uint32_t ts, 3369 uint8_t pass, struct rack_sendmap *hintrsm) 3370 { 3371 struct tcp_rack *rack; 3372 struct rack_sendmap *rsm, *nrsm; 3373 register uint32_t snd_max, snd_una; 3374 int32_t idx; 3375 3376 /* 3377 * Add to the RACK log of packets in flight or retransmitted. If 3378 * there is a TS option we will use the TS echoed, if not we will 3379 * grab a TS. 3380 * 3381 * Retransmissions will increment the count and move the ts to its 3382 * proper place. Note that if options do not include TS's then we 3383 * won't be able to effectively use the ACK for an RTT on a retran. 3384 * 3385 * Notes about r_start and r_end. Lets consider a send starting at 3386 * sequence 1 for 10 bytes. In such an example the r_start would be 3387 * 1 (starting sequence) but the r_end would be r_start+len i.e. 11. 3388 * This means that r_end is actually the first sequence for the next 3389 * slot (11). 3390 * 3391 */ 3392 /* 3393 * If err is set what do we do XXXrrs? should we not add the thing? 3394 * -- i.e. return if err != 0 or should we pretend we sent it? -- 3395 * i.e. proceed with add ** do this for now. 3396 */ 3397 INP_WLOCK_ASSERT(tp->t_inpcb); 3398 if (err) 3399 /* 3400 * We don't log errors -- we could but snd_max does not 3401 * advance in this case either. 3402 */ 3403 return; 3404 3405 if (th_flags & TH_RST) { 3406 /* 3407 * We don't log resets and we return immediately from 3408 * sending 3409 */ 3410 return; 3411 } 3412 rack = (struct tcp_rack *)tp->t_fb_ptr; 3413 snd_una = tp->snd_una; 3414 if (SEQ_LEQ((seq_out + len), snd_una)) { 3415 /* Are sending an old segment to induce an ack (keep-alive)? */ 3416 return; 3417 } 3418 if (SEQ_LT(seq_out, snd_una)) { 3419 /* huh? should we panic? */ 3420 uint32_t end; 3421 3422 end = seq_out + len; 3423 seq_out = snd_una; 3424 len = end - seq_out; 3425 } 3426 snd_max = tp->snd_max; 3427 if (th_flags & (TH_SYN | TH_FIN)) { 3428 /* 3429 * The call to rack_log_output is made before bumping 3430 * snd_max. This means we can record one extra byte on a SYN 3431 * or FIN if seq_out is adding more on and a FIN is present 3432 * (and we are not resending). 3433 */ 3434 if (th_flags & TH_SYN) 3435 len++; 3436 if (th_flags & TH_FIN) 3437 len++; 3438 if (SEQ_LT(snd_max, tp->snd_nxt)) { 3439 /* 3440 * The add/update as not been done for the FIN/SYN 3441 * yet. 3442 */ 3443 snd_max = tp->snd_nxt; 3444 } 3445 } 3446 if (len == 0) { 3447 /* We don't log zero window probes */ 3448 return; 3449 } 3450 rack->r_ctl.rc_time_last_sent = ts; 3451 if (IN_RECOVERY(tp->t_flags)) { 3452 rack->r_ctl.rc_prr_out += len; 3453 } 3454 /* First question is it a retransmission? */ 3455 if (seq_out == snd_max) { 3456 again: 3457 rsm = rack_alloc(rack); 3458 if (rsm == NULL) { 3459 /* 3460 * Hmm out of memory and the tcb got destroyed while 3461 * we tried to wait. 3462 */ 3463 return; 3464 } 3465 if (th_flags & TH_FIN) { 3466 rsm->r_flags = RACK_HAS_FIN; 3467 } else { 3468 rsm->r_flags = 0; 3469 } 3470 rsm->r_tim_lastsent[0] = ts; 3471 rsm->r_rtr_cnt = 1; 3472 rsm->r_rtr_bytes = 0; 3473 rsm->r_start = seq_out; 3474 rsm->r_end = rsm->r_start + len; 3475 rsm->r_sndcnt = 0; 3476 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_map, rsm, r_next); 3477 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext); 3478 rsm->r_in_tmap = 1; 3479 return; 3480 } 3481 /* 3482 * If we reach here its a retransmission and we need to find it. 3483 */ 3484 more: 3485 if (hintrsm && (hintrsm->r_start == seq_out)) { 3486 rsm = hintrsm; 3487 hintrsm = NULL; 3488 } else if (rack->r_ctl.rc_next) { 3489 /* We have a hint from a previous run */ 3490 rsm = rack->r_ctl.rc_next; 3491 } else { 3492 /* No hints sorry */ 3493 rsm = NULL; 3494 } 3495 if ((rsm) && (rsm->r_start == seq_out)) { 3496 /* 3497 * We used rc_next or hintrsm to retransmit, hopefully the 3498 * likely case. 3499 */ 3500 seq_out = rack_update_entry(tp, rack, rsm, ts, &len); 3501 if (len == 0) { 3502 return; 3503 } else { 3504 goto more; 3505 } 3506 } 3507 /* Ok it was not the last pointer go through it the hard way. */ 3508 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_map, r_next) { 3509 if (rsm->r_start == seq_out) { 3510 seq_out = rack_update_entry(tp, rack, rsm, ts, &len); 3511 rack->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next); 3512 if (len == 0) { 3513 return; 3514 } else { 3515 continue; 3516 } 3517 } 3518 if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) { 3519 /* Transmitted within this piece */ 3520 /* 3521 * Ok we must split off the front and then let the 3522 * update do the rest 3523 */ 3524 nrsm = rack_alloc_full_limit(rack); 3525 if (nrsm == NULL) { 3526 rack_update_rsm(tp, rack, rsm, ts); 3527 return; 3528 } 3529 /* 3530 * copy rsm to nrsm and then trim the front of rsm 3531 * to not include this part. 3532 */ 3533 nrsm->r_start = seq_out; 3534 nrsm->r_end = rsm->r_end; 3535 nrsm->r_rtr_cnt = rsm->r_rtr_cnt; 3536 nrsm->r_flags = rsm->r_flags; 3537 nrsm->r_sndcnt = rsm->r_sndcnt; 3538 nrsm->r_rtr_bytes = 0; 3539 for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { 3540 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; 3541 } 3542 rsm->r_end = nrsm->r_start; 3543 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_map, rsm, nrsm, r_next); 3544 if (rsm->r_in_tmap) { 3545 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 3546 nrsm->r_in_tmap = 1; 3547 } 3548 rsm->r_flags &= (~RACK_HAS_FIN); 3549 seq_out = rack_update_entry(tp, rack, nrsm, ts, &len); 3550 if (len == 0) { 3551 return; 3552 } 3553 } 3554 } 3555 /* 3556 * Hmm not found in map did they retransmit both old and on into the 3557 * new? 3558 */ 3559 if (seq_out == tp->snd_max) { 3560 goto again; 3561 } else if (SEQ_LT(seq_out, tp->snd_max)) { 3562 #ifdef INVARIANTS 3563 printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n", 3564 seq_out, len, tp->snd_una, tp->snd_max); 3565 printf("Starting Dump of all rack entries\n"); 3566 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_map, r_next) { 3567 printf("rsm:%p start:%u end:%u\n", 3568 rsm, rsm->r_start, rsm->r_end); 3569 } 3570 printf("Dump complete\n"); 3571 panic("seq_out not found rack:%p tp:%p", 3572 rack, tp); 3573 #endif 3574 } else { 3575 #ifdef INVARIANTS 3576 /* 3577 * Hmm beyond sndmax? (only if we are using the new rtt-pack 3578 * flag) 3579 */ 3580 panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p", 3581 seq_out, len, tp->snd_max, tp); 3582 #endif 3583 } 3584 } 3585 3586 /* 3587 * Record one of the RTT updates from an ack into 3588 * our sample structure. 3589 */ 3590 static void 3591 tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt) 3592 { 3593 if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) || 3594 (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) { 3595 rack->r_ctl.rack_rs.rs_rtt_lowest = rtt; 3596 } 3597 if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) || 3598 (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) { 3599 rack->r_ctl.rack_rs.rs_rtt_highest = rtt; 3600 } 3601 rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID; 3602 rack->r_ctl.rack_rs.rs_rtt_tot += rtt; 3603 rack->r_ctl.rack_rs.rs_rtt_cnt++; 3604 } 3605 3606 /* 3607 * Collect new round-trip time estimate 3608 * and update averages and current timeout. 3609 */ 3610 static void 3611 tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp) 3612 { 3613 int32_t delta; 3614 uint32_t o_srtt, o_var; 3615 int32_t rtt; 3616 3617 if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) 3618 /* No valid sample */ 3619 return; 3620 if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) { 3621 /* We are to use the lowest RTT seen in a single ack */ 3622 rtt = rack->r_ctl.rack_rs.rs_rtt_lowest; 3623 } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) { 3624 /* We are to use the highest RTT seen in a single ack */ 3625 rtt = rack->r_ctl.rack_rs.rs_rtt_highest; 3626 } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) { 3627 /* We are to use the average RTT seen in a single ack */ 3628 rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot / 3629 (uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt); 3630 } else { 3631 #ifdef INVARIANTS 3632 panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method); 3633 #endif 3634 return; 3635 } 3636 if (rtt == 0) 3637 rtt = 1; 3638 rack_log_rtt_sample(rack, rtt); 3639 o_srtt = tp->t_srtt; 3640 o_var = tp->t_rttvar; 3641 rack = (struct tcp_rack *)tp->t_fb_ptr; 3642 if (tp->t_srtt != 0) { 3643 /* 3644 * srtt is stored as fixed point with 5 bits after the 3645 * binary point (i.e., scaled by 8). The following magic is 3646 * equivalent to the smoothing algorithm in rfc793 with an 3647 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point). 3648 * Adjust rtt to origin 0. 3649 */ 3650 delta = ((rtt - 1) << TCP_DELTA_SHIFT) 3651 - (tp->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT)); 3652 3653 tp->t_srtt += delta; 3654 if (tp->t_srtt <= 0) 3655 tp->t_srtt = 1; 3656 3657 /* 3658 * We accumulate a smoothed rtt variance (actually, a 3659 * smoothed mean difference), then set the retransmit timer 3660 * to smoothed rtt + 4 times the smoothed variance. rttvar 3661 * is stored as fixed point with 4 bits after the binary 3662 * point (scaled by 16). The following is equivalent to 3663 * rfc793 smoothing with an alpha of .75 (rttvar = 3664 * rttvar*3/4 + |delta| / 4). This replaces rfc793's 3665 * wired-in beta. 3666 */ 3667 if (delta < 0) 3668 delta = -delta; 3669 delta -= tp->t_rttvar >> (TCP_RTTVAR_SHIFT - TCP_DELTA_SHIFT); 3670 tp->t_rttvar += delta; 3671 if (tp->t_rttvar <= 0) 3672 tp->t_rttvar = 1; 3673 if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar) 3674 tp->t_rttbest = tp->t_srtt + tp->t_rttvar; 3675 } else { 3676 /* 3677 * No rtt measurement yet - use the unsmoothed rtt. Set the 3678 * variance to half the rtt (so our first retransmit happens 3679 * at 3*rtt). 3680 */ 3681 tp->t_srtt = rtt << TCP_RTT_SHIFT; 3682 tp->t_rttvar = rtt << (TCP_RTTVAR_SHIFT - 1); 3683 tp->t_rttbest = tp->t_srtt + tp->t_rttvar; 3684 } 3685 TCPSTAT_INC(tcps_rttupdated); 3686 rack_log_rtt_upd(tp, rack, rtt, o_srtt, o_var); 3687 tp->t_rttupdated++; 3688 #ifdef NETFLIX_STATS 3689 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt)); 3690 #endif 3691 tp->t_rxtshift = 0; 3692 3693 /* 3694 * the retransmit should happen at rtt + 4 * rttvar. Because of the 3695 * way we do the smoothing, srtt and rttvar will each average +1/2 3696 * tick of bias. When we compute the retransmit timer, we want 1/2 3697 * tick of rounding and 1 extra tick because of +-1/2 tick 3698 * uncertainty in the firing of the timer. The bias will give us 3699 * exactly the 1.5 tick we need. But, because the bias is 3700 * statistical, we have to test that we don't drop below the minimum 3701 * feasible timer (which is 2 ticks). 3702 */ 3703 TCPT_RANGESET(tp->t_rxtcur, TCP_REXMTVAL(tp), 3704 max(MSEC_2_TICKS(rack_rto_min), rtt + 2), MSEC_2_TICKS(rack_rto_max)); 3705 tp->t_softerror = 0; 3706 } 3707 3708 static void 3709 rack_earlier_retran(struct tcpcb *tp, struct rack_sendmap *rsm, 3710 uint32_t t, uint32_t cts) 3711 { 3712 /* 3713 * For this RSM, we acknowledged the data from a previous 3714 * transmission, not the last one we made. This means we did a false 3715 * retransmit. 3716 */ 3717 struct tcp_rack *rack; 3718 3719 if (rsm->r_flags & RACK_HAS_FIN) { 3720 /* 3721 * The sending of the FIN often is multiple sent when we 3722 * have everything outstanding ack'd. We ignore this case 3723 * since its over now. 3724 */ 3725 return; 3726 } 3727 if (rsm->r_flags & RACK_TLP) { 3728 /* 3729 * We expect TLP's to have this occur. 3730 */ 3731 return; 3732 } 3733 rack = (struct tcp_rack *)tp->t_fb_ptr; 3734 /* should we undo cc changes and exit recovery? */ 3735 if (IN_RECOVERY(tp->t_flags)) { 3736 if (rack->r_ctl.rc_rsm_start == rsm->r_start) { 3737 /* 3738 * Undo what we ratched down and exit recovery if 3739 * possible 3740 */ 3741 EXIT_RECOVERY(tp->t_flags); 3742 tp->snd_recover = tp->snd_una; 3743 if (rack->r_ctl.rc_cwnd_at > tp->snd_cwnd) 3744 tp->snd_cwnd = rack->r_ctl.rc_cwnd_at; 3745 if (rack->r_ctl.rc_ssthresh_at > tp->snd_ssthresh) 3746 tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at; 3747 } 3748 } 3749 if (rsm->r_flags & RACK_WAS_SACKPASS) { 3750 /* 3751 * We retransmitted based on a sack and the earlier 3752 * retransmission ack'd it - re-ordering is occuring. 3753 */ 3754 counter_u64_add(rack_reorder_seen, 1); 3755 rack->r_ctl.rc_reorder_ts = cts; 3756 } 3757 counter_u64_add(rack_badfr, 1); 3758 counter_u64_add(rack_badfr_bytes, (rsm->r_end - rsm->r_start)); 3759 } 3760 3761 3762 static int 3763 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack, 3764 struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type) 3765 { 3766 int32_t i; 3767 uint32_t t; 3768 3769 if (rsm->r_flags & RACK_ACKED) 3770 /* Already done */ 3771 return (0); 3772 3773 3774 if ((rsm->r_rtr_cnt == 1) || 3775 ((ack_type == CUM_ACKED) && 3776 (to->to_flags & TOF_TS) && 3777 (to->to_tsecr) && 3778 (rsm->r_tim_lastsent[rsm->r_rtr_cnt - 1] == to->to_tsecr)) 3779 ) { 3780 /* 3781 * We will only find a matching timestamp if its cum-acked. 3782 * But if its only one retransmission its for-sure matching 3783 * :-) 3784 */ 3785 t = cts - rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]; 3786 if ((int)t <= 0) 3787 t = 1; 3788 if (!tp->t_rttlow || tp->t_rttlow > t) 3789 tp->t_rttlow = t; 3790 if (!rack->r_ctl.rc_rack_min_rtt || 3791 SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) { 3792 rack->r_ctl.rc_rack_min_rtt = t; 3793 if (rack->r_ctl.rc_rack_min_rtt == 0) { 3794 rack->r_ctl.rc_rack_min_rtt = 1; 3795 } 3796 } 3797 tcp_rack_xmit_timer(rack, TCP_TS_TO_TICKS(t) + 1); 3798 if ((rsm->r_flags & RACK_TLP) && 3799 (!IN_RECOVERY(tp->t_flags))) { 3800 /* Segment was a TLP and our retrans matched */ 3801 if (rack->r_ctl.rc_tlp_cwnd_reduce) { 3802 rack->r_ctl.rc_rsm_start = tp->snd_max; 3803 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd; 3804 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh; 3805 rack_cong_signal(tp, NULL, CC_NDUPACK); 3806 /* 3807 * When we enter recovery we need to assure 3808 * we send one packet. 3809 */ 3810 rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; 3811 } else 3812 rack->r_ctl.rc_tlp_rtx_out = 0; 3813 } 3814 if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) { 3815 /* New more recent rack_tmit_time */ 3816 rack->r_ctl.rc_rack_tmit_time = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]; 3817 rack->rc_rack_rtt = t; 3818 } 3819 return (1); 3820 } 3821 /* 3822 * We clear the soft/rxtshift since we got an ack. 3823 * There is no assurance we will call the commit() function 3824 * so we need to clear these to avoid incorrect handling. 3825 */ 3826 tp->t_rxtshift = 0; 3827 tp->t_softerror = 0; 3828 if ((to->to_flags & TOF_TS) && 3829 (ack_type == CUM_ACKED) && 3830 (to->to_tsecr) && 3831 ((rsm->r_flags & (RACK_DEFERRED | RACK_OVERMAX)) == 0)) { 3832 /* 3833 * Now which timestamp does it match? In this block the ACK 3834 * must be coming from a previous transmission. 3835 */ 3836 for (i = 0; i < rsm->r_rtr_cnt; i++) { 3837 if (rsm->r_tim_lastsent[i] == to->to_tsecr) { 3838 t = cts - rsm->r_tim_lastsent[i]; 3839 if ((int)t <= 0) 3840 t = 1; 3841 if ((i + 1) < rsm->r_rtr_cnt) { 3842 /* Likely */ 3843 rack_earlier_retran(tp, rsm, t, cts); 3844 } 3845 if (!tp->t_rttlow || tp->t_rttlow > t) 3846 tp->t_rttlow = t; 3847 if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) { 3848 rack->r_ctl.rc_rack_min_rtt = t; 3849 if (rack->r_ctl.rc_rack_min_rtt == 0) { 3850 rack->r_ctl.rc_rack_min_rtt = 1; 3851 } 3852 } 3853 /* 3854 * Note the following calls to 3855 * tcp_rack_xmit_timer() are being commented 3856 * out for now. They give us no more accuracy 3857 * and often lead to a wrong choice. We have 3858 * enough samples that have not been 3859 * retransmitted. I leave the commented out 3860 * code in here in case in the future we 3861 * decide to add it back (though I can't forsee 3862 * doing that). That way we will easily see 3863 * where they need to be placed. 3864 */ 3865 if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, 3866 rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) { 3867 /* New more recent rack_tmit_time */ 3868 rack->r_ctl.rc_rack_tmit_time = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]; 3869 rack->rc_rack_rtt = t; 3870 } 3871 return (1); 3872 } 3873 } 3874 goto ts_not_found; 3875 } else { 3876 /* 3877 * Ok its a SACK block that we retransmitted. or a windows 3878 * machine without timestamps. We can tell nothing from the 3879 * time-stamp since its not there or the time the peer last 3880 * recieved a segment that moved forward its cum-ack point. 3881 */ 3882 ts_not_found: 3883 i = rsm->r_rtr_cnt - 1; 3884 t = cts - rsm->r_tim_lastsent[i]; 3885 if ((int)t <= 0) 3886 t = 1; 3887 if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) { 3888 /* 3889 * We retransmitted and the ack came back in less 3890 * than the smallest rtt we have observed. We most 3891 * likey did an improper retransmit as outlined in 3892 * 4.2 Step 3 point 2 in the rack-draft. 3893 */ 3894 i = rsm->r_rtr_cnt - 2; 3895 t = cts - rsm->r_tim_lastsent[i]; 3896 rack_earlier_retran(tp, rsm, t, cts); 3897 } else if (rack->r_ctl.rc_rack_min_rtt) { 3898 /* 3899 * We retransmitted it and the retransmit did the 3900 * job. 3901 */ 3902 if (!rack->r_ctl.rc_rack_min_rtt || 3903 SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) { 3904 rack->r_ctl.rc_rack_min_rtt = t; 3905 if (rack->r_ctl.rc_rack_min_rtt == 0) { 3906 rack->r_ctl.rc_rack_min_rtt = 1; 3907 } 3908 } 3909 if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, rsm->r_tim_lastsent[i])) { 3910 /* New more recent rack_tmit_time */ 3911 rack->r_ctl.rc_rack_tmit_time = rsm->r_tim_lastsent[i]; 3912 rack->rc_rack_rtt = t; 3913 } 3914 return (1); 3915 } 3916 } 3917 return (0); 3918 } 3919 3920 /* 3921 * Mark the SACK_PASSED flag on all entries prior to rsm send wise. 3922 */ 3923 static void 3924 rack_log_sack_passed(struct tcpcb *tp, 3925 struct tcp_rack *rack, struct rack_sendmap *rsm) 3926 { 3927 struct rack_sendmap *nrsm; 3928 uint32_t ts; 3929 int32_t idx; 3930 3931 idx = rsm->r_rtr_cnt - 1; 3932 ts = rsm->r_tim_lastsent[idx]; 3933 nrsm = rsm; 3934 TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap, 3935 rack_head, r_tnext) { 3936 if (nrsm == rsm) { 3937 /* Skip orginal segment he is acked */ 3938 continue; 3939 } 3940 if (nrsm->r_flags & RACK_ACKED) { 3941 /* Skip ack'd segments */ 3942 continue; 3943 } 3944 if (nrsm->r_flags & RACK_SACK_PASSED) { 3945 /* 3946 * We found one that is already marked 3947 * passed, we have been here before and 3948 * so all others below this are marked. 3949 */ 3950 break; 3951 } 3952 idx = nrsm->r_rtr_cnt - 1; 3953 if (ts == nrsm->r_tim_lastsent[idx]) { 3954 /* 3955 * For this case lets use seq no, if we sent in a 3956 * big block (TSO) we would have a bunch of segments 3957 * sent at the same time. 3958 * 3959 * We would only get a report if its SEQ is earlier. 3960 * If we have done multiple retransmits the times 3961 * would not be equal. 3962 */ 3963 if (SEQ_LT(nrsm->r_start, rsm->r_start)) { 3964 nrsm->r_flags |= RACK_SACK_PASSED; 3965 nrsm->r_flags &= ~RACK_WAS_SACKPASS; 3966 } 3967 } else { 3968 /* 3969 * Here they were sent at different times, not a big 3970 * block. Since we transmitted this one later and 3971 * see it sack'd then this must also be missing (or 3972 * we would have gotten a sack block for it) 3973 */ 3974 nrsm->r_flags |= RACK_SACK_PASSED; 3975 nrsm->r_flags &= ~RACK_WAS_SACKPASS; 3976 } 3977 } 3978 } 3979 3980 static uint32_t 3981 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack, 3982 struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts) 3983 { 3984 int32_t idx; 3985 int32_t times = 0; 3986 uint32_t start, end, changed = 0; 3987 struct rack_sendmap *rsm, *nrsm; 3988 int32_t used_ref = 1; 3989 3990 start = sack->start; 3991 end = sack->end; 3992 rsm = *prsm; 3993 if (rsm && SEQ_LT(start, rsm->r_start)) { 3994 TAILQ_FOREACH_REVERSE_FROM(rsm, &rack->r_ctl.rc_map, rack_head, r_next) { 3995 if (SEQ_GEQ(start, rsm->r_start) && 3996 SEQ_LT(start, rsm->r_end)) { 3997 goto do_rest_ofb; 3998 } 3999 } 4000 } 4001 if (rsm == NULL) { 4002 start_at_beginning: 4003 rsm = NULL; 4004 used_ref = 0; 4005 } 4006 /* First lets locate the block where this guy is */ 4007 TAILQ_FOREACH_FROM(rsm, &rack->r_ctl.rc_map, r_next) { 4008 if (SEQ_GEQ(start, rsm->r_start) && 4009 SEQ_LT(start, rsm->r_end)) { 4010 break; 4011 } 4012 } 4013 do_rest_ofb: 4014 if (rsm == NULL) { 4015 /* 4016 * This happens when we get duplicate sack blocks with the 4017 * same end. For example SACK 4: 100 SACK 3: 100 The sort 4018 * will not change there location so we would just start at 4019 * the end of the first one and get lost. 4020 */ 4021 if (tp->t_flags & TF_SENTFIN) { 4022 /* 4023 * Check to see if we have not logged the FIN that 4024 * went out. 4025 */ 4026 nrsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_map, rack_sendmap, r_next); 4027 if (nrsm && (nrsm->r_end + 1) == tp->snd_max) { 4028 /* 4029 * Ok we did not get the FIN logged. 4030 */ 4031 nrsm->r_end++; 4032 rsm = nrsm; 4033 goto do_rest_ofb; 4034 } 4035 } 4036 if (times == 1) { 4037 #ifdef INVARIANTS 4038 panic("tp:%p rack:%p sack:%p to:%p prsm:%p", 4039 tp, rack, sack, to, prsm); 4040 #else 4041 goto out; 4042 #endif 4043 } 4044 times++; 4045 counter_u64_add(rack_sack_proc_restart, 1); 4046 goto start_at_beginning; 4047 } 4048 /* Ok we have an ACK for some piece of rsm */ 4049 if (rsm->r_start != start) { 4050 /* 4051 * Need to split this in two pieces the before and after. 4052 */ 4053 nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT); 4054 if (nrsm == NULL) { 4055 /* 4056 * failed XXXrrs what can we do but loose the sack 4057 * info? 4058 */ 4059 goto out; 4060 } 4061 nrsm->r_start = start; 4062 nrsm->r_rtr_bytes = 0; 4063 nrsm->r_end = rsm->r_end; 4064 nrsm->r_rtr_cnt = rsm->r_rtr_cnt; 4065 nrsm->r_flags = rsm->r_flags; 4066 nrsm->r_sndcnt = rsm->r_sndcnt; 4067 for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { 4068 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; 4069 } 4070 rsm->r_end = nrsm->r_start; 4071 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_map, rsm, nrsm, r_next); 4072 if (rsm->r_in_tmap) { 4073 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 4074 nrsm->r_in_tmap = 1; 4075 } 4076 rsm->r_flags &= (~RACK_HAS_FIN); 4077 rsm = nrsm; 4078 } 4079 if (SEQ_GEQ(end, rsm->r_end)) { 4080 /* 4081 * The end of this block is either beyond this guy or right 4082 * at this guy. 4083 */ 4084 4085 if ((rsm->r_flags & RACK_ACKED) == 0) { 4086 rack_update_rtt(tp, rack, rsm, to, cts, SACKED); 4087 changed += (rsm->r_end - rsm->r_start); 4088 rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start); 4089 rack_log_sack_passed(tp, rack, rsm); 4090 /* Is Reordering occuring? */ 4091 if (rsm->r_flags & RACK_SACK_PASSED) { 4092 counter_u64_add(rack_reorder_seen, 1); 4093 rack->r_ctl.rc_reorder_ts = cts; 4094 } 4095 rsm->r_flags |= RACK_ACKED; 4096 rsm->r_flags &= ~RACK_TLP; 4097 if (rsm->r_in_tmap) { 4098 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext); 4099 rsm->r_in_tmap = 0; 4100 } 4101 } 4102 if (end == rsm->r_end) { 4103 /* This block only - done */ 4104 goto out; 4105 } 4106 /* There is more not coverend by this rsm move on */ 4107 start = rsm->r_end; 4108 nrsm = TAILQ_NEXT(rsm, r_next); 4109 rsm = nrsm; 4110 times = 0; 4111 goto do_rest_ofb; 4112 } 4113 /* Ok we need to split off this one at the tail */ 4114 nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT); 4115 if (nrsm == NULL) { 4116 /* failed rrs what can we do but loose the sack info? */ 4117 goto out; 4118 } 4119 /* Clone it */ 4120 nrsm->r_start = end; 4121 nrsm->r_end = rsm->r_end; 4122 nrsm->r_rtr_bytes = 0; 4123 nrsm->r_rtr_cnt = rsm->r_rtr_cnt; 4124 nrsm->r_flags = rsm->r_flags; 4125 nrsm->r_sndcnt = rsm->r_sndcnt; 4126 for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { 4127 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; 4128 } 4129 /* The sack block does not cover this guy fully */ 4130 rsm->r_flags &= (~RACK_HAS_FIN); 4131 rsm->r_end = end; 4132 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_map, rsm, nrsm, r_next); 4133 if (rsm->r_in_tmap) { 4134 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 4135 nrsm->r_in_tmap = 1; 4136 } 4137 if (rsm->r_flags & RACK_ACKED) { 4138 /* Been here done that */ 4139 goto out; 4140 } 4141 rack_update_rtt(tp, rack, rsm, to, cts, SACKED); 4142 changed += (rsm->r_end - rsm->r_start); 4143 rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start); 4144 rack_log_sack_passed(tp, rack, rsm); 4145 /* Is Reordering occuring? */ 4146 if (rsm->r_flags & RACK_SACK_PASSED) { 4147 counter_u64_add(rack_reorder_seen, 1); 4148 rack->r_ctl.rc_reorder_ts = cts; 4149 } 4150 rsm->r_flags |= RACK_ACKED; 4151 rsm->r_flags &= ~RACK_TLP; 4152 if (rsm->r_in_tmap) { 4153 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext); 4154 rsm->r_in_tmap = 0; 4155 } 4156 out: 4157 if (rsm && (rsm->r_flags & RACK_ACKED)) { 4158 /* 4159 * Now can we merge this newly acked 4160 * block with either the previous or 4161 * next block? 