1 /*- 2 * Copyright (c) 2016-2019 Netflix, Inc. 3 * 4 * Redistribution and use in source and binary forms, with or without 5 * modification, are permitted provided that the following conditions 6 * are met: 7 * 1. Redistributions of source code must retain the above copyright 8 * notice, this list of conditions and the following disclaimer. 9 * 2. Redistributions in binary form must reproduce the above copyright 10 * notice, this list of conditions and the following disclaimer in the 11 * documentation and/or other materials provided with the distribution. 12 * 13 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 14 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 15 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 16 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 17 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 18 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 19 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 20 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 21 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 22 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 23 * SUCH DAMAGE. 24 * 25 */ 26 27 #include <sys/cdefs.h> 28 __FBSDID("$FreeBSD$"); 29 30 #include "opt_inet.h" 31 #include "opt_inet6.h" 32 #include "opt_ipsec.h" 33 #include "opt_tcpdebug.h" 34 35 #include <sys/param.h> 36 #include <sys/module.h> 37 #include <sys/kernel.h> 38 #ifdef TCP_HHOOK 39 #include <sys/hhook.h> 40 #endif 41 #include <sys/lock.h> 42 #include <sys/malloc.h> 43 #include <sys/lock.h> 44 #include <sys/mutex.h> 45 #include <sys/mbuf.h> 46 #include <sys/proc.h> /* for proc0 declaration */ 47 #include <sys/socket.h> 48 #include <sys/socketvar.h> 49 #include <sys/sysctl.h> 50 #include <sys/systm.h> 51 #ifdef NETFLIX_STATS 52 #include <sys/stats.h> 53 #endif 54 #include <sys/refcount.h> 55 #include <sys/queue.h> 56 #include <sys/smp.h> 57 #include <sys/kthread.h> 58 #include <sys/kern_prefetch.h> 59 60 #include <vm/uma.h> 61 62 #include <net/route.h> 63 #include <net/vnet.h> 64 65 #define TCPSTATES /* for logging */ 66 67 #include <netinet/in.h> 68 #include <netinet/in_kdtrace.h> 69 #include <netinet/in_pcb.h> 70 #include <netinet/ip.h> 71 #include <netinet/ip_icmp.h> /* required for icmp_var.h */ 72 #include <netinet/icmp_var.h> /* for ICMP_BANDLIM */ 73 #include <netinet/ip_var.h> 74 #include <netinet/ip6.h> 75 #include <netinet6/in6_pcb.h> 76 #include <netinet6/ip6_var.h> 77 #include <netinet/tcp.h> 78 #define TCPOUTFLAGS 79 #include <netinet/tcp_fsm.h> 80 #include <netinet/tcp_log_buf.h> 81 #include <netinet/tcp_seq.h> 82 #include <netinet/tcp_timer.h> 83 #include <netinet/tcp_var.h> 84 #include <netinet/tcp_hpts.h> 85 #include <netinet/tcpip.h> 86 #include <netinet/cc/cc.h> 87 #ifdef NETFLIX_CWV 88 #include <netinet/tcp_newcwv.h> 89 #endif 90 #include <netinet/tcp_fastopen.h> 91 #ifdef TCPDEBUG 92 #include <netinet/tcp_debug.h> 93 #endif /* TCPDEBUG */ 94 #ifdef TCP_OFFLOAD 95 #include <netinet/tcp_offload.h> 96 #endif 97 #ifdef INET6 98 #include <netinet6/tcp6_var.h> 99 #endif 100 101 #include <netipsec/ipsec_support.h> 102 103 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 104 #include <netipsec/ipsec.h> 105 #include <netipsec/ipsec6.h> 106 #endif /* IPSEC */ 107 108 #include <netinet/udp.h> 109 #include <netinet/udp_var.h> 110 #include <machine/in_cksum.h> 111 112 #ifdef MAC 113 #include <security/mac/mac_framework.h> 114 #endif 115 #include "sack_filter.h" 116 #include "tcp_rack.h" 117 #include "rack_bbr_common.h" 118 119 uma_zone_t rack_zone; 120 uma_zone_t rack_pcb_zone; 121 122 #ifndef TICKS2SBT 123 #define TICKS2SBT(__t) (tick_sbt * ((sbintime_t)(__t))) 124 #endif 125 126 struct sysctl_ctx_list rack_sysctl_ctx; 127 struct sysctl_oid *rack_sysctl_root; 128 129 #define CUM_ACKED 1 130 #define SACKED 2 131 132 /* 133 * The RACK module incorporates a number of 134 * TCP ideas that have been put out into the IETF 135 * over the last few years: 136 * - Matt Mathis's Rate Halving which slowly drops 137 * the congestion window so that the ack clock can 138 * be maintained during a recovery. 139 * - Yuchung Cheng's RACK TCP (for which its named) that 140 * will stop us using the number of dup acks and instead 141 * use time as the gage of when we retransmit. 142 * - Reorder Detection of RFC4737 and the Tail-Loss probe draft 143 * of Dukkipati et.al. 144 * RACK depends on SACK, so if an endpoint arrives that 145 * cannot do SACK the state machine below will shuttle the 146 * connection back to using the "default" TCP stack that is 147 * in FreeBSD. 148 * 149 * To implement RACK the original TCP stack was first decomposed 150 * into a functional state machine with individual states 151 * for each of the possible TCP connection states. The do_segement 152 * functions role in life is to mandate the connection supports SACK 153 * initially and then assure that the RACK state matches the conenction 154 * state before calling the states do_segment function. Each 155 * state is simplified due to the fact that the original do_segment 156 * has been decomposed and we *know* what state we are in (no 157 * switches on the state) and all tests for SACK are gone. This 158 * greatly simplifies what each state does. 159 * 160 * TCP output is also over-written with a new version since it 161 * must maintain the new rack scoreboard. 162 * 163 */ 164 static int32_t rack_precache = 1; 165 static int32_t rack_tlp_thresh = 1; 166 static int32_t rack_reorder_thresh = 2; 167 static int32_t rack_reorder_fade = 60000; /* 0 - never fade, def 60,000 168 * - 60 seconds */ 169 static int32_t rack_pkt_delay = 1; 170 static int32_t rack_inc_var = 0;/* For TLP */ 171 static int32_t rack_reduce_largest_on_idle = 0; 172 static int32_t rack_min_pace_time = 0; 173 static int32_t rack_min_pace_time_seg_req=6; 174 static int32_t rack_early_recovery = 1; 175 static int32_t rack_early_recovery_max_seg = 6; 176 static int32_t rack_send_a_lot_in_prr = 1; 177 static int32_t rack_min_to = 1; /* Number of ms minimum timeout */ 178 static int32_t rack_tlp_in_recovery = 1; /* Can we do TLP in recovery? */ 179 static int32_t rack_verbose_logging = 0; 180 static int32_t rack_ignore_data_after_close = 1; 181 /* 182 * Currently regular tcp has a rto_min of 30ms 183 * the backoff goes 12 times so that ends up 184 * being a total of 122.850 seconds before a 185 * connection is killed. 186 */ 187 static int32_t rack_tlp_min = 10; 188 static int32_t rack_rto_min = 30; /* 30ms same as main freebsd */ 189 static int32_t rack_rto_max = 30000; /* 30 seconds */ 190 static const int32_t rack_free_cache = 2; 191 static int32_t rack_hptsi_segments = 40; 192 static int32_t rack_rate_sample_method = USE_RTT_LOW; 193 static int32_t rack_pace_every_seg = 1; 194 static int32_t rack_delayed_ack_time = 200; /* 200ms */ 195 static int32_t rack_slot_reduction = 4; 196 static int32_t rack_lower_cwnd_at_tlp = 0; 197 static int32_t rack_use_proportional_reduce = 0; 198 static int32_t rack_proportional_rate = 10; 199 static int32_t rack_tlp_max_resend = 2; 200 static int32_t rack_limited_retran = 0; 201 static int32_t rack_always_send_oldest = 0; 202 static int32_t rack_sack_block_limit = 128; 203 static int32_t rack_use_sack_filter = 1; 204 static int32_t rack_tlp_threshold_use = TLP_USE_TWO_ONE; 205 static uint32_t rack_map_split_limit = 0; /* unlimited by default */ 206 207 /* Rack specific counters */ 208 counter_u64_t rack_badfr; 209 counter_u64_t rack_badfr_bytes; 210 counter_u64_t rack_rtm_prr_retran; 211 counter_u64_t rack_rtm_prr_newdata; 212 counter_u64_t rack_timestamp_mismatch; 213 counter_u64_t rack_reorder_seen; 214 counter_u64_t rack_paced_segments; 215 counter_u64_t rack_unpaced_segments; 216 counter_u64_t rack_saw_enobuf; 217 counter_u64_t rack_saw_enetunreach; 218 219 /* Tail loss probe counters */ 220 counter_u64_t rack_tlp_tot; 221 counter_u64_t rack_tlp_newdata; 222 counter_u64_t rack_tlp_retran; 223 counter_u64_t rack_tlp_retran_bytes; 224 counter_u64_t rack_tlp_retran_fail; 225 counter_u64_t rack_to_tot; 226 counter_u64_t rack_to_arm_rack; 227 counter_u64_t rack_to_arm_tlp; 228 counter_u64_t rack_to_alloc; 229 counter_u64_t rack_to_alloc_hard; 230 counter_u64_t rack_to_alloc_emerg; 231 counter_u64_t rack_alloc_limited_conns; 232 counter_u64_t rack_split_limited; 233 234 counter_u64_t rack_sack_proc_all; 235 counter_u64_t rack_sack_proc_short; 236 counter_u64_t rack_sack_proc_restart; 237 counter_u64_t rack_runt_sacks; 238 counter_u64_t rack_used_tlpmethod; 239 counter_u64_t rack_used_tlpmethod2; 240 counter_u64_t rack_enter_tlp_calc; 241 counter_u64_t rack_input_idle_reduces; 242 counter_u64_t rack_tlp_does_nada; 243 244 /* Temp CPU counters */ 245 counter_u64_t rack_find_high; 246 247 counter_u64_t rack_progress_drops; 248 counter_u64_t rack_out_size[TCP_MSS_ACCT_SIZE]; 249 counter_u64_t rack_opts_arry[RACK_OPTS_SIZE]; 250 251 static void 252 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick, int event, int line); 253 254 static int 255 rack_process_ack(struct mbuf *m, struct tcphdr *th, 256 struct socket *so, struct tcpcb *tp, struct tcpopt *to, 257 uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t * ret_val); 258 static int 259 rack_process_data(struct mbuf *m, struct tcphdr *th, 260 struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, 261 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); 262 static void 263 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, 264 struct tcphdr *th, uint16_t nsegs, uint16_t type, int32_t recovery); 265 static struct rack_sendmap *rack_alloc(struct tcp_rack *rack); 266 static struct rack_sendmap *rack_alloc_limit(struct tcp_rack *rack, 267 uint8_t limit_type); 268 static struct rack_sendmap * 269 rack_check_recovery_mode(struct tcpcb *tp, 270 uint32_t tsused); 271 static void 272 rack_cong_signal(struct tcpcb *tp, struct tcphdr *th, 273 uint32_t type); 274 static void rack_counter_destroy(void); 275 static int 276 rack_ctloutput(struct socket *so, struct sockopt *sopt, 277 struct inpcb *inp, struct tcpcb *tp); 278 static int32_t rack_ctor(void *mem, int32_t size, void *arg, int32_t how); 279 static void 280 rack_do_segment(struct mbuf *m, struct tcphdr *th, 281 struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, 282 uint8_t iptos); 283 static void rack_dtor(void *mem, int32_t size, void *arg); 284 static void 285 rack_earlier_retran(struct tcpcb *tp, struct rack_sendmap *rsm, 286 uint32_t t, uint32_t cts); 287 static struct rack_sendmap * 288 rack_find_high_nonack(struct tcp_rack *rack, 289 struct rack_sendmap *rsm); 290 static struct rack_sendmap *rack_find_lowest_rsm(struct tcp_rack *rack); 291 static void rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm); 292 static void rack_fini(struct tcpcb *tp, int32_t tcb_is_purged); 293 static int 294 rack_get_sockopt(struct socket *so, struct sockopt *sopt, 295 struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack); 296 static int32_t rack_handoff_ok(struct tcpcb *tp); 297 static int32_t rack_init(struct tcpcb *tp); 298 static void rack_init_sysctls(void); 299 static void 300 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, 301 struct tcphdr *th); 302 static void 303 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len, 304 uint32_t seq_out, uint8_t th_flags, int32_t err, uint32_t ts, 305 uint8_t pass, struct rack_sendmap *hintrsm); 306 static void 307 rack_log_sack_passed(struct tcpcb *tp, struct tcp_rack *rack, 308 struct rack_sendmap *rsm); 309 static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num); 310 static int32_t rack_output(struct tcpcb *tp); 311 static void 312 rack_hpts_do_segment(struct mbuf *m, struct tcphdr *th, 313 struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, 314 uint8_t iptos, int32_t nxt_pkt, struct timeval *tv); 315 316 static uint32_t 317 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, 318 struct sackblk *sack, struct tcpopt *to, struct rack_sendmap **prsm, 319 uint32_t cts); 320 static void rack_post_recovery(struct tcpcb *tp, struct tcphdr *th); 321 static void rack_remxt_tmr(struct tcpcb *tp); 322 static int 323 rack_set_sockopt(struct socket *so, struct sockopt *sopt, 324 struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack); 325 static void rack_set_state(struct tcpcb *tp, struct tcp_rack *rack); 326 static int32_t rack_stopall(struct tcpcb *tp); 327 static void 328 rack_timer_activate(struct tcpcb *tp, uint32_t timer_type, 329 uint32_t delta); 330 static int32_t rack_timer_active(struct tcpcb *tp, uint32_t timer_type); 331 static void rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line); 332 static void rack_timer_stop(struct tcpcb *tp, uint32_t timer_type); 333 static uint32_t 334 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack, 335 struct rack_sendmap *rsm, uint32_t ts, int32_t * lenp); 336 static void 337 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack, 338 struct rack_sendmap *rsm, uint32_t ts); 339 static int 340 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack, 341 struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type); 342 static int32_t tcp_addrack(module_t mod, int32_t type, void *data); 343 static void 344 rack_challenge_ack(struct mbuf *m, struct tcphdr *th, 345 struct tcpcb *tp, int32_t * ret_val); 346 static int 347 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, 348 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, 349 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); 350 static int 351 rack_do_closing(struct mbuf *m, struct tcphdr *th, 352 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, 353 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); 354 static void 355 rack_do_drop(struct mbuf *m, struct tcpcb *tp); 356 static void 357 rack_do_dropafterack(struct mbuf *m, struct tcpcb *tp, 358 struct tcphdr *th, int32_t thflags, int32_t tlen, int32_t * ret_val); 359 static void 360 rack_do_dropwithreset(struct mbuf *m, struct tcpcb *tp, 361 struct tcphdr *th, int32_t rstreason, int32_t tlen); 362 