4162 */ 4163 nrsm = TAILQ_NEXT(rsm, r_next); 4164 if (nrsm && 4165 (nrsm->r_flags & RACK_ACKED)) { 4166 /* yep this and next can be merged */ 4167 rsm = rack_merge_rsm(rack, rsm, nrsm); 4168 } 4169 /* Now what about the previous? */ 4170 nrsm = TAILQ_PREV(rsm, rack_head, r_next); 4171 if (nrsm && 4172 (nrsm->r_flags & RACK_ACKED)) { 4173 /* yep the previous and this can be merged */ 4174 rsm = rack_merge_rsm(rack, nrsm, rsm); 4175 } 4176 } 4177 if (used_ref == 0) { 4178 counter_u64_add(rack_sack_proc_all, 1); 4179 } else { 4180 counter_u64_add(rack_sack_proc_short, 1); 4181 } 4182 /* Save off where we last were */ 4183 if (rsm) 4184 rack->r_ctl.rc_sacklast = TAILQ_NEXT(rsm, r_next); 4185 else 4186 rack->r_ctl.rc_sacklast = NULL; 4187 *prsm = rsm; 4188 return (changed); 4189 } 4190 4191 static void inline 4192 rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack) 4193 { 4194 struct rack_sendmap *tmap; 4195 4196 tmap = NULL; 4197 while (rsm && (rsm->r_flags & RACK_ACKED)) { 4198 /* Its no longer sacked, mark it so */ 4199 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); 4200 #ifdef INVARIANTS 4201 if (rsm->r_in_tmap) { 4202 panic("rack:%p rsm:%p flags:0x%x in tmap?", 4203 rack, rsm, rsm->r_flags); 4204 } 4205 #endif 4206 rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS); 4207 /* Rebuild it into our tmap */ 4208 if (tmap == NULL) { 4209 TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext); 4210 tmap = rsm; 4211 } else { 4212 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext); 4213 tmap = rsm; 4214 } 4215 tmap->r_in_tmap = 1; 4216 rsm = TAILQ_NEXT(rsm, r_next); 4217 } 4218 /* 4219 * Now lets possibly clear the sack filter so we start 4220 * recognizing sacks that cover this area. 4221 */ 4222 if (rack_use_sack_filter) 4223 sack_filter_clear(&rack->r_ctl.rack_sf, th_ack); 4224 4225 } 4226 4227 static void 4228 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th) 4229 { 4230 uint32_t changed, last_seq, entered_recovery = 0; 4231 struct tcp_rack *rack; 4232 struct rack_sendmap *rsm; 4233 struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1]; 4234 register uint32_t th_ack; 4235 int32_t i, j, k, num_sack_blks = 0; 4236 uint32_t cts, acked, ack_point, sack_changed = 0; 4237 4238 INP_WLOCK_ASSERT(tp->t_inpcb); 4239 if (th->th_flags & TH_RST) { 4240 /* We don't log resets */ 4241 return; 4242 } 4243 rack = (struct tcp_rack *)tp->t_fb_ptr; 4244 cts = tcp_ts_getticks(); 4245 rsm = TAILQ_FIRST(&rack->r_ctl.rc_map); 4246 changed = 0; 4247 th_ack = th->th_ack; 4248 4249 if (SEQ_GT(th_ack, tp->snd_una)) { 4250 rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__); 4251 tp->t_acktime = ticks; 4252 } 4253 if (rsm && SEQ_GT(th_ack, rsm->r_start)) 4254 changed = th_ack - rsm->r_start; 4255 if (changed) { 4256 /* 4257 * The ACK point is advancing to th_ack, we must drop off 4258 * the packets in the rack log and calculate any eligble 4259 * RTT's. 4260 */ 4261 rack->r_wanted_output++; 4262 more: 4263 rsm = TAILQ_FIRST(&rack->r_ctl.rc_map); 4264 if (rsm == NULL) { 4265 if ((th_ack - 1) == tp->iss) { 4266 /* 4267 * For the SYN incoming case we will not 4268 * have called tcp_output for the sending of 4269 * the SYN, so there will be no map. All 4270 * other cases should probably be a panic. 4271 */ 4272 goto proc_sack; 4273 } 4274 if (tp->t_flags & TF_SENTFIN) { 4275 /* if we send a FIN we will not hav a map */ 4276 goto proc_sack; 4277 } 4278 #ifdef INVARIANTS 4279 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", 4280 tp, 4281 th, tp->t_state, rack, 4282 tp->snd_una, tp->snd_max, tp->snd_nxt, changed); 4283 #endif 4284 goto proc_sack; 4285 } 4286 if (SEQ_LT(th_ack, rsm->r_start)) { 4287 /* Huh map is missing this */ 4288 #ifdef INVARIANTS 4289 printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n", 4290 rsm->r_start, 4291 th_ack, tp->t_state, rack->r_state); 4292 #endif 4293 goto proc_sack; 4294 } 4295 rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED); 4296 /* Now do we consume the whole thing? */ 4297 if (SEQ_GEQ(th_ack, rsm->r_end)) { 4298 /* Its all consumed. */ 4299 uint32_t left; 4300 4301 rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; 4302 rsm->r_rtr_bytes = 0; 4303 TAILQ_REMOVE(&rack->r_ctl.rc_map, rsm, r_next); 4304 if (rsm->r_in_tmap) { 4305 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext); 4306 rsm->r_in_tmap = 0; 4307 } 4308 if (rack->r_ctl.rc_next == rsm) { 4309 /* scoot along the marker */ 4310 rack->r_ctl.rc_next = TAILQ_FIRST(&rack->r_ctl.rc_map); 4311 } 4312 if (rsm->r_flags & RACK_ACKED) { 4313 /* 4314 * It was acked on the scoreboard -- remove 4315 * it from total 4316 */ 4317 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); 4318 } else if (rsm->r_flags & RACK_SACK_PASSED) { 4319 /* 4320 * There are acked segments ACKED on the 4321 * scoreboard further up. We are seeing 4322 * reordering. 4323 */ 4324 counter_u64_add(rack_reorder_seen, 1); 4325 rsm->r_flags |= RACK_ACKED; 4326 rack->r_ctl.rc_reorder_ts = cts; 4327 } 4328 left = th_ack - rsm->r_end; 4329 if (rsm->r_rtr_cnt > 1) { 4330 /* 4331 * Technically we should make r_rtr_cnt be 4332 * monotonicly increasing and just mod it to 4333 * the timestamp it is replacing.. that way 4334 * we would have the last 3 retransmits. Now 4335 * rc_loss_count will be wrong if we 4336 * retransmit something more than 2 times in 4337 * recovery :( 4338 */ 4339 rack->r_ctl.rc_loss_count += (rsm->r_rtr_cnt - 1); 4340 } 4341 /* Free back to zone */ 4342 rack_free(rack, rsm); 4343 if (left) { 4344 goto more; 4345 } 4346 goto proc_sack; 4347 } 4348 if (rsm->r_flags & RACK_ACKED) { 4349 /* 4350 * It was acked on the scoreboard -- remove it from 4351 * total for the part being cum-acked. 4352 */ 4353 rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start); 4354 } 4355 rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; 4356 rsm->r_rtr_bytes = 0; 4357 rsm->r_start = th_ack; 4358 } 4359 proc_sack: 4360 /* Check for reneging */ 4361 rsm = TAILQ_FIRST(&rack->r_ctl.rc_map); 4362 if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) { 4363 /* 4364 * The peer has moved snd_una up to 4365 * the edge of this send, i.e. one 4366 * that it had previously acked. The only 4367 * way that can be true if the peer threw 4368 * away data (space issues) that it had 4369 * previously sacked (else it would have 4370 * given us snd_una up to (rsm->r_end). 4371 * We need to undo the acked markings here. 4372 * 4373 * Note we have to look to make sure th_ack is 4374 * our rsm->r_start in case we get an old ack 4375 * where th_ack is behind snd_una. 4376 */ 4377 rack_peer_reneges(rack, rsm, th->th_ack); 4378 } 4379 if ((to->to_flags & TOF_SACK) == 0) { 4380 /* We are done nothing left to log */ 4381 goto out; 4382 } 4383 rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_map, rack_sendmap, r_next); 4384 if (rsm) { 4385 last_seq = rsm->r_end; 4386 } else { 4387 last_seq = tp->snd_max; 4388 } 4389 /* Sack block processing */ 4390 if (SEQ_GT(th_ack, tp->snd_una)) 4391 ack_point = th_ack; 4392 else 4393 ack_point = tp->snd_una; 4394 for (i = 0; i < to->to_nsacks; i++) { 4395 bcopy((to->to_sacks + i * TCPOLEN_SACK), 4396 &sack, sizeof(sack)); 4397 sack.start = ntohl(sack.start); 4398 sack.end = ntohl(sack.end); 4399 if (SEQ_GT(sack.end, sack.start) && 4400 SEQ_GT(sack.start, ack_point) && 4401 SEQ_LT(sack.start, tp->snd_max) && 4402 SEQ_GT(sack.end, ack_point) && 4403 SEQ_LEQ(sack.end, tp->snd_max)) { 4404 if ((rack->r_ctl.rc_num_maps_alloced > rack_sack_block_limit) && 4405 (SEQ_LT(sack.end, last_seq)) && 4406 ((sack.end - sack.start) < (tp->t_maxseg / 8))) { 4407 /* 4408 * Not the last piece and its smaller than 4409 * 1/8th of a MSS. We ignore this. 4410 */ 4411 counter_u64_add(rack_runt_sacks, 1); 4412 continue; 4413 } 4414 sack_blocks[num_sack_blks] = sack; 4415 num_sack_blks++; 4416 } else if (SEQ_LEQ(sack.start, th_ack) && 4417 SEQ_LEQ(sack.end, th_ack)) { 4418 /* 4419 * Its a D-SACK block. 4420 */ 4421 /* tcp_record_dsack(sack.start, sack.end); */ 4422 } 4423 } 4424 if (num_sack_blks == 0) 4425 goto out; 4426 /* 4427 * Sort the SACK blocks so we can update the rack scoreboard with 4428 * just one pass. 4429 */ 4430 if (rack_use_sack_filter) { 4431 num_sack_blks = sack_filter_blks(&rack->r_ctl.rack_sf, sack_blocks, 4432 num_sack_blks, th->th_ack); 4433 ctf_log_sack_filter(rack->rc_tp, num_sack_blks, sack_blocks); 4434 } 4435 if (num_sack_blks < 2) { 4436 goto do_sack_work; 4437 } 4438 /* Sort the sacks */ 4439 for (i = 0; i < num_sack_blks; i++) { 4440 for (j = i + 1; j < num_sack_blks; j++) { 4441 if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) { 4442 sack = sack_blocks[i]; 4443 sack_blocks[i] = sack_blocks[j]; 4444 sack_blocks[j] = sack; 4445 } 4446 } 4447 } 4448 /* 4449 * Now are any of the sack block ends the same (yes some 4450 * implememtations send these)? 4451 */ 4452 again: 4453 if (num_sack_blks > 1) { 4454 for (i = 0; i < num_sack_blks; i++) { 4455 for (j = i + 1; j < num_sack_blks; j++) { 4456 if (sack_blocks[i].end == sack_blocks[j].end) { 4457 /* 4458 * Ok these two have the same end we 4459 * want the smallest end and then 4460 * throw away the larger and start 4461 * again. 4462 */ 4463 if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) { 4464 /* 4465 * The second block covers 4466 * more area use that 4467 */ 4468 sack_blocks[i].start = sack_blocks[j].start; 4469 } 4470 /* 4471 * Now collapse out the dup-sack and 4472 * lower the count 4473 */ 4474 for (k = (j + 1); k < num_sack_blks; k++) { 4475 sack_blocks[j].start = sack_blocks[k].start; 4476 sack_blocks[j].end = sack_blocks[k].end; 4477 j++; 4478 } 4479 num_sack_blks--; 4480 goto again; 4481 } 4482 } 4483 } 4484 } 4485 do_sack_work: 4486 rsm = rack->r_ctl.rc_sacklast; 4487 for (i = 0; i < num_sack_blks; i++) { 4488 acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts); 4489 if (acked) { 4490 rack->r_wanted_output++; 4491 changed += acked; 4492 sack_changed += acked; 4493 } 4494 } 4495 out: 4496 if (changed) { 4497 /* Something changed cancel the rack timer */ 4498 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 4499 } 4500 if ((sack_changed) && (!IN_RECOVERY(tp->t_flags))) { 4501 /* 4502 * Ok we have a high probability that we need to go in to 4503 * recovery since we have data sack'd 4504 */ 4505 struct rack_sendmap *rsm; 4506 uint32_t tsused; 4507 4508 tsused = tcp_ts_getticks(); 4509 rsm = tcp_rack_output(tp, rack, tsused); 4510 if (rsm) { 4511 /* Enter recovery */ 4512 rack->r_ctl.rc_rsm_start = rsm->r_start; 4513 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd; 4514 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh; 4515 entered_recovery = 1; 4516 rack_cong_signal(tp, NULL, CC_NDUPACK); 4517 /* 4518 * When we enter recovery we need to assure we send 4519 * one packet. 4520 */ 4521 rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; 4522 rack->r_timer_override = 1; 4523 } 4524 } 4525 if (IN_RECOVERY(tp->t_flags) && (entered_recovery == 0)) { 4526 /* Deal with changed an PRR here (in recovery only) */ 4527 uint32_t pipe, snd_una; 4528 4529 rack->r_ctl.rc_prr_delivered += changed; 4530 /* Compute prr_sndcnt */ 4531 if (SEQ_GT(tp->snd_una, th_ack)) { 4532 snd_una = tp->snd_una; 4533 } else { 4534 snd_una = th_ack; 4535 } 4536 pipe = ((tp->snd_max - snd_una) - rack->r_ctl.rc_sacked) + rack->r_ctl.rc_holes_rxt; 4537 if (pipe > tp->snd_ssthresh) { 4538 long sndcnt; 4539 4540 sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh; 4541 if (rack->r_ctl.rc_prr_recovery_fs > 0) 4542 sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs; 4543 else { 4544 rack->r_ctl.rc_prr_sndcnt = 0; 4545 sndcnt = 0; 4546 } 4547 sndcnt++; 4548 if (sndcnt > (long)rack->r_ctl.rc_prr_out) 4549 sndcnt -= rack->r_ctl.rc_prr_out; 4550 else 4551 sndcnt = 0; 4552 rack->r_ctl.rc_prr_sndcnt = sndcnt; 4553 } else { 4554 uint32_t limit; 4555 4556 if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out) 4557 limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out); 4558 else 4559 limit = 0; 4560 if (changed > limit) 4561 limit = changed; 4562 limit += tp->t_maxseg; 4563 if (tp->snd_ssthresh > pipe) { 4564 rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit); 4565 } else { 4566 rack->r_ctl.rc_prr_sndcnt = min(0, limit); 4567 } 4568 } 4569 if (rack->r_ctl.rc_prr_sndcnt >= tp->t_maxseg) { 4570 rack->r_timer_override = 1; 4571 } 4572 } 4573 } 4574 4575 /* 4576 * Return value of 1, we do not need to call rack_process_data(). 4577 * return value of 0, rack_process_data can be called. 4578 * For ret_val if its 0 the TCP is locked, if its non-zero 4579 * its unlocked and probably unsafe to touch the TCB. 4580 */ 4581 static int 4582 rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so, 4583 struct tcpcb *tp, struct tcpopt *to, 4584 uint32_t tiwin, int32_t tlen, 4585 int32_t * ofia, int32_t thflags, int32_t * ret_val) 4586 { 4587 int32_t ourfinisacked = 0; 4588 int32_t nsegs, acked_amount; 4589 int32_t acked; 4590 struct mbuf *mfree; 4591 struct tcp_rack *rack; 4592 int32_t recovery = 0; 4593 4594 rack = (struct tcp_rack *)tp->t_fb_ptr; 4595 if (SEQ_GT(th->th_ack, tp->snd_max)) { 4596 rack_do_dropafterack(m, tp, th, thflags, tlen, ret_val); 4597 return (1); 4598 } 4599 if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) { 4600 rack_log_ack(tp, to, th); 4601 } 4602 if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) { 4603 /* 4604 * Old ack, behind (or duplicate to) the last one rcv'd 4605 * Note: Should mark reordering is occuring! We should also 4606 * look for sack blocks arriving e.g. ack 1, 4-4 then ack 1, 4607 * 3-3, 4-4 would be reording. As well as ack 1, 3-3 <no 4608 * retran and> ack 3 4609 */ 4610 return (0); 4611 } 4612 /* 4613 * If we reach this point, ACK is not a duplicate, i.e., it ACKs 4614 * something we sent. 4615 */ 4616 if (tp->t_flags & TF_NEEDSYN) { 4617 /* 4618 * T/TCP: Connection was half-synchronized, and our SYN has 4619 * been ACK'd (so connection is now fully synchronized). Go 4620 * to non-starred state, increment snd_una for ACK of SYN, 4621 * and check if we can do window scaling. 4622 */ 4623 tp->t_flags &= ~TF_NEEDSYN; 4624 tp->snd_una++; 4625 /* Do window scaling? */ 4626 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == 4627 (TF_RCVD_SCALE | TF_REQ_SCALE)) { 4628 tp->rcv_scale = tp->request_r_scale; 4629 /* Send window already scaled. */ 4630 } 4631 } 4632 nsegs = max(1, m->m_pkthdr.lro_nsegs); 4633 INP_WLOCK_ASSERT(tp->t_inpcb); 4634 4635 acked = BYTES_THIS_ACK(tp, th); 4636 TCPSTAT_ADD(tcps_rcvackpack, nsegs); 4637 TCPSTAT_ADD(tcps_rcvackbyte, acked); 4638 4639 /* 4640 * If we just performed our first retransmit, and the ACK arrives 4641 * within our recovery window, then it was a mistake to do the 4642 * retransmit in the first place. Recover our original cwnd and 4643 * ssthresh, and proceed to transmit where we left off. 4644 */ 4645 if (tp->t_flags & TF_PREVVALID) { 4646 tp->t_flags &= ~TF_PREVVALID; 4647 if (tp->t_rxtshift == 1 && 4648 (int)(ticks - tp->t_badrxtwin) < 0) 4649 rack_cong_signal(tp, th, CC_RTO_ERR); 4650 } 4651 /* 4652 * If we have a timestamp reply, update smoothed round trip time. If 4653 * no timestamp is present but transmit timer is running and timed 4654 * sequence number was acked, update smoothed round trip time. Since 4655 * we now have an rtt measurement, cancel the timer backoff (cf., 4656 * Phil Karn's retransmit alg.). Recompute the initial retransmit 4657 * timer. 4658 * 4659 * Some boxes send broken timestamp replies during the SYN+ACK 4660 * phase, ignore timestamps of 0 or we could calculate a huge RTT 4661 * and blow up the retransmit timer. 4662 */ 4663 /* 4664 * If all outstanding data is acked, stop retransmit timer and 4665 * remember to restart (more output or persist). If there is more 4666 * data to be acked, restart retransmit timer, using current 4667 * (possibly backed-off) value. 4668 */ 4669 if (th->th_ack == tp->snd_max) { 4670 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 4671 rack->r_wanted_output++; 4672 } 4673 /* 4674 * If no data (only SYN) was ACK'd, skip rest of ACK processing. 4675 */ 4676 if (acked == 0) { 4677 if (ofia) 4678 *ofia = ourfinisacked; 4679 return (0); 4680 } 4681 if (rack->r_ctl.rc_early_recovery) { 4682 if (IN_RECOVERY(tp->t_flags)) { 4683 if (SEQ_LT(th->th_ack, tp->snd_recover) && 4684 (SEQ_LT(th->th_ack, tp->snd_max))) { 4685 tcp_rack_partialack(tp, th); 4686 } else { 4687 rack_post_recovery(tp, th); 4688 recovery = 1; 4689 } 4690 } 4691 } 4692 /* 4693 * Let the congestion control algorithm update congestion control 4694 * related information. This typically means increasing the 4695 * congestion window. 4696 */ 4697 rack_ack_received(tp, rack, th, nsegs, CC_ACK, recovery); 4698 SOCKBUF_LOCK(&so->so_snd); 4699 acked_amount = min(acked, (int)sbavail(&so->so_snd)); 4700 tp->snd_wnd -= acked_amount; 4701 mfree = sbcut_locked(&so->so_snd, acked_amount); 4702 if ((sbused(&so->so_snd) == 0) && 4703 (acked > acked_amount) && 4704 (tp->t_state >= TCPS_FIN_WAIT_1)) { 4705 ourfinisacked = 1; 4706 } 4707 /* NB: sowwakeup_locked() does an implicit unlock. */ 4708 sowwakeup_locked(so); 4709 m_freem(mfree); 4710 if (rack->r_ctl.rc_early_recovery == 0) { 4711 if (IN_RECOVERY(tp->t_flags)) { 4712 if (SEQ_LT(th->th_ack, tp->snd_recover) && 4713 (SEQ_LT(th->th_ack, tp->snd_max))) { 4714 tcp_rack_partialack(tp, th); 4715 } else { 4716 rack_post_recovery(tp, th); 4717 } 4718 } 4719 } 4720 tp->snd_una = th->th_ack; 4721 if (SEQ_GT(tp->snd_una, tp->snd_recover)) 4722 tp->snd_recover = tp->snd_una; 4723 4724 if (SEQ_LT(tp->snd_nxt, tp->snd_una)) { 4725 tp->snd_nxt = tp->snd_una; 4726 } 4727 if (tp->snd_una == tp->snd_max) { 4728 /* Nothing left outstanding */ 4729 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__); 4730 tp->t_acktime = 0; 4731 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 4732 /* Set need output so persist might get set */ 4733 rack->r_wanted_output++; 4734 if (rack_use_sack_filter) 4735 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una); 4736 if ((tp->t_state >= TCPS_FIN_WAIT_1) && 4737 (sbavail(&so->so_snd) == 0) && 4738 (tp->t_flags2 & TF2_DROP_AF_DATA)) { 4739 /* 4740 * The socket was gone and the 4741 * peer sent data, time to 4742 * reset him. 4743 */ 4744 *ret_val = 1; 4745 tp = tcp_close(tp); 4746 rack_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen); 4747 return (1); 4748 } 4749 } 4750 if (ofia) 4751 *ofia = ourfinisacked; 4752 return (0); 4753 } 4754 4755 4756 /* 4757 * Return value of 1, the TCB is unlocked and most 4758 * likely gone, return value of 0, the TCP is still 4759 * locked. 4760 */ 4761 static int 4762 rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so, 4763 struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, 4764 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 4765 { 4766 /* 4767 * Update window information. Don't look at window if no ACK: TAC's 4768 * send garbage on first SYN. 4769 */ 4770 int32_t nsegs; 4771 #ifdef TCP_RFC7413 4772 int32_t tfo_syn; 4773 #else 4774 #define tfo_syn (FALSE) 4775 #endif 4776 struct tcp_rack *rack; 4777 4778 rack = (struct tcp_rack *)tp->t_fb_ptr; 4779 INP_WLOCK_ASSERT(tp->t_inpcb); 4780 nsegs = max(1, m->m_pkthdr.lro_nsegs); 4781 if ((thflags & TH_ACK) && 4782 (SEQ_LT(tp->snd_wl1, th->th_seq) || 4783 (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) || 4784 (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) { 4785 /* keep track of pure window updates */ 4786 if (tlen == 0 && 4787 tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd) 4788 TCPSTAT_INC(tcps_rcvwinupd); 4789 tp->snd_wnd = tiwin; 4790 tp->snd_wl1 = th->th_seq; 4791 tp->snd_wl2 = th->th_ack; 4792 if (tp->snd_wnd > tp->max_sndwnd) 4793 tp->max_sndwnd = tp->snd_wnd; 4794 rack->r_wanted_output++; 4795 } else if (thflags & TH_ACK) { 4796 if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) { 4797 tp->snd_wnd = tiwin; 4798 tp->snd_wl1 = th->th_seq; 4799 tp->snd_wl2 = th->th_ack; 4800 } 4801 } 4802 /* Was persist timer active and now we have window space? */ 4803 if ((rack->rc_in_persist != 0) && tp->snd_wnd) { 4804 rack_exit_persist(tp, rack); 4805 tp->snd_nxt = tp->snd_max; 4806 /* Make sure we output to start the timer */ 4807 rack->r_wanted_output++; 4808 } 4809 if (tp->t_flags2 & TF2_DROP_AF_DATA) { 4810 m_freem(m); 4811 return (0); 4812 } 4813 /* 4814 * Process segments with URG. 4815 */ 4816 if ((thflags & TH_URG) && th->th_urp && 4817 TCPS_HAVERCVDFIN(tp->t_state) == 0) { 4818 /* 4819 * This is a kludge, but if we receive and accept random 4820 * urgent pointers, we'll crash in soreceive. It's hard to 4821 * imagine someone actually wanting to send this much urgent 4822 * data. 4823 */ 4824 SOCKBUF_LOCK(&so->so_rcv); 4825 if (th->th_urp + sbavail(&so->so_rcv) > sb_max) { 4826 th->th_urp = 0; /* XXX */ 4827 thflags &= ~TH_URG; /* XXX */ 4828 SOCKBUF_UNLOCK(&so->so_rcv); /* XXX */ 4829 goto dodata; /* XXX */ 4830 } 4831 /* 4832 * If this segment advances the known urgent pointer, then 4833 * mark the data stream. This should not happen in 4834 * CLOSE_WAIT, CLOSING, LAST_ACK or TIME_WAIT STATES since a 4835 * FIN has been received from the remote side. In these 4836 * states we ignore the URG. 4837 * 4838 * According to RFC961 (Assigned Protocols), the urgent 4839 * pointer points to the last octet of urgent data. We 4840 * continue, however, to consider it to indicate the first 4841 * octet of data past the urgent section as the original 4842 * spec states (in one of two places). 4843 */ 4844 if (SEQ_GT(th->th_seq + th->th_urp, tp->rcv_up)) { 4845 tp->rcv_up = th->th_seq + th->th_urp; 4846 so->so_oobmark = sbavail(&so->so_rcv) + 4847 (tp->rcv_up - tp->rcv_nxt) - 1; 4848 if (so->so_oobmark == 0) 4849 so->so_rcv.sb_state |= SBS_RCVATMARK; 4850 sohasoutofband(so); 4851 tp->t_oobflags &= ~(TCPOOB_HAVEDATA | TCPOOB_HADDATA); 4852 } 4853 SOCKBUF_UNLOCK(&so->so_rcv); 4854 /* 4855 * Remove out of band data so doesn't get presented to user. 4856 * This can happen independent of advancing the URG pointer, 4857 * but if two URG's are pending at once, some out-of-band 4858 * data may creep in... ick. 4859 */ 4860 if (th->th_urp <= (uint32_t) tlen && 4861 !(so->so_options & SO_OOBINLINE)) { 4862 /* hdr drop is delayed */ 4863 tcp_pulloutofband(so, th, m, drop_hdrlen); 4864 } 4865 } else { 4866 /* 4867 * If no out of band data is expected, pull receive urgent 4868 * pointer along with the receive window. 4869 */ 4870 if (SEQ_GT(tp->rcv_nxt, tp->rcv_up)) 4871 tp->rcv_up = tp->rcv_nxt; 4872 } 4873 dodata: /* XXX */ 4874 INP_WLOCK_ASSERT(tp->t_inpcb); 4875 4876 /* 4877 * Process the segment text, merging it into the TCP sequencing 4878 * queue, and arranging for acknowledgment of receipt if necessary. 4879 * This process logically involves adjusting tp->rcv_wnd as data is 4880 * presented to the user (this happens in tcp_usrreq.c, case 4881 * PRU_RCVD). If a FIN has already been received on this connection 4882 * then we just ignore the text. 4883 */ 4884 #ifdef TCP_RFC7413 4885 tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) && 4886 (tp->t_flags & TF_FASTOPEN)); 4887 #endif 4888 if ((tlen || (thflags & TH_FIN) || tfo_syn) && 4889 TCPS_HAVERCVDFIN(tp->t_state) == 0) { 4890 tcp_seq save_start = th->th_seq; 4891 tcp_seq save_rnxt = tp->rcv_nxt; 4892 int save_tlen = tlen; 4893 4894 m_adj(m, drop_hdrlen); /* delayed header drop */ 4895 /* 4896 * Insert segment which includes th into TCP reassembly 4897 * queue with control block tp. Set thflags to whether 4898 * reassembly now includes a segment with FIN. This handles 4899 * the common case inline (segment is the next to be 4900 * received on an established connection, and the queue is 4901 * empty), avoiding linkage into and removal from the queue 4902 * and repetition of various conversions. Set DELACK for 4903 * segments received in order, but ack immediately when 4904 * segments are out of order (so fast retransmit can work). 4905 */ 4906 if (th->th_seq == tp->rcv_nxt && 4907 SEGQ_EMPTY(tp) && 4908 (TCPS_HAVEESTABLISHED(tp->t_state) || 4909 tfo_syn)) { 4910 if (DELAY_ACK(tp, tlen) || tfo_syn) { 4911 rack_timer_cancel(tp, rack, 4912 rack->r_ctl.rc_rcvtime, __LINE__); 4913 tp->t_flags |= TF_DELACK; 4914 } else { 4915 rack->r_wanted_output++; 4916 tp->t_flags |= TF_ACKNOW; 4917 } 4918 tp->rcv_nxt += tlen; 4919 thflags = th->th_flags & TH_FIN; 4920 TCPSTAT_ADD(tcps_rcvpack, nsegs); 4921 TCPSTAT_ADD(tcps_rcvbyte, tlen); 4922 SOCKBUF_LOCK(&so->so_rcv); 4923 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) 4924 m_freem(m); 4925 else 4926 sbappendstream_locked(&so->so_rcv, m, 0); 4927 /* NB: sorwakeup_locked() does an implicit unlock. */ 4928 sorwakeup_locked(so); 4929 } else { 4930 /* 4931 * XXX: Due to the header drop above "th" is 4932 * theoretically invalid by now. Fortunately 4933 * m_adj() doesn't actually frees any mbufs when 4934 * trimming from the head. 4935 */ 4936 tcp_seq temp = save_start; 4937 thflags = tcp_reass(tp, th, &temp, &tlen, m); 4938 tp->t_flags |= TF_ACKNOW; 4939 } 4940 if (((tlen == 0) && (save_tlen > 0) && 4941 (SEQ_LT(save_start, save_rnxt)))) { 4942 /* 4943 * DSACK actually handled in the fastpath 4944 * above. 4945 */ 4946 tcp_update_sack_list(tp, save_start, 4947 save_start + save_tlen); 4948 } else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) { 4949 /* 4950 * Cleaning sackblks by using zero length 4951 * update. 4952 */ 4953 if ((tp->rcv_numsacks >= 1) && 4954 (tp->sackblks[0].end == save_start)) { 4955 /* partial overlap, recorded at todrop above */ 4956 tcp_update_sack_list(tp, tp->sackblks[0].start, 4957 tp->sackblks[0].end); 4958 } else { 4959 tcp_update_dsack_list(tp, save_start, 4960 save_start + save_tlen); 4961 } 4962 } else if ((tlen > 0) && (tlen >= save_tlen)) { 4963 /* Update of sackblks. */ 4964 tcp_update_dsack_list(tp, save_start, 4965 save_start + save_tlen); 4966 } else if (tlen > 0) { 4967 tcp_update_dsack_list(tp, save_start, 4968 save_start + tlen); 4969 } 4970 } else { 4971 m_freem(m); 4972 thflags &= ~TH_FIN; 4973 } 4974 4975 /* 4976 * If FIN is received ACK the FIN and let the user know that the 4977 * connection is closing. 4978 */ 4979 if (thflags & TH_FIN) { 4980 if (TCPS_HAVERCVDFIN(tp->t_state) == 0) { 4981 socantrcvmore(so); 4982 /* 4983 * If connection is half-synchronized (ie NEEDSYN 4984 * flag on) then delay ACK, so it may be piggybacked 4985 * when SYN is sent. Otherwise, since we received a 4986 * FIN then no more input can be expected, send ACK 4987 * now. 4988 */ 4989 if (tp->t_flags & TF_NEEDSYN) { 4990 rack_timer_cancel(tp, rack, 4991 rack->r_ctl.rc_rcvtime, __LINE__); 4992 tp->t_flags |= TF_DELACK; 4993 } else { 4994 tp->t_flags |= TF_ACKNOW; 4995 } 4996 tp->rcv_nxt++; 4997 } 4998 switch (tp->t_state) { 4999 5000 /* 5001 * In SYN_RECEIVED and ESTABLISHED STATES enter the 5002 * CLOSE_WAIT state. 5003 */ 5004 case TCPS_SYN_RECEIVED: 5005 tp->t_starttime = ticks; 5006 /* FALLTHROUGH */ 5007 case TCPS_ESTABLISHED: 5008 rack_timer_cancel(tp, rack, 5009 rack->r_ctl.rc_rcvtime, __LINE__); 5010 tcp_state_change(tp, TCPS_CLOSE_WAIT); 5011 break; 5012 5013 /* 5014 * If still in FIN_WAIT_1 STATE FIN has not been 5015 * acked so enter the CLOSING state. 5016 */ 5017 case TCPS_FIN_WAIT_1: 5018 rack_timer_cancel(tp, rack, 5019 rack->r_ctl.rc_rcvtime, __LINE__); 5020 tcp_state_change(tp, TCPS_CLOSING); 5021 break; 5022 5023 /* 5024 * In FIN_WAIT_2 state enter the TIME_WAIT state, 5025 * starting the time-wait timer, turning off the 5026 * other standard timers. 5027 */ 5028 case TCPS_FIN_WAIT_2: 5029 rack_timer_cancel(tp, rack, 5030 rack->r_ctl.rc_rcvtime, __LINE__); 5031 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 5032 tcp_twstart(tp); 5033 return (1); 5034 } 5035 } 5036 /* 5037 * Return any desired output. 5038 */ 5039 if ((tp->t_flags & TF_ACKNOW) || 5040 (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) { 5041 rack->r_wanted_output++; 5042 } 5043 INP_WLOCK_ASSERT(tp->t_inpcb); 5044 return (0); 5045 } 5046 5047 /* 5048 * Here nothing is really faster, its just that we 5049 * have broken out the fast-data path also just like 5050 * the fast-ack. 5051 */ 5052 static int 5053 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so, 5054 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 5055 uint32_t tiwin, int32_t nxt_pkt) 5056 { 5057 int32_t nsegs; 5058 int32_t newsize = 0; /* automatic sockbuf scaling */ 5059 struct tcp_rack *rack; 5060 #ifdef TCPDEBUG 5061 /* 5062 * The size of tcp_saveipgen must be the size of the max ip header, 5063 * now IPv6. 5064 */ 5065 u_char tcp_saveipgen[IP6_HDR_LEN]; 5066 struct tcphdr tcp_savetcp; 5067 short ostate = 0; 5068 5069 #endif 5070 /* 5071 * If last ACK falls within this segment's sequence numbers, record 5072 * the timestamp. NOTE that the test is modified according to the 5073 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26). 5074 */ 5075 if (__predict_false(th->th_seq != tp->rcv_nxt)) { 5076 return (0); 5077 } 5078 if (__predict_false(tp->snd_nxt != tp->snd_max)) { 5079 return (0); 5080 } 5081 if (tiwin && tiwin != tp->snd_wnd) { 5082 return (0); 5083 } 5084 if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) { 5085 return (0); 5086 } 5087 if (__predict_false((to->to_flags & TOF_TS) && 5088 (TSTMP_LT(to->to_tsval, tp->ts_recent)))) { 5089 return (0); 5090 } 5091 if (__predict_false((th->th_ack != tp->snd_una))) { 5092 return (0); 5093 } 5094 if (__predict_false(tlen > sbspace(&so->so_rcv))) { 5095 return (0); 5096 } 5097 if ((to->to_flags & TOF_TS) != 0 && 5098 SEQ_LEQ(th->th_seq, tp->last_ack_sent)) { 5099 tp->ts_recent_age = tcp_ts_getticks(); 5100 tp->ts_recent = to->to_tsval; 5101 } 5102 rack = (struct tcp_rack *)tp->t_fb_ptr; 5103 /* 5104 * This is a pure, in-sequence data packet with nothing on the 5105 * reassembly queue and we have enough buffer space to take it. 5106 */ 5107 nsegs = max(1, m->m_pkthdr.lro_nsegs); 5108 5109 5110 /* Clean receiver SACK report if present */ 5111 if (tp->rcv_numsacks) 5112 tcp_clean_sackreport(tp); 5113 TCPSTAT_INC(tcps_preddat); 5114 tp->rcv_nxt += tlen; 5115 /* 5116 * Pull snd_wl1 up to prevent seq wrap relative to th_seq. 5117 */ 5118 tp->snd_wl1 = th->th_seq; 5119 /* 5120 * Pull rcv_up up to prevent seq wrap relative to rcv_nxt. 5121 */ 5122 tp->rcv_up = tp->rcv_nxt; 5123 TCPSTAT_ADD(tcps_rcvpack, nsegs); 5124 TCPSTAT_ADD(tcps_rcvbyte, tlen); 5125 #ifdef TCPDEBUG 5126 if (so->so_options & SO_DEBUG) 5127 tcp_trace(TA_INPUT, ostate, tp, 5128 (void *)tcp_saveipgen, &tcp_savetcp, 0); 5129 #endif 5130 newsize = tcp_autorcvbuf(m, th, so, tp, tlen); 5131 5132 /* Add data to socket buffer. */ 5133 SOCKBUF_LOCK(&so->so_rcv); 5134 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 5135 m_freem(m); 5136 } else { 5137 /* 5138 * Set new socket buffer size. Give up when limit is 5139 * reached. 5140 */ 5141 if (newsize) 5142 if (!sbreserve_locked(&so->so_rcv, 5143 newsize, so, NULL)) 5144 so->so_rcv.sb_flags &= ~SB_AUTOSIZE; 5145 m_adj(m, drop_hdrlen); /* delayed header drop */ 5146 sbappendstream_locked(&so->so_rcv, m, 0); 5147 rack_calc_rwin(so, tp); 5148 } 5149 /* NB: sorwakeup_locked() does an implicit unlock. */ 5150 sorwakeup_locked(so); 5151 if (DELAY_ACK(tp, tlen)) { 5152 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 5153 tp->t_flags |= TF_DELACK; 5154 } else { 5155 tp->t_flags |= TF_ACKNOW; 5156 rack->r_wanted_output++; 5157 } 5158 if ((tp->snd_una == tp->snd_max) && rack_use_sack_filter) 5159 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una); 5160 return (1); 5161 } 5162 5163 /* 5164 * This subfunction is used to try to highly optimize the 5165 * fast path. We again allow window updates that are 5166 * in sequence to remain in the fast-path. We also add 5167 * in the __predict's to attempt to help the compiler. 5168 * Note that if we return a 0, then we can *not* process 5169 * it and the caller should push the packet into the 5170 * slow-path. 5171 */ 5172 static int 5173 rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so, 5174 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 5175 uint32_t tiwin, int32_t nxt_pkt, uint32_t cts) 5176 { 5177 int32_t acked; 5178 int32_t nsegs; 5179 5180 #ifdef TCPDEBUG 5181 /* 5182 * The size of tcp_saveipgen must be the size of the max ip header, 5183 * now IPv6. 5184 */ 5185 u_char tcp_saveipgen[IP6_HDR_LEN]; 5186 struct tcphdr tcp_savetcp; 5187 short ostate = 0; 5188 5189 #endif 5190 struct tcp_rack *rack; 5191 5192 if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) { 5193 /* Old ack, behind (or duplicate to) the last one rcv'd */ 5194 return (0); 5195 } 5196 if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) { 5197 /* Above what we have sent? */ 5198 return (0); 5199 } 5200 if (__predict_false(tp->snd_nxt != tp->snd_max)) { 5201 /* We are retransmitting */ 5202 return (0); 5203 } 5204 if (__predict_false(tiwin == 0)) { 5205 /* zero window */ 5206 return (0); 5207 } 5208 if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) { 5209 /* We need a SYN or a FIN, unlikely.. */ 5210 return (0); 5211 } 5212 if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) { 5213 /* Timestamp is behind .. old ack with seq wrap? */ 5214 return (0); 5215 } 5216 if (__predict_false(IN_RECOVERY(tp->t_flags))) { 5217 /* Still recovering */ 5218 return (0); 5219 } 5220 rack = (struct tcp_rack *)tp->t_fb_ptr; 5221 if (rack->r_ctl.rc_sacked) { 5222 /* We have sack holes on our scoreboard */ 5223 return (0); 5224 } 5225 /* Ok if we reach here, we can process a fast-ack */ 5226 nsegs = max(1, m->m_pkthdr.lro_nsegs); 5227 rack_log_ack(tp, to, th); 5228 /* Did the window get updated? */ 5229 if (tiwin != tp->snd_wnd) { 5230 tp->snd_wnd = tiwin; 5231 tp->snd_wl1 = th->th_seq; 5232 if (tp->snd_wnd > tp->max_sndwnd) 5233 tp->max_sndwnd = tp->snd_wnd; 5234 } 5235 if ((rack->rc_in_persist != 0) && (tp->snd_wnd >= tp->t_maxseg)) { 5236 rack_exit_persist(tp, rack); 5237 } 5238 /* 5239 * If last ACK falls within this segment's sequence numbers, record 5240 * the timestamp. NOTE that the test is modified according to the 5241 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26). 5242 */ 5243 if ((to->to_flags & TOF_TS) != 0 && 5244 SEQ_LEQ(th->th_seq, tp->last_ack_sent)) { 5245 tp->ts_recent_age = tcp_ts_getticks(); 5246 tp->ts_recent = to->to_tsval; 5247 } 5248 /* 5249 * This is a pure ack for outstanding data. 5250 */ 5251 TCPSTAT_INC(tcps_predack); 5252 5253 /* 5254 * "bad retransmit" recovery. 5255 */ 5256 if (tp->t_flags & TF_PREVVALID) { 5257 tp->t_flags &= ~TF_PREVVALID; 5258 if (tp->t_rxtshift == 1 && 5259 (int)(ticks - tp->t_badrxtwin) < 0) 5260 rack_cong_signal(tp, th, CC_RTO_ERR); 5261 } 5262 /* 5263 * Recalculate the transmit timer / rtt. 5264 * 5265 * Some boxes send broken timestamp replies during the SYN+ACK 5266 * phase, ignore timestamps of 0 or we could calculate a huge RTT 5267 * and blow up the retransmit timer. 5268 */ 5269 acked = BYTES_THIS_ACK(tp, th); 5270 5271 #ifdef TCP_HHOOK 5272 /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */ 5273 hhook_run_tcp_est_in(tp, th, to); 5274 #endif 5275 5276 TCPSTAT_ADD(tcps_rcvackpack, nsegs); 5277 TCPSTAT_ADD(tcps_rcvackbyte, acked); 5278 sbdrop(&so->so_snd, acked); 5279 /* 5280 * Let the congestion control algorithm update congestion control 5281 * related information. This typically means increasing the 5282 * congestion window. 5283 */ 5284 rack_ack_received(tp, rack, th, nsegs, CC_ACK, 0); 5285 5286 tp->snd_una = th->th_ack; 5287 /* 5288 * Pull snd_wl2 up to prevent seq wrap relative to th_ack. 5289 */ 5290 tp->snd_wl2 = th->th_ack; 5291 tp->t_dupacks = 0; 5292 m_freem(m); 5293 /* ND6_HINT(tp); *//* Some progress has been made. */ 5294 5295 /* 5296 * If all outstanding data are acked, stop retransmit timer, 5297 * otherwise restart timer using current (possibly backed-off) 5298 * value. If process is waiting for space, wakeup/selwakeup/signal. 5299 * If data are ready to send, let tcp_output decide between more 5300 * output or persist. 5301 */ 5302 #ifdef TCPDEBUG 5303 if (so->so_options & SO_DEBUG) 5304 tcp_trace(TA_INPUT, ostate, tp, 5305 (void *)tcp_saveipgen, 5306 &tcp_savetcp, 0); 5307 #endif 5308 if (tp->snd_una == tp->snd_max) { 5309 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__); 5310 tp->t_acktime = 0; 5311 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 5312 } 5313 /* Wake up the socket if we have room to write more */ 5314 sowwakeup(so); 5315 if (sbavail(&so->so_snd)) { 5316 rack->r_wanted_output++; 5317 } 5318 return (1); 5319 } 5320 5321 /* 5322 * Return value of 1, the TCB is unlocked and most 5323 * likely gone, return value of 0, the TCP is still 5324 * locked. 5325 */ 5326 static int 5327 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so, 5328 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 5329 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 5330 { 5331 int32_t ret_val = 0; 5332 int32_t todrop; 5333 int32_t ourfinisacked = 0; 5334 5335 rack_calc_rwin(so, tp); 5336 /* 5337 * If the state is SYN_SENT: if seg contains an ACK, but not for our 5338 * SYN, drop the input. if seg contains a RST, then drop the 5339 * connection. if seg does not contain SYN, then drop it. Otherwise 5340 * this is an acceptable SYN segment initialize tp->rcv_nxt and 5341 * tp->irs if seg contains ack then advance tp->snd_una if seg 5342 * contains an ECE and ECN support is enabled, the stream is ECN 5343 * capable. if SYN has been acked change to ESTABLISHED else 5344 * SYN_RCVD state arrange for segment to be acked (eventually) 5345 * continue processing rest of data/controls, beginning with URG 5346 */ 5347 if ((thflags & TH_ACK) && 5348 (SEQ_LEQ(th->th_ack, tp->iss) || 5349 SEQ_GT(th->th_ack, tp->snd_max))) { 5350 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 5351 return (1); 5352 } 5353 if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) { 5354 TCP_PROBE5(connect__refused, NULL, tp, 5355 mtod(m, const char *), tp, th); 5356 tp = tcp_drop(tp, ECONNREFUSED); 5357 rack_do_drop(m, tp); 5358 return (1); 5359 } 5360 if (thflags & TH_RST) { 5361 rack_do_drop(m, tp); 5362 return (1); 5363 } 5364 if (!(thflags & TH_SYN)) { 5365 rack_do_drop(m, tp); 5366 return (1); 5367 } 5368 tp->irs = th->th_seq; 5369 tcp_rcvseqinit(tp); 5370 if (thflags & TH_ACK) { 5371 TCPSTAT_INC(tcps_connects); 5372 soisconnected(so); 5373 #ifdef MAC 5374 mac_socketpeer_set_from_mbuf(m, so); 5375 #endif 5376 /* Do window scaling on this connection? */ 5377 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == 5378 (TF_RCVD_SCALE | TF_REQ_SCALE)) { 5379 tp->rcv_scale = tp->request_r_scale; 5380 } 5381 tp->rcv_adv += min(tp->rcv_wnd, 5382 TCP_MAXWIN << tp->rcv_scale); 5383 /* 5384 * If there's data, delay ACK; if there's also a FIN ACKNOW 5385 * will be turned on later. 5386 */ 5387 if (DELAY_ACK(tp, tlen) && tlen != 0) { 5388 rack_timer_cancel(tp, (struct tcp_rack *)tp->t_fb_ptr, 5389 ((struct tcp_rack *)tp->t_fb_ptr)->r_ctl.rc_rcvtime, __LINE__); 5390 tp->t_flags |= TF_DELACK; 5391 } else { 5392 ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output++; 5393 tp->t_flags |= TF_ACKNOW; 5394 } 5395 5396 if ((thflags & TH_ECE) && V_tcp_do_ecn) { 5397 tp->t_flags |= TF_ECN_PERMIT; 5398 TCPSTAT_INC(tcps_ecn_shs); 5399 } 5400 /* 5401 * Received <SYN,ACK> in SYN_SENT[*] state. Transitions: 5402 * SYN_SENT --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1 5403 */ 5404 tp->t_starttime = ticks; 5405 if (tp->t_flags & TF_NEEDFIN) { 5406 tcp_state_change(tp, TCPS_FIN_WAIT_1); 5407 tp->t_flags &= ~TF_NEEDFIN; 5408 thflags &= ~TH_SYN; 5409 } else { 5410 tcp_state_change(tp, TCPS_ESTABLISHED); 5411 TCP_PROBE5(connect__established, NULL, tp, 5412 mtod(m, const char *), tp, th); 5413 cc_conn_init(tp); 5414 } 5415 } else { 5416 /* 5417 * Received initial SYN in SYN-SENT[*] state => simultaneous 5418 * open. If segment contains CC option and there is a 5419 * cached CC, apply TAO test. If it succeeds, connection is * 5420 * half-synchronized. Otherwise, do 3-way handshake: 5421 * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If 5422 * there was no CC option, clear cached CC value. 5423 */ 5424 tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN); 5425 tcp_state_change(tp, TCPS_SYN_RECEIVED); 5426 } 5427 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 5428 INP_WLOCK_ASSERT(tp->t_inpcb); 5429 /* 5430 * Advance th->th_seq to correspond to first data byte. If data, 5431 * trim to stay within window, dropping FIN if necessary. 5432 */ 5433 th->th_seq++; 5434 if (tlen > tp->rcv_wnd) { 5435 todrop = tlen - tp->rcv_wnd; 5436 m_adj(m, -todrop); 5437 tlen = tp->rcv_wnd; 5438 thflags &= ~TH_FIN; 5439 TCPSTAT_INC(tcps_rcvpackafterwin); 5440 TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop); 5441 } 5442 tp->snd_wl1 = th->th_seq - 1; 5443 tp->rcv_up = th->th_seq; 5444 /* 5445 * Client side of transaction: already sent SYN and data. If the 5446 * remote host used T/TCP to validate the SYN, our data will be 5447 * ACK'd; if so, enter normal data segment processing in the middle 5448 * of step 5, ack processing. Otherwise, goto step 6. 5449 */ 5450 if (thflags & TH_ACK) { 5451 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) 5452 return (ret_val); 5453 /* We may have changed to FIN_WAIT_1 above */ 5454 if (tp->t_state == TCPS_FIN_WAIT_1) { 5455 /* 5456 * In FIN_WAIT_1 STATE in addition to the processing 5457 * for the ESTABLISHED state if our FIN is now 5458 * acknowledged then enter FIN_WAIT_2. 5459 */ 5460 if (ourfinisacked) { 5461 /* 5462 * If we can't receive any more data, then 5463 * closing user can proceed. Starting the 5464 * timer is contrary to the specification, 5465 * but if we don't get a FIN we'll hang 5466 * forever. 5467 * 5468 * XXXjl: we should release the tp also, and 5469 * use a compressed state. 5470 */ 5471 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 5472 soisdisconnected(so); 5473 tcp_timer_activate(tp, TT_2MSL, 5474 (tcp_fast_finwait2_recycle ? 5475 tcp_finwait2_timeout : 5476 TP_MAXIDLE(tp))); 5477 } 5478 tcp_state_change(tp, TCPS_FIN_WAIT_2); 5479 } 5480 } 5481 } 5482 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5483 tiwin, thflags, nxt_pkt)); 5484 } 5485 5486 /* 5487 * Return value of 1, the TCB is unlocked and most 5488 * likely gone, return value of 0, the TCP is still 5489 * locked. 5490 */ 5491 static int 5492 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so, 5493 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 5494 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 5495 { 5496 int32_t ret_val = 0; 5497 int32_t ourfinisacked = 0; 5498 5499 rack_calc_rwin(so, tp); 5500 5501 if ((thflags & TH_ACK) && 5502 (SEQ_LEQ(th->th_ack, tp->snd_una) || 5503 SEQ_GT(th->th_ack, tp->snd_max))) { 5504 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 5505 return (1); 5506 } 5507 #ifdef TCP_RFC7413 5508 if (tp->t_flags & TF_FASTOPEN) { 5509 /* 5510 * When a TFO connection is in SYN_RECEIVED, the only valid 5511 * packets are the initial SYN, a retransmit/copy of the 5512 * initial SYN (possibly with a subset of the original 5513 * data), a valid ACK, a FIN, or a RST. 5514 */ 5515 if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) { 5516 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 5517 return (1); 5518 } else if (thflags & TH_SYN) { 5519 /* non-initial SYN is ignored */ 5520 struct tcp_rack *rack; 5521 5522 rack = (struct tcp_rack *)tp->t_fb_ptr; 5523 if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) || 5524 (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) || 5525 (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) { 5526 rack_do_drop(m, NULL); 5527 return (0); 5528 } 5529 } else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) { 5530 rack_do_drop(m, NULL); 5531 return (0); 5532 } 5533 } 5534 #endif 5535 if (thflags & TH_RST) 5536 return (rack_process_rst(m, th, so, tp)); 5537 /* 5538 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 5539 * synchronized state. 5540 */ 5541 if (thflags & TH_SYN) { 5542 rack_challenge_ack(m, th, tp, &ret_val); 5543 return (ret_val); 5544 } 5545 /* 5546 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 5547 * it's less than ts_recent, drop it. 5548 */ 5549 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 5550 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 5551 if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val)) 5552 return (ret_val); 5553 } 5554 /* 5555 * In the SYN-RECEIVED state, validate that the packet belongs to 5556 * this connection before trimming the data to fit the receive 5557 * window. Check the sequence number versus IRS since we know the 5558 * sequence numbers haven't wrapped. This is a partial fix for the 5559 * "LAND" DoS attack. 5560 */ 5561 if (SEQ_LT(th->th_seq, tp->irs)) { 5562 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 5563 return (1); 5564 } 5565 if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 5566 return (ret_val); 5567 } 5568 /* 5569 * If last ACK falls within this segment's sequence numbers, record 5570 * its timestamp. NOTE: 1) That the test incorporates suggestions 5571 * from the latest proposal of the tcplw@cray.com list (Braden 5572 * 1993/04/26). 