static int 363 rack_do_established(struct mbuf *m, struct tcphdr *th, 364 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, 365 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); 366 static int 367 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, 368 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, 369 int32_t tlen, uint32_t tiwin, int32_t nxt_pkt); 370 static int 371 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, 372 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, 373 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); 374 static int 375 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, 376 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, 377 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); 378 static int 379 rack_do_lastack(struct mbuf *m, struct tcphdr *th, 380 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, 381 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); 382 static int 383 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, 384 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, 385 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); 386 static int 387 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, 388 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, 389 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); 390 static int 391 rack_drop_checks(struct tcpopt *to, struct mbuf *m, 392 struct tcphdr *th, struct tcpcb *tp, int32_t * tlenp, int32_t * thf, 393 int32_t * drop_hdrlen, int32_t * ret_val); 394 static int 395 rack_process_rst(struct mbuf *m, struct tcphdr *th, 396 struct socket *so, struct tcpcb *tp); 397 struct rack_sendmap * 398 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, 399 uint32_t tsused); 400 static void tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt); 401 static void 402 tcp_rack_partialack(struct tcpcb *tp, struct tcphdr *th); 403 404 static int 405 rack_ts_check(struct mbuf *m, struct tcphdr *th, 406 struct tcpcb *tp, int32_t tlen, int32_t thflags, int32_t * ret_val); 407 408 int32_t rack_clear_counter=0; 409 410 411 static int 412 sysctl_rack_clear(SYSCTL_HANDLER_ARGS) 413 { 414 uint32_t stat; 415 int32_t error; 416 417 error = SYSCTL_OUT(req, &rack_clear_counter, sizeof(uint32_t)); 418 if (error || req->newptr == NULL) 419 return error; 420 421 error = SYSCTL_IN(req, &stat, sizeof(uint32_t)); 422 if (error) 423 return (error); 424 if (stat == 1) { 425 #ifdef INVARIANTS 426 printf("Clearing RACK counters\n"); 427 #endif 428 counter_u64_zero(rack_badfr); 429 counter_u64_zero(rack_badfr_bytes); 430 counter_u64_zero(rack_rtm_prr_retran); 431 counter_u64_zero(rack_rtm_prr_newdata); 432 counter_u64_zero(rack_timestamp_mismatch); 433 counter_u64_zero(rack_reorder_seen); 434 counter_u64_zero(rack_tlp_tot); 435 counter_u64_zero(rack_tlp_newdata); 436 counter_u64_zero(rack_tlp_retran); 437 counter_u64_zero(rack_tlp_retran_bytes); 438 counter_u64_zero(rack_tlp_retran_fail); 439 counter_u64_zero(rack_to_tot); 440 counter_u64_zero(rack_to_arm_rack); 441 counter_u64_zero(rack_to_arm_tlp); 442 counter_u64_zero(rack_paced_segments); 443 counter_u64_zero(rack_unpaced_segments); 444 counter_u64_zero(rack_saw_enobuf); 445 counter_u64_zero(rack_saw_enetunreach); 446 counter_u64_zero(rack_to_alloc_hard); 447 counter_u64_zero(rack_to_alloc_emerg); 448 counter_u64_zero(rack_sack_proc_all); 449 counter_u64_zero(rack_sack_proc_short); 450 counter_u64_zero(rack_sack_proc_restart); 451 counter_u64_zero(rack_to_alloc); 452 counter_u64_zero(rack_alloc_limited_conns); 453 counter_u64_zero(rack_split_limited); 454 counter_u64_zero(rack_find_high); 455 counter_u64_zero(rack_runt_sacks); 456 counter_u64_zero(rack_used_tlpmethod); 457 counter_u64_zero(rack_used_tlpmethod2); 458 counter_u64_zero(rack_enter_tlp_calc); 459 counter_u64_zero(rack_progress_drops); 460 counter_u64_zero(rack_tlp_does_nada); 461 } 462 rack_clear_counter = 0; 463 return (0); 464 } 465 466 467 468 static void 469 rack_init_sysctls() 470 { 471 SYSCTL_ADD_S32(&rack_sysctl_ctx, 472 SYSCTL_CHILDREN(rack_sysctl_root), 473 OID_AUTO, "rate_sample_method", CTLFLAG_RW, 474 &rack_rate_sample_method , USE_RTT_LOW, 475 "What method should we use for rate sampling 0=high, 1=low "); 476 SYSCTL_ADD_S32(&rack_sysctl_ctx, 477 SYSCTL_CHILDREN(rack_sysctl_root), 478 OID_AUTO, "data_after_close", CTLFLAG_RW, 479 &rack_ignore_data_after_close, 0, 480 "Do we hold off sending a RST until all pending data is ack'd"); 481 SYSCTL_ADD_S32(&rack_sysctl_ctx, 482 SYSCTL_CHILDREN(rack_sysctl_root), 483 OID_AUTO, "tlpmethod", CTLFLAG_RW, 484 &rack_tlp_threshold_use, TLP_USE_TWO_ONE, 485 "What method do we do for TLP time calc 0=no-de-ack-comp, 1=ID, 2=2.1, 3=2.2"); 486 SYSCTL_ADD_S32(&rack_sysctl_ctx, 487 SYSCTL_CHILDREN(rack_sysctl_root), 488 OID_AUTO, "min_pace_time", CTLFLAG_RW, 489 &rack_min_pace_time, 0, 490 "Should we enforce a minimum pace time of 1ms"); 491 SYSCTL_ADD_S32(&rack_sysctl_ctx, 492 SYSCTL_CHILDREN(rack_sysctl_root), 493 OID_AUTO, "min_pace_segs", CTLFLAG_RW, 494 &rack_min_pace_time_seg_req, 6, 495 "How many segments have to be in the len to enforce min-pace-time"); 496 SYSCTL_ADD_S32(&rack_sysctl_ctx, 497 SYSCTL_CHILDREN(rack_sysctl_root), 498 OID_AUTO, "idle_reduce_high", CTLFLAG_RW, 499 &rack_reduce_largest_on_idle, 0, 500 "Should we reduce the largest cwnd seen to IW on idle reduction"); 501 SYSCTL_ADD_S32(&rack_sysctl_ctx, 502 SYSCTL_CHILDREN(rack_sysctl_root), 503 OID_AUTO, "bb_verbose", CTLFLAG_RW, 504 &rack_verbose_logging, 0, 505 "Should RACK black box logging be verbose"); 506 SYSCTL_ADD_S32(&rack_sysctl_ctx, 507 SYSCTL_CHILDREN(rack_sysctl_root), 508 OID_AUTO, "sackfiltering", CTLFLAG_RW, 509 &rack_use_sack_filter, 1, 510 "Do we use sack filtering?"); 511 SYSCTL_ADD_S32(&rack_sysctl_ctx, 512 SYSCTL_CHILDREN(rack_sysctl_root), 513 OID_AUTO, "delayed_ack", CTLFLAG_RW, 514 &rack_delayed_ack_time, 200, 515 "Delayed ack time (200ms)"); 516 SYSCTL_ADD_S32(&rack_sysctl_ctx, 517 SYSCTL_CHILDREN(rack_sysctl_root), 518 OID_AUTO, "tlpminto", CTLFLAG_RW, 519 &rack_tlp_min, 10, 520 "TLP minimum timeout per the specification (10ms)"); 521 SYSCTL_ADD_S32(&rack_sysctl_ctx, 522 SYSCTL_CHILDREN(rack_sysctl_root), 523 OID_AUTO, "precache", CTLFLAG_RW, 524 &rack_precache, 0, 525 "Where should we precache the mcopy (0 is not at all)"); 526 SYSCTL_ADD_S32(&rack_sysctl_ctx, 527 SYSCTL_CHILDREN(rack_sysctl_root), 528 OID_AUTO, "sblklimit", CTLFLAG_RW, 529 &rack_sack_block_limit, 128, 530 "When do we start paying attention to small sack blocks"); 531 SYSCTL_ADD_S32(&rack_sysctl_ctx, 532 SYSCTL_CHILDREN(rack_sysctl_root), 533 OID_AUTO, "send_oldest", CTLFLAG_RW, 534 &rack_always_send_oldest, 1, 535 "Should we always send the oldest TLP and RACK-TLP"); 536 SYSCTL_ADD_S32(&rack_sysctl_ctx, 537 SYSCTL_CHILDREN(rack_sysctl_root), 538 OID_AUTO, "rack_tlp_in_recovery", CTLFLAG_RW, 539 &rack_tlp_in_recovery, 1, 540 "Can we do a TLP during recovery?"); 541 SYSCTL_ADD_S32(&rack_sysctl_ctx, 542 SYSCTL_CHILDREN(rack_sysctl_root), 543 OID_AUTO, "rack_tlimit", CTLFLAG_RW, 544 &rack_limited_retran, 0, 545 "How many times can a rack timeout drive out sends"); 546 SYSCTL_ADD_S32(&rack_sysctl_ctx, 547 SYSCTL_CHILDREN(rack_sysctl_root), 548 OID_AUTO, "minrto", CTLFLAG_RW, 549 &rack_rto_min, 0, 550 "Minimum RTO in ms -- set with caution below 1000 due to TLP"); 551 SYSCTL_ADD_S32(&rack_sysctl_ctx, 552 SYSCTL_CHILDREN(rack_sysctl_root), 553 OID_AUTO, "maxrto", CTLFLAG_RW, 554 &rack_rto_max, 0, 555 "Maxiumum RTO in ms -- should be at least as large as min_rto"); 556 SYSCTL_ADD_S32(&rack_sysctl_ctx, 557 SYSCTL_CHILDREN(rack_sysctl_root), 558 OID_AUTO, "tlp_retry", CTLFLAG_RW, 559 &rack_tlp_max_resend, 2, 560 "How many times does TLP retry a single segment or multiple with no ACK"); 561 SYSCTL_ADD_S32(&rack_sysctl_ctx, 562 SYSCTL_CHILDREN(rack_sysctl_root), 563 OID_AUTO, "recovery_loss_prop", CTLFLAG_RW, 564 &rack_use_proportional_reduce, 0, 565 "Should we proportionaly reduce cwnd based on the number of losses "); 566 SYSCTL_ADD_S32(&rack_sysctl_ctx, 567 SYSCTL_CHILDREN(rack_sysctl_root), 568 OID_AUTO, "recovery_prop", CTLFLAG_RW, 569 &rack_proportional_rate, 10, 570 "What percent reduction per loss"); 571 SYSCTL_ADD_S32(&rack_sysctl_ctx, 572 SYSCTL_CHILDREN(rack_sysctl_root), 573 OID_AUTO, "tlp_cwnd_flag", CTLFLAG_RW, 574 &rack_lower_cwnd_at_tlp, 0, 575 "When a TLP completes a retran should we enter recovery?"); 576 SYSCTL_ADD_S32(&rack_sysctl_ctx, 577 SYSCTL_CHILDREN(rack_sysctl_root), 578 OID_AUTO, "hptsi_reduces", CTLFLAG_RW, 579 &rack_slot_reduction, 4, 580 "When setting a slot should we reduce by divisor"); 581 SYSCTL_ADD_S32(&rack_sysctl_ctx, 582 SYSCTL_CHILDREN(rack_sysctl_root), 583 OID_AUTO, "hptsi_every_seg", CTLFLAG_RW, 584 &rack_pace_every_seg, 1, 585 "Should we pace out every segment hptsi"); 586 SYSCTL_ADD_S32(&rack_sysctl_ctx, 587 SYSCTL_CHILDREN(rack_sysctl_root), 588 OID_AUTO, "hptsi_seg_max", CTLFLAG_RW, 589 &rack_hptsi_segments, 6, 590 "Should we pace out only a limited size of segments"); 591 SYSCTL_ADD_S32(&rack_sysctl_ctx, 592 SYSCTL_CHILDREN(rack_sysctl_root), 593 OID_AUTO, "prr_sendalot", CTLFLAG_RW, 594 &rack_send_a_lot_in_prr, 1, 595 "Send a lot in prr"); 596 SYSCTL_ADD_S32(&rack_sysctl_ctx, 597 SYSCTL_CHILDREN(rack_sysctl_root), 598 OID_AUTO, "minto", CTLFLAG_RW, 599 &rack_min_to, 1, 600 "Minimum rack timeout in milliseconds"); 601 SYSCTL_ADD_S32(&rack_sysctl_ctx, 602 SYSCTL_CHILDREN(rack_sysctl_root), 603 OID_AUTO, "earlyrecoveryseg", CTLFLAG_RW, 604 &rack_early_recovery_max_seg, 6, 605 "Max segments in early recovery"); 606 SYSCTL_ADD_S32(&rack_sysctl_ctx, 607 SYSCTL_CHILDREN(rack_sysctl_root), 608 OID_AUTO, "earlyrecovery", CTLFLAG_RW, 609 &rack_early_recovery, 1, 610 "Do we do early recovery with rack"); 611 SYSCTL_ADD_S32(&rack_sysctl_ctx, 612 SYSCTL_CHILDREN(rack_sysctl_root), 613 OID_AUTO, "reorder_thresh", CTLFLAG_RW, 614 &rack_reorder_thresh, 2, 615 "What factor for rack will be added when seeing reordering (shift right)"); 616 SYSCTL_ADD_S32(&rack_sysctl_ctx, 617 SYSCTL_CHILDREN(rack_sysctl_root), 618 OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW, 619 &rack_tlp_thresh, 1, 620 "what divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)"); 621 SYSCTL_ADD_S32(&rack_sysctl_ctx, 622 SYSCTL_CHILDREN(rack_sysctl_root), 623 OID_AUTO, "reorder_fade", CTLFLAG_RW, 624 &rack_reorder_fade, 0, 625 "Does reorder detection fade, if so how many ms (0 means never)"); 626 SYSCTL_ADD_S32(&rack_sysctl_ctx, 627 SYSCTL_CHILDREN(rack_sysctl_root), 628 OID_AUTO, "pktdelay", CTLFLAG_RW, 629 &rack_pkt_delay, 1, 630 "Extra RACK time (in ms) besides reordering thresh"); 631 SYSCTL_ADD_U32(&rack_sysctl_ctx, 632 SYSCTL_CHILDREN(rack_sysctl_root), 633 OID_AUTO, "split_limit", CTLFLAG_RW, 634 &rack_map_split_limit, 0, 635 "Is there a limit on the number of map split entries (0=unlimited)"); 636 SYSCTL_ADD_S32(&rack_sysctl_ctx, 637 SYSCTL_CHILDREN(rack_sysctl_root), 638 OID_AUTO, "inc_var", CTLFLAG_RW, 639 &rack_inc_var, 0, 640 "Should rack add to the TLP timer the variance in rtt calculation"); 641 rack_badfr = counter_u64_alloc(M_WAITOK); 642 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 643 SYSCTL_CHILDREN(rack_sysctl_root), 644 OID_AUTO, "badfr", CTLFLAG_RD, 645 &rack_badfr, "Total number of bad FRs"); 646 rack_badfr_bytes = counter_u64_alloc(M_WAITOK); 647 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 648 SYSCTL_CHILDREN(rack_sysctl_root), 649 OID_AUTO, "badfr_bytes", CTLFLAG_RD, 650 &rack_badfr_bytes, "Total number of bad FRs"); 651 rack_rtm_prr_retran = counter_u64_alloc(M_WAITOK); 652 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 653 SYSCTL_CHILDREN(rack_sysctl_root), 654 OID_AUTO, "prrsndret", CTLFLAG_RD, 655 &rack_rtm_prr_retran, 656 "Total number of prr based retransmits"); 657 rack_rtm_prr_newdata = counter_u64_alloc(M_WAITOK); 658 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 659 SYSCTL_CHILDREN(rack_sysctl_root), 660 OID_AUTO, "prrsndnew", CTLFLAG_RD, 661 &rack_rtm_prr_newdata, 662 "Total number of prr based new transmits"); 663 rack_timestamp_mismatch = counter_u64_alloc(M_WAITOK); 664 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 665 SYSCTL_CHILDREN(rack_sysctl_root), 666 OID_AUTO, "tsnf", CTLFLAG_RD, 667 &rack_timestamp_mismatch, 668 "Total number of timestamps that we could not find the reported ts"); 669 rack_find_high = counter_u64_alloc(M_WAITOK); 670 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 671 SYSCTL_CHILDREN(rack_sysctl_root), 672 OID_AUTO, "findhigh", CTLFLAG_RD, 673 &rack_find_high, 674 "Total number of FIN causing find-high"); 675 rack_reorder_seen = counter_u64_alloc(M_WAITOK); 676 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 677 SYSCTL_CHILDREN(rack_sysctl_root), 678 OID_AUTO, "reordering", CTLFLAG_RD, 679 &rack_reorder_seen, 680 "Total number of times we added delay due to reordering"); 681 rack_tlp_tot = counter_u64_alloc(M_WAITOK); 682 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 683 SYSCTL_CHILDREN(rack_sysctl_root), 684 OID_AUTO, "tlp_to_total", CTLFLAG_RD, 685 &rack_tlp_tot, 686 "Total number of tail loss probe expirations"); 687 rack_tlp_newdata = counter_u64_alloc(M_WAITOK); 688 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 689 SYSCTL_CHILDREN(rack_sysctl_root), 690 OID_AUTO, "tlp_new", CTLFLAG_RD, 691 &rack_tlp_newdata, 692 "Total number of tail loss probe sending new data"); 693 694 rack_tlp_retran = counter_u64_alloc(M_WAITOK); 695 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 696 SYSCTL_CHILDREN(rack_sysctl_root), 697 OID_AUTO, "tlp_retran", CTLFLAG_RD, 698 &rack_tlp_retran, 699 "Total number of tail loss probe sending retransmitted data"); 700 rack_tlp_retran_bytes = counter_u64_alloc(M_WAITOK); 701 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 702 SYSCTL_CHILDREN(rack_sysctl_root), 703 OID_AUTO, "tlp_retran_bytes", CTLFLAG_RD, 704 &rack_tlp_retran_bytes, 705 "Total bytes of tail loss probe sending retransmitted data"); 706 rack_tlp_retran_fail = counter_u64_alloc(M_WAITOK); 