2) That updating only on newer timestamps interferes 5573 * with our earlier PAWS tests, so this check should be solely 5574 * predicated on the sequence space of this segment. 3) That we 5575 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 5576 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 5577 * SEG.Len, This modified check allows us to overcome RFC1323's 5578 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 5579 * p.869. In such cases, we can still calculate the RTT correctly 5580 * when RCV.NXT == Last.ACK.Sent. 5581 */ 5582 if ((to->to_flags & TOF_TS) != 0 && 5583 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 5584 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 5585 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 5586 tp->ts_recent_age = tcp_ts_getticks(); 5587 tp->ts_recent = to->to_tsval; 5588 } 5589 /* 5590 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 5591 * is on (half-synchronized state), then queue data for later 5592 * processing; else drop segment and return. 5593 */ 5594 if ((thflags & TH_ACK) == 0) { 5595 #ifdef TCP_RFC7413 5596 if (tp->t_flags & TF_FASTOPEN) { 5597 tp->snd_wnd = tiwin; 5598 cc_conn_init(tp); 5599 } 5600 #endif 5601 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5602 tiwin, thflags, nxt_pkt)); 5603 } 5604 TCPSTAT_INC(tcps_connects); 5605 soisconnected(so); 5606 /* Do window scaling? */ 5607 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == 5608 (TF_RCVD_SCALE | TF_REQ_SCALE)) { 5609 tp->rcv_scale = tp->request_r_scale; 5610 tp->snd_wnd = tiwin; 5611 } 5612 /* 5613 * Make transitions: SYN-RECEIVED -> ESTABLISHED SYN-RECEIVED* -> 5614 * FIN-WAIT-1 5615 */ 5616 tp->t_starttime = ticks; 5617 if (tp->t_flags & TF_NEEDFIN) { 5618 tcp_state_change(tp, TCPS_FIN_WAIT_1); 5619 tp->t_flags &= ~TF_NEEDFIN; 5620 } else { 5621 tcp_state_change(tp, TCPS_ESTABLISHED); 5622 TCP_PROBE5(accept__established, NULL, tp, 5623 mtod(m, const char *), tp, th); 5624 #ifdef TCP_RFC7413 5625 if (tp->t_tfo_pending) { 5626 tcp_fastopen_decrement_counter(tp->t_tfo_pending); 5627 tp->t_tfo_pending = NULL; 5628 5629 /* 5630 * Account for the ACK of our SYN prior to regular 5631 * ACK processing below. 5632 */ 5633 tp->snd_una++; 5634 } 5635 /* 5636 * TFO connections call cc_conn_init() during SYN 5637 * processing. Calling it again here for such connections 5638 * is not harmless as it would undo the snd_cwnd reduction 5639 * that occurs when a TFO SYN|ACK is retransmitted. 5640 */ 5641 if (!(tp->t_flags & TF_FASTOPEN)) 5642 #endif 5643 cc_conn_init(tp); 5644 } 5645 /* 5646 * If segment contains data or ACK, will call tcp_reass() later; if 5647 * not, do so now to pass queued data to user. 5648 */ 5649 if (tlen == 0 && (thflags & TH_FIN) == 0) 5650 (void)tcp_reass(tp, (struct tcphdr *)0, NULL, 0, 5651 (struct mbuf *)0); 5652 tp->snd_wl1 = th->th_seq - 1; 5653 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 5654 return (ret_val); 5655 } 5656 if (tp->t_state == TCPS_FIN_WAIT_1) { 5657 /* We could have went to FIN_WAIT_1 (or EST) above */ 5658 /* 5659 * In FIN_WAIT_1 STATE in addition to the processing for the 5660 * ESTABLISHED state if our FIN is now acknowledged then 5661 * enter FIN_WAIT_2. 5662 */ 5663 if (ourfinisacked) { 5664 /* 5665 * If we can't receive any more data, then closing 5666 * user can proceed. Starting the timer is contrary 5667 * to the specification, but if we don't get a FIN 5668 * we'll hang forever. 5669 * 5670 * XXXjl: we should release the tp also, and use a 5671 * compressed state. 5672 */ 5673 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 5674 soisdisconnected(so); 5675 tcp_timer_activate(tp, TT_2MSL, 5676 (tcp_fast_finwait2_recycle ? 5677 tcp_finwait2_timeout : 5678 TP_MAXIDLE(tp))); 5679 } 5680 tcp_state_change(tp, TCPS_FIN_WAIT_2); 5681 } 5682 } 5683 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5684 tiwin, thflags, nxt_pkt)); 5685 } 5686 5687 /* 5688 * Return value of 1, the TCB is unlocked and most 5689 * likely gone, return value of 0, the TCP is still 5690 * locked. 5691 */ 5692 static int 5693 rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so, 5694 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 5695 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 5696 { 5697 int32_t ret_val = 0; 5698 5699 /* 5700 * Header prediction: check for the two common cases of a 5701 * uni-directional data xfer. If the packet has no control flags, 5702 * is in-sequence, the window didn't change and we're not 5703 * retransmitting, it's a candidate. If the length is zero and the 5704 * ack moved forward, we're the sender side of the xfer. Just free 5705 * the data acked & wake any higher level process that was blocked 5706 * waiting for space. If the length is non-zero and the ack didn't 5707 * move, we're the receiver side. If we're getting packets in-order 5708 * (the reassembly queue is empty), add the data toc The socket 5709 * buffer and note that we need a delayed ack. Make sure that the 5710 * hidden state-flags are also off. Since we check for 5711 * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN. 5712 */ 5713 if (__predict_true(((to->to_flags & TOF_SACK) == 0)) && 5714 __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_URG | TH_ACK)) == TH_ACK) && 5715 __predict_true(SEGQ_EMPTY(tp)) && 5716 __predict_true(th->th_seq == tp->rcv_nxt)) { 5717 struct tcp_rack *rack; 5718 5719 rack = (struct tcp_rack *)tp->t_fb_ptr; 5720 if (tlen == 0) { 5721 if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen, 5722 tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) { 5723 return (0); 5724 } 5725 } else { 5726 if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen, 5727 tiwin, nxt_pkt)) { 5728 return (0); 5729 } 5730 } 5731 } 5732 rack_calc_rwin(so, tp); 5733 5734 if (thflags & TH_RST) 5735 return (rack_process_rst(m, th, so, tp)); 5736 5737 /* 5738 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 5739 * synchronized state. 5740 */ 5741 if (thflags & TH_SYN) { 5742 rack_challenge_ack(m, th, tp, &ret_val); 5743 return (ret_val); 5744 } 5745 /* 5746 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 5747 * it's less than ts_recent, drop it. 5748 */ 5749 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 5750 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 5751 if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val)) 5752 return (ret_val); 5753 } 5754 if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 5755 return (ret_val); 5756 } 5757 /* 5758 * If last ACK falls within this segment's sequence numbers, record 5759 * its timestamp. NOTE: 1) That the test incorporates suggestions 5760 * from the latest proposal of the tcplw@cray.com list (Braden 5761 * 1993/04/26). 2) That updating only on newer timestamps interferes 5762 * with our earlier PAWS tests, so this check should be solely 5763 * predicated on the sequence space of this segment. 3) That we 5764 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 5765 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 5766 * SEG.Len, This modified check allows us to overcome RFC1323's 5767 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 5768 * p.869. In such cases, we can still calculate the RTT correctly 5769 * when RCV.NXT == Last.ACK.Sent. 5770 */ 5771 if ((to->to_flags & TOF_TS) != 0 && 5772 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 5773 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 5774 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 5775 tp->ts_recent_age = tcp_ts_getticks(); 5776 tp->ts_recent = to->to_tsval; 5777 } 5778 /* 5779 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 5780 * is on (half-synchronized state), then queue data for later 5781 * processing; else drop segment and return. 5782 */ 5783 if ((thflags & TH_ACK) == 0) { 5784 if (tp->t_flags & TF_NEEDSYN) { 5785 5786 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5787 tiwin, thflags, nxt_pkt)); 5788 5789 } else if (tp->t_flags & TF_ACKNOW) { 5790 rack_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 5791 return (ret_val); 5792 } else { 5793 rack_do_drop(m, NULL); 5794 return (0); 5795 } 5796 } 5797 /* 5798 * Ack processing. 5799 */ 5800 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) { 5801 return (ret_val); 5802 } 5803 if (sbavail(&so->so_snd)) { 5804 if (rack_progress_timeout_check(tp)) { 5805 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); 5806 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 5807 return (1); 5808 } 5809 } 5810 /* State changes only happen in rack_process_data() */ 5811 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5812 tiwin, thflags, nxt_pkt)); 5813 } 5814 5815 /* 5816 * Return value of 1, the TCB is unlocked and most 5817 * likely gone, return value of 0, the TCP is still 5818 * locked. 5819 */ 5820 static int 5821 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so, 5822 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 5823 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 5824 { 5825 int32_t ret_val = 0; 5826 5827 rack_calc_rwin(so, tp); 5828 if (thflags & TH_RST) 5829 return (rack_process_rst(m, th, so, tp)); 5830 /* 5831 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 5832 * synchronized state. 5833 */ 5834 if (thflags & TH_SYN) { 5835 rack_challenge_ack(m, th, tp, &ret_val); 5836 return (ret_val); 5837 } 5838 /* 5839 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 5840 * it's less than ts_recent, drop it. 5841 */ 5842 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 5843 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 5844 if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val)) 5845 return (ret_val); 5846 } 5847 if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 5848 return (ret_val); 5849 } 5850 /* 5851 * If last ACK falls within this segment's sequence numbers, record 5852 * its timestamp. NOTE: 1) That the test incorporates suggestions 5853 * from the latest proposal of the tcplw@cray.com list (Braden 5854 * 1993/04/26). 2) That updating only on newer timestamps interferes 5855 * with our earlier PAWS tests, so this check should be solely 5856 * predicated on the sequence space of this segment. 3) That we 5857 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 5858 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 5859 * SEG.Len, This modified check allows us to overcome RFC1323's 5860 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 5861 * p.869. In such cases, we can still calculate the RTT correctly 5862 * when RCV.NXT == Last.ACK.Sent. 5863 */ 5864 if ((to->to_flags & TOF_TS) != 0 && 5865 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 5866 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 5867 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 5868 tp->ts_recent_age = tcp_ts_getticks(); 5869 tp->ts_recent = to->to_tsval; 5870 } 5871 /* 5872 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 5873 * is on (half-synchronized state), then queue data for later 5874 * processing; else drop segment and return. 5875 */ 5876 if ((thflags & TH_ACK) == 0) { 5877 if (tp->t_flags & TF_NEEDSYN) { 5878 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5879 tiwin, thflags, nxt_pkt)); 5880 5881 } else if (tp->t_flags & TF_ACKNOW) { 5882 rack_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 5883 return (ret_val); 5884 } else { 5885 rack_do_drop(m, NULL); 5886 return (0); 5887 } 5888 } 5889 /* 5890 * Ack processing. 5891 */ 5892 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) { 5893 return (ret_val); 5894 } 5895 if (sbavail(&so->so_snd)) { 5896 if (rack_progress_timeout_check(tp)) { 5897 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); 5898 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 5899 return (1); 5900 } 5901 } 5902 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5903 tiwin, thflags, nxt_pkt)); 5904 } 5905 5906 static int 5907 rack_check_data_after_close(struct mbuf *m, 5908 struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so) 5909 { 5910 struct tcp_rack *rack; 5911 5912 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 5913 rack = (struct tcp_rack *)tp->t_fb_ptr; 5914 if (rack->rc_allow_data_af_clo == 0) { 5915 close_now: 5916 tp = tcp_close(tp); 5917 TCPSTAT_INC(tcps_rcvafterclose); 5918 rack_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen)); 5919 return (1); 5920 } 5921 if (sbavail(&so->so_snd) == 0) 5922 goto close_now; 5923 /* Ok we allow data that is ignored and a followup reset */ 5924 tp->rcv_nxt = th->th_seq + *tlen; 5925 tp->t_flags2 |= TF2_DROP_AF_DATA; 5926 rack->r_wanted_output = 1; 5927 *tlen = 0; 5928 return (0); 5929 } 5930 5931 /* 5932 * Return value of 1, the TCB is unlocked and most 5933 * likely gone, return value of 0, the TCP is still 5934 * locked. 5935 */ 5936 static int 5937 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so, 5938 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 5939 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 5940 { 5941 int32_t ret_val = 0; 5942 int32_t ourfinisacked = 0; 5943 5944 rack_calc_rwin(so, tp); 5945 5946 if (thflags & TH_RST) 5947 return (rack_process_rst(m, th, so, tp)); 5948 /* 5949 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 5950 * synchronized state. 5951 */ 5952 if (thflags & TH_SYN) { 5953 rack_challenge_ack(m, th, tp, &ret_val); 5954 return (ret_val); 5955 } 5956 /* 5957 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 5958 * it's less than ts_recent, drop it. 5959 */ 5960 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 5961 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 5962 if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val)) 5963 return (ret_val); 5964 } 5965 if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 5966 return (ret_val); 5967 } 5968 /* 5969 * If new data are received on a connection after the user processes 5970 * are gone, then RST the other end. 5971 */ 5972 if ((so->so_state & SS_NOFDREF) && tlen) { 5973 if (rack_check_data_after_close(m, tp, &tlen, th, so)) 5974 return (1); 5975 } 5976 /* 5977 * If last ACK falls within this segment's sequence numbers, record 5978 * its timestamp. NOTE: 1) That the test incorporates suggestions 5979 * from the latest proposal of the tcplw@cray.com list (Braden 5980 * 1993/04/26). 2) That updating only on newer timestamps interferes 5981 * with our earlier PAWS tests, so this check should be solely 5982 * predicated on the sequence space of this segment. 3) That we 5983 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 5984 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 5985 * SEG.Len, This modified check allows us to overcome RFC1323's 5986 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 5987 * p.869. In such cases, we can still calculate the RTT correctly 5988 * when RCV.NXT == Last.ACK.Sent. 5989 */ 5990 if ((to->to_flags & TOF_TS) != 0 && 5991 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 5992 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 5993 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 5994 tp->ts_recent_age = tcp_ts_getticks(); 5995 tp->ts_recent = to->to_tsval; 5996 } 5997 /* 5998 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 5999 * is on (half-synchronized state), then queue data for later 6000 * processing; else drop segment and return. 6001 */ 6002 if ((thflags & TH_ACK) == 0) { 6003 if (tp->t_flags & TF_NEEDSYN) { 6004 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 6005 tiwin, thflags, nxt_pkt)); 6006 } else if (tp->t_flags & TF_ACKNOW) { 6007 rack_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 6008 return (ret_val); 6009 } else { 6010 rack_do_drop(m, NULL); 6011 return (0); 6012 } 6013 } 6014 /* 6015 * Ack processing. 6016 */ 6017 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 6018 return (ret_val); 6019 } 6020 if (ourfinisacked) { 6021 /* 6022 * If we can't receive any more data, then closing user can 6023 * proceed. Starting the timer is contrary to the 6024 * specification, but if we don't get a FIN we'll hang 6025 * forever. 6026 * 6027 * XXXjl: we should release the tp also, and use a 6028 * compressed state. 6029 */ 6030 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 6031 soisdisconnected(so); 6032 tcp_timer_activate(tp, TT_2MSL, 6033 (tcp_fast_finwait2_recycle ? 6034 tcp_finwait2_timeout : 6035 TP_MAXIDLE(tp))); 6036 } 6037 tcp_state_change(tp, TCPS_FIN_WAIT_2); 6038 } 6039 if (sbavail(&so->so_snd)) { 6040 if (rack_progress_timeout_check(tp)) { 6041 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); 6042 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 6043 return (1); 6044 } 6045 } 6046 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 6047 tiwin, thflags, nxt_pkt)); 6048 } 6049 6050 /* 6051 * Return value of 1, the TCB is unlocked and most 6052 * likely gone, return value of 0, the TCP is still 6053 * locked. 6054 */ 6055 static int 6056 rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so, 6057 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 6058 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 6059 { 6060 int32_t ret_val = 0; 6061 int32_t ourfinisacked = 0; 6062 6063 rack_calc_rwin(so, tp); 6064 6065 if (thflags & TH_RST) 6066 return (rack_process_rst(m, th, so, tp)); 6067 /* 6068 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 6069 * synchronized state. 6070 */ 6071 if (thflags & TH_SYN) { 6072 rack_challenge_ack(m, th, tp, &ret_val); 6073 return (ret_val); 6074 } 6075 /* 6076 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 6077 * it's less than ts_recent, drop it. 6078 */ 6079 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 6080 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 6081 if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val)) 6082 return (ret_val); 6083 } 6084 if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 6085 return (ret_val); 6086 } 6087 /* 6088 * If new data are received on a connection after the user processes 6089 * are gone, then RST the other end. 6090 */ 6091 if ((so->so_state & SS_NOFDREF) && tlen) { 6092 if (rack_check_data_after_close(m, tp, &tlen, th, so)) 6093 return (1); 6094 } 6095 /* 6096 * If last ACK falls within this segment's sequence numbers, record 6097 * its timestamp. NOTE: 1) That the test incorporates suggestions 6098 * from the latest proposal of the tcplw@cray.com list (Braden 6099 * 1993/04/26). 2) That updating only on newer timestamps interferes 6100 * with our earlier PAWS tests, so this check should be solely 6101 * predicated on the sequence space of this segment. 3) That we 6102 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 6103 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 6104 * SEG.Len, This modified check allows us to overcome RFC1323's 6105 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 6106 * p.869. In such cases, we can still calculate the RTT correctly 6107 * when RCV.NXT == Last.ACK.Sent. 6108 */ 6109 if ((to->to_flags & TOF_TS) != 0 && 6110 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 6111 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 6112 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 6113 tp->ts_recent_age = tcp_ts_getticks(); 6114 tp->ts_recent = to->to_tsval; 6115 } 6116 /* 6117 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 6118 * is on (half-synchronized state), then queue data for later 6119 * processing; else drop segment and return. 6120 */ 6121 if ((thflags & TH_ACK) == 0) { 6122 if (tp->t_flags & TF_NEEDSYN) { 6123 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 6124 tiwin, thflags, nxt_pkt)); 6125 } else if (tp->t_flags & TF_ACKNOW) { 6126 rack_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 6127 return (ret_val); 6128 } else { 6129 rack_do_drop(m, NULL); 6130 return (0); 6131 } 6132 } 6133 /* 6134 * Ack processing. 6135 */ 6136 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 6137 return (ret_val); 6138 } 6139 if (ourfinisacked) { 6140 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 6141 tcp_twstart(tp); 6142 m_freem(m); 6143 return (1); 6144 } 6145 if (sbavail(&so->so_snd)) { 6146 if (rack_progress_timeout_check(tp)) { 6147 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); 6148 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 6149 return (1); 6150 } 6151 } 6152 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 6153 tiwin, thflags, nxt_pkt)); 6154 } 6155 6156 /* 6157 * Return value of 1, the TCB is unlocked and most 6158 * likely gone, return value of 0, the TCP is still 6159 * locked. 6160 */ 6161 static int 6162 rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so, 6163 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 6164 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 6165 { 6166 int32_t ret_val = 0; 6167 int32_t ourfinisacked = 0; 6168 6169 rack_calc_rwin(so, tp); 6170 6171 if (thflags & TH_RST) 6172 return (rack_process_rst(m, th, so, tp)); 6173 /* 6174 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 6175 * synchronized state. 6176 */ 6177 if (thflags & TH_SYN) { 6178 rack_challenge_ack(m, th, tp, &ret_val); 6179 return (ret_val); 6180 } 6181 /* 6182 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 6183 * it's less than ts_recent, drop it. 6184 */ 6185 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 6186 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 6187 if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val)) 6188 return (ret_val); 6189 } 6190 if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 6191 return (ret_val); 6192 } 6193 /* 6194 * If new data are received on a connection after the user processes 6195 * are gone, then RST the other end. 6196 */ 6197 if ((so->so_state & SS_NOFDREF) && tlen) { 6198 if (rack_check_data_after_close(m, tp, &tlen, th, so)) 6199 return (1); 6200 } 6201 /* 6202 * If last ACK falls within this segment's sequence numbers, record 6203 * its timestamp. NOTE: 1) That the test incorporates suggestions 6204 * from the latest proposal of the tcplw@cray.com list (Braden 6205 * 1993/04/26). 2) That updating only on newer timestamps interferes 6206 * with our earlier PAWS tests, so this check should be solely 6207 * predicated on the sequence space of this segment. 3) That we 6208 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 6209 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 6210 * SEG.Len, This modified check allows us to overcome RFC1323's 6211 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 6212 * p.869. In such cases, we can still calculate the RTT correctly 6213 * when RCV.NXT == Last.ACK.Sent. 6214 */ 6215 if ((to->to_flags & TOF_TS) != 0 && 6216 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 6217 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 6218 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 6219 tp->ts_recent_age = tcp_ts_getticks(); 6220 tp->ts_recent = to->to_tsval; 6221 } 6222 /* 6223 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 6224 * is on (half-synchronized state), then queue data for later 6225 * processing; else drop segment and return. 6226 */ 6227 if ((thflags & TH_ACK) == 0) { 6228 if (tp->t_flags & TF_NEEDSYN) { 6229 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 6230 tiwin, thflags, nxt_pkt)); 6231 } else if (tp->t_flags & TF_ACKNOW) { 6232 rack_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 6233 return (ret_val); 6234 } else { 6235 rack_do_drop(m, NULL); 6236 return (0); 6237 } 6238 } 6239 /* 6240 * case TCPS_LAST_ACK: Ack processing. 