707 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 708 SYSCTL_CHILDREN(rack_sysctl_root), 709 OID_AUTO, "tlp_retran_fail", CTLFLAG_RD, 710 &rack_tlp_retran_fail, 711 "Total number of tail loss probe sending retransmitted data that failed (wait for t3)"); 712 rack_to_tot = counter_u64_alloc(M_WAITOK); 713 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 714 SYSCTL_CHILDREN(rack_sysctl_root), 715 OID_AUTO, "rack_to_tot", CTLFLAG_RD, 716 &rack_to_tot, 717 "Total number of times the rack to expired?"); 718 rack_to_arm_rack = counter_u64_alloc(M_WAITOK); 719 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 720 SYSCTL_CHILDREN(rack_sysctl_root), 721 OID_AUTO, "arm_rack", CTLFLAG_RD, 722 &rack_to_arm_rack, 723 "Total number of times the rack timer armed?"); 724 rack_to_arm_tlp = counter_u64_alloc(M_WAITOK); 725 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 726 SYSCTL_CHILDREN(rack_sysctl_root), 727 OID_AUTO, "arm_tlp", CTLFLAG_RD, 728 &rack_to_arm_tlp, 729 "Total number of times the tlp timer armed?"); 730 rack_paced_segments = counter_u64_alloc(M_WAITOK); 731 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 732 SYSCTL_CHILDREN(rack_sysctl_root), 733 OID_AUTO, "paced", CTLFLAG_RD, 734 &rack_paced_segments, 735 "Total number of times a segment send caused hptsi"); 736 rack_unpaced_segments = counter_u64_alloc(M_WAITOK); 737 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 738 SYSCTL_CHILDREN(rack_sysctl_root), 739 OID_AUTO, "unpaced", CTLFLAG_RD, 740 &rack_unpaced_segments, 741 "Total number of times a segment did not cause hptsi"); 742 rack_saw_enobuf = counter_u64_alloc(M_WAITOK); 743 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 744 SYSCTL_CHILDREN(rack_sysctl_root), 745 OID_AUTO, "saw_enobufs", CTLFLAG_RD, 746 &rack_saw_enobuf, 747 "Total number of times a segment did not cause hptsi"); 748 rack_saw_enetunreach = counter_u64_alloc(M_WAITOK); 749 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 750 SYSCTL_CHILDREN(rack_sysctl_root), 751 OID_AUTO, "saw_enetunreach", CTLFLAG_RD, 752 &rack_saw_enetunreach, 753 "Total number of times a segment did not cause hptsi"); 754 rack_to_alloc = counter_u64_alloc(M_WAITOK); 755 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 756 SYSCTL_CHILDREN(rack_sysctl_root), 757 OID_AUTO, "allocs", CTLFLAG_RD, 758 &rack_to_alloc, 759 "Total allocations of tracking structures"); 760 rack_to_alloc_hard = counter_u64_alloc(M_WAITOK); 761 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 762 SYSCTL_CHILDREN(rack_sysctl_root), 763 OID_AUTO, "allochard", CTLFLAG_RD, 764 &rack_to_alloc_hard, 765 "Total allocations done with sleeping the hard way"); 766 rack_to_alloc_emerg = counter_u64_alloc(M_WAITOK); 767 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 768 SYSCTL_CHILDREN(rack_sysctl_root), 769 OID_AUTO, "allocemerg", CTLFLAG_RD, 770 &rack_to_alloc_emerg, 771 "Total allocations done from emergency cache"); 772 rack_alloc_limited_conns = counter_u64_alloc(M_WAITOK); 773 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 774 SYSCTL_CHILDREN(rack_sysctl_root), 775 OID_AUTO, "alloc_limited_conns", CTLFLAG_RD, 776 &rack_alloc_limited_conns, 777 "Connections with allocations dropped due to limit"); 778 rack_split_limited = counter_u64_alloc(M_WAITOK); 779 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 780 SYSCTL_CHILDREN(rack_sysctl_root), 781 OID_AUTO, "split_limited", CTLFLAG_RD, 782 &rack_split_limited, 783 "Split allocations dropped due to limit"); 784 rack_sack_proc_all = counter_u64_alloc(M_WAITOK); 785 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 786 SYSCTL_CHILDREN(rack_sysctl_root), 787 OID_AUTO, "sack_long", CTLFLAG_RD, 788 &rack_sack_proc_all, 789 "Total times we had to walk whole list for sack processing"); 790 791 rack_sack_proc_restart = counter_u64_alloc(M_WAITOK); 792 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 793 SYSCTL_CHILDREN(rack_sysctl_root), 794 OID_AUTO, "sack_restart", CTLFLAG_RD, 795 &rack_sack_proc_restart, 796 "Total times we had to walk whole list due to a restart"); 797 rack_sack_proc_short = counter_u64_alloc(M_WAITOK); 798 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 799 SYSCTL_CHILDREN(rack_sysctl_root), 800 OID_AUTO, "sack_short", CTLFLAG_RD, 801 &rack_sack_proc_short, 802 "Total times we took shortcut for sack processing"); 803 rack_enter_tlp_calc = counter_u64_alloc(M_WAITOK); 804 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 805 SYSCTL_CHILDREN(rack_sysctl_root), 806 OID_AUTO, "tlp_calc_entered", CTLFLAG_RD, 807 &rack_enter_tlp_calc, 808 "Total times we called calc-tlp"); 809 rack_used_tlpmethod = counter_u64_alloc(M_WAITOK); 810 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 811 SYSCTL_CHILDREN(rack_sysctl_root), 812 OID_AUTO, "hit_tlp_method", CTLFLAG_RD, 813 &rack_used_tlpmethod, 814 "Total number of runt sacks"); 815 rack_used_tlpmethod2 = counter_u64_alloc(M_WAITOK); 816 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 817 SYSCTL_CHILDREN(rack_sysctl_root), 818 OID_AUTO, "hit_tlp_method2", CTLFLAG_RD, 819 &rack_used_tlpmethod2, 820 "Total number of runt sacks 2"); 821 rack_runt_sacks = counter_u64_alloc(M_WAITOK); 822 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 823 SYSCTL_CHILDREN(rack_sysctl_root), 824 OID_AUTO, "runtsacks", CTLFLAG_RD, 825 &rack_runt_sacks, 826 "Total number of runt sacks"); 827 rack_progress_drops = counter_u64_alloc(M_WAITOK); 828 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 829 SYSCTL_CHILDREN(rack_sysctl_root), 830 OID_AUTO, "prog_drops", CTLFLAG_RD, 831 &rack_progress_drops, 832 "Total number of progress drops"); 833 rack_input_idle_reduces = counter_u64_alloc(M_WAITOK); 834 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 835 SYSCTL_CHILDREN(rack_sysctl_root), 836 OID_AUTO, "idle_reduce_oninput", CTLFLAG_RD, 837 &rack_input_idle_reduces, 838 "Total number of idle reductions on input"); 839 rack_tlp_does_nada = counter_u64_alloc(M_WAITOK); 840 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 841 SYSCTL_CHILDREN(rack_sysctl_root), 842 OID_AUTO, "tlp_nada", CTLFLAG_RD, 843 &rack_tlp_does_nada, 844 "Total number of nada tlp calls"); 845 COUNTER_ARRAY_ALLOC(rack_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK); 846 SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), 847 OID_AUTO, "outsize", CTLFLAG_RD, 848 rack_out_size, TCP_MSS_ACCT_SIZE, "MSS send sizes"); 849 COUNTER_ARRAY_ALLOC(rack_opts_arry, RACK_OPTS_SIZE, M_WAITOK); 850 SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), 851 OID_AUTO, "opts", CTLFLAG_RD, 852 rack_opts_arry, RACK_OPTS_SIZE, "RACK Option Stats"); 853 SYSCTL_ADD_PROC(&rack_sysctl_ctx, 854 SYSCTL_CHILDREN(rack_sysctl_root), 855 OID_AUTO, "clear", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE, 856 &rack_clear_counter, 0, sysctl_rack_clear, "IU", "Clear counters"); 857 } 858 859 static inline int32_t 860 rack_progress_timeout_check(struct tcpcb *tp) 861 { 862 if (tp->t_maxunacktime && tp->t_acktime && TSTMP_GT(ticks, tp->t_acktime)) { 863 if ((ticks - tp->t_acktime) >= tp->t_maxunacktime) { 864 /* 865 * There is an assumption that the caller 866 * will drop the connection so we will 867 * increment the counters here. 868 */ 869 struct tcp_rack *rack; 870 rack = (struct tcp_rack *)tp->t_fb_ptr; 871 counter_u64_add(rack_progress_drops, 1); 872 #ifdef NETFLIX_STATS 873 TCPSTAT_INC(tcps_progdrops); 874 #endif 875 rack_log_progress_event(rack, tp, ticks, PROGRESS_DROP, __LINE__); 876 return (1); 877 } 878 } 879 return (0); 880 } 881 882 883 static void 884 rack_log_to_start(struct tcp_rack *rack, uint32_t cts, uint32_t to, int32_t slot, uint8_t which) 885 { 886 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 887 union tcp_log_stackspecific log; 888 889 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 890 log.u_bbr.flex1 = TICKS_2_MSEC(rack->rc_tp->t_srtt >> TCP_RTT_SHIFT); 891 log.u_bbr.flex2 = to; 892 log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags; 893 log.u_bbr.flex4 = slot; 894 log.u_bbr.flex5 = rack->rc_inp->inp_hptsslot; 895 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur; 896 log.u_bbr.flex8 = which; 897 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; 898 log.u_bbr.ininput = rack->rc_inp->inp_in_input; 899 TCP_LOG_EVENT(rack->rc_tp, NULL, 900 &rack->rc_inp->inp_socket->so_rcv, 901 &rack->rc_inp->inp_socket->so_snd, 902 BBR_LOG_TIMERSTAR, 0, 903 0, &log, false); 904 } 905 } 906 907 static void 908 rack_log_to_event(struct tcp_rack *rack, int32_t to_num) 909 { 910 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 911 union tcp_log_stackspecific log; 912 913 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 914 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; 915 log.u_bbr.ininput = rack->rc_inp->inp_in_input; 916 log.u_bbr.flex8 = to_num; 917 log.u_bbr.flex1 = rack->r_ctl.rc_rack_min_rtt; 918 log.u_bbr.flex2 = rack->rc_rack_rtt; 919 TCP_LOG_EVENT(rack->rc_tp, NULL, 920 &rack->rc_inp->inp_socket->so_rcv, 921 &rack->rc_inp->inp_socket->so_snd, 922 BBR_LOG_RTO, 0, 923 0, &log, false); 924 } 925 } 926 927 static void 928 rack_log_rtt_upd(struct tcpcb *tp, struct tcp_rack *rack, int32_t t, 929 uint32_t o_srtt, uint32_t o_var) 930 { 931 if (tp->t_logstate != TCP_LOG_STATE_OFF) { 932 union tcp_log_stackspecific log; 933 934 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 935 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; 936 log.u_bbr.ininput = rack->rc_inp->inp_in_input; 937 log.u_bbr.flex1 = t; 938 log.u_bbr.flex2 = o_srtt; 939 log.u_bbr.flex3 = o_var; 940 log.u_bbr.flex4 = rack->r_ctl.rack_rs.rs_rtt_lowest; 941 log.u_bbr.flex5 = rack->r_ctl.rack_rs.rs_rtt_highest; 942 log.u_bbr.flex6 = rack->r_ctl.rack_rs.rs_rtt_cnt; 943 log.u_bbr.rttProp = rack->r_ctl.rack_rs.rs_rtt_tot; 944 log.u_bbr.flex8 = rack->r_ctl.rc_rate_sample_method; 945 TCP_LOG_EVENT(tp, NULL, 946 &rack->rc_inp->inp_socket->so_rcv, 947 &rack->rc_inp->inp_socket->so_snd, 948 BBR_LOG_BBRRTT, 0, 949 0, &log, false); 950 } 951 } 952 953 static void 954 rack_log_rtt_sample(struct tcp_rack *rack, uint32_t rtt) 955 { 956 /* 957 * Log the rtt sample we are 958 * applying to the srtt algorithm in 959 * useconds. 960 */ 961 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 962 union tcp_log_stackspecific log; 963 struct timeval tv; 964 965 /* Convert our ms to a microsecond */ 966 log.u_bbr.flex1 = rtt * 1000; 967 log.u_bbr.timeStamp = tcp_get_usecs(&tv); 968 TCP_LOG_EVENTP(rack->rc_tp, NULL, 969 &rack->rc_inp->inp_socket->so_rcv, 970 &rack->rc_inp->inp_socket->so_snd, 971 TCP_LOG_RTT, 0, 972 0, &log, false, &tv); 973 } 974 } 975 976 977 static inline void 978 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick, int event, int line) 979 { 980 if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) { 981 union tcp_log_stackspecific log; 982 983 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 984 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; 985 log.u_bbr.ininput = rack->rc_inp->inp_in_input; 986 log.u_bbr.flex1 = line; 987 log.u_bbr.flex2 = tick; 988 log.u_bbr.flex3 = tp->t_maxunacktime; 989 log.u_bbr.flex4 = tp->t_acktime; 990 log.u_bbr.flex8 = event; 991 TCP_LOG_EVENT(tp, NULL, 992 &rack->rc_inp->inp_socket->so_rcv, 993 &rack->rc_inp->inp_socket->so_snd, 994 BBR_LOG_PROGRESS, 0, 995 0, &log, false); 996 } 997 } 998 999 static void 1000 rack_log_type_bbrsnd(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint32_t cts) 1001 { 1002 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 1003 union tcp_log_stackspecific log; 1004 1005 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 1006 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; 1007 log.u_bbr.ininput = rack->rc_inp->inp_in_input; 1008 log.u_bbr.flex1 = slot; 1009 log.u_bbr.flex7 = (0x0000ffff & rack->r_ctl.rc_hpts_flags); 1010 log.u_bbr.flex8 = rack->rc_in_persist; 1011 TCP_LOG_EVENT(rack->rc_tp, NULL, 1012 &rack->rc_inp->inp_socket->so_rcv, 1013 &rack->rc_inp->inp_socket->so_snd, 1014 BBR_LOG_BBRSND, 0, 1015 0, &log, false); 1016 } 1017 } 1018 1019 static void 1020 rack_log_doseg_done(struct tcp_rack *rack, uint32_t cts, int32_t nxt_pkt, int32_t did_out, int way_out) 1021 { 1022 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 1023 union tcp_log_stackspecific log; 1024 log.u_bbr.flex1 = did_out; 1025 log.u_bbr.flex2 = nxt_pkt; 1026 log.u_bbr.flex3 = way_out; 1027 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags; 1028 log.u_bbr.flex7 = rack->r_wanted_output; 1029 log.u_bbr.flex8 = rack->rc_in_persist; 1030 TCP_LOG_EVENT(rack->rc_tp, NULL, 1031 &rack->rc_inp->inp_socket->so_rcv, 1032 &rack->rc_inp->inp_socket->so_snd, 1033 BBR_LOG_DOSEG_DONE, 0, 1034 0, &log, false); 1035 } 1036 } 1037 1038 1039 static void 1040 rack_log_type_just_return(struct tcp_rack *rack, uint32_t cts, uint32_t tlen, uint32_t slot, uint8_t hpts_calling) 1041 { 1042 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 1043 union tcp_log_stackspecific log; 1044 1045 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 1046 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; 1047 log.u_bbr.ininput = rack->rc_inp->inp_in_input; 1048 log.u_bbr.flex1 = slot; 1049 log.u_bbr.flex2 = rack->r_ctl.rc_hpts_flags; 1050 log.u_bbr.flex7 = hpts_calling; 1051 log.u_bbr.flex8 = rack->rc_in_persist; 1052 TCP_LOG_EVENT(rack->rc_tp, NULL, 1053 &rack->rc_inp->inp_socket->so_rcv, 1054 &rack->rc_inp->inp_socket->so_snd, 1055 BBR_LOG_JUSTRET, 0, 1056 tlen, &log, false); 1057 } 1058 } 1059 1060 static void 1061 rack_log_to_cancel(struct tcp_rack *rack, int32_t hpts_removed, int line) 1062 { 1063 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 1064 union tcp_log_stackspecific log; 1065 1066 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 1067 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; 1068 log.u_bbr.ininput = rack->rc_inp->inp_in_input; 1069 log.u_bbr.flex1 = line; 1070 log.u_bbr.flex2 = 0; 1071 log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags; 1072 log.u_bbr.flex4 = 0; 1073 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur; 1074 log.u_bbr.flex8 = hpts_removed; 1075 TCP_LOG_EVENT(rack->rc_tp, NULL, 1076 &rack->rc_inp->inp_socket->so_rcv, 1077 &rack->rc_inp->inp_socket->so_snd, 1078 BBR_LOG_TIMERCANC, 0, 1079 0, &log, false); 1080 } 1081 } 1082 1083 static void 1084 rack_log_to_processing(struct tcp_rack *rack, uint32_t cts, int32_t ret, int32_t timers) 1085 { 1086 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 1087 union tcp_log_stackspecific log; 1088 1089 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 1090 log.u_bbr.flex1 = timers; 1091 log.u_bbr.flex2 = ret; 1092 log.u_bbr.flex3 = rack->r_ctl.rc_timer_exp; 1093 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags; 1094 log.u_bbr.flex5 = cts; 1095 TCP_LOG_EVENT(rack->rc_tp, NULL, 1096 &rack->rc_inp->inp_socket->so_rcv, 1097 &rack->rc_inp->inp_socket->so_snd, 1098 BBR_LOG_TO_PROCESS, 0, 1099 0, &log, false); 1100 } 1101 } 1102 1103 static void 1104 rack_counter_destroy() 1105 { 1106 counter_u64_free(rack_badfr); 1107 counter_u64_free(rack_badfr_bytes); 1108 counter_u64_free(rack_rtm_prr_retran); 1109 counter_u64_free(rack_rtm_prr_newdata); 1110 counter_u64_free(rack_timestamp_mismatch); 1111 counter_u64_free(rack_reorder_seen); 1112 counter_u64_free(rack_tlp_tot); 1113 counter_u64_free(rack_tlp_newdata); 1114 counter_u64_free(rack_tlp_retran); 1115 counter_u64_free(rack_tlp_retran_bytes); 1116 counter_u64_free(rack_tlp_retran_fail); 1117 counter_u64_free(rack_to_tot); 1118 counter_u64_free(rack_to_arm_rack); 1119 counter_u64_free(rack_to_arm_tlp); 1120 counter_u64_free(rack_paced_segments); 1121 counter_u64_free(rack_unpaced_segments); 1122 counter_u64_free(rack_saw_enobuf); 1123 counter_u64_free(rack_saw_enetunreach); 1124 counter_u64_free(rack_to_alloc_hard); 1125 counter_u64_free(rack_to_alloc_emerg); 1126 counter_u64_free(rack_sack_proc_all); 1127 counter_u64_free(rack_sack_proc_short); 1128 counter_u64_free(rack_sack_proc_restart); 1129 counter_u64_free(rack_to_alloc); 1130 counter_u64_free(rack_find_high); 1131 counter_u64_free(rack_runt_sacks); 1132 counter_u64_free(rack_enter_tlp_calc); 1133 counter_u64_free(rack_used_tlpmethod); 1134 counter_u64_free(rack_used_tlpmethod2); 1135 counter_u64_free(rack_progress_drops); 1136 counter_u64_free(rack_input_idle_reduces); 1137 counter_u64_free(rack_tlp_does_nada); 1138 COUNTER_ARRAY_FREE(rack_out_size, TCP_MSS_ACCT_SIZE); 1139 COUNTER_ARRAY_FREE(rack_opts_arry, RACK_OPTS_SIZE); 1140 } 1141 1142 static struct rack_sendmap * 1143 rack_alloc(struct tcp_rack *rack) 1144 { 1145 struct rack_sendmap *rsm; 1146 1147 rsm = uma_zalloc(rack_zone, M_NOWAIT); 1148 if (rsm) { 1149 alloc_done: 1150 counter_u64_add(rack_to_alloc, 1); 1151 rack->r_ctl.rc_num_maps_alloced++; 1152 return (rsm); 1153 } 1154 if (rack->rc_free_cnt) { 1155 counter_u64_add(rack_to_alloc_emerg, 1); 1156 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free); 1157 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_next); 1158 rack->rc_free_cnt--; 1159 goto alloc_done; 1160 } 1161 return (NULL); 1162 } 1163 1164 /* wrapper to allocate a sendmap entry, subject to a specific limit */ 1165 static struct rack_sendmap * 1166 rack_alloc_limit(struct tcp_rack *rack, uint8_t limit_type) 1167 { 1168 struct rack_sendmap *rsm; 1169 1170 if (limit_type) { 1171 /* currently there is only one limit type */ 1172 if (rack_map_split_limit > 0 && 1173 rack->r_ctl.rc_num_split_allocs >= rack_map_split_limit) { 1174 counter_u64_add(rack_split_limited, 1); 1175 if (!rack->alloc_limit_reported) { 1176 rack->alloc_limit_reported = 1; 1177 counter_u64_add(rack_alloc_limited_conns, 1); 1178 } 1179 return (NULL); 1180 } 1181 } 1182 1183 /* allocate and mark in the limit type, if set */ 1184 rsm = rack_alloc(rack); 1185 if (rsm != NULL && limit_type) { 1186 rsm->r_limit_type = limit_type; 1187 rack->r_ctl.rc_num_split_allocs++; 1188 } 1189 return (rsm); 1190 } 1191 1192 static void 1193 rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm) 1194 { 1195 if (rsm->r_limit_type) { 1196 /* currently there is only one limit type */ 1197 rack->r_ctl.rc_num_split_allocs--; 1198 } 1199 rack->r_ctl.rc_num_maps_alloced--; 1200 if (rack->r_ctl.rc_tlpsend == rsm) 1201 rack->r_ctl.rc_tlpsend = NULL; 1202 if (rack->r_ctl.rc_next == rsm) 1203 rack->r_ctl.rc_next = NULL; 1204 if (rack->r_ctl.rc_sacklast == rsm) 1205 rack->r_ctl.rc_sacklast = NULL; 1206 if (rack->rc_free_cnt < rack_free_cache) { 1207 memset(rsm, 0, sizeof(struct rack_sendmap)); 1208 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_next); 1209 rack->rc_free_cnt++; 1210 return; 1211 } 1212 uma_zfree(rack_zone, rsm); 1213 } 1214 1215 /* 1216 * CC wrapper hook functions 1217 */ 1218 static void 1219 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, struct tcphdr *th, uint16_t nsegs, 1220 uint16_t type, int32_t recovery) 1221 { 1222 #ifdef NETFLIX_STATS 1223 int32_t gput; 1224 #endif 1225 #ifdef NETFLIX_CWV 1226 u_long old_cwnd = tp->snd_cwnd; 1227 #endif 1228 1229 INP_WLOCK_ASSERT(tp->t_inpcb); 1230 tp->ccv->nsegs = nsegs; 1231 tp->ccv->bytes_this_ack = BYTES_THIS_ACK(tp, th); 1232 if ((recovery) && (rack->r_ctl.rc_early_recovery_segs)) { 1233 uint32_t max; 1234 1235 max = rack->r_ctl.rc_early_recovery_segs * tp->t_maxseg; 1236 if (tp->ccv->bytes_this_ack > max) { 1237 tp->ccv->bytes_this_ack = max; 1238 } 1239 } 1240 if (tp->snd_cwnd <= tp->snd_wnd) 1241 tp->ccv->flags |= CCF_CWND_LIMITED; 1242 else 1243 tp->ccv->flags &= ~CCF_CWND_LIMITED; 1244 1245 if (type == CC_ACK) { 1246 #ifdef NETFLIX_STATS 1247 stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF, 1248 ((int32_t) tp->snd_cwnd) - tp->snd_wnd); 1249 if ((tp->t_flags & TF_GPUTINPROG) && 1250 SEQ_GEQ(th->th_ack, tp->gput_ack)) { 1251 gput = (((int64_t) (th->th_ack - tp->gput_seq)) << 3) / 1252 max(1, tcp_ts_getticks() - tp->gput_ts); 1253 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT, 1254 gput); 1255 /* 1256 * XXXLAS: This is a temporary hack, and should be 1257 * chained off VOI_TCP_GPUT when stats(9) grows an 1258 * API to deal with chained VOIs. 1259 */ 1260 if (tp->t_stats_gput_prev > 0) 1261 stats_voi_update_abs_s32(tp->t_stats, 1262 VOI_TCP_GPUT_ND, 1263 ((gput - tp->t_stats_gput_prev) * 100) / 1264 tp->t_stats_gput_prev); 1265 tp->t_flags &= ~TF_GPUTINPROG; 1266 tp->t_stats_gput_prev = gput; 1267 #ifdef NETFLIX_CWV 1268 if (tp->t_maxpeakrate) { 1269 /* 1270 * We update t_peakrate_thr. This gives us roughly 1271 * one update per round trip time. 1272 */ 1273 tcp_update_peakrate_thr(tp); 1274 } 1275 #endif 1276 } 1277 #endif 1278 if (tp->snd_cwnd > tp->snd_ssthresh) { 1279 tp->t_bytes_acked += min(tp->ccv->bytes_this_ack, 1280 nsegs * V_tcp_abc_l_var * tp->t_maxseg); 1281 if (tp->t_bytes_acked >= tp->snd_cwnd) { 1282 tp->t_bytes_acked -= tp->snd_cwnd; 1283 tp->ccv->flags |= CCF_ABC_SENTAWND; 1284 } 1285 } else { 1286 tp->ccv->flags &= ~CCF_ABC_SENTAWND; 1287 tp->t_bytes_acked = 0; 1288 } 1289 } 1290 if (CC_ALGO(tp)->ack_received != NULL) { 1291 /* XXXLAS: Find a way to live without this */ 1292 tp->ccv->curack = th->th_ack; 1293 CC_ALGO(tp)->ack_received(tp->ccv, type); 1294 } 1295 #ifdef NETFLIX_STATS 1296 stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, tp->snd_cwnd); 1297 #endif 1298 if (rack->r_ctl.rc_rack_largest_cwnd < tp->snd_cwnd) { 1299 rack->r_ctl.rc_rack_largest_cwnd = tp->snd_cwnd; 1300 } 1301 #ifdef NETFLIX_CWV 1302 if (tp->cwv_enabled) { 1303 /* 1304 * Per RFC 7661: The behaviour in the non-validated phase is 1305 * specified as: o A sender determines whether to increase 1306 * the cwnd based upon whether it is cwnd-limited (see 1307 * Section 4.5.3): * A sender that is cwnd-limited MAY use 1308 * the standard TCP method to increase cwnd (i.e., the 1309 * standard method permits a TCP sender that fully utilises 1310 * the cwnd to increase the cwnd each time it receives an 1311 * ACK). * A sender that is not cwnd-limited MUST NOT 1312 * increase the cwnd when ACK packets are received in this 1313 * phase (i.e., needs to avoid growing the cwnd when it has 1314 * not recently sent using the current size of cwnd). 1315 */ 1316 if ((tp->snd_cwnd > old_cwnd) && 1317 (tp->cwv_cwnd_valid == 0) && 1318 (!(tp->ccv->flags & CCF_CWND_LIMITED))) { 1319 tp->snd_cwnd = old_cwnd; 1320 } 1321 /* Try to update pipeAck and NCWV state */ 1322 if (TCPS_HAVEESTABLISHED(tp->t_state) && 1323 !IN_RECOVERY(tp->t_flags)) { 1324 uint32_t data = sbavail(&(tp->t_inpcb->inp_socket->so_snd)); 1325 1326 tcp_newcwv_update_pipeack(tp, data); 1327 } 1328 } 1329 /* we enforce max peak rate if it is set. */ 1330 if (tp->t_peakrate_thr && tp->snd_cwnd > tp->t_peakrate_thr) { 1331 tp->snd_cwnd = tp->t_peakrate_thr; 1332 } 1333 #endif 1334 } 1335 1336 static void 1337 tcp_rack_partialack(struct tcpcb *tp, struct tcphdr *th) 1338 { 1339 struct tcp_rack *rack; 1340 1341 rack = (struct tcp_rack *)tp->t_fb_ptr; 1342 INP_WLOCK_ASSERT(tp->t_inpcb); 1343 if (rack->r_ctl.rc_prr_sndcnt > 0) 1344 rack->r_wanted_output++; 1345 } 1346 1347 static void 1348 rack_post_recovery(struct tcpcb *tp, struct tcphdr *th) 1349 { 1350 struct tcp_rack *rack; 1351 1352 INP_WLOCK_ASSERT(tp->t_inpcb); 1353 rack = (struct tcp_rack *)tp->t_fb_ptr; 1354 if (CC_ALGO(tp)->post_recovery != NULL) { 1355 tp->ccv->curack = th->th_ack; 1356 CC_ALGO(tp)->post_recovery(tp->ccv); 1357 } 1358 /* 1359 * Here we can in theory adjust cwnd to be based on the number of 1360 * losses in the window (rack->r_ctl.rc_loss_count). This is done 1361 * based on the rack_use_proportional flag. 1362 */ 1363 if (rack->r_ctl.rc_prop_reduce && rack->r_ctl.rc_prop_rate) { 1364 int32_t reduce; 1365 1366 reduce = (rack->r_ctl.rc_loss_count * rack->r_ctl.rc_prop_rate); 1367 if (reduce > 50) { 1368 reduce = 50; 1369 } 1370 tp->snd_cwnd -= ((reduce * tp->snd_cwnd) / 100); 1371 } else { 1372 if (tp->snd_cwnd > tp->snd_ssthresh) { 1373 /* Drop us down to the ssthresh (1/2 cwnd at loss) */ 1374 tp->snd_cwnd = tp->snd_ssthresh; 1375 } 1376 } 1377 if (rack->r_ctl.rc_prr_sndcnt > 0) { 1378 /* Suck the next prr cnt back into cwnd */ 1379 tp->snd_cwnd += rack->r_ctl.rc_prr_sndcnt; 1380 rack->r_ctl.rc_prr_sndcnt = 0; 1381 } 1382 EXIT_RECOVERY(tp->t_flags); 1383 1384 1385 #ifdef NETFLIX_CWV 1386 if (tp->cwv_enabled) { 1387 if ((tp->cwv_cwnd_valid == 0) && 1388 (tp->snd_cwv.in_recovery)) 1389 tcp_newcwv_end_recovery(tp); 1390 } 1391 #endif 1392 } 1393 1394 static void 1395 rack_cong_signal(struct tcpcb *tp, struct tcphdr *th, uint32_t type) 1396 { 1397 struct tcp_rack *rack; 1398 1399 INP_WLOCK_ASSERT(tp->t_inpcb); 1400 1401 rack = (struct tcp_rack *)tp->t_fb_ptr; 1402 switch (type) { 1403 case CC_NDUPACK: 1404 /* rack->r_ctl.rc_ssthresh_set = 1;*/ 1405 if (!IN_FASTRECOVERY(tp->t_flags)) { 1406 rack->r_ctl.rc_tlp_rtx_out = 0; 1407 rack->r_ctl.rc_prr_delivered = 0; 1408 rack->r_ctl.rc_prr_out = 0; 1409 rack->r_ctl.rc_loss_count = 0; 1410 rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; 1411 rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una; 1412 tp->snd_recover = tp->snd_max; 1413 if (tp->t_flags & TF_ECN_PERMIT) 1414 tp->t_flags |= TF_ECN_SND_CWR; 1415 } 1416 break; 1417 case CC_ECN: 1418 if (!IN_CONGRECOVERY(tp->t_flags)) { 1419 TCPSTAT_INC(tcps_ecn_rcwnd); 1420 tp->snd_recover = tp->snd_max; 1421 if (tp->t_flags & TF_ECN_PERMIT) 1422 tp->t_flags |= TF_ECN_SND_CWR; 1423 } 1424 break; 1425 case CC_RTO: 1426 tp->t_dupacks = 0; 1427 tp->t_bytes_acked = 0; 1428 EXIT_RECOVERY(tp->t_flags); 1429 tp->snd_ssthresh = max(2, min(tp->snd_wnd, tp->snd_cwnd) / 2 / 1430 tp->t_maxseg) * tp->t_maxseg; 1431 tp->snd_cwnd = tp->t_maxseg; 1432 break; 1433 case CC_RTO_ERR: 1434 TCPSTAT_INC(tcps_sndrexmitbad); 1435 /* RTO was unnecessary, so reset everything. */ 1436 tp->snd_cwnd = tp->snd_cwnd_prev; 1437 tp->snd_ssthresh = tp->snd_ssthresh_prev; 1438 tp->snd_recover = tp->snd_recover_prev; 1439 if (tp->t_flags & TF_WASFRECOVERY) 1440 ENTER_FASTRECOVERY(tp->t_flags); 1441 if (tp->t_flags & TF_WASCRECOVERY) 1442 ENTER_CONGRECOVERY(tp->t_flags); 1443 tp->snd_nxt = tp->snd_max; 1444 tp->t_badrxtwin = 0; 1445 break; 1446 } 1447 1448 if (CC_ALGO(tp)->cong_signal != NULL) { 1449 if (th != NULL) 1450 tp->ccv->curack = th->th_ack; 1451 CC_ALGO(tp)->cong_signal(tp->ccv, type); 1452 } 1453 #ifdef NETFLIX_CWV 1454 if (tp->cwv_enabled) { 1455 if (tp->snd_cwv.in_recovery == 0 && IN_RECOVERY(tp->t_flags)) { 1456 tcp_newcwv_enter_recovery(tp); 1457 } 1458 if (type == CC_RTO) { 1459 tcp_newcwv_reset(tp); 1460 } 1461 } 1462 #endif 1463 } 1464 1465 1466 1467 static inline void 1468 rack_cc_after_idle(struct tcpcb *tp, int reduce_largest) 1469 { 1470 uint32_t i_cwnd; 1471 1472 INP_WLOCK_ASSERT(tp->t_inpcb); 1473 1474 #ifdef NETFLIX_STATS 1475 TCPSTAT_INC(tcps_idle_restarts); 1476 if (tp->t_state == TCPS_ESTABLISHED) 1477 TCPSTAT_INC(tcps_idle_estrestarts); 1478 #endif 1479 if (CC_ALGO(tp)->after_idle != NULL) 1480 CC_ALGO(tp)->after_idle(tp->ccv); 1481 1482 if (tp->snd_cwnd == 1) 1483 i_cwnd = tp->t_maxseg; /* SYN(-ACK) lost */ 1484 else 1485 i_cwnd = tcp_compute_initwnd(tcp_maxseg(tp)); 1486 1487 if (reduce_largest) { 1488 /* 1489 * Do we reduce the largest cwnd to make 1490 * rack play nice on restart hptsi wise? 1491 */ 1492 if (((struct tcp_rack *)tp->t_fb_ptr)->r_ctl.rc_rack_largest_cwnd > i_cwnd) 1493 ((struct tcp_rack *)tp->t_fb_ptr)->r_ctl.rc_rack_largest_cwnd = i_cwnd; 1494 } 1495 /* 1496 * Being idle is no differnt than the initial window. If the cc 1497 * clamps it down below the initial window raise it to the initial 1498 * window. 1499 */ 1500 if (tp->snd_cwnd < i_cwnd) { 1501 tp->snd_cwnd = i_cwnd; 1502 } 1503 } 1504 1505 1506 /* 1507 * Indicate whether this ack should be delayed. We can delay the ack if 1508 * following conditions are met: 1509 * - There is no delayed ack timer in progress. 1510 * - Our last ack wasn't a 0-sized window. We never want to delay 1511 * the ack that opens up a 0-sized window. 1512 * - LRO wasn't used for this segment. We make sure by checking that the 1513 * segment size is not larger than the MSS. 1514 * - Delayed acks are enabled or this is a half-synchronized T/TCP 1515 * connection. 