6241 */ 6242 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 6243 return (ret_val); 6244 } 6245 if (ourfinisacked) { 6246 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 6247 tp = tcp_close(tp); 6248 rack_do_drop(m, tp); 6249 return (1); 6250 } 6251 if (sbavail(&so->so_snd)) { 6252 if (rack_progress_timeout_check(tp)) { 6253 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); 6254 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 6255 return (1); 6256 } 6257 } 6258 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 6259 tiwin, thflags, nxt_pkt)); 6260 } 6261 6262 6263 /* 6264 * Return value of 1, the TCB is unlocked and most 6265 * likely gone, return value of 0, the TCP is still 6266 * locked. 6267 */ 6268 static int 6269 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so, 6270 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 6271 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 6272 { 6273 int32_t ret_val = 0; 6274 int32_t ourfinisacked = 0; 6275 6276 rack_calc_rwin(so, tp); 6277 6278 /* Reset receive buffer auto scaling when not in bulk receive mode. */ 6279 if (thflags & TH_RST) 6280 return (rack_process_rst(m, th, so, tp)); 6281 /* 6282 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 6283 * synchronized state. 6284 */ 6285 if (thflags & TH_SYN) { 6286 rack_challenge_ack(m, th, tp, &ret_val); 6287 return (ret_val); 6288 } 6289 /* 6290 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 6291 * it's less than ts_recent, drop it. 6292 */ 6293 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 6294 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 6295 if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val)) 6296 return (ret_val); 6297 } 6298 if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 6299 return (ret_val); 6300 } 6301 /* 6302 * If new data are received on a connection after the user processes 6303 * are gone, then RST the other end. 6304 */ 6305 if ((so->so_state & SS_NOFDREF) && 6306 tlen) { 6307 if (rack_check_data_after_close(m, tp, &tlen, th, so)) 6308 return (1); 6309 } 6310 /* 6311 * If last ACK falls within this segment's sequence numbers, record 6312 * its timestamp. NOTE: 1) That the test incorporates suggestions 6313 * from the latest proposal of the tcplw@cray.com list (Braden 6314 * 1993/04/26). 2) That updating only on newer timestamps interferes 6315 * with our earlier PAWS tests, so this check should be solely 6316 * predicated on the sequence space of this segment. 3) That we 6317 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 6318 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 6319 * SEG.Len, This modified check allows us to overcome RFC1323's 6320 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 6321 * p.869. In such cases, we can still calculate the RTT correctly 6322 * when RCV.NXT == Last.ACK.Sent. 6323 */ 6324 if ((to->to_flags & TOF_TS) != 0 && 6325 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 6326 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 6327 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 6328 tp->ts_recent_age = tcp_ts_getticks(); 6329 tp->ts_recent = to->to_tsval; 6330 } 6331 /* 6332 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 6333 * is on (half-synchronized state), then queue data for later 6334 * processing; else drop segment and return. 6335 */ 6336 if ((thflags & TH_ACK) == 0) { 6337 if (tp->t_flags & TF_NEEDSYN) { 6338 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 6339 tiwin, thflags, nxt_pkt)); 6340 } else if (tp->t_flags & TF_ACKNOW) { 6341 rack_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 6342 return (ret_val); 6343 } else { 6344 rack_do_drop(m, NULL); 6345 return (0); 6346 } 6347 } 6348 /* 6349 * Ack processing. 6350 */ 6351 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 6352 return (ret_val); 6353 } 6354 if (sbavail(&so->so_snd)) { 6355 if (rack_progress_timeout_check(tp)) { 6356 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); 6357 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 6358 return (1); 6359 } 6360 } 6361 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 6362 tiwin, thflags, nxt_pkt)); 6363 } 6364 6365 6366 static void inline 6367 rack_clear_rate_sample(struct tcp_rack *rack) 6368 { 6369 rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY; 6370 rack->r_ctl.rack_rs.rs_rtt_cnt = 0; 6371 rack->r_ctl.rack_rs.rs_rtt_tot = 0; 6372 } 6373 6374 static int 6375 rack_init(struct tcpcb *tp) 6376 { 6377 struct tcp_rack *rack = NULL; 6378 6379 tp->t_fb_ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT); 6380 if (tp->t_fb_ptr == NULL) { 6381 /* 6382 * We need to allocate memory but cant. The INP and INP_INFO 6383 * locks and they are recusive (happens during setup. So a 6384 * scheme to drop the locks fails :( 6385 * 6386 */ 6387 return (ENOMEM); 6388 } 6389 memset(tp->t_fb_ptr, 0, sizeof(struct tcp_rack)); 6390 6391 rack = (struct tcp_rack *)tp->t_fb_ptr; 6392 TAILQ_INIT(&rack->r_ctl.rc_map); 6393 TAILQ_INIT(&rack->r_ctl.rc_free); 6394 TAILQ_INIT(&rack->r_ctl.rc_tmap); 6395 rack->rc_tp = tp; 6396 if (tp->t_inpcb) { 6397 rack->rc_inp = tp->t_inpcb; 6398 } 6399 /* Probably not needed but lets be sure */ 6400 rack_clear_rate_sample(rack); 6401 rack->r_cpu = 0; 6402 rack->r_ctl.rc_reorder_fade = rack_reorder_fade; 6403 rack->rc_allow_data_af_clo = rack_ignore_data_after_close; 6404 rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh; 6405 rack->rc_pace_reduce = rack_slot_reduction; 6406 if (V_tcp_delack_enabled) 6407 tp->t_delayed_ack = 1; 6408 else 6409 tp->t_delayed_ack = 0; 6410 rack->rc_pace_max_segs = rack_hptsi_segments; 6411 rack->r_ctl.rc_early_recovery_segs = rack_early_recovery_max_seg; 6412 rack->r_ctl.rc_reorder_shift = rack_reorder_thresh; 6413 rack->r_ctl.rc_pkt_delay = rack_pkt_delay; 6414 rack->r_ctl.rc_prop_reduce = rack_use_proportional_reduce; 6415 rack->r_idle_reduce_largest = rack_reduce_largest_on_idle; 6416 rack->r_enforce_min_pace = rack_min_pace_time; 6417 rack->r_min_pace_seg_thresh = rack_min_pace_time_seg_req; 6418 rack->r_ctl.rc_prop_rate = rack_proportional_rate; 6419 rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp; 6420 rack->r_ctl.rc_early_recovery = rack_early_recovery; 6421 rack->rc_always_pace = rack_pace_every_seg; 6422 rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method; 6423 rack->rack_tlp_threshold_use = rack_tlp_threshold_use; 6424 rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr; 6425 rack->r_ctl.rc_min_to = rack_min_to; 6426 rack->r_ctl.rc_prr_inc_var = rack_inc_var; 6427 if (tp->snd_una != tp->snd_max) { 6428 /* Create a send map for the current outstanding data */ 6429 struct rack_sendmap *rsm; 6430 6431 rsm = rack_alloc(rack); 6432 if (rsm == NULL) { 6433 uma_zfree(rack_pcb_zone, tp->t_fb_ptr); 6434 tp->t_fb_ptr = NULL; 6435 return (ENOMEM); 6436 } 6437 rsm->r_flags = RACK_OVERMAX; 6438 rsm->r_tim_lastsent[0] = tcp_ts_getticks(); 6439 rsm->r_rtr_cnt = 1; 6440 rsm->r_rtr_bytes = 0; 6441 rsm->r_start = tp->snd_una; 6442 rsm->r_end = tp->snd_max; 6443 rsm->r_sndcnt = 0; 6444 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_map, rsm, r_next); 6445 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext); 6446 rsm->r_in_tmap = 1; 6447 } 6448 rack_stop_all_timers(tp); 6449 rack_start_hpts_timer(rack, tp, tcp_ts_getticks(), __LINE__, 0, 0, 0); 6450 return (0); 6451 } 6452 6453 static int 6454 rack_handoff_ok(struct tcpcb *tp) 6455 { 6456 if ((tp->t_state == TCPS_CLOSED) || 6457 (tp->t_state == TCPS_LISTEN)) { 6458 /* Sure no problem though it may not stick */ 6459 return (0); 6460 } 6461 if ((tp->t_state == TCPS_SYN_SENT) || 6462 (tp->t_state == TCPS_SYN_RECEIVED)) { 6463 /* 6464 * We really don't know you have to get to ESTAB or beyond 6465 * to tell. 6466 */ 6467 return (EAGAIN); 6468 } 6469 if (tp->t_flags & TF_SACK_PERMIT) { 6470 return (0); 6471 } 6472 /* 6473 * If we reach here we don't do SACK on this connection so we can 6474 * never do rack. 6475 */ 6476 return (EINVAL); 6477 } 6478 6479 static void 6480 rack_fini(struct tcpcb *tp, int32_t tcb_is_purged) 6481 { 6482 if (tp->t_fb_ptr) { 6483 struct tcp_rack *rack; 6484 struct rack_sendmap *rsm; 6485 6486 rack = (struct tcp_rack *)tp->t_fb_ptr; 6487 #ifdef TCP_BLACKBOX 6488 tcp_log_flowend(tp); 6489 #endif 6490 rsm = TAILQ_FIRST(&rack->r_ctl.rc_map); 6491 while (rsm) { 6492 TAILQ_REMOVE(&rack->r_ctl.rc_map, rsm, r_next); 6493 uma_zfree(rack_zone, rsm); 6494 rsm = TAILQ_FIRST(&rack->r_ctl.rc_map); 6495 } 6496 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free); 6497 while (rsm) { 6498 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_next); 6499 uma_zfree(rack_zone, rsm); 6500 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free); 6501 } 6502 rack->rc_free_cnt = 0; 6503 uma_zfree(rack_pcb_zone, tp->t_fb_ptr); 6504 tp->t_fb_ptr = NULL; 6505 } 6506 /* Make sure snd_nxt is correctly set */ 6507 tp->snd_nxt = tp->snd_max; 6508 } 6509 6510 static void 6511 rack_set_state(struct tcpcb *tp, struct tcp_rack *rack) 6512 { 6513 switch (tp->t_state) { 6514 case TCPS_SYN_SENT: 6515 rack->r_state = TCPS_SYN_SENT; 6516 rack->r_substate = rack_do_syn_sent; 6517 break; 6518 case TCPS_SYN_RECEIVED: 6519 rack->r_state = TCPS_SYN_RECEIVED; 6520 rack->r_substate = rack_do_syn_recv; 6521 break; 6522 case TCPS_ESTABLISHED: 6523 rack->r_state = TCPS_ESTABLISHED; 6524 rack->r_substate = rack_do_established; 6525 break; 6526 case TCPS_CLOSE_WAIT: 6527 rack->r_state = TCPS_CLOSE_WAIT; 6528 rack->r_substate = rack_do_close_wait; 6529 break; 6530 case TCPS_FIN_WAIT_1: 6531 rack->r_state = TCPS_FIN_WAIT_1; 6532 rack->r_substate = rack_do_fin_wait_1; 6533 break; 6534 case TCPS_CLOSING: 6535 rack->r_state = TCPS_CLOSING; 6536 rack->r_substate = rack_do_closing; 6537 break; 6538 case TCPS_LAST_ACK: 6539 rack->r_state = TCPS_LAST_ACK; 6540 rack->r_substate = rack_do_lastack; 6541 break; 6542 case TCPS_FIN_WAIT_2: 6543 rack->r_state = TCPS_FIN_WAIT_2; 6544 rack->r_substate = rack_do_fin_wait_2; 6545 break; 6546 case TCPS_LISTEN: 6547 case TCPS_CLOSED: 6548 case TCPS_TIME_WAIT: 6549 default: 6550 break; 6551 }; 6552 } 6553 6554 6555 static void 6556 rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb) 6557 { 6558 /* 6559 * We received an ack, and then did not 6560 * call send or were bounced out due to the 6561 * hpts was running. Now a timer is up as well, is 6562 * it the right timer? 6563 */ 6564 struct rack_sendmap *rsm; 6565 int tmr_up; 6566 6567 tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK; 6568 if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT)) 6569 return; 6570 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); 6571 if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) && 6572 (tmr_up == PACE_TMR_RXT)) { 6573 /* Should be an RXT */ 6574 return; 6575 } 6576 if (rsm == NULL) { 6577 /* Nothing outstanding? */ 6578 if (tp->t_flags & TF_DELACK) { 6579 if (tmr_up == PACE_TMR_DELACK) 6580 /* We are supposed to have delayed ack up and we do */ 6581 return; 6582 } else if (sbavail(&tp->t_inpcb->inp_socket->so_snd) && (tmr_up == PACE_TMR_RXT)) { 6583 /* 6584 * if we hit enobufs then we would expect the possiblity 6585 * of nothing outstanding and the RXT up (and the hptsi timer). 6586 */ 6587 return; 6588 } else if (((tcp_always_keepalive || 6589 rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) && 6590 (tp->t_state <= TCPS_CLOSING)) && 6591 (tmr_up == PACE_TMR_KEEP) && 6592 (tp->snd_max == tp->snd_una)) { 6593 /* We should have keep alive up and we do */ 6594 return; 6595 } 6596 } 6597 if (rsm && (rsm->r_flags & RACK_SACK_PASSED)) { 6598 if ((tp->t_flags & TF_SENTFIN) && 6599 ((tp->snd_max - tp->snd_una) == 1) && 6600 (rsm->r_flags & RACK_HAS_FIN)) { 6601 /* needs to be a RXT */ 6602 if (tmr_up == PACE_TMR_RXT) 6603 return; 6604 } else if (tmr_up == PACE_TMR_RACK) 6605 return; 6606 } else if (SEQ_GT(tp->snd_max,tp->snd_una) && 6607 ((tmr_up == PACE_TMR_TLP) || 6608 (tmr_up == PACE_TMR_RXT))) { 6609 /* 6610 * Either a TLP or RXT is fine if no sack-passed 6611 * is in place and data is outstanding. 6612 */ 6613 return; 6614 } else if (tmr_up == PACE_TMR_DELACK) { 6615 /* 6616 * If the delayed ack was going to go off 6617 * before the rtx/tlp/rack timer were going to 6618 * expire, then that would be the timer in control. 6619 * Note we don't check the time here trusting the 6620 * code is correct. 6621 */ 6622 return; 6623 } 6624 /* 6625 * Ok the timer originally started is not what we want now. 6626 * We will force the hpts to be stopped if any, and restart 6627 * with the slot set to what was in the saved slot. 6628 */ 6629 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 6630 rack_start_hpts_timer(rack, tp, tcp_ts_getticks(), __LINE__, 0, 0, 0); 6631 } 6632 6633 static void 6634 rack_hpts_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so, 6635 struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos, 6636 int32_t nxt_pkt, struct timeval *tv) 6637 { 6638 int32_t thflags, retval, did_out = 0; 6639 int32_t way_out = 0; 6640 uint32_t cts; 6641 uint32_t tiwin; 6642 struct tcpopt to; 6643 struct tcp_rack *rack; 6644 struct rack_sendmap *rsm; 6645 int32_t prev_state = 0; 6646 6647 cts = tcp_tv_to_mssectick(tv); 6648 rack = (struct tcp_rack *)tp->t_fb_ptr; 6649 6650 kern_prefetch(rack, &prev_state); 6651 prev_state = 0; 6652 thflags = th->th_flags; 6653 /* 6654 * If this is either a state-changing packet or current state isn't 6655 * established, we require a read lock on tcbinfo. Otherwise, we 6656 * allow the tcbinfo to be in either locked or unlocked, as the 6657 * caller may have unnecessarily acquired a lock due to a race. 6658 */ 6659 INP_WLOCK_ASSERT(tp->t_inpcb); 6660 KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN", 6661 __func__)); 6662 KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT", 6663 __func__)); 6664 { 6665 union tcp_log_stackspecific log; 6666 6667 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 6668 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; 6669 log.u_bbr.ininput = rack->rc_inp->inp_in_input; 6670 log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced; 6671 TCP_LOG_EVENT(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0, 6672 tlen, &log, true); 6673 } 6674 /* 6675 * Segment received on connection. Reset idle time and keep-alive 6676 * timer. XXX: This should be done after segment validation to 6677 * ignore broken/spoofed segs. 6678 */ 6679 if (tp->t_idle_reduce && (tp->snd_max == tp->snd_una)) { 6680 if ((ticks - tp->t_rcvtime) >= tp->t_rxtcur) { 6681 counter_u64_add(rack_input_idle_reduces, 1); 6682 rack_cc_after_idle(tp, 6683 (rack->r_idle_reduce_largest ? 1 :0)); 6684 } 6685 } 6686 rack->r_ctl.rc_rcvtime = cts; 6687 tp->t_rcvtime = ticks; 6688 6689 /* 6690 * Unscale the window into a 32-bit value. For the SYN_SENT state 6691 * the scale is zero. 6692 */ 6693 tiwin = th->th_win << tp->snd_scale; 6694 #ifdef NETFLIX_STATS 6695 stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin); 6696 #endif 6697 /* 6698 * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move 6699 * this to occur after we've validated the segment. 6700 */ 6701 if (tp->t_flags & TF_ECN_PERMIT) { 6702 if (thflags & TH_CWR) 6703 tp->t_flags &= ~TF_ECN_SND_ECE; 6704 switch (iptos & IPTOS_ECN_MASK) { 6705 case IPTOS_ECN_CE: 6706 tp->t_flags |= TF_ECN_SND_ECE; 6707 TCPSTAT_INC(tcps_ecn_ce); 6708 break; 6709 case IPTOS_ECN_ECT0: 6710 TCPSTAT_INC(tcps_ecn_ect0); 6711 break; 6712 case IPTOS_ECN_ECT1: 6713 TCPSTAT_INC(tcps_ecn_ect1); 6714 break; 6715 } 6716 /* Congestion experienced. */ 6717 if (thflags & TH_ECE) { 6718 rack_cong_signal(tp, th, CC_ECN); 6719 } 6720 } 6721 /* 6722 * Parse options on any incoming segment. 6723 */ 6724 tcp_dooptions(&to, (u_char *)(th + 1), 6725 (th->th_off << 2) - sizeof(struct tcphdr), 6726 (thflags & TH_SYN) ? TO_SYN : 0); 6727 6728 /* 6729 * If echoed timestamp is later than the current time, fall back to 6730 * non RFC1323 RTT calculation. Normalize timestamp if syncookies 6731 * were used when this connection was established. 6732 */ 6733 if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) { 6734 to.to_tsecr -= tp->ts_offset; 6735 if (TSTMP_GT(to.to_tsecr, cts)) 6736 to.to_tsecr = 0; 6737 } 6738 /* 6739 * If its the first time in we need to take care of options and 6740 * verify we can do SACK for rack! 6741 */ 6742 if (rack->r_state == 0) { 6743 /* Should be init'd by rack_init() */ 6744 KASSERT(rack->rc_inp != NULL, 6745 ("%s: rack->rc_inp unexpectedly NULL", __func__)); 6746 if (rack->rc_inp == NULL) { 6747 rack->rc_inp = tp->t_inpcb; 6748 } 6749 6750 /* 6751 * Process options only when we get SYN/ACK back. The SYN 6752 * case for incoming connections is handled in tcp_syncache. 6753 * According to RFC1323 the window field in a SYN (i.e., a 6754 * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX 6755 * this is traditional behavior, may need to be cleaned up. 6756 */ 6757 rack->r_cpu = inp_to_cpuid(tp->t_inpcb); 6758 if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) { 6759 if ((to.to_flags & TOF_SCALE) && 6760 (tp->t_flags & TF_REQ_SCALE)) { 6761 tp->t_flags |= TF_RCVD_SCALE; 6762 tp->snd_scale = to.to_wscale; 6763 } 6764 /* 6765 * Initial send window. It will be updated with the 6766 * next incoming segment to the scaled value. 6767 */ 6768 tp->snd_wnd = th->th_win; 6769 if (to.to_flags & TOF_TS) { 6770 tp->t_flags |= TF_RCVD_TSTMP; 6771 tp->ts_recent = to.to_tsval; 6772 tp->ts_recent_age = cts; 6773 } 6774 if (to.to_flags & TOF_MSS) 6775 tcp_mss(tp, to.to_mss); 6776 if ((tp->t_flags & TF_SACK_PERMIT) && 6777 (to.to_flags & TOF_SACKPERM) == 0) 6778 tp->t_flags &= ~TF_SACK_PERMIT; 6779 } 6780 /* 6781 * At this point we are at the initial call. Here we decide 6782 * if we are doing RACK or not. We do this by seeing if 6783 * TF_SACK_PERMIT is set, if not rack is *not* possible and 6784 * we switch to the default code. 6785 */ 6786 if ((tp->t_flags & TF_SACK_PERMIT) == 0) { 6787 tcp_switch_back_to_default(tp); 6788 (*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen, 6789 tlen, iptos); 6790 return; 6791 } 6792 /* Set the flag */ 6793 rack->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0; 6794 tcp_set_hpts(tp->t_inpcb); 6795 sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack); 6796 } 6797 /* 6798 * This is the one exception case where we set the rack state 6799 * always. All other times (timers etc) we must have a rack-state 6800 * set (so we assure we have done the checks above for SACK). 6801 */ 6802 if (rack->r_state != tp->t_state) 6803 rack_set_state(tp, rack); 6804 if (SEQ_GT(th->th_ack, tp->snd_una) && (rsm = TAILQ_FIRST(&rack->r_ctl.rc_map)) != NULL) 6805 kern_prefetch(rsm, &prev_state); 6806 prev_state = rack->r_state; 6807 rack->r_ctl.rc_tlp_send_cnt = 0; 6808 rack_clear_rate_sample(rack); 6809 retval = (*rack->r_substate) (m, th, so, 6810 tp, &to, drop_hdrlen, 6811 tlen, tiwin, thflags, nxt_pkt); 6812 #ifdef INVARIANTS 6813 if ((retval == 0) && 6814 (tp->t_inpcb == NULL)) { 6815 panic("retval:%d tp:%p t_inpcb:NULL state:%d", 6816 retval, tp, prev_state); 6817 } 6818 #endif 6819 if (retval == 0) { 6820 /* 6821 * If retval is 1 the tcb is unlocked and most likely the tp 6822 * is gone. 6823 */ 6824 INP_WLOCK_ASSERT(tp->t_inpcb); 6825 tcp_rack_xmit_timer_commit(rack, tp); 6826 if (nxt_pkt == 0) { 6827 if (rack->r_wanted_output != 0) { 6828 did_out = 1; 6829 (void)tp->t_fb->tfb_tcp_output(tp); 6830 } 6831 rack_start_hpts_timer(rack, tp, cts, __LINE__, 0, 0, 0); 6832 } 6833 if (((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) && 6834 (SEQ_GT(tp->snd_max, tp->snd_una) || 6835 (tp->t_flags & TF_DELACK) || 6836 ((tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) && 6837 (tp->t_state <= TCPS_CLOSING)))) { 6838 /* We could not send (probably in the hpts but stopped the timer earlier)? */ 6839 if ((tp->snd_max == tp->snd_una) && 6840 ((tp->t_flags & TF_DELACK) == 0) && 6841 (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) { 6842 /* keep alive not needed if we are hptsi output yet */ 6843 ; 6844 } else { 6845 if (rack->rc_inp->inp_in_hpts) 6846 tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT); 6847 rack_start_hpts_timer(rack, tp, tcp_ts_getticks(), __LINE__, 0, 0, 0); 6848 } 6849 way_out = 1; 6850 } else { 6851 /* Do we have the correct timer running? */ 6852 rack_timer_audit(tp, rack, &so->so_snd); 6853 way_out = 2; 6854 } 6855 rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out); 6856 if (did_out) 6857 rack->r_wanted_output = 0; 6858 #ifdef INVARIANTS 6859 if (tp->t_inpcb == NULL) { 6860 panic("OP:%d retval:%d tp:%p t_inpcb:NULL state:%d", 6861 did_out, 6862 retval, tp, prev_state); 6863 } 6864 #endif 6865 INP_WUNLOCK(tp->t_inpcb); 6866 } 6867 } 6868 6869 void 6870 rack_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so, 6871 struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos) 6872 { 6873 struct timeval tv; 6874 #ifdef RSS 6875 struct tcp_function_block *tfb; 6876 struct tcp_rack *rack; 6877 struct inpcb *inp; 6878 6879 rack = (struct tcp_rack *)tp->t_fb_ptr; 6880 if (rack->r_state == 0) { 6881 /* 6882 * Initial input (ACK to SYN-ACK etc)lets go ahead and get 6883 * it processed 6884 */ 6885 tcp_get_usecs(&tv); 6886 rack_hpts_do_segment(m, th, so, tp, drop_hdrlen, 6887 tlen, iptos, 0, &tv); 6888 return; 6889 } 6890 tcp_queue_to_input(tp, m, th, tlen, drop_hdrlen, iptos); 6891 INP_WUNLOCK(tp->t_inpcb); 6892 #else 6893 tcp_get_usecs(&tv); 6894 rack_hpts_do_segment(m, th, so, tp, drop_hdrlen, 6895 tlen, iptos, 0, &tv); 6896 #endif 6897 } 6898 6899 struct rack_sendmap * 6900 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused) 6901 { 6902 struct rack_sendmap *rsm = NULL; 6903 int32_t idx; 6904 uint32_t srtt_cur, srtt = 0, thresh = 0, ts_low = 0; 6905 6906 /* Return the next guy to be re-transmitted */ 6907 if (TAILQ_EMPTY(&rack->r_ctl.rc_map)) { 6908 return (NULL); 6909 } 6910 if (tp->t_flags & TF_SENTFIN) { 6911 /* retran the end FIN? */ 6912 return (NULL); 6913 } 6914 /* ok lets look at this one */ 6915 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); 6916 if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) { 6917 goto check_it; 6918 } 6919 rsm = rack_find_lowest_rsm(rack); 6920 if (rsm == NULL) { 6921 return (NULL); 6922 } 6923 check_it: 6924 srtt_cur = tp->t_srtt >> TCP_RTT_SHIFT; 6925 srtt = TICKS_2_MSEC(srtt_cur); 6926 if (rack->rc_rack_rtt && (srtt > rack->rc_rack_rtt)) 6927 srtt = rack->rc_rack_rtt; 6928 if (rsm->r_flags & RACK_ACKED) { 6929 return (NULL); 6930 } 6931 if ((rsm->r_flags & RACK_SACK_PASSED) == 0) { 6932 /* Its not yet ready */ 6933 return (NULL); 6934 } 6935 idx = rsm->r_rtr_cnt - 1; 6936 ts_low = rsm->r_tim_lastsent[idx]; 6937 thresh = rack_calc_thresh_rack(rack, srtt, tsused); 6938 if (tsused <= ts_low) { 6939 return (NULL); 6940 } 6941 if ((tsused - ts_low) >= thresh) { 6942 return (rsm); 6943 } 6944 return (NULL); 6945 } 6946 6947 static int 6948 rack_output(struct tcpcb *tp) 6949 { 6950 struct socket *so; 6951 uint32_t recwin, sendwin; 6952 uint32_t sb_offset; 6953 int32_t len, flags, error = 0; 6954 struct mbuf *m; 6955 struct mbuf *mb; 6956 uint32_t if_hw_tsomaxsegcount = 0; 6957 uint32_t if_hw_tsomaxsegsize; 6958 long tot_len_this_send = 0; 6959 struct ip *ip = NULL; 6960 #ifdef TCPDEBUG 6961 struct ipovly *ipov = NULL; 6962 #endif 6963 #ifdef NETFLIX_TCP_O_UDP 6964 struct udphdr *udp = NULL; 6965 #endif 6966 struct tcp_rack *rack; 6967 struct tcphdr *th; 6968 uint8_t pass = 0; 6969 u_char opt[TCP_MAXOLEN]; 6970 unsigned ipoptlen, optlen, hdrlen; 6971 #ifdef NETFLIX_TCP_O_UDP 6972 unsigned ulen; 6973 #endif 6974 uint32_t rack_seq; 6975 6976 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 6977 unsigned ipsec_optlen = 0; 6978 6979 #endif 6980 int32_t idle, sendalot; 6981 int32_t sub_from_prr = 0; 6982 volatile int32_t sack_rxmit; 6983 struct rack_sendmap *rsm = NULL; 6984 int32_t tso, mtu, would_have_fin = 0; 6985 struct tcpopt to; 6986 int32_t slot = 0; 6987 uint32_t cts; 6988 uint8_t hpts_calling, doing_tlp = 0; 6989 int32_t do_a_prefetch; 6990 int32_t prefetch_rsm = 0; 6991 int32_t prefetch_so_done = 0; 6992 struct tcp_log_buffer *lgb = NULL; 6993 struct inpcb *inp; 6994 struct sockbuf *sb; 6995 #ifdef INET6 6996 struct ip6_hdr *ip6 = NULL; 6997 int32_t isipv6; 6998 #endif 6999 #ifdef KERN_TLS 7000 const bool hw_tls = (so->so_snd.sb_flags & SB_TLS_IFNET) != 0; 7001 #else 7002 const bool hw_tls = false; 7003 #endif 7004 7005 /* setup and take the cache hits here */ 7006 rack = (struct tcp_rack *)tp->t_fb_ptr; 7007 inp = rack->rc_inp; 7008 so = inp->inp_socket; 7009 sb = &so->so_snd; 7010 kern_prefetch(sb, &do_a_prefetch); 7011 do_a_prefetch = 1; 7012 7013 INP_WLOCK_ASSERT(inp); 7014 #ifdef TCP_OFFLOAD 7015 if (tp->t_flags & TF_TOE) 7016 return (tcp_offload_output(tp)); 7017 #endif 7018 7019 #ifdef TCP_RFC7413 7020 /* 7021 * For TFO connections in SYN_RECEIVED, only allow the initial 7022 * SYN|ACK and those sent by the retransmit timer. 