1516 */ 1517 #define DELAY_ACK(tp, tlen) \ 1518 (((tp->t_flags & TF_RXWIN0SENT) == 0) && \ 1519 ((tp->t_flags & TF_DELACK) == 0) && \ 1520 (tlen <= tp->t_maxseg) && \ 1521 (tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN))) 1522 1523 static inline void 1524 rack_calc_rwin(struct socket *so, struct tcpcb *tp) 1525 { 1526 int32_t win; 1527 1528 /* 1529 * Calculate amount of space in receive window, and then do TCP 1530 * input processing. Receive window is amount of space in rcv queue, 1531 * but not less than advertised window. 1532 */ 1533 win = sbspace(&so->so_rcv); 1534 if (win < 0) 1535 win = 0; 1536 tp->rcv_wnd = imax(win, (int)(tp->rcv_adv - tp->rcv_nxt)); 1537 } 1538 1539 static void 1540 rack_do_drop(struct mbuf *m, struct tcpcb *tp) 1541 { 1542 /* 1543 * Drop space held by incoming segment and return. 1544 */ 1545 if (tp != NULL) 1546 INP_WUNLOCK(tp->t_inpcb); 1547 if (m) 1548 m_freem(m); 1549 } 1550 1551 static void 1552 rack_do_dropwithreset(struct mbuf *m, struct tcpcb *tp, struct tcphdr *th, 1553 int32_t rstreason, int32_t tlen) 1554 { 1555 if (tp != NULL) { 1556 tcp_dropwithreset(m, th, tp, tlen, rstreason); 1557 INP_WUNLOCK(tp->t_inpcb); 1558 } else 1559 tcp_dropwithreset(m, th, NULL, tlen, rstreason); 1560 } 1561 1562 /* 1563 * The value in ret_val informs the caller 1564 * if we dropped the tcb (and lock) or not. 1565 * 1 = we dropped it, 0 = the TCB is still locked 1566 * and valid. 1567 */ 1568 static void 1569 rack_do_dropafterack(struct mbuf *m, struct tcpcb *tp, struct tcphdr *th, int32_t thflags, int32_t tlen, int32_t * ret_val) 1570 { 1571 /* 1572 * Generate an ACK dropping incoming segment if it occupies sequence 1573 * space, where the ACK reflects our state. 1574 * 1575 * We can now skip the test for the RST flag since all paths to this 1576 * code happen after packets containing RST have been dropped. 1577 * 1578 * In the SYN-RECEIVED state, don't send an ACK unless the segment 1579 * we received passes the SYN-RECEIVED ACK test. If it fails send a 1580 * RST. This breaks the loop in the "LAND" DoS attack, and also 1581 * prevents an ACK storm between two listening ports that have been 1582 * sent forged SYN segments, each with the source address of the 1583 * other. 1584 */ 1585 struct tcp_rack *rack; 1586 1587 if (tp->t_state == TCPS_SYN_RECEIVED && (thflags & TH_ACK) && 1588 (SEQ_GT(tp->snd_una, th->th_ack) || 1589 SEQ_GT(th->th_ack, tp->snd_max))) { 1590 *ret_val = 1; 1591 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 1592 return; 1593 } else 1594 *ret_val = 0; 1595 rack = (struct tcp_rack *)tp->t_fb_ptr; 1596 rack->r_wanted_output++; 1597 tp->t_flags |= TF_ACKNOW; 1598 if (m) 1599 m_freem(m); 1600 } 1601 1602 1603 static int 1604 rack_process_rst(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp) 1605 { 1606 /* 1607 * RFC5961 Section 3.2 1608 * 1609 * - RST drops connection only if SEG.SEQ == RCV.NXT. - If RST is in 1610 * window, we send challenge ACK. 1611 * 1612 * Note: to take into account delayed ACKs, we should test against 1613 * last_ack_sent instead of rcv_nxt. Note 2: we handle special case 1614 * of closed window, not covered by the RFC. 1615 */ 1616 int dropped = 0; 1617 1618 if ((SEQ_GEQ(th->th_seq, (tp->last_ack_sent - 1)) && 1619 SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) || 1620 (tp->rcv_wnd == 0 && tp->last_ack_sent == th->th_seq)) { 1621 1622 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 1623 KASSERT(tp->t_state != TCPS_SYN_SENT, 1624 ("%s: TH_RST for TCPS_SYN_SENT th %p tp %p", 1625 __func__, th, tp)); 1626 1627 if (V_tcp_insecure_rst || 1628 (tp->last_ack_sent == th->th_seq) || 1629 (tp->rcv_nxt == th->th_seq) || 1630 ((tp->last_ack_sent - 1) == th->th_seq)) { 1631 TCPSTAT_INC(tcps_drops); 1632 /* Drop the connection. */ 1633 switch (tp->t_state) { 1634 case TCPS_SYN_RECEIVED: 1635 so->so_error = ECONNREFUSED; 1636 goto close; 1637 case TCPS_ESTABLISHED: 1638 case TCPS_FIN_WAIT_1: 1639 case TCPS_FIN_WAIT_2: 1640 case TCPS_CLOSE_WAIT: 1641 case TCPS_CLOSING: 1642 case TCPS_LAST_ACK: 1643 so->so_error = ECONNRESET; 1644 close: 1645 tcp_state_change(tp, TCPS_CLOSED); 1646 /* FALLTHROUGH */ 1647 default: 1648 tp = tcp_close(tp); 1649 } 1650 dropped = 1; 1651 rack_do_drop(m, tp); 1652 } else { 1653 TCPSTAT_INC(tcps_badrst); 1654 /* Send challenge ACK. */ 1655 tcp_respond(tp, mtod(m, void *), th, m, 1656 tp->rcv_nxt, tp->snd_nxt, TH_ACK); 1657 tp->last_ack_sent = tp->rcv_nxt; 1658 } 1659 } else { 1660 m_freem(m); 1661 } 1662 return (dropped); 1663 } 1664 1665 /* 1666 * The value in ret_val informs the caller 1667 * if we dropped the tcb (and lock) or not. 1668 * 1 = we dropped it, 0 = the TCB is still locked 1669 * and valid. 1670 */ 1671 static void 1672 rack_challenge_ack(struct mbuf *m, struct tcphdr *th, struct tcpcb *tp, int32_t * ret_val) 1673 { 1674 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 1675 1676 TCPSTAT_INC(tcps_badsyn); 1677 if (V_tcp_insecure_syn && 1678 SEQ_GEQ(th->th_seq, tp->last_ack_sent) && 1679 SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) { 1680 tp = tcp_drop(tp, ECONNRESET); 1681 *ret_val = 1; 1682 rack_do_drop(m, tp); 1683 } else { 1684 /* Send challenge ACK. */ 1685 tcp_respond(tp, mtod(m, void *), th, m, tp->rcv_nxt, 1686 tp->snd_nxt, TH_ACK); 1687 tp->last_ack_sent = tp->rcv_nxt; 1688 m = NULL; 1689 *ret_val = 0; 1690 rack_do_drop(m, NULL); 1691 } 1692 } 1693 1694 /* 1695 * rack_ts_check returns 1 for you should not proceed. It places 1696 * in ret_val what should be returned 1/0 by the caller. The 1 indicates 1697 * that the TCB is unlocked and probably dropped. The 0 indicates the 1698 * TCB is still valid and locked. 1699 */ 1700 static int 1701 rack_ts_check(struct mbuf *m, struct tcphdr *th, struct tcpcb *tp, int32_t tlen, int32_t thflags, int32_t * ret_val) 1702 { 1703 1704 /* Check to see if ts_recent is over 24 days old. */ 1705 if (tcp_ts_getticks() - tp->ts_recent_age > TCP_PAWS_IDLE) { 1706 /* 1707 * Invalidate ts_recent. If this segment updates ts_recent, 1708 * the age will be reset later and ts_recent will get a 1709 * valid value. If it does not, setting ts_recent to zero 1710 * will at least satisfy the requirement that zero be placed 1711 * in the timestamp echo reply when ts_recent isn't valid. 1712 * The age isn't reset until we get a valid ts_recent 1713 * because we don't want out-of-order segments to be dropped 1714 * when ts_recent is old. 1715 */ 1716 tp->ts_recent = 0; 1717 } else { 1718 TCPSTAT_INC(tcps_rcvduppack); 1719 TCPSTAT_ADD(tcps_rcvdupbyte, tlen); 1720 TCPSTAT_INC(tcps_pawsdrop); 1721 *ret_val = 0; 1722 if (tlen) { 1723 rack_do_dropafterack(m, tp, th, thflags, tlen, ret_val); 1724 } else { 1725 rack_do_drop(m, NULL); 1726 } 1727 return (1); 1728 } 1729 return (0); 1730 } 1731 1732 /* 1733 * rack_drop_checks returns 1 for you should not proceed. It places 1734 * in ret_val what should be returned 1/0 by the caller. The 1 indicates 1735 * that the TCB is unlocked and probably dropped. The 0 indicates the 1736 * TCB is still valid and locked. 1737 */ 1738 static int 1739 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) 1740 { 1741 int32_t todrop; 1742 int32_t thflags; 1743 int32_t tlen; 1744 1745 thflags = *thf; 1746 tlen = *tlenp; 1747 todrop = tp->rcv_nxt - th->th_seq; 1748 if (todrop > 0) { 1749 if (thflags & TH_SYN) { 1750 thflags &= ~TH_SYN; 1751 th->th_seq++; 1752 if (th->th_urp > 1) 1753 th->th_urp--; 1754 else 1755 thflags &= ~TH_URG; 1756 todrop--; 1757 } 1758 /* 1759 * Following if statement from Stevens, vol. 2, p. 960. 1760 */ 1761 if (todrop > tlen 1762 || (todrop == tlen && (thflags & TH_FIN) == 0)) { 1763 /* 1764 * Any valid FIN must be to the left of the window. 1765 * At this point the FIN must be a duplicate or out 1766 * of sequence; drop it. 1767 */ 1768 thflags &= ~TH_FIN; 1769 /* 1770 * Send an ACK to resynchronize and drop any data. 1771 * But keep on processing for RST or ACK. 1772 */ 1773 tp->t_flags |= TF_ACKNOW; 1774 todrop = tlen; 1775 TCPSTAT_INC(tcps_rcvduppack); 1776 TCPSTAT_ADD(tcps_rcvdupbyte, todrop); 1777 } else { 1778 TCPSTAT_INC(tcps_rcvpartduppack); 1779 TCPSTAT_ADD(tcps_rcvpartdupbyte, todrop); 1780 } 1781 /* 1782 * DSACK - add SACK block for dropped range 1783 */ 1784 if (tp->t_flags & TF_SACK_PERMIT) { 1785 tcp_update_sack_list(tp, th->th_seq, th->th_seq + tlen); 1786 /* 1787 * ACK now, as the next in-sequence segment 1788 * will clear the DSACK block again 1789 */ 1790 tp->t_flags |= TF_ACKNOW; 1791 } 1792 *drop_hdrlen += todrop; /* drop from the top afterwards */ 1793 th->th_seq += todrop; 1794 tlen -= todrop; 1795 if (th->th_urp > todrop) 1796 th->th_urp -= todrop; 1797 else { 1798 thflags &= ~TH_URG; 1799 th->th_urp = 0; 1800 } 1801 } 1802 /* 1803 * If segment ends after window, drop trailing data (and PUSH and 1804 * FIN); if nothing left, just ACK. 1805 */ 1806 todrop = (th->th_seq + tlen) - (tp->rcv_nxt + tp->rcv_wnd); 1807 if (todrop > 0) { 1808 TCPSTAT_INC(tcps_rcvpackafterwin); 1809 if (todrop >= tlen) { 1810 TCPSTAT_ADD(tcps_rcvbyteafterwin, tlen); 1811 /* 1812 * If window is closed can only take segments at 1813 * window edge, and have to drop data and PUSH from 1814 * incoming segments. Continue processing, but 1815 * remember to ack. Otherwise, drop segment and 1816 * ack. 1817 */ 1818 if (tp->rcv_wnd == 0 && th->th_seq == tp->rcv_nxt) { 1819 tp->t_flags |= TF_ACKNOW; 1820 TCPSTAT_INC(tcps_rcvwinprobe); 1821 } else { 1822 rack_do_dropafterack(m, tp, th, thflags, tlen, ret_val); 1823 return (1); 1824 } 1825 } else 1826 TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop); 1827 m_adj(m, -todrop); 1828 tlen -= todrop; 1829 thflags &= ~(TH_PUSH | TH_FIN); 1830 } 1831 *thf = thflags; 1832 *tlenp = tlen; 1833 return (0); 1834 } 1835 1836 static struct rack_sendmap * 1837 rack_find_lowest_rsm(struct tcp_rack *rack) 1838 { 1839 struct rack_sendmap *rsm; 1840 1841 /* 1842 * Walk the time-order transmitted list looking for an rsm that is 1843 * not acked. This will be the one that was sent the longest time 1844 * ago that is still outstanding. 1845 */ 1846 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) { 1847 if (rsm->r_flags & RACK_ACKED) { 1848 continue; 1849 } 1850 goto finish; 1851 } 1852 finish: 1853 return (rsm); 1854 } 1855 1856 static struct rack_sendmap * 1857 rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm) 1858 { 1859 struct rack_sendmap *prsm; 1860 1861 /* 1862 * Walk the sequence order list backward until we hit and arrive at 1863 * the highest seq not acked. In theory when this is called it 1864 * should be the last segment (which it was not). 1865 */ 1866 counter_u64_add(rack_find_high, 1); 1867 prsm = rsm; 1868 TAILQ_FOREACH_REVERSE_FROM(prsm, &rack->r_ctl.rc_map, rack_head, r_next) { 1869 if (prsm->r_flags & (RACK_ACKED | RACK_HAS_FIN)) { 1870 continue; 1871 } 1872 return (prsm); 1873 } 1874 return (NULL); 1875 } 1876 1877 1878 static uint32_t 1879 rack_calc_thresh_rack(struct tcp_rack *rack, uint32_t srtt, uint32_t cts) 1880 { 1881 int32_t lro; 1882 uint32_t thresh; 1883 1884 /* 1885 * lro is the flag we use to determine if we have seen reordering. 1886 * If it gets set we have seen reordering. The reorder logic either 1887 * works in one of two ways: 1888 * 1889 * If reorder-fade is configured, then we track the last time we saw 1890 * re-ordering occur. If we reach the point where enough time as 1891 * passed we no longer consider reordering has occuring. 1892 * 1893 * Or if reorder-face is 0, then once we see reordering we consider 1894 * the connection to alway be subject to reordering and just set lro 1895 * to 1. 1896 * 1897 * In the end if lro is non-zero we add the extra time for 1898 * reordering in. 1899 */ 1900 if (srtt == 0) 1901 srtt = 1; 1902 if (rack->r_ctl.rc_reorder_ts) { 1903 if (rack->r_ctl.rc_reorder_fade) { 1904 if (SEQ_GEQ(cts, rack->r_ctl.rc_reorder_ts)) { 1905 lro = cts - rack->r_ctl.rc_reorder_ts; 1906 if (lro == 0) { 1907 /* 1908 * No time as passed since the last 1909 * reorder, mark it as reordering. 1910 */ 1911 lro = 1; 1912 } 1913 } else { 1914 /* Negative time? */ 1915 lro = 0; 1916 } 1917 if (lro > rack->r_ctl.rc_reorder_fade) { 1918 /* Turn off reordering seen too */ 1919 rack->r_ctl.rc_reorder_ts = 0; 1920 lro = 0; 1921 } 1922 } else { 1923 /* Reodering does not fade */ 1924 lro = 1; 1925 } 1926 } else { 1927 lro = 0; 1928 } 1929 thresh = srtt + rack->r_ctl.rc_pkt_delay; 1930 if (lro) { 1931 /* It must be set, if not you get 1/4 rtt */ 1932 if (rack->r_ctl.rc_reorder_shift) 1933 thresh += (srtt >> rack->r_ctl.rc_reorder_shift); 1934 else 1935 thresh += (srtt >> 2); 1936 } else { 1937 thresh += 1; 1938 } 1939 /* We don't let the rack timeout be above a RTO */ 1940 1941 if (thresh > TICKS_2_MSEC(rack->rc_tp->t_rxtcur)) { 1942 thresh = TICKS_2_MSEC(rack->rc_tp->t_rxtcur); 1943 } 1944 /* And we don't want it above the RTO max either */ 1945 if (thresh > rack_rto_max) { 1946 thresh = rack_rto_max; 1947 } 1948 return (thresh); 1949 } 1950 1951 static uint32_t 1952 rack_calc_thresh_tlp(struct tcpcb *tp, struct tcp_rack *rack, 1953 struct rack_sendmap *rsm, uint32_t srtt) 1954 { 1955 struct rack_sendmap *prsm; 1956 uint32_t thresh, len; 1957 int maxseg; 1958 1959 if (srtt == 0) 1960 srtt = 1; 1961 if (rack->r_ctl.rc_tlp_threshold) 1962 thresh = srtt + (srtt / rack->r_ctl.rc_tlp_threshold); 1963 else 1964 thresh = (srtt * 2); 1965 1966 /* Get the previous sent packet, if any */ 1967 maxseg = tcp_maxseg(tp); 1968 counter_u64_add(rack_enter_tlp_calc, 1); 1969 len = rsm->r_end - rsm->r_start; 1970 if (rack->rack_tlp_threshold_use == TLP_USE_ID) { 1971 /* Exactly like the ID */ 1972 if (((tp->snd_max - tp->snd_una) - rack->r_ctl.rc_sacked + rack->r_ctl.rc_holes_rxt) <= maxseg) { 1973 uint32_t alt_thresh; 1974 /* 1975 * Compensate for delayed-ack with the d-ack time. 1976 */ 1977 counter_u64_add(rack_used_tlpmethod, 1); 1978 alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time; 1979 if (alt_thresh > thresh) 1980 thresh = alt_thresh; 1981 } 1982 } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_ONE) { 1983 /* 2.1 behavior */ 1984 prsm = TAILQ_PREV(rsm, rack_head, r_tnext); 1985 if (prsm && (len <= maxseg)) { 1986 /* 1987 * Two packets outstanding, thresh should be (2*srtt) + 1988 * possible inter-packet delay (if any). 1989 */ 1990 uint32_t inter_gap = 0; 1991 int idx, nidx; 1992 1993 counter_u64_add(rack_used_tlpmethod, 1); 1994 idx = rsm->r_rtr_cnt - 1; 1995 nidx = prsm->r_rtr_cnt - 1; 1996 if (TSTMP_GEQ(rsm->r_tim_lastsent[nidx], prsm->r_tim_lastsent[idx])) { 1997 /* Yes it was sent later (or at the same time) */ 1998 inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx]; 1999 } 2000 thresh += inter_gap; 2001 } else if (len <= maxseg) { 2002 /* 2003 * Possibly compensate for delayed-ack. 2004 */ 2005 uint32_t alt_thresh; 2006 2007 counter_u64_add(rack_used_tlpmethod2, 1); 2008 alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time; 2009 if (alt_thresh > thresh) 2010 thresh = alt_thresh; 2011 } 2012 } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_TWO) { 2013 /* 2.2 behavior */ 2014 if (len <= maxseg) { 2015 uint32_t alt_thresh; 2016 /* 2017 * Compensate for delayed-ack with the d-ack time. 2018 */ 2019 counter_u64_add(rack_used_tlpmethod, 1); 2020 alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time; 2021 if (alt_thresh > thresh) 2022 thresh = alt_thresh; 2023 } 2024 } 2025 /* Not above an RTO */ 2026 if (thresh > TICKS_2_MSEC(tp->t_rxtcur)) { 2027 thresh = TICKS_2_MSEC(tp->t_rxtcur); 2028 } 2029 /* Not above a RTO max */ 2030 if (thresh > rack_rto_max) { 2031 thresh = rack_rto_max; 2032 } 2033 /* Apply user supplied min TLP */ 2034 if (thresh < rack_tlp_min) { 2035 thresh = rack_tlp_min; 2036 } 2037 return (thresh); 2038 } 2039 2040 static struct rack_sendmap * 2041 rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused) 2042 { 2043 /* 2044 * Check to see that we don't need to fall into recovery. We will 2045 * need to do so if our oldest transmit is past the time we should 2046 * have had an ack. 2047 */ 2048 struct tcp_rack *rack; 2049 struct rack_sendmap *rsm; 2050 int32_t idx; 2051 uint32_t srtt_cur, srtt, thresh; 2052 2053 rack = (struct tcp_rack *)tp->t_fb_ptr; 2054 if (TAILQ_EMPTY(&rack->r_ctl.rc_map)) { 2055 return (NULL); 2056 } 2057 srtt_cur = tp->t_srtt >> TCP_RTT_SHIFT; 2058 srtt = TICKS_2_MSEC(srtt_cur); 2059 if (rack->rc_rack_rtt && (srtt > rack->rc_rack_rtt)) 2060 srtt = rack->rc_rack_rtt; 2061 2062 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); 2063 if (rsm == NULL) 2064 return (NULL); 2065 2066 if (rsm->r_flags & RACK_ACKED) { 2067 rsm = rack_find_lowest_rsm(rack); 2068 if (rsm == NULL) 2069 return (NULL); 2070 } 2071 idx = rsm->r_rtr_cnt - 1; 2072 thresh = rack_calc_thresh_rack(rack, srtt, tsused); 2073 if (tsused < rsm->r_tim_lastsent[idx]) { 2074 return (NULL); 2075 } 2076 if ((tsused - rsm->r_tim_lastsent[idx]) < thresh) { 2077 return (NULL); 2078 } 2079 /* Ok if we reach here we are over-due */ 2080 rack->r_ctl.rc_rsm_start = rsm->r_start; 2081 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd; 2082 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh; 2083 rack_cong_signal(tp, NULL, CC_NDUPACK); 2084 return (rsm); 2085 } 2086 2087 static uint32_t 2088 rack_get_persists_timer_val(struct tcpcb *tp, struct tcp_rack *rack) 2089 { 2090 int32_t t; 2091 int32_t tt; 2092 uint32_t ret_val; 2093 2094 t = TICKS_2_MSEC((tp->t_srtt >> TCP_RTT_SHIFT) + ((tp->t_rttvar * 4) >> TCP_RTT_SHIFT)); 2095 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], 2096 tcp_persmin, tcp_persmax); 2097 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 2098 tp->t_rxtshift++; 2099 rack->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT; 2100 ret_val = (uint32_t)tt; 2101 return (ret_val); 2102 } 2103 2104 static uint32_t 2105 rack_timer_start(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) 2106 { 2107 /* 2108 * Start the FR timer, we do this based on getting the first one in 2109 * the rc_tmap. Note that if its NULL we must stop the timer. in all 2110 * events we need to stop the running timer (if its running) before 2111 * starting the new one. 2112 */ 2113 uint32_t thresh, exp, to, srtt, time_since_sent; 2114 uint32_t srtt_cur; 2115 int32_t idx; 2116 int32_t is_tlp_timer = 0; 2117 struct rack_sendmap *rsm; 2118 2119 if (rack->t_timers_stopped) { 2120 /* All timers have been stopped none are to run */ 2121 return (0); 2122 } 2123 if (rack->rc_in_persist) { 2124 /* We can't start any timer in persists */ 2125 return (rack_get_persists_timer_val(tp, rack)); 2126 } 2127 if (tp->t_state < TCPS_ESTABLISHED) 2128 goto activate_rxt; 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 rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext); 2188 if (rsm == NULL) { 2189 /* We found no rsm to TLP with. */ 2190 goto activate_rxt; 2191 } 2192 if (rsm->r_flags & RACK_HAS_FIN) { 2193 /* If its a FIN we dont do TLP */ 2194 rsm = NULL; 2195 goto activate_rxt; 2196 } 2197 idx = rsm->r_rtr_cnt - 1; 2198 if (TSTMP_GT(cts, rsm->r_tim_lastsent[idx])) 2199 time_since_sent = cts - rsm->r_tim_lastsent[idx]; 2200 else 2201 time_since_sent = 0; 2202 is_tlp_timer = 1; 2203 if (tp->t_srtt) { 2204 srtt_cur = (tp->t_srtt >> TCP_RTT_SHIFT); 2205 srtt = TICKS_2_MSEC(srtt_cur); 2206 } else 2207 srtt = RACK_INITIAL_RTO; 2208 thresh = rack_calc_thresh_tlp(tp, rack, rsm, srtt); 2209 if (thresh > time_since_sent) 2210 to = thresh - time_since_sent; 2211 else 2212 to = rack->r_ctl.rc_min_to; 2213 if (to > TCPTV_REXMTMAX) { 2214 /* 2215 * If the TLP time works out to larger than the max 2216 * RTO lets not do TLP.. just RTO. 2217 */ 2218 goto activate_rxt; 2219 } 2220 if (rsm->r_start != rack->r_ctl.rc_last_tlp_seq) { 2221 /* 2222 * The tail is no longer the last one I did a probe 2223 * on 2224 */ 2225 rack->r_ctl.rc_tlp_seg_send_cnt = 0; 2226 rack->r_ctl.rc_last_tlp_seq = rsm->r_start; 2227 } 2228 } 2229 if (is_tlp_timer == 0) { 2230 rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK; 2231 } else { 2232 if ((rack->r_ctl.rc_tlp_send_cnt > rack_tlp_max_resend) || 2233 (rack->r_ctl.rc_tlp_seg_send_cnt > rack_tlp_max_resend)) { 2234 /* 2235 * We have exceeded how many times we can retran the 2236 * current TLP timer, switch to the RTO timer. 2237 */ 2238 goto activate_rxt; 2239 } else { 2240 rack->r_ctl.rc_hpts_flags |= PACE_TMR_TLP; 2241 } 2242 } 2243 if (to == 0) 2244 to = 1; 2245 return (to); 2246 } 2247 2248 static void 2249 rack_enter_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) 2250 { 2251 if (rack->rc_in_persist == 0) { 2252 if (((tp->t_flags & TF_SENTFIN) == 0) && 2253 (tp->snd_max - tp->snd_una) >= sbavail(&rack->rc_inp->inp_socket->so_snd)) 2254 /* Must need to send more data to enter persist */ 2255 return; 2256 rack->r_ctl.rc_went_idle_time = cts; 2257 rack_timer_cancel(tp, rack, cts, __LINE__); 2258 tp->t_rxtshift = 0; 2259 rack->rc_in_persist = 1; 2260 } 2261 } 2262 2263 static void 2264 rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack) 2265 { 2266 if (rack->rc_inp->inp_in_hpts) { 2267 tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT); 2268 rack->r_ctl.rc_hpts_flags = 0; 2269 } 2270 rack->rc_in_persist = 0; 2271 rack->r_ctl.rc_went_idle_time = 0; 2272 tp->t_flags &= ~TF_FORCEDATA; 2273 tp->t_rxtshift = 0; 2274 } 2275 2276 static void 2277 rack_start_hpts_timer(struct tcp_rack *rack, struct tcpcb *tp, uint32_t cts, int32_t line, 2278 int32_t slot, uint32_t tot_len_this_send, int32_t frm_out_sbavail) 2279 { 2280 struct inpcb *inp; 2281 uint32_t delayed_ack = 0; 2282 uint32_t hpts_timeout; 2283 uint8_t stopped; 2284 uint32_t left = 0; 2285 2286 inp = tp->t_inpcb; 2287 if (inp->inp_in_hpts) { 2288 /* A previous call is already set up */ 2289 return; 2290 } 2291 if (tp->t_state == TCPS_CLOSED) { 2292 return; 2293 } 2294 stopped = rack->rc_tmr_stopped; 2295 if (stopped && TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) { 2296 left = rack->r_ctl.rc_timer_exp - cts; 2297 } 2298 rack->r_ctl.rc_timer_exp = 0; 2299 if (rack->rc_inp->inp_in_hpts == 0) { 2300 rack->r_ctl.rc_hpts_flags = 0; 2301 } 2302 if (slot) { 2303 /* We are hptsi too */ 2304 rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT; 2305 } else if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) { 2306 /* 2307 * We are still left on the hpts when the to goes 2308 * it will be for output. 2309 */ 2310 if (TSTMP_GT(cts, rack->r_ctl.rc_last_output_to)) 2311 slot = cts - rack->r_ctl.rc_last_output_to; 2312 else 2313 slot = 1; 2314 } 2315 if ((tp->snd_wnd == 0) && TCPS_HAVEESTABLISHED(tp->t_state)) { 2316 /* No send window.. we must enter persist */ 2317 rack_enter_persist(tp, rack, cts); 2318 } else if ((frm_out_sbavail && 2319 (frm_out_sbavail > (tp->snd_max - tp->snd_una)) && 2320 (tp->snd_wnd < tp->t_maxseg)) && 2321 TCPS_HAVEESTABLISHED(tp->t_state)) { 2322 /* 2323 * If we have no window or we can't send a segment (and have 2324 * data to send.. we cheat here and frm_out_sbavail is 2325 * passed in with the sbavail(sb) only from bbr_output) and 2326 * we are established, then we must enter persits (if not 2327 * already in persits). 2328 */ 2329 rack_enter_persist(tp, rack, cts); 2330 } 2331 hpts_timeout = rack_timer_start(tp, rack, cts); 2332 if (tp->t_flags & TF_DELACK) { 2333 delayed_ack = TICKS_2_MSEC(tcp_delacktime); 2334 rack->r_ctl.rc_hpts_flags |= PACE_TMR_DELACK; 2335 } 2336 if (delayed_ack && ((hpts_timeout == 0) || 2337 (delayed_ack < hpts_timeout))) 2338 hpts_timeout = delayed_ack; 2339 else 2340 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK; 2341 /* 2342 * If no timers are going to run and we will fall off the hptsi 2343 * wheel, we resort to a keep-alive timer if its configured. 2344 */ 2345 if ((hpts_timeout == 0) && 2346 (slot == 0)) { 2347 if ((tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && 2348 (tp->t_state <= TCPS_CLOSING)) { 2349 /* 2350 * Ok we have no timer (persists, rack, tlp, rxt or 2351 * del-ack), we don't have segments being paced. So 2352 * all that is left is the keepalive timer. 2353 */ 2354 if (TCPS_HAVEESTABLISHED(tp->t_state)) { 2355 /* Get the established keep-alive time */ 2356 hpts_timeout = TP_KEEPIDLE(tp); 2357 } else { 2358 /* Get the initial setup keep-alive time */ 2359 hpts_timeout = TP_KEEPINIT(tp); 2360 } 2361 rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP; 2362 } 2363 } 2364 if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) == 2365 (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) { 2366 /* 2367 * RACK, TLP, persists and RXT timers all are restartable 2368 * based on actions input .. i.e we received a packet (ack 2369 * or sack) and that changes things (rw, or snd_una etc). 2370 * Thus we can restart them with a new value. For 2371 * keep-alive, delayed_ack we keep track of what was left 2372 * and restart the timer with a smaller value. 2373 */ 2374 if (left < hpts_timeout) 2375 hpts_timeout = left; 2376 } 2377 if (hpts_timeout) { 2378 /* 2379 * Hack alert for now we can't time-out over 2,147,483 2380 * seconds (a bit more than 596 hours), which is probably ok 2381 * :). 2382 */ 2383 if (hpts_timeout > 0x7ffffffe) 2384 hpts_timeout = 0x7ffffffe; 2385 rack->r_ctl.rc_timer_exp = cts + hpts_timeout; 2386 } 2387 if (slot) { 2388 rack->r_ctl.rc_last_output_to = cts + slot; 2389 if ((hpts_timeout == 0) || (hpts_timeout > slot)) { 2390 if (rack->rc_inp->inp_in_hpts == 0) 2391 tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(slot)); 2392 rack_log_to_start(rack, cts, hpts_timeout, slot, 1); 2393 } else { 2394 /* 2395 * Arrange for the hpts to kick back in after the 2396 * t-o if the t-o does not cause a send. 2397 */ 2398 if (rack->rc_inp->inp_in_hpts == 0) 2399 tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(hpts_timeout)); 2400 rack_log_to_start(rack, cts, hpts_timeout, slot, 0); 2401 } 2402 } else if (hpts_timeout) { 2403 if (rack->rc_inp->inp_in_hpts == 0) 2404 tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(hpts_timeout)); 2405 rack_log_to_start(rack, cts, hpts_timeout, slot, 0); 2406 } else { 2407 /* No timer starting */ 2408 #ifdef INVARIANTS 2409 if (SEQ_GT(tp->snd_max, tp->snd_una)) { 2410 panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?", 2411 tp, rack, tot_len_this_send, cts, slot, hpts_timeout); 2412 } 2413 #endif 2414 } 2415 rack->rc_tmr_stopped = 0; 2416 if (slot) 2417 rack_log_type_bbrsnd(rack, tot_len_this_send, slot, cts); 2418 } 2419 2420 /* 2421 * RACK Timer, here we simply do logging and house keeping. 2422 * the normal rack_output() function will call the 2423 * appropriate thing to check if we need to do a RACK retransmit. 2424 * We return 1, saying don't proceed with rack_output only 2425 * when all timers have been stopped (destroyed PCB?). 2426 */ 2427 static int 2428 rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) 2429 { 2430 /* 2431 * This timer simply provides an internal trigger to send out data. 2432 * The check_recovery_mode call will see if there are needed 2433 * retransmissions, if so we will enter fast-recovery. The output 2434 * call may or may not do the same thing depending on sysctl 2435 * settings. 2436 */ 2437 struct rack_sendmap *rsm; 2438 int32_t recovery; 2439 2440 if (tp->t_timers->tt_flags & TT_STOPPED) { 2441 return (1); 2442 } 2443 if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) { 2444 /* Its not time yet */ 2445 return (0); 2446 } 2447 rack_log_to_event(rack, RACK_TO_FRM_RACK); 2448 recovery = IN_RECOVERY(tp->t_flags); 2449 counter_u64_add(rack_to_tot, 1); 2450 if (rack->r_state && (rack->r_state != tp->t_state)) 2451 rack_set_state(tp, rack); 2452 rsm = rack_check_recovery_mode(tp, cts); 2453 if (rsm) { 2454 uint32_t rtt; 2455 2456 rtt = rack->rc_rack_rtt; 2457 if (rtt == 0) 2458 rtt = 1; 2459 if ((recovery == 0) && 2460 (rack->r_ctl.rc_prr_sndcnt < tp->t_maxseg)) { 2461 /* 2462 * The rack-timeout that enter's us into recovery 2463 * will force out one MSS and set us up so that we 2464 * can do one more send in 2*rtt (transitioning the 2465 * rack timeout into a rack-tlp). 