7023 */ 7024 if ((tp->t_flags & TF_FASTOPEN) && 7025 (tp->t_state == TCPS_SYN_RECEIVED) && 7026 SEQ_GT(tp->snd_max, tp->snd_una) && /* inital SYN|ACK sent */ 7027 (tp->snd_nxt != tp->snd_una)) /* not a retransmit */ 7028 return (0); 7029 #endif 7030 #ifdef INET6 7031 if (rack->r_state) { 7032 /* Use the cache line loaded if possible */ 7033 isipv6 = rack->r_is_v6; 7034 } else { 7035 isipv6 = (inp->inp_vflag & INP_IPV6) != 0; 7036 } 7037 #endif 7038 cts = tcp_ts_getticks(); 7039 if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) && 7040 inp->inp_in_hpts) { 7041 /* 7042 * We are on the hpts for some timer but not hptsi output. 7043 * Remove from the hpts unconditionally. 7044 */ 7045 rack_timer_cancel(tp, rack, cts, __LINE__); 7046 } 7047 /* Mark that we have called rack_output(). */ 7048 if ((rack->r_timer_override) || 7049 (tp->t_flags & TF_FORCEDATA) || 7050 (tp->t_state < TCPS_ESTABLISHED)) { 7051 if (tp->t_inpcb->inp_in_hpts) 7052 tcp_hpts_remove(tp->t_inpcb, HPTS_REMOVE_OUTPUT); 7053 } else if (tp->t_inpcb->inp_in_hpts) { 7054 /* 7055 * On the hpts you can't pass even if ACKNOW is on, we will 7056 * when the hpts fires. 7057 */ 7058 counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1); 7059 return (0); 7060 } 7061 hpts_calling = inp->inp_hpts_calls; 7062 inp->inp_hpts_calls = 0; 7063 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) { 7064 if (rack_process_timers(tp, rack, cts, hpts_calling)) { 7065 counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1); 7066 return (0); 7067 } 7068 } 7069 rack->r_wanted_output = 0; 7070 rack->r_timer_override = 0; 7071 /* 7072 * Determine length of data that should be transmitted, and flags 7073 * that will be used. If there is some data or critical controls 7074 * (SYN, RST) to send, then transmit; otherwise, investigate 7075 * further. 7076 */ 7077 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una); 7078 if (tp->t_idle_reduce) { 7079 if (idle && ((ticks - tp->t_rcvtime) >= tp->t_rxtcur)) 7080 rack_cc_after_idle(tp, 7081 (rack->r_idle_reduce_largest ? 1 :0)); 7082 } 7083 tp->t_flags &= ~TF_LASTIDLE; 7084 if (idle) { 7085 if (tp->t_flags & TF_MORETOCOME) { 7086 tp->t_flags |= TF_LASTIDLE; 7087 idle = 0; 7088 } 7089 } 7090 again: 7091 /* 7092 * If we've recently taken a timeout, snd_max will be greater than 7093 * snd_nxt. There may be SACK information that allows us to avoid 7094 * resending already delivered data. Adjust snd_nxt accordingly. 7095 */ 7096 sendalot = 0; 7097 cts = tcp_ts_getticks(); 7098 tso = 0; 7099 mtu = 0; 7100 sb_offset = tp->snd_max - tp->snd_una; 7101 sendwin = min(tp->snd_wnd, tp->snd_cwnd); 7102 7103 flags = tcp_outflags[tp->t_state]; 7104 /* 7105 * Send any SACK-generated retransmissions. If we're explicitly 7106 * trying to send out new data (when sendalot is 1), bypass this 7107 * function. If we retransmit in fast recovery mode, decrement 7108 * snd_cwnd, since we're replacing a (future) new transmission with 7109 * a retransmission now, and we previously incremented snd_cwnd in 7110 * tcp_input(). 7111 */ 7112 /* 7113 * Still in sack recovery , reset rxmit flag to zero. 7114 */ 7115 while (rack->rc_free_cnt < rack_free_cache) { 7116 rsm = rack_alloc(rack); 7117 if (rsm == NULL) { 7118 if (inp->inp_hpts_calls) 7119 /* Retry in a ms */ 7120 slot = 1; 7121 goto just_return_nolock; 7122 } 7123 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_next); 7124 rack->rc_free_cnt++; 7125 rsm = NULL; 7126 } 7127 if (inp->inp_hpts_calls) 7128 inp->inp_hpts_calls = 0; 7129 sack_rxmit = 0; 7130 len = 0; 7131 rsm = NULL; 7132 if (flags & TH_RST) { 7133 SOCKBUF_LOCK(sb); 7134 goto send; 7135 } 7136 if (rack->r_ctl.rc_tlpsend) { 7137 /* Tail loss probe */ 7138 long cwin; 7139 long tlen; 7140 7141 doing_tlp = 1; 7142 rsm = rack->r_ctl.rc_tlpsend; 7143 rack->r_ctl.rc_tlpsend = NULL; 7144 sack_rxmit = 1; 7145 tlen = rsm->r_end - rsm->r_start; 7146 if (tlen > tp->t_maxseg) 7147 tlen = tp->t_maxseg; 7148 #ifdef INVARIANTS 7149 if (SEQ_GT(tp->snd_una, rsm->r_start)) { 7150 panic("tp:%p rack:%p snd_una:%u rsm:%p r_start:%u", 7151 tp, rack, tp->snd_una, rsm, rsm->r_start); 7152 } 7153 #endif 7154 sb_offset = rsm->r_start - tp->snd_una; 7155 cwin = min(tp->snd_wnd, tlen); 7156 len = cwin; 7157 } else if (rack->r_ctl.rc_resend) { 7158 /* Retransmit timer */ 7159 rsm = rack->r_ctl.rc_resend; 7160 rack->r_ctl.rc_resend = NULL; 7161 len = rsm->r_end - rsm->r_start; 7162 sack_rxmit = 1; 7163 sendalot = 0; 7164 sb_offset = rsm->r_start - tp->snd_una; 7165 if (len >= tp->t_maxseg) { 7166 len = tp->t_maxseg; 7167 } 7168 KASSERT(sb_offset >= 0, ("%s: sack block to the left of una : %d", 7169 __func__, sb_offset)); 7170 } else if ((rack->rc_in_persist == 0) && 7171 ((rsm = tcp_rack_output(tp, rack, cts)) != NULL)) { 7172 long tlen; 7173 7174 if ((!IN_RECOVERY(tp->t_flags)) && 7175 ((tp->t_flags & (TF_WASFRECOVERY | TF_WASCRECOVERY)) == 0)) { 7176 /* Enter recovery if not induced by a time-out */ 7177 rack->r_ctl.rc_rsm_start = rsm->r_start; 7178 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd; 7179 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh; 7180 rack_cong_signal(tp, NULL, CC_NDUPACK); 7181 /* 7182 * When we enter recovery we need to assure we send 7183 * one packet. 7184 */ 7185 rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; 7186 } 7187 #ifdef INVARIANTS 7188 if (SEQ_LT(rsm->r_start, tp->snd_una)) { 7189 panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n", 7190 tp, rack, rsm, rsm->r_start, tp->snd_una); 7191 } 7192 #endif 7193 tlen = rsm->r_end - rsm->r_start; 7194 sb_offset = rsm->r_start - tp->snd_una; 7195 if (tlen > rack->r_ctl.rc_prr_sndcnt) { 7196 len = rack->r_ctl.rc_prr_sndcnt; 7197 } else { 7198 len = tlen; 7199 } 7200 if (len >= tp->t_maxseg) { 7201 sendalot = 1; 7202 len = tp->t_maxseg; 7203 } else { 7204 sendalot = 0; 7205 if ((rack->rc_timer_up == 0) && 7206 (len < tlen)) { 7207 /* 7208 * If its not a timer don't send a partial 7209 * segment. 7210 */ 7211 len = 0; 7212 goto just_return_nolock; 7213 } 7214 } 7215 KASSERT(sb_offset >= 0, ("%s: sack block to the left of una : %d", 7216 __func__, sb_offset)); 7217 if (len > 0) { 7218 sub_from_prr = 1; 7219 sack_rxmit = 1; 7220 TCPSTAT_INC(tcps_sack_rexmits); 7221 TCPSTAT_ADD(tcps_sack_rexmit_bytes, 7222 min(len, tp->t_maxseg)); 7223 counter_u64_add(rack_rtm_prr_retran, 1); 7224 } 7225 } 7226 if (rsm && (rsm->r_flags & RACK_HAS_FIN)) { 7227 /* we are retransmitting the fin */ 7228 len--; 7229 if (len) { 7230 /* 7231 * When retransmitting data do *not* include the 7232 * FIN. This could happen from a TLP probe. 7233 */ 7234 flags &= ~TH_FIN; 7235 } 7236 } 7237 #ifdef INVARIANTS 7238 /* For debugging */ 7239 rack->r_ctl.rc_rsm_at_retran = rsm; 7240 #endif 7241 /* 7242 * Enforce a connection sendmap count limit if set 7243 * as long as we are not retransmiting. 7244 */ 7245 if ((rsm == NULL) && 7246 (rack_map_entries_limit > 0) && 7247 (rack->r_ctl.rc_num_maps_alloced >= rack_map_entries_limit)) { 7248 counter_u64_add(rack_to_alloc_limited, 1); 7249 if (!rack->alloc_limit_reported) { 7250 rack->alloc_limit_reported = 1; 7251 counter_u64_add(rack_alloc_limited_conns, 1); 7252 } 7253 goto just_return_nolock; 7254 } 7255 /* 7256 * Get standard flags, and add SYN or FIN if requested by 'hidden' 7257 * state flags. 7258 */ 7259 if (tp->t_flags & TF_NEEDFIN) 7260 flags |= TH_FIN; 7261 if (tp->t_flags & TF_NEEDSYN) 7262 flags |= TH_SYN; 7263 if ((sack_rxmit == 0) && (prefetch_rsm == 0)) { 7264 void *end_rsm; 7265 end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext); 7266 if (end_rsm) 7267 kern_prefetch(end_rsm, &prefetch_rsm); 7268 prefetch_rsm = 1; 7269 } 7270 SOCKBUF_LOCK(sb); 7271 /* 7272 * If in persist timeout with window of 0, send 1 byte. Otherwise, 7273 * if window is small but nonzero and time TF_SENTFIN expired, we 7274 * will send what we can and go to transmit state. 7275 */ 7276 if (tp->t_flags & TF_FORCEDATA) { 7277 if (sendwin == 0) { 7278 /* 7279 * If we still have some data to send, then clear 7280 * the FIN bit. Usually this would happen below 7281 * when it realizes that we aren't sending all the 7282 * data. However, if we have exactly 1 byte of 7283 * unsent data, then it won't clear the FIN bit 7284 * below, and if we are in persist state, we wind up 7285 * sending the packet without recording that we sent 7286 * the FIN bit. 7287 * 7288 * We can't just blindly clear the FIN bit, because 7289 * if we don't have any more data to send then the 7290 * probe will be the FIN itself. 7291 */ 7292 if (sb_offset < sbused(sb)) 7293 flags &= ~TH_FIN; 7294 sendwin = 1; 7295 } else { 7296 if (rack->rc_in_persist) 7297 rack_exit_persist(tp, rack); 7298 /* 7299 * If we are dropping persist mode then we need to 7300 * correct snd_nxt/snd_max and off. 7301 */ 7302 tp->snd_nxt = tp->snd_max; 7303 sb_offset = tp->snd_nxt - tp->snd_una; 7304 } 7305 } 7306 /* 7307 * If snd_nxt == snd_max and we have transmitted a FIN, the 7308 * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a 7309 * negative length. This can also occur when TCP opens up its 7310 * congestion window while receiving additional duplicate acks after 7311 * fast-retransmit because TCP will reset snd_nxt to snd_max after 7312 * the fast-retransmit. 7313 * 7314 * In the normal retransmit-FIN-only case, however, snd_nxt will be 7315 * set to snd_una, the sb_offset will be 0, and the length may wind 7316 * up 0. 7317 * 7318 * If sack_rxmit is true we are retransmitting from the scoreboard 7319 * in which case len is already set. 7320 */ 7321 if (sack_rxmit == 0) { 7322 uint32_t avail; 7323 7324 avail = sbavail(sb); 7325 if (SEQ_GT(tp->snd_nxt, tp->snd_una)) 7326 sb_offset = tp->snd_nxt - tp->snd_una; 7327 else 7328 sb_offset = 0; 7329 if (IN_RECOVERY(tp->t_flags) == 0) { 7330 if (rack->r_ctl.rc_tlp_new_data) { 7331 /* TLP is forcing out new data */ 7332 if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) { 7333 rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset); 7334 } 7335 if (rack->r_ctl.rc_tlp_new_data > tp->snd_wnd) 7336 len = tp->snd_wnd; 7337 else 7338 len = rack->r_ctl.rc_tlp_new_data; 7339 rack->r_ctl.rc_tlp_new_data = 0; 7340 doing_tlp = 1; 7341 } else { 7342 if (sendwin > avail) { 7343 /* use the available */ 7344 if (avail > sb_offset) { 7345 len = (int32_t)(avail - sb_offset); 7346 } else { 7347 len = 0; 7348 } 7349 } else { 7350 if (sendwin > sb_offset) { 7351 len = (int32_t)(sendwin - sb_offset); 7352 } else { 7353 len = 0; 7354 } 7355 } 7356 } 7357 } else { 7358 uint32_t outstanding; 7359 7360 /* 7361 * We are inside of a SACK recovery episode and are 7362 * sending new data, having retransmitted all the 7363 * data possible so far in the scoreboard. 7364 */ 7365 outstanding = tp->snd_max - tp->snd_una; 7366 if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) { 7367 if (tp->snd_wnd > outstanding) { 7368 len = tp->snd_wnd - outstanding; 7369 /* Check to see if we have the data */ 7370 if (((sb_offset + len) > avail) && 7371 (avail > sb_offset)) 7372 len = avail - sb_offset; 7373 else 7374 len = 0; 7375 } else 7376 len = 0; 7377 } else if (avail > sb_offset) 7378 len = avail - sb_offset; 7379 else 7380 len = 0; 7381 if (len > 0) { 7382 if (len > rack->r_ctl.rc_prr_sndcnt) 7383 len = rack->r_ctl.rc_prr_sndcnt; 7384 7385 if (len > 0) { 7386 sub_from_prr = 1; 7387 counter_u64_add(rack_rtm_prr_newdata, 1); 7388 } 7389 } 7390 if (len > tp->t_maxseg) { 7391 /* 7392 * We should never send more than a MSS when 7393 * retransmitting or sending new data in prr 7394 * mode unless the override flag is on. Most 7395 * likely the PRR algorithm is not going to 7396 * let us send a lot as well :-) 7397 */ 7398 if (rack->r_ctl.rc_prr_sendalot == 0) 7399 len = tp->t_maxseg; 7400 } else if (len < tp->t_maxseg) { 7401 /* 7402 * Do we send any? The idea here is if the 7403 * send empty's the socket buffer we want to 7404 * do it. However if not then lets just wait 7405 * for our prr_sndcnt to get bigger. 7406 */ 7407 long leftinsb; 7408 7409 leftinsb = sbavail(sb) - sb_offset; 7410 if (leftinsb > len) { 7411 /* This send does not empty the sb */ 7412 len = 0; 7413 } 7414 } 7415 } 7416 } 7417 if (prefetch_so_done == 0) { 7418 kern_prefetch(so, &prefetch_so_done); 7419 prefetch_so_done = 1; 7420 } 7421 /* 7422 * Lop off SYN bit if it has already been sent. However, if this is 7423 * SYN-SENT state and if segment contains data and if we don't know 7424 * that foreign host supports TAO, suppress sending segment. 7425 */ 7426 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) { 7427 if ((tp->t_state != TCPS_SYN_RECEIVED) && 7428 (tp->t_state != TCPS_SYN_SENT)) 7429 flags &= ~TH_SYN; 7430 #ifdef TCP_RFC7413 7431 /* 7432 * When sending additional segments following a TFO SYN|ACK, 7433 * do not include the SYN bit. 7434 */ 7435 if ((tp->t_flags & TF_FASTOPEN) && 7436 (tp->t_state == TCPS_SYN_RECEIVED)) 7437 flags &= ~TH_SYN; 7438 #endif 7439 } 7440 /* 7441 * Be careful not to send data and/or FIN on SYN segments. This 7442 * measure is needed to prevent interoperability problems with not 7443 * fully conformant TCP implementations. 7444 */ 7445 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) { 7446 len = 0; 7447 flags &= ~TH_FIN; 7448 } 7449 #ifdef TCP_RFC7413 7450 /* 7451 * When retransmitting SYN|ACK on a passively-created TFO socket, 7452 * don't include data, as the presence of data may have caused the 7453 * original SYN|ACK to have been dropped by a middlebox. 7454 */ 7455 if ((tp->t_flags & TF_FASTOPEN) && 7456 ((tp->t_state == TCPS_SYN_RECEIVED) && (tp->t_rxtshift > 0))) 7457 len = 0; 7458 #endif 7459 if (len <= 0) { 7460 /* 7461 * If FIN has been sent but not acked, but we haven't been 7462 * called to retransmit, len will be < 0. Otherwise, window 7463 * shrank after we sent into it. If window shrank to 0, 7464 * cancel pending retransmit, pull snd_nxt back to (closed) 7465 * window, and set the persist timer if it isn't already 7466 * going. If the window didn't close completely, just wait 7467 * for an ACK. 7468 * 7469 * We also do a general check here to ensure that we will 7470 * set the persist timer when we have data to send, but a 7471 * 0-byte window. This makes sure the persist timer is set 7472 * even if the packet hits one of the "goto send" lines 7473 * below. 7474 */ 7475 len = 0; 7476 if ((tp->snd_wnd == 0) && 7477 (TCPS_HAVEESTABLISHED(tp->t_state)) && 7478 (sb_offset < (int)sbavail(sb))) { 7479 tp->snd_nxt = tp->snd_una; 7480 rack_enter_persist(tp, rack, cts); 7481 } 7482 } 7483 /* len will be >= 0 after this point. */ 7484 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 7485 tcp_sndbuf_autoscale(tp, so, sendwin); 7486 /* 7487 * Decide if we can use TCP Segmentation Offloading (if supported by 7488 * hardware). 7489 * 7490 * TSO may only be used if we are in a pure bulk sending state. The 7491 * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP 7492 * options prevent using TSO. With TSO the TCP header is the same 7493 * (except for the sequence number) for all generated packets. This 7494 * makes it impossible to transmit any options which vary per 7495 * generated segment or packet. 7496 * 7497 * IPv4 handling has a clear separation of ip options and ip header 7498 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does 7499 * the right thing below to provide length of just ip options and thus 7500 * checking for ipoptlen is enough to decide if ip options are present. 7501 */ 7502 7503 #ifdef INET6 7504 if (isipv6) 7505 ipoptlen = ip6_optlen(tp->t_inpcb); 7506 else 7507 #endif 7508 if (tp->t_inpcb->inp_options) 7509 ipoptlen = tp->t_inpcb->inp_options->m_len - 7510 offsetof(struct ipoption, ipopt_list); 7511 else 7512 ipoptlen = 0; 7513 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 7514 /* 7515 * Pre-calculate here as we save another lookup into the darknesses 7516 * of IPsec that way and can actually decide if TSO is ok. 7517 */ 7518 #ifdef INET6 7519 if (isipv6 && IPSEC_ENABLED(ipv6)) 7520 ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb); 7521 #ifdef INET 7522 else 7523 #endif 7524 #endif /* INET6 */ 7525 #ifdef INET 7526 if (IPSEC_ENABLED(ipv4)) 7527 ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb); 7528 #endif /* INET */ 7529 #endif 7530 7531 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 7532 ipoptlen += ipsec_optlen; 7533 #endif 7534 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg && 7535 #ifdef NETFLIX_TCP_O_UDP 7536 (tp->t_port == 0) && 7537 #endif 7538 ((tp->t_flags & TF_SIGNATURE) == 0) && 7539 tp->rcv_numsacks == 0 && sack_rxmit == 0 && 7540 ipoptlen == 0) 7541 tso = 1; 7542 { 7543 uint32_t outstanding; 7544 7545 outstanding = tp->snd_max - tp->snd_una; 7546 if (tp->t_flags & TF_SENTFIN) { 7547 /* 7548 * If we sent a fin, snd_max is 1 higher than 7549 * snd_una 7550 */ 7551 outstanding--; 7552 } 7553 if (outstanding > 0) { 7554 /* 7555 * This is sub-optimal. We only send a stand alone 7556 * FIN on its own segment. 7557 */ 7558 if (flags & TH_FIN) { 7559 flags &= ~TH_FIN; 7560 would_have_fin = 1; 7561 } 7562 } else if (sack_rxmit) { 7563 if ((rsm->r_flags & RACK_HAS_FIN) == 0) 7564 flags &= ~TH_FIN; 7565 } else { 7566 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + 7567 sbused(sb))) 7568 flags &= ~TH_FIN; 7569 } 7570 } 7571 recwin = sbspace(&so->so_rcv); 7572 7573 /* 7574 * Sender silly window avoidance. We transmit under the following 7575 * conditions when len is non-zero: 7576 * 7577 * - We have a full segment (or more with TSO) - This is the last 7578 * buffer in a write()/send() and we are either idle or running 7579 * NODELAY - we've timed out (e.g. persist timer) - we have more 7580 * then 1/2 the maximum send window's worth of data (receiver may be 7581 * limited the window size) - we need to retransmit 7582 */ 7583 if (len) { 7584 if (len >= tp->t_maxseg) { 7585 pass = 1; 7586 goto send; 7587 } 7588 /* 7589 * NOTE! on localhost connections an 'ack' from the remote 7590 * end may occur synchronously with the output and cause us 7591 * to flush a buffer queued with moretocome. XXX 7592 * 7593 */ 7594 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */ 7595 (idle || (tp->t_flags & TF_NODELAY)) && 7596 ((uint32_t)len + (uint32_t)sb_offset >= sbavail(&so->so_snd)) && 7597 (tp->t_flags & TF_NOPUSH) == 0) { 7598 pass = 2; 7599 goto send; 7600 } 7601 if (tp->t_flags & TF_FORCEDATA) { /* typ. timeout case */ 7602 pass = 3; 7603 goto send; 7604 } 7605 if ((tp->snd_una == tp->snd_max) && len) { /* Nothing outstanding */ 7606 goto send; 7607 } 7608 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) { 7609 pass = 4; 7610 goto send; 7611 } 7612 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { /* retransmit case */ 7613 pass = 5; 7614 goto send; 7615 } 7616 if (sack_rxmit) { 7617 pass = 6; 7618 goto send; 7619 } 7620 } 7621 /* 7622 * Sending of standalone window updates. 7623 * 7624 * Window updates are important when we close our window due to a 7625 * full socket buffer and are opening it again after the application 7626 * reads data from it. Once the window has opened again and the 7627 * remote end starts to send again the ACK clock takes over and 7628 * provides the most current window information. 7629 * 7630 * We must avoid the silly window syndrome whereas every read from 7631 * the receive buffer, no matter how small, causes a window update 7632 * to be sent. We also should avoid sending a flurry of window 7633 * updates when the socket buffer had queued a lot of data and the 7634 * application is doing small reads. 7635 * 7636 * Prevent a flurry of pointless window updates by only sending an 7637 * update when we can increase the advertized window by more than 7638 * 1/4th of the socket buffer capacity. When the buffer is getting 7639 * full or is very small be more aggressive and send an update 7640 * whenever we can increase by two mss sized segments. In all other 7641 * situations the ACK's to new incoming data will carry further 7642 * window increases. 7643 * 7644 * Don't send an independent window update if a delayed ACK is 7645 * pending (it will get piggy-backed on it) or the remote side 7646 * already has done a half-close and won't send more data. Skip 7647 * this if the connection is in T/TCP half-open state. 7648 */ 7649 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) && 7650 !(tp->t_flags & TF_DELACK) && 7651 !TCPS_HAVERCVDFIN(tp->t_state)) { 7652 /* 7653 * "adv" is the amount we could increase the window, taking 7654 * into account that we are limited by TCP_MAXWIN << 7655 * tp->rcv_scale. 7656 */ 7657 int32_t adv; 7658 int oldwin; 7659 7660 adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale); 7661 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) { 7662 oldwin = (tp->rcv_adv - tp->rcv_nxt); 7663 adv -= oldwin; 7664 } else 7665 oldwin = 0; 7666 7667 /* 7668 * If the new window size ends up being the same as the old 7669 * size when it is scaled, then don't force a window update. 7670 */ 7671 if (oldwin >> tp->rcv_scale == (adv + oldwin) >> tp->rcv_scale) 7672 goto dontupdate; 7673 7674 if (adv >= (int32_t)(2 * tp->t_maxseg) && 7675 (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) || 7676 recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) || 7677 so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg)) { 7678 pass = 7; 7679 goto send; 7680 } 7681 if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) 7682 goto send; 7683 } 7684 dontupdate: 7685 7686 /* 7687 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW 7688 * is also a catch-all for the retransmit timer timeout case. 7689 */ 7690 if (tp->t_flags & TF_ACKNOW) { 7691 pass = 8; 7692 goto send; 7693 } 7694 if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) { 7695 pass = 9; 7696 goto send; 7697 } 7698 if (SEQ_GT(tp->snd_up, tp->snd_una)) { 7699 pass = 10; 7700 goto send; 7701 } 7702 /* 7703 * If our state indicates that FIN should be sent and we have not 7704 * yet done so, then we need to send. 7705 */ 7706 if (flags & TH_FIN) { 7707 if ((tp->t_flags & TF_SENTFIN) || 7708 (((tp->t_flags & TF_SENTFIN) == 0) && 7709 (tp->snd_nxt == tp->snd_una))) { 7710 pass = 11; 7711 goto send; 7712 } 7713 } 7714 /* 7715 * No reason to send a segment, just return. 