2466 */ 2467 rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; 2468 } else if ((rack->r_ctl.rc_prr_sndcnt < tp->t_maxseg) && 2469 ((rsm->r_end - rsm->r_start) > rack->r_ctl.rc_prr_sndcnt)) { 2470 /* 2471 * When a rack timer goes, we have to send at 2472 * least one segment. They will be paced a min of 1ms 2473 * apart via the next rack timer (or further 2474 * if the rack timer dictates it). 2475 */ 2476 rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; 2477 } 2478 } else { 2479 /* This is a case that should happen rarely if ever */ 2480 counter_u64_add(rack_tlp_does_nada, 1); 2481 #ifdef TCP_BLACKBOX 2482 tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true); 2483 #endif 2484 rack->r_ctl.rc_resend = TAILQ_FIRST(&rack->r_ctl.rc_tmap); 2485 } 2486 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK; 2487 return (0); 2488 } 2489 2490 /* 2491 * TLP Timer, here we simply setup what segment we want to 2492 * have the TLP expire on, the normal rack_output() will then 2493 * send it out. 2494 * 2495 * We return 1, saying don't proceed with rack_output only 2496 * when all timers have been stopped (destroyed PCB?). 2497 */ 2498 static int 2499 rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) 2500 { 2501 /* 2502 * Tail Loss Probe. 2503 */ 2504 struct rack_sendmap *rsm = NULL; 2505 struct socket *so; 2506 uint32_t amm, old_prr_snd = 0; 2507 uint32_t out, avail; 2508 2509 if (tp->t_timers->tt_flags & TT_STOPPED) { 2510 return (1); 2511 } 2512 if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) { 2513 /* Its not time yet */ 2514 return (0); 2515 } 2516 if (rack_progress_timeout_check(tp)) { 2517 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); 2518 return (1); 2519 } 2520 /* 2521 * A TLP timer has expired. We have been idle for 2 rtts. So we now 2522 * need to figure out how to force a full MSS segment out. 2523 */ 2524 rack_log_to_event(rack, RACK_TO_FRM_TLP); 2525 counter_u64_add(rack_tlp_tot, 1); 2526 if (rack->r_state && (rack->r_state != tp->t_state)) 2527 rack_set_state(tp, rack); 2528 so = tp->t_inpcb->inp_socket; 2529 avail = sbavail(&so->so_snd); 2530 out = tp->snd_max - tp->snd_una; 2531 rack->rc_timer_up = 1; 2532 /* 2533 * If we are in recovery we can jazz out a segment if new data is 2534 * present simply by setting rc_prr_sndcnt to a segment. 2535 */ 2536 if ((avail > out) && 2537 ((rack_always_send_oldest == 0) || (TAILQ_EMPTY(&rack->r_ctl.rc_tmap)))) { 2538 /* New data is available */ 2539 amm = avail - out; 2540 if (amm > tp->t_maxseg) { 2541 amm = tp->t_maxseg; 2542 } else if ((amm < tp->t_maxseg) && ((tp->t_flags & TF_NODELAY) == 0)) { 2543 /* not enough to fill a MTU and no-delay is off */ 2544 goto need_retran; 2545 } 2546 if (IN_RECOVERY(tp->t_flags)) { 2547 /* Unlikely */ 2548 old_prr_snd = rack->r_ctl.rc_prr_sndcnt; 2549 if (out + amm <= tp->snd_wnd) 2550 rack->r_ctl.rc_prr_sndcnt = amm; 2551 else 2552 goto need_retran; 2553 } else { 2554 /* Set the send-new override */ 2555 if (out + amm <= tp->snd_wnd) 2556 rack->r_ctl.rc_tlp_new_data = amm; 2557 else 2558 goto need_retran; 2559 } 2560 rack->r_ctl.rc_tlp_seg_send_cnt = 0; 2561 rack->r_ctl.rc_last_tlp_seq = tp->snd_max; 2562 rack->r_ctl.rc_tlpsend = NULL; 2563 counter_u64_add(rack_tlp_newdata, 1); 2564 goto send; 2565 } 2566 need_retran: 2567 /* 2568 * Ok we need to arrange the last un-acked segment to be re-sent, or 2569 * optionally the first un-acked segment. 2570 */ 2571 if (rack_always_send_oldest) 2572 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); 2573 else { 2574 rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_map, rack_sendmap, r_next); 2575 if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) { 2576 rsm = rack_find_high_nonack(rack, rsm); 2577 } 2578 } 2579 if (rsm == NULL) { 2580 counter_u64_add(rack_tlp_does_nada, 1); 2581 #ifdef TCP_BLACKBOX 2582 tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true); 2583 #endif 2584 goto out; 2585 } 2586 if ((rsm->r_end - rsm->r_start) > tp->t_maxseg) { 2587 /* 2588 * We need to split this the last segment in two. 2589 */ 2590 int32_t idx; 2591 struct rack_sendmap *nrsm; 2592 2593 nrsm = rack_alloc(rack); 2594 if (nrsm == NULL) { 2595 /* 2596 * No memory to split, we will just exit and punt 2597 * off to the RXT timer. 2598 */ 2599 counter_u64_add(rack_tlp_does_nada, 1); 2600 goto out; 2601 } 2602 nrsm->r_start = (rsm->r_end - tp->t_maxseg); 2603 nrsm->r_end = rsm->r_end; 2604 nrsm->r_rtr_cnt = rsm->r_rtr_cnt; 2605 nrsm->r_flags = rsm->r_flags; 2606 nrsm->r_sndcnt = rsm->r_sndcnt; 2607 nrsm->r_rtr_bytes = 0; 2608 rsm->r_end = nrsm->r_start; 2609 for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { 2610 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; 2611 } 2612 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_map, rsm, nrsm, r_next); 2613 if (rsm->r_in_tmap) { 2614 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 2615 nrsm->r_in_tmap = 1; 2616 } 2617 rsm->r_flags &= (~RACK_HAS_FIN); 2618 rsm = nrsm; 2619 } 2620 rack->r_ctl.rc_tlpsend = rsm; 2621 rack->r_ctl.rc_tlp_rtx_out = 1; 2622 if (rsm->r_start == rack->r_ctl.rc_last_tlp_seq) { 2623 rack->r_ctl.rc_tlp_seg_send_cnt++; 2624 tp->t_rxtshift++; 2625 } else { 2626 rack->r_ctl.rc_last_tlp_seq = rsm->r_start; 2627 rack->r_ctl.rc_tlp_seg_send_cnt = 1; 2628 } 2629 send: 2630 rack->r_ctl.rc_tlp_send_cnt++; 2631 if (rack->r_ctl.rc_tlp_send_cnt > rack_tlp_max_resend) { 2632 /* 2633 * Can't [re]/transmit a segment we have not heard from the 2634 * peer in max times. We need the retransmit timer to take 2635 * over. 2636 */ 2637 restore: 2638 rack->r_ctl.rc_tlpsend = NULL; 2639 if (rsm) 2640 rsm->r_flags &= ~RACK_TLP; 2641 rack->r_ctl.rc_prr_sndcnt = old_prr_snd; 2642 counter_u64_add(rack_tlp_retran_fail, 1); 2643 goto out; 2644 } else if (rsm) { 2645 rsm->r_flags |= RACK_TLP; 2646 } 2647 if (rsm && (rsm->r_start == rack->r_ctl.rc_last_tlp_seq) && 2648 (rack->r_ctl.rc_tlp_seg_send_cnt > rack_tlp_max_resend)) { 2649 /* 2650 * We don't want to send a single segment more than the max 2651 * either. 2652 */ 2653 goto restore; 2654 } 2655 rack->r_timer_override = 1; 2656 rack->r_tlp_running = 1; 2657 rack->rc_tlp_in_progress = 1; 2658 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP; 2659 return (0); 2660 out: 2661 rack->rc_timer_up = 0; 2662 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP; 2663 return (0); 2664 } 2665 2666 /* 2667 * Delayed ack Timer, here we simply need to setup the 2668 * ACK_NOW flag and remove the DELACK flag. From there 2669 * the output routine will send the ack out. 2670 * 2671 * We only return 1, saying don't proceed, if all timers 2672 * are stopped (destroyed PCB?). 2673 */ 2674 static int 2675 rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) 2676 { 2677 if (tp->t_timers->tt_flags & TT_STOPPED) { 2678 return (1); 2679 } 2680 rack_log_to_event(rack, RACK_TO_FRM_DELACK); 2681 tp->t_flags &= ~TF_DELACK; 2682 tp->t_flags |= TF_ACKNOW; 2683 TCPSTAT_INC(tcps_delack); 2684 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK; 2685 return (0); 2686 } 2687 2688 /* 2689 * Persists timer, here we simply need to setup the 2690 * FORCE-DATA flag the output routine will send 2691 * the one byte send. 2692 * 2693 * We only return 1, saying don't proceed, if all timers 2694 * are stopped (destroyed PCB?). 2695 */ 2696 static int 2697 rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) 2698 { 2699 struct inpcb *inp; 2700 int32_t retval = 0; 2701 2702 inp = tp->t_inpcb; 2703 2704 if (tp->t_timers->tt_flags & TT_STOPPED) { 2705 return (1); 2706 } 2707 if (rack->rc_in_persist == 0) 2708 return (0); 2709 if (rack_progress_timeout_check(tp)) { 2710 tcp_set_inp_to_drop(inp, ETIMEDOUT); 2711 return (1); 2712 } 2713 KASSERT(inp != NULL, ("%s: tp %p tp->t_inpcb == NULL", __func__, tp)); 2714 /* 2715 * Persistence timer into zero window. Force a byte to be output, if 2716 * possible. 2717 */ 2718 TCPSTAT_INC(tcps_persisttimeo); 2719 /* 2720 * Hack: if the peer is dead/unreachable, we do not time out if the 2721 * window is closed. After a full backoff, drop the connection if 2722 * the idle time (no responses to probes) reaches the maximum 2723 * backoff that we would use if retransmitting. 2724 */ 2725 if (tp->t_rxtshift == TCP_MAXRXTSHIFT && 2726 (ticks - tp->t_rcvtime >= tcp_maxpersistidle || 2727 ticks - tp->t_rcvtime >= TCP_REXMTVAL(tp) * tcp_totbackoff)) { 2728 TCPSTAT_INC(tcps_persistdrop); 2729 retval = 1; 2730 tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT); 2731 goto out; 2732 } 2733 if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) && 2734 tp->snd_una == tp->snd_max) 2735 rack_exit_persist(tp, rack); 2736 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT; 2737 /* 2738 * If the user has closed the socket then drop a persisting 2739 * connection after a much reduced timeout. 2740 */ 2741 if (tp->t_state > TCPS_CLOSE_WAIT && 2742 (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) { 2743 retval = 1; 2744 TCPSTAT_INC(tcps_persistdrop); 2745 tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT); 2746 goto out; 2747 } 2748 tp->t_flags |= TF_FORCEDATA; 2749 out: 2750 rack_log_to_event(rack, RACK_TO_FRM_PERSIST); 2751 return (retval); 2752 } 2753 2754 /* 2755 * If a keepalive goes off, we had no other timers 2756 * happening. We always return 1 here since this 2757 * routine either drops the connection or sends 2758 * out a segment with respond. 2759 */ 2760 static int 2761 rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) 2762 { 2763 struct tcptemp *t_template; 2764 struct inpcb *inp; 2765 2766 if (tp->t_timers->tt_flags & TT_STOPPED) { 2767 return (1); 2768 } 2769 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP; 2770 inp = tp->t_inpcb; 2771 rack_log_to_event(rack, RACK_TO_FRM_KEEP); 2772 /* 2773 * Keep-alive timer went off; send something or drop connection if 2774 * idle for too long. 2775 */ 2776 TCPSTAT_INC(tcps_keeptimeo); 2777 if (tp->t_state < TCPS_ESTABLISHED) 2778 goto dropit; 2779 if ((tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && 2780 tp->t_state <= TCPS_CLOSING) { 2781 if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp)) 2782 goto dropit; 2783 /* 2784 * Send a packet designed to force a response if the peer is 2785 * up and reachable: either an ACK if the connection is 2786 * still alive, or an RST if the peer has closed the 2787 * connection due to timeout or reboot. Using sequence 2788 * number tp->snd_una-1 causes the transmitted zero-length 2789 * segment to lie outside the receive window; by the 2790 * protocol spec, this requires the correspondent TCP to 2791 * respond. 2792 */ 2793 TCPSTAT_INC(tcps_keepprobe); 2794 t_template = tcpip_maketemplate(inp); 2795 if (t_template) { 2796 tcp_respond(tp, t_template->tt_ipgen, 2797 &t_template->tt_t, (struct mbuf *)NULL, 2798 tp->rcv_nxt, tp->snd_una - 1, 0); 2799 free(t_template, M_TEMP); 2800 } 2801 } 2802 rack_start_hpts_timer(rack, tp, cts, __LINE__, 0, 0, 0); 2803 return (1); 2804 dropit: 2805 TCPSTAT_INC(tcps_keepdrops); 2806 tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT); 2807 return (1); 2808 } 2809 2810 /* 2811 * Retransmit helper function, clear up all the ack 2812 * flags and take care of important book keeping. 2813 */ 2814 static void 2815 rack_remxt_tmr(struct tcpcb *tp) 2816 { 2817 /* 2818 * The retransmit timer went off, all sack'd blocks must be 2819 * un-acked. 2820 */ 2821 struct rack_sendmap *rsm, *trsm = NULL; 2822 struct tcp_rack *rack; 2823 int32_t cnt = 0; 2824 2825 rack = (struct tcp_rack *)tp->t_fb_ptr; 2826 rack_timer_cancel(tp, rack, tcp_ts_getticks(), __LINE__); 2827 rack_log_to_event(rack, RACK_TO_FRM_TMR); 2828 if (rack->r_state && (rack->r_state != tp->t_state)) 2829 rack_set_state(tp, rack); 2830 /* 2831 * Ideally we would like to be able to 2832 * mark SACK-PASS on anything not acked here. 2833 * However, if we do that we would burst out 2834 * all that data 1ms apart. This would be unwise, 2835 * so for now we will just let the normal rxt timer 2836 * and tlp timer take care of it. 2837 */ 2838 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_map, r_next) { 2839 if (rsm->r_flags & RACK_ACKED) { 2840 cnt++; 2841 rsm->r_sndcnt = 0; 2842 if (rsm->r_in_tmap == 0) { 2843 /* We must re-add it back to the tlist */ 2844 if (trsm == NULL) { 2845 TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext); 2846 } else { 2847 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext); 2848 } 2849 rsm->r_in_tmap = 1; 2850 trsm = rsm; 2851 } 2852 } 2853 rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS); 2854 } 2855 /* Clear the count (we just un-acked them) */ 2856 rack->r_ctl.rc_sacked = 0; 2857 /* Clear the tlp rtx mark */ 2858 rack->r_ctl.rc_tlp_rtx_out = 0; 2859 rack->r_ctl.rc_tlp_seg_send_cnt = 0; 2860 rack->r_ctl.rc_resend = TAILQ_FIRST(&rack->r_ctl.rc_map); 2861 /* Setup so we send one segment */ 2862 if (rack->r_ctl.rc_prr_sndcnt < tp->t_maxseg) 2863 rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; 2864 rack->r_timer_override = 1; 2865 } 2866 2867 /* 2868 * Re-transmit timeout! If we drop the PCB we will return 1, otherwise 2869 * we will setup to retransmit the lowest seq number outstanding. 2870 */ 2871 static int 2872 rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) 2873 { 2874 int32_t rexmt; 2875 struct inpcb *inp; 2876 int32_t retval = 0; 2877 2878 inp = tp->t_inpcb; 2879 if (tp->t_timers->tt_flags & TT_STOPPED) { 2880 return (1); 2881 } 2882 if (rack_progress_timeout_check(tp)) { 2883 tcp_set_inp_to_drop(inp, ETIMEDOUT); 2884 return (1); 2885 } 2886 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT; 2887 if (TCPS_HAVEESTABLISHED(tp->t_state) && 2888 (tp->snd_una == tp->snd_max)) { 2889 /* Nothing outstanding .. nothing to do */ 2890 return (0); 2891 } 2892 /* 2893 * Retransmission timer went off. Message has not been acked within 2894 * retransmit interval. Back off to a longer retransmit interval 2895 * and retransmit one segment. 