7716 */ 7717 just_return: 7718 SOCKBUF_UNLOCK(sb); 7719 just_return_nolock: 7720 if (tot_len_this_send == 0) 7721 counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1); 7722 rack_start_hpts_timer(rack, tp, cts, __LINE__, slot, tot_len_this_send, 1); 7723 rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling); 7724 tp->t_flags &= ~TF_FORCEDATA; 7725 return (0); 7726 7727 send: 7728 if (doing_tlp == 0) { 7729 /* 7730 * Data not a TLP, and its not the rxt firing. If it is the 7731 * rxt firing, we want to leave the tlp_in_progress flag on 7732 * so we don't send another TLP. It has to be a rack timer 7733 * or normal send (response to acked data) to clear the tlp 7734 * in progress flag. 7735 */ 7736 rack->rc_tlp_in_progress = 0; 7737 } 7738 SOCKBUF_LOCK_ASSERT(sb); 7739 if (len > 0) { 7740 if (len >= tp->t_maxseg) 7741 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT; 7742 else 7743 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT; 7744 } 7745 /* 7746 * Before ESTABLISHED, force sending of initial options unless TCP 7747 * set not to do any options. NOTE: we assume that the IP/TCP header 7748 * plus TCP options always fit in a single mbuf, leaving room for a 7749 * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr) 7750 * + optlen <= MCLBYTES 7751 */ 7752 optlen = 0; 7753 #ifdef INET6 7754 if (isipv6) 7755 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 7756 else 7757 #endif 7758 hdrlen = sizeof(struct tcpiphdr); 7759 7760 /* 7761 * Compute options for segment. We only have to care about SYN and 7762 * established connection segments. Options for SYN-ACK segments 7763 * are handled in TCP syncache. 7764 */ 7765 to.to_flags = 0; 7766 if ((tp->t_flags & TF_NOOPT) == 0) { 7767 /* Maximum segment size. */ 7768 if (flags & TH_SYN) { 7769 tp->snd_nxt = tp->iss; 7770 to.to_mss = tcp_mssopt(&inp->inp_inc); 7771 #ifdef NETFLIX_TCP_O_UDP 7772 if (tp->t_port) 7773 to.to_mss -= V_tcp_udp_tunneling_overhead; 7774 #endif 7775 to.to_flags |= TOF_MSS; 7776 #ifdef TCP_RFC7413 7777 /* 7778 * Only include the TFO option on the first 7779 * transmission of the SYN|ACK on a 7780 * passively-created TFO socket, as the presence of 7781 * the TFO option may have caused the original 7782 * SYN|ACK to have been dropped by a middlebox. 7783 */ 7784 if ((tp->t_flags & TF_FASTOPEN) && 7785 (tp->t_state == TCPS_SYN_RECEIVED) && 7786 (tp->t_rxtshift == 0)) { 7787 to.to_tfo_len = TCP_FASTOPEN_MAX_COOKIE_LEN; 7788 to.to_tfo_cookie = (u_char *)&tp->t_tfo_cookie; 7789 to.to_flags |= TOF_FASTOPEN; 7790 } 7791 #endif 7792 } 7793 /* Window scaling. */ 7794 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) { 7795 to.to_wscale = tp->request_r_scale; 7796 to.to_flags |= TOF_SCALE; 7797 } 7798 /* Timestamps. */ 7799 if ((tp->t_flags & TF_RCVD_TSTMP) || 7800 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) { 7801 to.to_tsval = cts + tp->ts_offset; 7802 to.to_tsecr = tp->ts_recent; 7803 to.to_flags |= TOF_TS; 7804 } 7805 /* Set receive buffer autosizing timestamp. */ 7806 if (tp->rfbuf_ts == 0 && 7807 (so->so_rcv.sb_flags & SB_AUTOSIZE)) 7808 tp->rfbuf_ts = tcp_ts_getticks(); 7809 /* Selective ACK's. */ 7810 if (flags & TH_SYN) 7811 to.to_flags |= TOF_SACKPERM; 7812 else if (TCPS_HAVEESTABLISHED(tp->t_state) && 7813 tp->rcv_numsacks > 0) { 7814 to.to_flags |= TOF_SACK; 7815 to.to_nsacks = tp->rcv_numsacks; 7816 to.to_sacks = (u_char *)tp->sackblks; 7817 } 7818 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 7819 /* TCP-MD5 (RFC2385). */ 7820 if (tp->t_flags & TF_SIGNATURE) 7821 to.to_flags |= TOF_SIGNATURE; 7822 #endif /* TCP_SIGNATURE */ 7823 7824 /* Processing the options. */ 7825 hdrlen += optlen = tcp_addoptions(&to, opt); 7826 } 7827 #ifdef NETFLIX_TCP_O_UDP 7828 if (tp->t_port) { 7829 if (V_tcp_udp_tunneling_port == 0) { 7830 /* The port was removed?? */ 7831 SOCKBUF_UNLOCK(&so->so_snd); 7832 return (EHOSTUNREACH); 7833 } 7834 hdrlen += sizeof(struct udphdr); 7835 } 7836 #endif 7837 ipoptlen = 0; 7838 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 7839 ipoptlen += ipsec_optlen; 7840 #endif 7841 7842 /* 7843 * Adjust data length if insertion of options will bump the packet 7844 * length beyond the t_maxseg length. Clear the FIN bit because we 7845 * cut off the tail of the segment. 7846 */ 7847 if (len + optlen + ipoptlen > tp->t_maxseg) { 7848 if (flags & TH_FIN) { 7849 would_have_fin = 1; 7850 flags &= ~TH_FIN; 7851 } 7852 if (tso) { 7853 uint32_t if_hw_tsomax; 7854 uint32_t moff; 7855 int32_t max_len; 7856 7857 /* extract TSO information */ 7858 if_hw_tsomax = tp->t_tsomax; 7859 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount; 7860 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize; 7861 KASSERT(ipoptlen == 0, 7862 ("%s: TSO can't do IP options", __func__)); 7863 7864 /* 7865 * Check if we should limit by maximum payload 7866 * length: 7867 */ 7868 if (if_hw_tsomax != 0) { 7869 /* compute maximum TSO length */ 7870 max_len = (if_hw_tsomax - hdrlen - 7871 max_linkhdr); 7872 if (max_len <= 0) { 7873 len = 0; 7874 } else if (len > max_len) { 7875 sendalot = 1; 7876 len = max_len; 7877 } 7878 } 7879 /* 7880 * Prevent the last segment from being fractional 7881 * unless the send sockbuf can be emptied: 7882 */ 7883 max_len = (tp->t_maxseg - optlen); 7884 if ((sb_offset + len) < sbavail(sb)) { 7885 moff = len % (u_int)max_len; 7886 if (moff != 0) { 7887 len -= moff; 7888 sendalot = 1; 7889 } 7890 } 7891 /* 7892 * In case there are too many small fragments don't 7893 * use TSO: 7894 */ 7895 if (len <= max_len) { 7896 len = max_len; 7897 sendalot = 1; 7898 tso = 0; 7899 } 7900 /* 7901 * Send the FIN in a separate segment after the bulk 7902 * sending is done. We don't trust the TSO 7903 * implementations to clear the FIN flag on all but 7904 * the last segment. 7905 */ 7906 if (tp->t_flags & TF_NEEDFIN) 7907 sendalot = 1; 7908 7909 } else { 7910 len = tp->t_maxseg - optlen - ipoptlen; 7911 sendalot = 1; 7912 } 7913 } else 7914 tso = 0; 7915 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET, 7916 ("%s: len > IP_MAXPACKET", __func__)); 7917 #ifdef DIAGNOSTIC 7918 #ifdef INET6 7919 if (max_linkhdr + hdrlen > MCLBYTES) 7920 #else 7921 if (max_linkhdr + hdrlen > MHLEN) 7922 #endif 7923 panic("tcphdr too big"); 7924 #endif 7925 7926 /* 7927 * This KASSERT is here to catch edge cases at a well defined place. 7928 * Before, those had triggered (random) panic conditions further 7929 * down. 7930 */ 7931 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 7932 if ((len == 0) && 7933 (flags & TH_FIN) && 7934 (sbused(sb))) { 7935 /* 7936 * We have outstanding data, don't send a fin by itself!. 7937 */ 7938 goto just_return; 7939 } 7940 /* 7941 * Grab a header mbuf, attaching a copy of data to be transmitted, 7942 * and initialize the header from the template for sends on this 7943 * connection. 7944 */ 7945 if (len) { 7946 uint32_t max_val; 7947 uint32_t moff; 7948 7949 if (rack->rc_pace_max_segs) 7950 max_val = rack->rc_pace_max_segs * tp->t_maxseg; 7951 else 7952 max_val = len; 7953 /* 7954 * We allow a limit on sending with hptsi. 7955 */ 7956 if (len > max_val) { 7957 len = max_val; 7958 } 7959 #ifdef INET6 7960 if (MHLEN < hdrlen + max_linkhdr) 7961 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); 7962 else 7963 #endif 7964 m = m_gethdr(M_NOWAIT, MT_DATA); 7965 7966 if (m == NULL) { 7967 SOCKBUF_UNLOCK(sb); 7968 error = ENOBUFS; 7969 sack_rxmit = 0; 7970 goto out; 7971 } 7972 m->m_data += max_linkhdr; 7973 m->m_len = hdrlen; 7974 7975 /* 7976 * Start the m_copy functions from the closest mbuf to the 7977 * sb_offset in the socket buffer chain. 7978 */ 7979 mb = sbsndptr_noadv(sb, sb_offset, &moff); 7980 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) { 7981 m_copydata(mb, moff, (int)len, 7982 mtod(m, caddr_t)+hdrlen); 7983 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 7984 sbsndptr_adv(sb, mb, len); 7985 m->m_len += len; 7986 } else { 7987 struct sockbuf *msb; 7988 7989 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 7990 msb = NULL; 7991 else 7992 msb = sb; 7993 m->m_next = tcp_m_copym(/*tp, */ mb, moff, &len, 7994 if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb, 7995 hw_tls /*, NULL */); 7996 if (len <= (tp->t_maxseg - optlen)) { 7997 /* 7998 * Must have ran out of mbufs for the copy 7999 * shorten it to no longer need tso. Lets 8000 * not put on sendalot since we are low on 8001 * mbufs. 8002 */ 8003 tso = 0; 8004 } 8005 if (m->m_next == NULL) { 8006 SOCKBUF_UNLOCK(sb); 8007 (void)m_free(m); 8008 error = ENOBUFS; 8009 sack_rxmit = 0; 8010 goto out; 8011 } 8012 } 8013 if ((tp->t_flags & TF_FORCEDATA) && len == 1) { 8014 TCPSTAT_INC(tcps_sndprobe); 8015 #ifdef NETFLIX_STATS 8016 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 8017 stats_voi_update_abs_u32(tp->t_stats, 8018 VOI_TCP_RETXPB, len); 8019 else 8020 stats_voi_update_abs_u64(tp->t_stats, 8021 VOI_TCP_TXPB, len); 8022 #endif 8023 } else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) { 8024 if (rsm && (rsm->r_flags & RACK_TLP)) { 8025 /* 8026 * TLP should not count in retran count, but 8027 * in its own bin 8028 */ 8029 /* tp->t_sndtlppack++;*/ 8030 /* tp->t_sndtlpbyte += len;*/ 8031 counter_u64_add(rack_tlp_retran, 1); 8032 counter_u64_add(rack_tlp_retran_bytes, len); 8033 } else { 8034 tp->t_sndrexmitpack++; 8035 TCPSTAT_INC(tcps_sndrexmitpack); 8036 TCPSTAT_ADD(tcps_sndrexmitbyte, len); 8037 } 8038 #ifdef NETFLIX_STATS 8039 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB, 8040 len); 8041 #endif 8042 } else { 8043 TCPSTAT_INC(tcps_sndpack); 8044 TCPSTAT_ADD(tcps_sndbyte, len); 8045 #ifdef NETFLIX_STATS 8046 stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB, 8047 len); 8048 #endif 8049 } 8050 /* 8051 * If we're sending everything we've got, set PUSH. (This 8052 * will keep happy those implementations which only give 8053 * data to the user when a buffer fills or a PUSH comes in.) 8054 */ 8055 if (sb_offset + len == sbused(sb) && 8056 sbused(sb) && 8057 !(flags & TH_SYN)) 8058 flags |= TH_PUSH; 8059 8060 /* 8061 * Are we doing hptsi, if so we must calculate the slot. We 8062 * only do hptsi in ESTABLISHED and with no RESET being 8063 * sent where we have data to send. 8064 */ 8065 if (((tp->t_state == TCPS_ESTABLISHED) || 8066 (tp->t_state == TCPS_CLOSE_WAIT) || 8067 ((tp->t_state == TCPS_FIN_WAIT_1) && 8068 ((tp->t_flags & TF_SENTFIN) == 0) && 8069 ((flags & TH_FIN) == 0))) && 8070 ((flags & TH_RST) == 0) && 8071 (rack->rc_always_pace)) { 8072 /* 8073 * We use the most optimistic possible cwnd/srtt for 8074 * sending calculations. This will make our 8075 * calculation anticipate getting more through 8076 * quicker then possible. But thats ok we don't want 8077 * the peer to have a gap in data sending. 8078 */ 8079 uint32_t srtt, cwnd, tr_perms = 0; 8080 8081 if (rack->r_ctl.rc_rack_min_rtt) 8082 srtt = rack->r_ctl.rc_rack_min_rtt; 8083 else 8084 srtt = TICKS_2_MSEC((tp->t_srtt >> TCP_RTT_SHIFT)); 8085 if (rack->r_ctl.rc_rack_largest_cwnd) 8086 cwnd = rack->r_ctl.rc_rack_largest_cwnd; 8087 else 8088 cwnd = tp->snd_cwnd; 8089 tr_perms = cwnd / srtt; 8090 if (tr_perms == 0) { 8091 tr_perms = tp->t_maxseg; 8092 } 8093 tot_len_this_send += len; 8094 /* 8095 * Calculate how long this will take to drain, if 8096 * the calculation comes out to zero, thats ok we 8097 * will use send_a_lot to possibly spin around for 8098 * more increasing tot_len_this_send to the point 8099 * that its going to require a pace, or we hit the 8100 * cwnd. Which in that case we are just waiting for 8101 * a ACK. 8102 */ 8103 slot = tot_len_this_send / tr_perms; 8104 /* Now do we reduce the time so we don't run dry? */ 8105 if (slot && rack->rc_pace_reduce) { 8106 int32_t reduce; 8107 8108 reduce = (slot / rack->rc_pace_reduce); 8109 if (reduce < slot) { 8110 slot -= reduce; 8111 } else 8112 slot = 0; 8113 } 8114 if (rack->r_enforce_min_pace && 8115 (slot == 0) && 8116 (tot_len_this_send >= (rack->r_min_pace_seg_thresh * tp->t_maxseg))) { 8117 /* We are enforcing a minimum pace time of 1ms */ 8118 slot = rack->r_enforce_min_pace; 8119 } 8120 } 8121 SOCKBUF_UNLOCK(sb); 8122 } else { 8123 SOCKBUF_UNLOCK(sb); 8124 if (tp->t_flags & TF_ACKNOW) 8125 TCPSTAT_INC(tcps_sndacks); 8126 else if (flags & (TH_SYN | TH_FIN | TH_RST)) 8127 TCPSTAT_INC(tcps_sndctrl); 8128 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 8129 TCPSTAT_INC(tcps_sndurg); 8130 else 8131 TCPSTAT_INC(tcps_sndwinup); 8132 8133 m = m_gethdr(M_NOWAIT, MT_DATA); 8134 if (m == NULL) { 8135 error = ENOBUFS; 8136 sack_rxmit = 0; 8137 goto out; 8138 } 8139 #ifdef INET6 8140 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && 8141 MHLEN >= hdrlen) { 8142 M_ALIGN(m, hdrlen); 8143 } else 8144 #endif 8145 m->m_data += max_linkhdr; 8146 m->m_len = hdrlen; 8147 } 8148 SOCKBUF_UNLOCK_ASSERT(sb); 8149 m->m_pkthdr.rcvif = (struct ifnet *)0; 8150 #ifdef MAC 8151 mac_inpcb_create_mbuf(inp, m); 8152 #endif 8153 #ifdef INET6 8154 if (isipv6) { 8155 ip6 = mtod(m, struct ip6_hdr *); 8156 #ifdef NETFLIX_TCP_O_UDP 8157 if (tp->t_port) { 8158 udp = (struct udphdr *)((caddr_t)ip6 + ipoptlen + sizeof(struct ip6_hdr)); 8159 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 8160 udp->uh_dport = tp->t_port; 8161 ulen = hdrlen + len - sizeof(struct ip6_hdr); 8162 udp->uh_ulen = htons(ulen); 8163 th = (struct tcphdr *)(udp + 1); 8164 } else 8165 #endif 8166 th = (struct tcphdr *)(ip6 + 1); 8167 tcpip_fillheaders(inp, /*tp->t_port, */ ip6, th); 8168 } else 8169 #endif /* INET6 */ 8170 { 8171 ip = mtod(m, struct ip *); 8172 #ifdef TCPDEBUG 8173 ipov = (struct ipovly *)ip; 8174 #endif 8175 #ifdef NETFLIX_TCP_O_UDP 8176 if (tp->t_port) { 8177 udp = (struct udphdr *)((caddr_t)ip + ipoptlen + sizeof(struct ip)); 8178 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 8179 udp->uh_dport = tp->t_port; 8180 ulen = hdrlen + len - sizeof(struct ip); 8181 udp->uh_ulen = htons(ulen); 8182 th = (struct tcphdr *)(udp + 1); 8183 } else 8184 #endif 8185 th = (struct tcphdr *)(ip + 1); 8186 tcpip_fillheaders(inp,/*tp->t_port, */ ip, th); 8187 } 8188 /* 8189 * Fill in fields, remembering maximum advertised window for use in 8190 * delaying messages about window sizes. If resending a FIN, be sure 8191 * not to use a new sequence number. 8192 */ 8193 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && 8194 tp->snd_nxt == tp->snd_max) 8195 tp->snd_nxt--; 8196 /* 8197 * If we are starting a connection, send ECN setup SYN packet. If we 8198 * are on a retransmit, we may resend those bits a number of times 8199 * as per RFC 3168. 8200 */ 8201 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) { 8202 if (tp->t_rxtshift >= 1) { 8203 if (tp->t_rxtshift <= V_tcp_ecn_maxretries) 8204 flags |= TH_ECE | TH_CWR; 8205 } else 8206 flags |= TH_ECE | TH_CWR; 8207 } 8208 if (tp->t_state == TCPS_ESTABLISHED && 8209 (tp->t_flags & TF_ECN_PERMIT)) { 8210 /* 8211 * If the peer has ECN, mark data packets with ECN capable 8212 * transmission (ECT). Ignore pure ack packets, 8213 * retransmissions and window probes. 8214 */ 8215 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) && 8216 !((tp->t_flags & TF_FORCEDATA) && len == 1)) { 8217 #ifdef INET6 8218 if (isipv6) 8219 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20); 8220 else 8221 #endif 8222 ip->ip_tos |= IPTOS_ECN_ECT0; 8223 TCPSTAT_INC(tcps_ecn_ect0); 8224 } 8225 /* 8226 * Reply with proper ECN notifications. 8227 */ 8228 if (tp->t_flags & TF_ECN_SND_CWR) { 8229 flags |= TH_CWR; 8230 tp->t_flags &= ~TF_ECN_SND_CWR; 8231 } 8232 if (tp->t_flags & TF_ECN_SND_ECE) 8233 flags |= TH_ECE; 8234 } 8235 /* 8236 * If we are doing retransmissions, then snd_nxt will not reflect 8237 * the first unsent octet. For ACK only packets, we do not want the 8238 * sequence number of the retransmitted packet, we want the sequence 8239 * number of the next unsent octet. So, if there is no data (and no 8240 * SYN or FIN), use snd_max instead of snd_nxt when filling in 8241 * ti_seq. But if we are in persist state, snd_max might reflect 8242 * one byte beyond the right edge of the window, so use snd_nxt in 8243 * that case, since we know we aren't doing a retransmission. 8244 * (retransmit and persist are mutually exclusive...) 8245 */ 8246 if (sack_rxmit == 0) { 8247 if (len || (flags & (TH_SYN | TH_FIN)) || 8248 rack->rc_in_persist) { 8249 th->th_seq = htonl(tp->snd_nxt); 8250 rack_seq = tp->snd_nxt; 8251 } else if (flags & TH_RST) { 8252 /* 8253 * For a Reset send the last cum ack in sequence 8254 * (this like any other choice may still generate a 8255 * challenge ack, if a ack-update packet is in 8256 * flight). 8257 */ 8258 th->th_seq = htonl(tp->snd_una); 8259 rack_seq = tp->snd_una; 8260 } else { 8261 th->th_seq = htonl(tp->snd_max); 8262 rack_seq = tp->snd_max; 8263 } 8264 } else { 8265 th->th_seq = htonl(rsm->r_start); 8266 rack_seq = rsm->r_start; 8267 } 8268 th->th_ack = htonl(tp->rcv_nxt); 8269 if (optlen) { 8270 bcopy(opt, th + 1, optlen); 8271 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 8272 } 8273 th->th_flags = flags; 8274 /* 8275 * Calculate receive window. Don't shrink window, but avoid silly 8276 * window syndrome. 8277 */ 8278 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) && 8279 recwin < (long)tp->t_maxseg) 8280 recwin = 0; 8281 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) && 8282 recwin < (long)(tp->rcv_adv - tp->rcv_nxt)) 8283 recwin = (long)(tp->rcv_adv - tp->rcv_nxt); 8284 if (recwin > (long)TCP_MAXWIN << tp->rcv_scale) 8285 recwin = (long)TCP_MAXWIN << tp->rcv_scale; 8286 8287 /* 8288 * According to RFC1323 the window field in a SYN (i.e., a <SYN> or 8289 * <SYN,ACK>) segment itself is never scaled. The <SYN,ACK> case is 8290 * handled in syncache. 8291 */ 8292 if (flags & TH_SYN) 8293 th->th_win = htons((u_short) 8294 (min(sbspace(&so->so_rcv), TCP_MAXWIN))); 8295 else 8296 th->th_win = htons((u_short)(recwin >> tp->rcv_scale)); 8297 /* 8298 * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0 8299 * window. This may cause the remote transmitter to stall. This 8300 * flag tells soreceive() to disable delayed acknowledgements when 8301 * draining the buffer. This can occur if the receiver is 8302 * attempting to read more data than can be buffered prior to 8303 * transmitting on the connection. 8304 */ 8305 if (th->th_win == 0) { 8306 tp->t_sndzerowin++; 8307 tp->t_flags |= TF_RXWIN0SENT; 8308 } else 8309 tp->t_flags &= ~TF_RXWIN0SENT; 8310 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 8311 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); 8312 th->th_flags |= TH_URG; 8313 } else 8314 /* 8315 * If no urgent pointer to send, then we pull the urgent 8316 * pointer to the left edge of the send window so that it 8317 * doesn't drift into the send window on sequence number 8318 * wraparound. 8319 */ 8320 tp->snd_up = tp->snd_una; /* drag it along */ 8321 8322 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 8323 if (to.to_flags & TOF_SIGNATURE) { 8324 /* 8325 * Calculate MD5 signature and put it into the place 8326 * determined before. 8327 * NOTE: since TCP options buffer doesn't point into 8328 * mbuf's data, calculate offset and use it. 8329 */ 8330 if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th, 8331 (u_char *)(th + 1) + (to.to_signature - opt)) != 0) { 8332 /* 8333 * Do not send segment if the calculation of MD5 8334 * digest has failed. 8335 */ 8336 goto out; 8337 } 8338 } 8339 #endif 8340 8341 /* 8342 * Put TCP length in extended header, and then checksum extended 8343 * header and data. 8344 */ 8345 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 8346 #ifdef INET6 8347 if (isipv6) { 8348 /* 8349 * ip6_plen is not need to be filled now, and will be filled 8350 * in ip6_output. 8351 */ 8352 #ifdef NETFLIX_TCP_O_UDP 8353 if (tp->t_port) { 8354 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6; 8355 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 8356 udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0); 8357 th->th_sum = htons(0); 8358 UDPSTAT_INC(udps_opackets); 8359 } else { 8360 #endif 8361 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 8362 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 8363 th->th_sum = in6_cksum_pseudo(ip6, 8364 sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP, 8365 0); 8366 #ifdef NETFLIX_TCP_O_UDP 8367 } 8368 #endif 8369 } 8370 #endif 8371 #if defined(INET6) && defined(INET) 8372 else 8373 #endif 8374 #ifdef INET 8375 { 8376 #ifdef NETFLIX_TCP_O_UDP 8377 if (tp->t_port) { 8378 m->m_pkthdr.csum_flags = CSUM_UDP; 8379 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 8380 udp->uh_sum = in_pseudo(ip->ip_src.s_addr, 8381 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP)); 8382 th->th_sum = htons(0); 8383 UDPSTAT_INC(udps_opackets); 8384 } else { 8385 #endif 8386 m->m_pkthdr.csum_flags = CSUM_TCP; 8387 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 8388 th->th_sum = in_pseudo(ip->ip_src.s_addr, 8389 ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) + 8390 IPPROTO_TCP + len + optlen)); 8391 #ifdef NETFLIX_TCP_O_UDP 8392 } 8393 #endif 8394 /* IP version must be set here for ipv4/ipv6 checking later */ 8395 KASSERT(ip->ip_v == IPVERSION, 8396 ("%s: IP version incorrect: %d", __func__, ip->ip_v)); 8397 } 8398 #endif 8399 8400 /* 8401 * Enable TSO and specify the size of the segments. The TCP pseudo 8402 * header checksum is always provided. XXX: Fixme: This is currently 8403 * not the case for IPv6. 8404 */ 8405 if (tso) { 8406 KASSERT(len > tp->t_maxseg - optlen, 8407 ("%s: len <= tso_segsz", __func__)); 8408 m->m_pkthdr.csum_flags |= CSUM_TSO; 8409 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen; 8410 } 8411 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 8412 KASSERT(len + hdrlen + ipoptlen - ipsec_optlen == m_length(m, NULL), 8413 ("%s: mbuf chain shorter than expected: %d + %u + %u - %u != %u", 8414 __func__, len, hdrlen, ipoptlen, ipsec_optlen, m_length(m, NULL))); 8415 #else 8416 KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL), 8417 ("%s: mbuf chain shorter than expected: %d + %u + %u != %u", 8418 __func__, len, hdrlen, ipoptlen, m_length(m, NULL))); 8419 #endif 8420 8421 #ifdef TCP_HHOOK 8422 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */ 8423 hhook_run_tcp_est_out(tp, th, &to, len, tso); 8424 #endif 8425 8426 #ifdef TCPDEBUG 8427 /* 8428 * Trace. 8429 */ 8430 if (so->so_options & SO_DEBUG) { 8431 u_short save = 0; 8432 8433 #ifdef INET6 8434 if (!isipv6) 8435 #endif 8436 { 8437 save = ipov->ih_len; 8438 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + 8439 * (th->th_off << 2) */ ); 8440 } 8441 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0); 8442 #ifdef INET6 8443 if (!isipv6) 8444 #endif 8445 ipov->ih_len = save; 8446 } 8447 #endif /* TCPDEBUG */ 8448 8449 /* We're getting ready to send; log now. */ 8450 if (tp->t_logstate != TCP_LOG_STATE_OFF) { 8451 union tcp_log_stackspecific log; 8452 8453 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 8454 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; 8455 log.u_bbr.ininput = rack->rc_inp->inp_in_input; 8456 log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt; 8457 if (rsm || sack_rxmit) { 8458 log.u_bbr.flex8 = 1; 8459 } else { 8460 log.u_bbr.flex8 = 0; 8461 } 8462 lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK, 8463 len, &log, false, NULL, NULL, 0, NULL); 8464 } else 8465 lgb = NULL; 8466 8467 /* 8468 * Fill in IP length and desired time to live and send to IP level. 8469 * There should be a better way to handle ttl and tos; we could keep 8470 * them in the template, but need a way to checksum without them. 8471 */ 8472 /* 8473 * m->m_pkthdr.len should have been set before cksum calcuration, 8474 * because in6_cksum() need it. 8475 */ 8476 #ifdef INET6 8477 if (isipv6) { 8478 /* 8479 * we separately set hoplimit for every segment, since the 8480 * user might want to change the value via setsockopt. Also, 8481 * desired default hop limit might be changed via Neighbor 8482 * Discovery. 8483 */ 8484 ip6->ip6_hlim = in6_selecthlim(inp, NULL); 8485 8486 /* 8487 * Set the packet size here for the benefit of DTrace 8488 * probes. ip6_output() will set it properly; it's supposed 8489 * to include the option header lengths as well. 8490 */ 8491 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6)); 8492 8493 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) 8494 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 8495 else 8496 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 8497 8498 if (tp->t_state == TCPS_SYN_SENT) 8499 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th); 8500 8501 TCP_PROBE5(send, NULL, tp, ip6, tp, th); 8502 /* TODO: IPv6 IP6TOS_ECT bit on */ 8503 error = ip6_output(m, tp->t_inpcb->in6p_outputopts, 8504 &inp->inp_route6, 8505 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 8506 NULL, NULL, inp); 8507 8508 if (error == EMSGSIZE && inp->inp_route6.ro_rt != NULL) 8509 mtu = inp->inp_route6.ro_rt->rt_mtu; 8510 } 8511 #endif /* INET6 */ 8512 #if defined(INET) && defined(INET6) 8513 else 8514 #endif 8515 #ifdef INET 8516 { 8517 ip->ip_len = htons(m->m_pkthdr.len); 8518 #ifdef INET6 8519 if (inp->inp_vflag & INP_IPV6PROTO) 8520 ip->ip_ttl = in6_selecthlim(inp, NULL); 8521 #endif /* INET6 */ 8522 /* 8523 * If we do path MTU discovery, then we set DF on every 8524 * packet. This might not be the best thing to do according 8525 * to RFC3390 Section 2. However the tcp hostcache migitates 8526 * the problem so it affects only the first tcp connection 8527 * with a host. 8528 * 8529 * NB: Don't set DF on small MTU/MSS to have a safe 8530 * fallback. 8531 */ 8532 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) { 8533 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 8534 if (tp->t_port == 0 || len < V_tcp_minmss) { 8535 ip->ip_off |= htons(IP_DF); 8536 } 8537 } else { 8538 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 8539 } 8540 8541 if (tp->t_state == TCPS_SYN_SENT) 8542 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th); 8543 8544 TCP_PROBE5(send, NULL, tp, ip, tp, th); 8545 8546 error = ip_output(m, tp->t_inpcb->inp_options, &inp->inp_route, 8547 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, 8548 inp); 8549 if (error == EMSGSIZE && inp->inp_route.ro_rt != NULL) 8550 mtu = inp->inp_route.ro_rt->rt_mtu; 8551 } 8552 #endif /* INET */ 8553 8554 out: 8555 if (lgb) { 8556 lgb->tlb_errno = error; 8557 lgb = NULL; 8558 } 8559 /* 8560 * In transmit state, time the transmission and arrange for the 8561 * retransmit. In persist state, just set snd_max. 8562 */ 8563 if (error == 0) { 8564 if (TCPS_HAVEESTABLISHED(tp->t_state) && 8565 (tp->t_flags & TF_SACK_PERMIT) && 8566 tp->rcv_numsacks > 0) 8567 tcp_clean_dsack_blocks(tp); 8568 if (len == 0) 8569 counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1); 8570 else if (len == 1) { 8571 counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1); 8572 } else if (len > 1) { 8573 int idx; 8574 8575 idx = (len / tp->t_maxseg) + 3; 8576 if (idx >= TCP_MSS_ACCT_ATIMER) 8577 counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1); 8578 else 8579 counter_u64_add(rack_out_size[idx], 1); 8580 } 8581 } 8582 if (sub_from_prr && (error == 0)) { 8583 if (rack->r_ctl.rc_prr_sndcnt >= len) 8584 rack->r_ctl.rc_prr_sndcnt -= len; 8585 else 8586 rack->r_ctl.rc_prr_sndcnt = 0; 8587 } 8588 sub_from_prr = 0; 8589 rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error, cts, 8590 pass, rsm); 8591 if ((tp->t_flags & TF_FORCEDATA) == 0 || 8592 (rack->rc_in_persist == 0)) { 8593 #ifdef NETFLIX_STATS 8594 tcp_seq startseq = tp->snd_nxt; 8595 #endif 8596 /* 8597 * Advance snd_nxt over sequence space of this segment. 8598 */ 8599 if (error) 8600 /* We don't log or do anything with errors */ 8601 goto timer; 8602 8603 if (flags & (TH_SYN | TH_FIN)) { 8604 if (flags & TH_SYN) 8605 tp->snd_nxt++; 8606 if (flags & TH_FIN) { 8607 tp->snd_nxt++; 8608 tp->t_flags |= TF_SENTFIN; 8609 } 8610 } 8611 /* In the ENOBUFS case we do *not* update snd_max */ 8612 if (sack_rxmit) 8613 goto timer; 8614 8615 tp->snd_nxt += len; 8616 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 8617 if (tp->snd_una == tp->snd_max) { 8618 /* 8619 * Update the time we just added data since 8620 * none was outstanding. 8621 */ 8622 rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__); 8623 tp->t_acktime = ticks; 8624 } 8625 tp->snd_max = tp->snd_nxt; 8626 #ifdef NETFLIX_STATS 8627 if (!(tp->t_flags & TF_GPUTINPROG) && len) { 8628 tp->t_flags |= TF_GPUTINPROG; 8629 tp->gput_seq = startseq; 8630 tp->gput_ack = startseq + 8631 ulmin(sbavail(sb) - sb_offset, sendwin); 8632 tp->gput_ts = tcp_ts_getticks(); 8633 } 8634 #endif 8635 } 8636 /* 8637 * Set retransmit timer if not currently set, and not doing 8638 * a pure ack or a keep-alive probe. Initial value for 8639 * retransmit timer is smoothed round-trip time + 2 * 8640 * round-trip time variance. Initialize shift counter which 8641 * is used for backoff of retransmit time. 8642 */ 8643 timer: 8644 if ((tp->snd_wnd == 0) && 8645 TCPS_HAVEESTABLISHED(tp->t_state)) { 8646 /* 8647 * If the persists timer was set above (right before 8648 * the goto send), and still needs to be on. Lets 8649 * make sure all is canceled. If the persist timer 8650 * is not running, we want to get it up. 8651 */ 8652 if (rack->rc_in_persist == 0) { 8653 rack_enter_persist(tp, rack, cts); 8654 } 8655 } 8656 } else { 8657 /* 8658 * Persist case, update snd_max but since we are in persist 8659 * mode (no window) we do not update snd_nxt. 8660 */ 8661 int32_t xlen = len; 8662 8663 if (error) 8664 goto nomore; 8665 8666 if (flags & TH_SYN) 8667 ++xlen; 8668 if (flags & TH_FIN) { 8669 ++xlen; 8670 tp->t_flags |= TF_SENTFIN; 8671 } 8672 /* In the ENOBUFS case we do *not* update snd_max */ 8673 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) { 8674 if (tp->snd_una == tp->snd_max) { 8675 /* 8676 * Update the time we just added data since 8677 * none was outstanding. 8678 */ 8679 rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__); 8680 tp->t_acktime = ticks; 8681 } 8682 tp->snd_max = tp->snd_nxt + len; 8683 } 8684 } 8685 nomore: 8686 if (error) { 8687 SOCKBUF_UNLOCK_ASSERT(sb); /* Check gotos. */ 8688 /* 8689 * Failures do not advance the seq counter above. For the 8690 * case of ENOBUFS we will fall out and retry in 1ms with 8691 * the hpts. Everything else will just have to retransmit 8692 * with the timer. 8693 * 8694 * In any case, we do not want to loop around for another 8695 * send without a good reason. 8696 */ 8697 sendalot = 0; 8698 switch (error) { 8699 case EPERM: 8700 tp->t_flags &= ~TF_FORCEDATA; 8701 tp->t_softerror = error; 8702 return (error); 8703 case ENOBUFS: 8704 if (slot == 0) { 8705 /* 8706 * Pace us right away to retry in a some 8707 * time 8708 */ 8709 slot = 1 + rack->rc_enobuf; 8710 if (rack->rc_enobuf < 255) 8711 rack->rc_enobuf++; 8712 if (slot > (rack->rc_rack_rtt / 2)) { 8713 slot = rack->rc_rack_rtt / 2; 8714 } 8715 if (slot < 10) 8716 slot = 10; 8717 } 8718 counter_u64_add(rack_saw_enobuf, 1); 8719 error = 0; 8720 goto enobufs; 8721 case EMSGSIZE: 8722 /* 8723 * For some reason the interface we used initially 8724 * to send segments changed to another or lowered 8725 * its MTU. If TSO was active we either got an 8726 * interface without TSO capabilits or TSO was 8727 * turned off. If we obtained mtu from ip_output() 8728 * then update it and try again. 8729 */ 8730 if (tso) 8731 tp->t_flags &= ~TF_TSO; 8732 if (mtu != 0) { 8733 tcp_mss_update(tp, -1, mtu, NULL, NULL); 8734 goto again; 8735 } 8736 slot = 10; 8737 rack_start_hpts_timer(rack, tp, cts, __LINE__, slot, 0, 1); 8738 tp->t_flags &= ~TF_FORCEDATA; 8739 return (error); 8740 case ENETUNREACH: 8741 counter_u64_add(rack_saw_enetunreach, 1); 8742 case EHOSTDOWN: 8743 case EHOSTUNREACH: 8744 case ENETDOWN: 8745 if (TCPS_HAVERCVDSYN(tp->t_state)) { 8746 tp->t_softerror = error; 8747 } 8748 /* FALLTHROUGH */ 8749 default: 8750 slot = 10; 8751 rack_start_hpts_timer(rack, tp, cts, __LINE__, slot, 0, 1); 8752 tp->t_flags &= ~TF_FORCEDATA; 8753 return (error); 8754 } 8755 } else { 8756 rack->rc_enobuf = 0; 8757 } 8758 TCPSTAT_INC(tcps_sndtotal); 8759 8760 /* 8761 * Data sent (as far as we can tell). If this advertises a larger 8762 * window than any other segment, then remember the size of the 8763 * advertised window. Any pending ACK has now been sent. 8764 */ 8765 if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) 8766 tp->rcv_adv = tp->rcv_nxt + recwin; 8767 tp->last_ack_sent = tp->rcv_nxt; 8768 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); 8769 enobufs: 8770 rack->r_tlp_running = 0; 8771 if ((flags & TH_RST) || (would_have_fin == 1)) { 8772 /* 8773 * We don't send again after a RST. We also do *not* send 8774 * again if we would have had a find, but now have 8775 * outstanding data. 8776 */ 8777 slot = 0; 8778 sendalot = 0; 8779 } 8780 if (slot) { 8781 /* set the rack tcb into the slot N */ 8782 counter_u64_add(rack_paced_segments, 1); 8783 } else if (sendalot) { 8784 if (len) 8785 counter_u64_add(rack_unpaced_segments, 1); 8786 sack_rxmit = 0; 8787 tp->t_flags &= ~TF_FORCEDATA; 8788 goto again; 8789 } else if (len) { 8790 counter_u64_add(rack_unpaced_segments, 1); 8791 } 8792 tp->t_flags &= ~TF_FORCEDATA; 8793 rack_start_hpts_timer(rack, tp, cts, __LINE__, slot, tot_len_this_send, 1); 8794 return (error); 8795 } 8796 8797 /* 8798 * rack_ctloutput() must drop the inpcb lock before performing copyin on 8799 * socket option arguments. When it re-acquires the lock after the copy, it 8800 * has to revalidate that the connection is still valid for the socket 8801 * option. 8802 */ 8803 static int 8804 rack_set_sockopt(struct socket *so, struct sockopt *sopt, 8805 struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack) 8806 { 8807 int32_t error = 0, optval; 8808 8809 switch (sopt->sopt_name) { 8810 case TCP_RACK_PROP_RATE: 8811 case TCP_RACK_PROP: 8812 case TCP_RACK_TLP_REDUCE: 8813 case TCP_RACK_EARLY_RECOV: 8814 case TCP_RACK_PACE_ALWAYS: 8815 case TCP_DELACK: 8816 case TCP_RACK_PACE_REDUCE: 8817 case TCP_RACK_PACE_MAX_SEG: 8818 case TCP_RACK_PRR_SENDALOT: 8819 case TCP_RACK_MIN_TO: 8820 case TCP_RACK_EARLY_SEG: 8821 case TCP_RACK_REORD_THRESH: 8822 case TCP_RACK_REORD_FADE: 8823 case TCP_RACK_TLP_THRESH: 8824 case TCP_RACK_PKT_DELAY: 8825 case TCP_RACK_TLP_USE: 8826 case TCP_RACK_TLP_INC_VAR: 8827 case TCP_RACK_IDLE_REDUCE_HIGH: 8828 case TCP_RACK_MIN_PACE: 8829 case TCP_RACK_MIN_PACE_SEG: 8830 case TCP_BBR_RACK_RTT_USE: 8831 case TCP_DATA_AFTER_CLOSE: 8832 break; 8833 default: 8834 return (tcp_default_ctloutput(so, sopt, inp, tp)); 8835 break; 8836 } 8837 INP_WUNLOCK(inp); 8838 error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval)); 8839 if (error) 8840 return (error); 8841 INP_WLOCK(inp); 8842 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { 8843 INP_WUNLOCK(inp); 8844 return (ECONNRESET); 8845 } 8846 tp = intotcpcb(inp); 8847 rack = (struct tcp_rack *)tp->t_fb_ptr; 8848 switch (sopt->sopt_name) { 8849 case TCP_RACK_PROP_RATE: 8850 if ((optval <= 0) || (optval >= 100)) { 8851 error = EINVAL; 8852 break; 8853 } 8854 RACK_OPTS_INC(tcp_rack_prop_rate); 8855 rack->r_ctl.rc_prop_rate = optval; 8856 break; 8857 case TCP_RACK_TLP_USE: 8858 if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) { 8859 error = EINVAL; 8860 break; 8861 } 8862 RACK_OPTS_INC(tcp_tlp_use); 8863 rack->rack_tlp_threshold_use = optval; 8864 break; 8865 case TCP_RACK_PROP: 8866 /* RACK proportional rate reduction (bool) */ 8867 RACK_OPTS_INC(tcp_rack_prop); 8868 rack->r_ctl.rc_prop_reduce = optval; 8869 break; 8870 case TCP_RACK_TLP_REDUCE: 8871 /* RACK TLP cwnd reduction (bool) */ 8872 RACK_OPTS_INC(tcp_rack_tlp_reduce); 8873 rack->r_ctl.rc_tlp_cwnd_reduce = optval; 8874 break; 8875 case TCP_RACK_EARLY_RECOV: 8876 /* Should recovery happen early (bool) */ 8877 RACK_OPTS_INC(tcp_rack_early_recov); 8878 rack->r_ctl.rc_early_recovery = optval; 8879 break; 8880 case TCP_RACK_PACE_ALWAYS: 8881 /* Use the always pace method (bool) */ 8882 RACK_OPTS_INC(tcp_rack_pace_always); 8883 if (optval > 0) 8884 rack->rc_always_pace = 1; 8885 else 8886 rack->rc_always_pace = 0; 8887 break; 8888 case TCP_RACK_PACE_REDUCE: 8889 /* RACK Hptsi reduction factor (divisor) */ 8890 RACK_OPTS_INC(tcp_rack_pace_reduce); 8891 if (optval) 8892 /* Must be non-zero */ 8893 rack->rc_pace_reduce = optval; 8894 else 8895 error = EINVAL; 8896 break; 8897 case TCP_RACK_PACE_MAX_SEG: 8898 /* Max segments in a pace */ 8899 RACK_OPTS_INC(tcp_rack_max_seg); 8900 rack->rc_pace_max_segs = optval; 8901 break; 8902 case TCP_RACK_PRR_SENDALOT: 8903 /* Allow PRR to send more than one seg */ 8904 RACK_OPTS_INC(tcp_rack_prr_sendalot); 8905 rack->r_ctl.rc_prr_sendalot = optval; 8906 break; 8907 case TCP_RACK_MIN_TO: 8908 /* Minimum time between rack t-o's in ms */ 8909 RACK_OPTS_INC(tcp_rack_min_to); 8910 rack->r_ctl.rc_min_to = optval; 8911 break; 8912 case TCP_RACK_EARLY_SEG: 8913 /* If early recovery max segments */ 8914 RACK_OPTS_INC(tcp_rack_early_seg); 8915 rack->r_ctl.rc_early_recovery_segs = optval; 8916 break; 8917 case TCP_RACK_REORD_THRESH: 8918 /* RACK reorder threshold (shift amount) */ 8919 RACK_OPTS_INC(tcp_rack_reord_thresh); 8920 if ((optval > 0) && (optval < 31)) 8921 rack->r_ctl.rc_reorder_shift = optval; 8922 else 8923 error = EINVAL; 8924 break; 8925 case TCP_RACK_REORD_FADE: 8926 /* Does reordering fade after ms time */ 8927 RACK_OPTS_INC(tcp_rack_reord_fade); 8928 rack->r_ctl.rc_reorder_fade = optval; 8929 break; 8930 case TCP_RACK_TLP_THRESH: 8931 /* RACK TLP theshold i.e. srtt+(srtt/N) */ 8932 RACK_OPTS_INC(tcp_rack_tlp_thresh); 8933 if (optval) 8934 rack->r_ctl.rc_tlp_threshold = optval; 8935 else 8936 error = EINVAL; 8937 break; 8938 case TCP_RACK_PKT_DELAY: 8939 /* RACK added ms i.e. rack-rtt + reord + N */ 8940 RACK_OPTS_INC(tcp_rack_pkt_delay); 8941 rack->r_ctl.rc_pkt_delay = optval; 8942 break; 8943 case TCP_RACK_TLP_INC_VAR: 8944 /* Does TLP include rtt variance in t-o */ 8945 RACK_OPTS_INC(tcp_rack_tlp_inc_var); 8946 rack->r_ctl.rc_prr_inc_var = optval; 8947 break; 8948 case TCP_RACK_IDLE_REDUCE_HIGH: 8949 RACK_OPTS_INC(tcp_rack_idle_reduce_high); 8950 if (optval) 8951 rack->r_idle_reduce_largest = 1; 8952 else 8953 rack->r_idle_reduce_largest = 0; 8954 break; 8955 case TCP_DELACK: 8956 if (optval == 0) 8957 tp->t_delayed_ack = 0; 8958 else 8959 tp->t_delayed_ack = 1; 8960 if (tp->t_flags & TF_DELACK) { 8961 tp->t_flags &= ~TF_DELACK; 8962 tp->t_flags |= TF_ACKNOW; 8963 rack_output(tp); 8964 } 8965 break; 8966 case TCP_RACK_MIN_PACE: 8967 RACK_OPTS_INC(tcp_rack_min_pace); 8968 if (optval > 3) 8969 rack->r_enforce_min_pace = 3; 8970 else 8971 rack->r_enforce_min_pace = optval; 8972 break; 8973 case TCP_RACK_MIN_PACE_SEG: 8974 RACK_OPTS_INC(tcp_rack_min_pace_seg); 8975 if (optval >= 16) 8976 rack->r_min_pace_seg_thresh = 15; 8977 else 8978 rack->r_min_pace_seg_thresh = optval; 8979 break; 8980 case TCP_BBR_RACK_RTT_USE: 8981 if ((optval != USE_RTT_HIGH) && 8982 (optval != USE_RTT_LOW) && 8983 (optval != USE_RTT_AVG)) 8984 error = EINVAL; 8985 else 8986 rack->r_ctl.rc_rate_sample_method = optval; 8987 break; 8988 case TCP_DATA_AFTER_CLOSE: 8989 if (optval) 8990 rack->rc_allow_data_af_clo = 1; 8991 else 8992 rack->rc_allow_data_af_clo = 0; 8993 break; 8994 default: 8995 return (tcp_default_ctloutput(so, sopt, inp, tp)); 8996 break; 8997 } 8998 /* tcp_log_socket_option(tp, sopt->sopt_name, optval, error);*/ 8999 INP_WUNLOCK(inp); 9000 return (error); 9001 } 9002 9003 static int 9004 rack_get_sockopt(struct socket *so, struct sockopt *sopt, 9005 struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack) 9006 { 9007 int32_t error, optval; 9008 9009 /* 9010 * Because all our options are either boolean or an int, we can just 9011 * pull everything into optval and then unlock and copy. If we ever 9012 * add a option that is not a int, then this will have quite an 9013 * impact to this routine. 9014 */ 9015 switch (sopt->sopt_name) { 9016 case TCP_RACK_PROP_RATE: 9017 optval = rack->r_ctl.rc_prop_rate; 9018 break; 9019 case TCP_RACK_PROP: 9020 /* RACK proportional rate reduction (bool) */ 9021 optval = rack->r_ctl.rc_prop_reduce; 9022 break; 9023 case TCP_RACK_TLP_REDUCE: 9024 /* RACK TLP cwnd reduction (bool) */ 9025 optval = rack->r_ctl.rc_tlp_cwnd_reduce; 9026 break; 9027 case TCP_RACK_EARLY_RECOV: 9028 /* Should recovery happen early (bool) */ 9029 optval = rack->r_ctl.rc_early_recovery; 9030 break; 9031 case TCP_RACK_PACE_REDUCE: 9032 /* RACK Hptsi reduction factor (divisor) */ 9033 optval = rack->rc_pace_reduce; 9034 break; 9035 case TCP_RACK_PACE_MAX_SEG: 9036 /* Max segments in a pace */ 9037 optval = rack->rc_pace_max_segs; 9038 break; 9039 case TCP_RACK_PACE_ALWAYS: 9040 /* Use the always pace method */ 9041 optval = rack->rc_always_pace; 9042 break; 9043 case TCP_RACK_PRR_SENDALOT: 9044 /* Allow PRR to send more than one seg */ 9045 optval = rack->r_ctl.rc_prr_sendalot; 9046 break; 9047 case TCP_RACK_MIN_TO: 9048 /* Minimum time between rack t-o's in ms */ 9049 optval = rack->r_ctl.rc_min_to; 9050 break; 9051 case TCP_RACK_EARLY_SEG: 9052 /* If early recovery max segments */ 9053 optval = rack->r_ctl.rc_early_recovery_segs; 9054 break; 9055 case TCP_RACK_REORD_THRESH: 9056 /* RACK reorder threshold (shift amount) */ 9057 optval = rack->r_ctl.rc_reorder_shift; 9058 break; 9059 case TCP_RACK_REORD_FADE: 9060 /* Does reordering fade after ms time */ 9061 optval = rack->r_ctl.rc_reorder_fade; 9062 break; 9063 case TCP_RACK_TLP_THRESH: 9064 /* RACK TLP theshold i.e. srtt+(srtt/N) */ 9065 optval = rack->r_ctl.rc_tlp_threshold; 9066 break; 9067 case TCP_RACK_PKT_DELAY: 9068 /* RACK added ms i.e. rack-rtt + reord + N */ 9069 optval = rack->r_ctl.rc_pkt_delay; 9070 break; 9071 case TCP_RACK_TLP_USE: 9072 optval = rack->rack_tlp_threshold_use; 9073 break; 9074 case TCP_RACK_TLP_INC_VAR: 9075 /* Does TLP include rtt variance in t-o */ 9076 optval = rack->r_ctl.rc_prr_inc_var; 9077 break; 9078 case TCP_RACK_IDLE_REDUCE_HIGH: 9079 optval = rack->r_idle_reduce_largest; 9080 break; 9081 case TCP_RACK_MIN_PACE: 9082 optval = rack->r_enforce_min_pace; 9083 break; 9084 case TCP_RACK_MIN_PACE_SEG: 9085 optval = rack->r_min_pace_seg_thresh; 9086 break; 9087 case TCP_BBR_RACK_RTT_USE: 9088 optval = rack->r_ctl.rc_rate_sample_method; 9089 break; 9090 case TCP_DELACK: 9091 optval = tp->t_delayed_ack; 9092 break; 9093 case TCP_DATA_AFTER_CLOSE: 9094 optval = rack->rc_allow_data_af_clo; 9095 break; 9096 default: 9097 return (tcp_default_ctloutput(so, sopt, inp, tp)); 9098 break; 9099 } 9100 INP_WUNLOCK(inp); 9101 error = sooptcopyout(sopt, &optval, sizeof optval); 9102 return (error); 9103 } 9104 9105 static int 9106 rack_ctloutput(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp) 9107 { 9108 int32_t error = EINVAL; 9109 struct tcp_rack *rack; 9110 9111 rack = (struct tcp_rack *)tp->t_fb_ptr; 9112 if (rack == NULL) { 9113 /* Huh? */ 9114 goto out; 9115 } 9116 if (sopt->sopt_dir == SOPT_SET) { 9117 return (rack_set_sockopt(so, sopt, inp, tp, rack)); 9118 } else if (sopt->sopt_dir == SOPT_GET) { 9119 return (rack_get_sockopt(so, sopt, inp, tp, rack)); 9120 } 9121 out: 9122 INP_WUNLOCK(inp); 9123 return (error); 9124 } 9125 9126 9127 struct tcp_function_block __tcp_rack = { 9128 .tfb_tcp_block_name = __XSTRING(STACKNAME), 9129 .tfb_tcp_output = rack_output, 9130 .tfb_tcp_do_segment = rack_do_segment, 9131 .tfb_tcp_ctloutput = rack_ctloutput, 9132 .tfb_tcp_fb_init = rack_init, 9133 .tfb_tcp_fb_fini = rack_fini, 9134 .tfb_tcp_timer_stop_all = rack_stopall, 9135 .tfb_tcp_timer_activate = rack_timer_activate, 9136 .tfb_tcp_timer_active = rack_timer_active, 9137 .tfb_tcp_timer_stop = rack_timer_stop, 9138 .tfb_tcp_rexmit_tmr = rack_remxt_tmr, 9139 .tfb_tcp_handoff_ok = rack_handoff_ok 9140 }; 9141 9142 static const char *rack_stack_names[] = { 9143 __XSTRING(STACKNAME), 9144 #ifdef STACKALIAS 9145 __XSTRING(STACKALIAS), 9146 #endif 9147 }; 9148 9149 static int 9150 rack_ctor(void *mem, int32_t size, void *arg, int32_t how) 9151 { 9152 memset(mem, 0, size); 9153 return (0); 9154 } 9155 9156 static void 9157 rack_dtor(void *mem, int32_t size, void *arg) 9158 { 9159 9160 } 9161 9162 static bool rack_mod_inited = false; 9163 9164 static int 9165 tcp_addrack(module_t mod, int32_t type, void *data) 9166 { 9167 int32_t err = 0; 9168 int num_stacks; 9169 9170 switch (type) { 9171 case MOD_LOAD: 9172 rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map", 9173 sizeof(struct rack_sendmap), 9174 rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0); 9175 9176 rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb", 9177 sizeof(struct tcp_rack), 9178 rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0); 9179 9180 sysctl_ctx_init(&rack_sysctl_ctx); 9181 rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx, 9182 SYSCTL_STATIC_CHILDREN(_net_inet_tcp), 9183 OID_AUTO, 9184 __XSTRING(STACKNAME), 9185 CTLFLAG_RW, 0, 9186 ""); 9187 if (rack_sysctl_root == NULL) { 9188 printf("Failed to add sysctl node\n"); 9189 err = EFAULT; 9190 goto free_uma; 9191 } 9192 rack_init_sysctls(); 9193 num_stacks = nitems(rack_stack_names); 9194 err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK, 9195 rack_stack_names, &num_stacks); 9196 if (err) { 9197 printf("Failed to register %s stack name for " 9198 "%s module\n", rack_stack_names[num_stacks], 9199 __XSTRING(MODNAME)); 9200 sysctl_ctx_free(&rack_sysctl_ctx); 9201 free_uma: 9202 uma_zdestroy(rack_zone); 9203 uma_zdestroy(rack_pcb_zone); 9204 rack_counter_destroy(); 9205 printf("Failed to register rack module -- err:%d\n", err); 9206 return (err); 9207 } 9208 rack_mod_inited = true; 9209 break; 9210 case MOD_QUIESCE: 9211 err = deregister_tcp_functions(&__tcp_rack, true, false); 9212 break; 9213 case MOD_UNLOAD: 9214 err = deregister_tcp_functions(&__tcp_rack, false, true); 9215 if (err == EBUSY) 9216 break; 9217 if (rack_mod_inited) { 9218 uma_zdestroy(rack_zone); 9219 uma_zdestroy(rack_pcb_zone); 9220 sysctl_ctx_free(&rack_sysctl_ctx); 9221 rack_counter_destroy(); 9222 rack_mod_inited = false; 9223 } 9224 err = 0; 9225 break; 9226 default: 9227 return (EOPNOTSUPP); 9228 } 9229 return (err); 9230 } 9231 9232 static moduledata_t tcp_rack = { 9233 .name = __XSTRING(MODNAME), 9234 .evhand = tcp_addrack, 9235 .priv = 0 9236 }; 9237 9238 MODULE_VERSION(MODNAME, 1); 9239 DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY); 9240 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1); 9241