2896 */ 2897 if (++tp->t_rxtshift > TCP_MAXRXTSHIFT) { 2898 tp->t_rxtshift = TCP_MAXRXTSHIFT; 2899 TCPSTAT_INC(tcps_timeoutdrop); 2900 retval = 1; 2901 tcp_set_inp_to_drop(rack->rc_inp, 2902 (tp->t_softerror ? (uint16_t) tp->t_softerror : ETIMEDOUT)); 2903 goto out; 2904 } 2905 rack_remxt_tmr(tp); 2906 if (tp->t_state == TCPS_SYN_SENT) { 2907 /* 2908 * If the SYN was retransmitted, indicate CWND to be limited 2909 * to 1 segment in cc_conn_init(). 2910 */ 2911 tp->snd_cwnd = 1; 2912 } else if (tp->t_rxtshift == 1) { 2913 /* 2914 * first retransmit; record ssthresh and cwnd so they can be 2915 * recovered if this turns out to be a "bad" retransmit. A 2916 * retransmit is considered "bad" if an ACK for this segment 2917 * is received within RTT/2 interval; the assumption here is 2918 * that the ACK was already in flight. See "On Estimating 2919 * End-to-End Network Path Properties" by Allman and Paxson 2920 * for more details. 2921 */ 2922 tp->snd_cwnd_prev = tp->snd_cwnd; 2923 tp->snd_ssthresh_prev = tp->snd_ssthresh; 2924 tp->snd_recover_prev = tp->snd_recover; 2925 if (IN_FASTRECOVERY(tp->t_flags)) 2926 tp->t_flags |= TF_WASFRECOVERY; 2927 else 2928 tp->t_flags &= ~TF_WASFRECOVERY; 2929 if (IN_CONGRECOVERY(tp->t_flags)) 2930 tp->t_flags |= TF_WASCRECOVERY; 2931 else 2932 tp->t_flags &= ~TF_WASCRECOVERY; 2933 tp->t_badrxtwin = ticks + (tp->t_srtt >> (TCP_RTT_SHIFT + 1)); 2934 tp->t_flags |= TF_PREVVALID; 2935 } else 2936 tp->t_flags &= ~TF_PREVVALID; 2937 TCPSTAT_INC(tcps_rexmttimeo); 2938 if ((tp->t_state == TCPS_SYN_SENT) || 2939 (tp->t_state == TCPS_SYN_RECEIVED)) 2940 rexmt = MSEC_2_TICKS(RACK_INITIAL_RTO * tcp_backoff[tp->t_rxtshift]); 2941 else 2942 rexmt = TCP_REXMTVAL(tp) * tcp_backoff[tp->t_rxtshift]; 2943 TCPT_RANGESET(tp->t_rxtcur, rexmt, 2944 max(MSEC_2_TICKS(rack_rto_min), rexmt), 2945 MSEC_2_TICKS(rack_rto_max)); 2946 /* 2947 * We enter the path for PLMTUD if connection is established or, if 2948 * connection is FIN_WAIT_1 status, reason for the last is that if 2949 * amount of data we send is very small, we could send it in couple 2950 * of packets and process straight to FIN. In that case we won't 2951 * catch ESTABLISHED state. 2952 */ 2953 if (V_tcp_pmtud_blackhole_detect && (((tp->t_state == TCPS_ESTABLISHED)) 2954 || (tp->t_state == TCPS_FIN_WAIT_1))) { 2955 #ifdef INET6 2956 int32_t isipv6; 2957 #endif 2958 2959 /* 2960 * Idea here is that at each stage of mtu probe (usually, 2961 * 1448 -> 1188 -> 524) should be given 2 chances to recover 2962 * before further clamping down. 'tp->t_rxtshift % 2 == 0' 2963 * should take care of that. 2964 */ 2965 if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) == 2966 (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) && 2967 (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 && 2968 tp->t_rxtshift % 2 == 0)) { 2969 /* 2970 * Enter Path MTU Black-hole Detection mechanism: - 2971 * Disable Path MTU Discovery (IP "DF" bit). - 2972 * Reduce MTU to lower value than what we negotiated 2973 * with peer. 2974 */ 2975 if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) { 2976 /* Record that we may have found a black hole. */ 2977 tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE; 2978 /* Keep track of previous MSS. */ 2979 tp->t_pmtud_saved_maxseg = tp->t_maxseg; 2980 } 2981 2982 /* 2983 * Reduce the MSS to blackhole value or to the 2984 * default in an attempt to retransmit. 2985 */ 2986 #ifdef INET6 2987 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) ? 1 : 0; 2988 if (isipv6 && 2989 tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) { 2990 /* Use the sysctl tuneable blackhole MSS. */ 2991 tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss; 2992 TCPSTAT_INC(tcps_pmtud_blackhole_activated); 2993 } else if (isipv6) { 2994 /* Use the default MSS. */ 2995 tp->t_maxseg = V_tcp_v6mssdflt; 2996 /* 2997 * Disable Path MTU Discovery when we switch 2998 * to minmss. 2999 */ 3000 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 3001 TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss); 3002 } 3003 #endif 3004 #if defined(INET6) && defined(INET) 3005 else 3006 #endif 3007 #ifdef INET 3008 if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) { 3009 /* Use the sysctl tuneable blackhole MSS. */ 3010 tp->t_maxseg = V_tcp_pmtud_blackhole_mss; 3011 TCPSTAT_INC(tcps_pmtud_blackhole_activated); 3012 } else { 3013 /* Use the default MSS. */ 3014 tp->t_maxseg = V_tcp_mssdflt; 3015 /* 3016 * Disable Path MTU Discovery when we switch 3017 * to minmss. 3018 */ 3019 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 3020 TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss); 3021 } 3022 #endif 3023 } else { 3024 /* 3025 * If further retransmissions are still unsuccessful 3026 * with a lowered MTU, maybe this isn't a blackhole 3027 * and we restore the previous MSS and blackhole 3028 * detection flags. The limit '6' is determined by 3029 * giving each probe stage (1448, 1188, 524) 2 3030 * chances to recover. 3031 */ 3032 if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) && 3033 (tp->t_rxtshift >= 6)) { 3034 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 3035 tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE; 3036 tp->t_maxseg = tp->t_pmtud_saved_maxseg; 3037 TCPSTAT_INC(tcps_pmtud_blackhole_failed); 3038 } 3039 } 3040 } 3041 /* 3042 * Disable RFC1323 and SACK if we haven't got any response to our 3043 * third SYN to work-around some broken terminal servers (most of 3044 * which have hopefully been retired) that have bad VJ header 3045 * compression code which trashes TCP segments containing 3046 * unknown-to-them TCP options. 3047 */ 3048 if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) && 3049 (tp->t_rxtshift == 3)) 3050 tp->t_flags &= ~(TF_REQ_SCALE | TF_REQ_TSTMP | TF_SACK_PERMIT); 3051 /* 3052 * If we backed off this far, our srtt estimate is probably bogus. 3053 * Clobber it so we'll take the next rtt measurement as our srtt; 3054 * move the current srtt into rttvar to keep the current retransmit 3055 * times until then. 3056 */ 3057 if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) { 3058 #ifdef INET6 3059 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0) 3060 in6_losing(tp->t_inpcb); 3061 else 3062 #endif 3063 in_losing(tp->t_inpcb); 3064 tp->t_rttvar += (tp->t_srtt >> TCP_RTT_SHIFT); 3065 tp->t_srtt = 0; 3066 } 3067 if (rack_use_sack_filter) 3068 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una); 3069 tp->snd_recover = tp->snd_max; 3070 tp->t_flags |= TF_ACKNOW; 3071 tp->t_rtttime = 0; 3072 rack_cong_signal(tp, NULL, CC_RTO); 3073 out: 3074 return (retval); 3075 } 3076 3077 static int 3078 rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling) 3079 { 3080 int32_t ret = 0; 3081 int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK); 3082 3083 if (timers == 0) { 3084 return (0); 3085 } 3086 if (tp->t_state == TCPS_LISTEN) { 3087 /* no timers on listen sockets */ 3088 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) 3089 return (0); 3090 return (1); 3091 } 3092 if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) { 3093 uint32_t left; 3094 3095 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) { 3096 ret = -1; 3097 rack_log_to_processing(rack, cts, ret, 0); 3098 return (0); 3099 } 3100 if (hpts_calling == 0) { 3101 ret = -2; 3102 rack_log_to_processing(rack, cts, ret, 0); 3103 return (0); 3104 } 3105 /* 3106 * Ok our timer went off early and we are not paced false 3107 * alarm, go back to sleep. 3108 */ 3109 ret = -3; 3110 left = rack->r_ctl.rc_timer_exp - cts; 3111 tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(left)); 3112 rack_log_to_processing(rack, cts, ret, left); 3113 rack->rc_last_pto_set = 0; 3114 return (1); 3115 } 3116 rack->rc_tmr_stopped = 0; 3117 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK; 3118 if (timers & PACE_TMR_DELACK) { 3119 ret = rack_timeout_delack(tp, rack, cts); 3120 } else if (timers & PACE_TMR_RACK) { 3121 ret = rack_timeout_rack(tp, rack, cts); 3122 } else if (timers & PACE_TMR_TLP) { 3123 ret = rack_timeout_tlp(tp, rack, cts); 3124 } else if (timers & PACE_TMR_RXT) { 3125 ret = rack_timeout_rxt(tp, rack, cts); 3126 } else if (timers & PACE_TMR_PERSIT) { 3127 ret = rack_timeout_persist(tp, rack, cts); 3128 } else if (timers & PACE_TMR_KEEP) { 3129 ret = rack_timeout_keepalive(tp, rack, cts); 3130 } 3131 rack_log_to_processing(rack, cts, ret, timers); 3132 return (ret); 3133 } 3134 3135 static void 3136 rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line) 3137 { 3138 uint8_t hpts_removed = 0; 3139 3140 if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) && 3141 TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to)) { 3142 tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT); 3143 hpts_removed = 1; 3144 } 3145 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) { 3146 rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK; 3147 if (rack->rc_inp->inp_in_hpts && 3148 ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) { 3149 /* 3150 * Canceling timer's when we have no output being 3151 * paced. We also must remove ourselves from the 3152 * hpts. 3153 */ 3154 tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT); 3155 hpts_removed = 1; 3156 } 3157 rack_log_to_cancel(rack, hpts_removed, line); 3158 rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK); 3159 } 3160 } 3161 3162 static void 3163 rack_timer_stop(struct tcpcb *tp, uint32_t timer_type) 3164 { 3165 return; 3166 } 3167 3168 static int 3169 rack_stopall(struct tcpcb *tp) 3170 { 3171 struct tcp_rack *rack; 3172 rack = (struct tcp_rack *)tp->t_fb_ptr; 3173 rack->t_timers_stopped = 1; 3174 return (0); 3175 } 3176 3177 static void 3178 rack_timer_activate(struct tcpcb *tp, uint32_t timer_type, uint32_t delta) 3179 { 3180 return; 3181 } 3182 3183 static int 3184 rack_timer_active(struct tcpcb *tp, uint32_t timer_type) 3185 { 3186 return (0); 3187 } 3188 3189 static void 3190 rack_stop_all_timers(struct tcpcb *tp) 3191 { 3192 struct tcp_rack *rack; 3193 3194 /* 3195 * Assure no timers are running. 3196 */ 3197 if (tcp_timer_active(tp, TT_PERSIST)) { 3198 /* We enter in persists, set the flag appropriately */ 3199 rack = (struct tcp_rack *)tp->t_fb_ptr; 3200 rack->rc_in_persist = 1; 3201 } 3202 tcp_timer_suspend(tp, TT_PERSIST); 3203 tcp_timer_suspend(tp, TT_REXMT); 3204 tcp_timer_suspend(tp, TT_KEEP); 3205 tcp_timer_suspend(tp, TT_DELACK); 3206 } 3207 3208 static void 3209 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack, 3210 struct rack_sendmap *rsm, uint32_t ts) 3211 { 3212 int32_t idx; 3213 3214 rsm->r_rtr_cnt++; 3215 rsm->r_sndcnt++; 3216 if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) { 3217 rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS; 3218 rsm->r_flags |= RACK_OVERMAX; 3219 } 3220 if ((rsm->r_rtr_cnt > 1) && (rack->r_tlp_running == 0)) { 3221 rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start); 3222 rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start); 3223 } 3224 idx = rsm->r_rtr_cnt - 1; 3225 rsm->r_tim_lastsent[idx] = ts; 3226 if (rsm->r_flags & RACK_ACKED) { 3227 /* Problably MTU discovery messing with us */ 3228 rsm->r_flags &= ~RACK_ACKED; 3229 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); 3230 } 3231 if (rsm->r_in_tmap) { 3232 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext); 3233 } 3234 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext); 3235 rsm->r_in_tmap = 1; 3236 if (rsm->r_flags & RACK_SACK_PASSED) { 3237 /* We have retransmitted due to the SACK pass */ 3238 rsm->r_flags &= ~RACK_SACK_PASSED; 3239 rsm->r_flags |= RACK_WAS_SACKPASS; 3240 } 3241 /* Update memory for next rtr */ 3242 rack->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next); 3243 } 3244 3245 3246 static uint32_t 3247 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack, 3248 struct rack_sendmap *rsm, uint32_t ts, int32_t * lenp) 3249 { 3250 /* 3251 * We (re-)transmitted starting at rsm->r_start for some length 3252 * (possibly less than r_end. 3253 */ 3254 struct rack_sendmap *nrsm; 3255 uint32_t c_end; 3256 int32_t len; 3257 int32_t idx; 3258 3259 len = *lenp; 3260 c_end = rsm->r_start + len; 3261 if (SEQ_GEQ(c_end, rsm->r_end)) { 3262 /* 3263 * We retransmitted the whole piece or more than the whole 3264 * slopping into the next rsm. 3265 */ 3266 rack_update_rsm(tp, rack, rsm, ts); 3267 if (c_end == rsm->r_end) { 3268 *lenp = 0; 3269 return (0); 3270 } else { 3271 int32_t act_len; 3272 3273 /* Hangs over the end return whats left */ 3274 act_len = rsm->r_end - rsm->r_start; 3275 *lenp = (len - act_len); 3276 return (rsm->r_end); 3277 } 3278 /* We don't get out of this block. */ 3279 } 3280 /* 3281 * Here we retransmitted less than the whole thing which means we 3282 * have to split this into what was transmitted and what was not. 3283 */ 3284 nrsm = rack_alloc(rack); 3285 if (nrsm == NULL) { 3286 /* 3287 * We can't get memory, so lets not proceed. 3288 */ 3289 *lenp = 0; 3290 return (0); 3291 } 3292 /* 3293 * So here we are going to take the original rsm and make it what we 3294 * retransmitted. nrsm will be the tail portion we did not 3295 * retransmit. For example say the chunk was 1, 11 (10 bytes). And 3296 * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to 3297 * 1, 6 and the new piece will be 6, 11. 3298 */ 3299 nrsm->r_start = c_end; 3300 nrsm->r_end = rsm->r_end; 3301 nrsm->r_rtr_cnt = rsm->r_rtr_cnt; 3302 nrsm->r_flags = rsm->r_flags; 3303 nrsm->r_sndcnt = rsm->r_sndcnt; 3304 nrsm->r_rtr_bytes = 0; 3305 rsm->r_end = c_end; 3306 for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { 3307 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; 3308 } 3309 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_map, rsm, nrsm, r_next); 3310 if (rsm->r_in_tmap) { 3311 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 3312 nrsm->r_in_tmap = 1; 3313 } 3314 rsm->r_flags &= (~RACK_HAS_FIN); 3315 rack_update_rsm(tp, rack, rsm, ts); 3316 *lenp = 0; 3317 return (0); 3318 } 3319 3320 3321 static void 3322 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len, 3323 uint32_t seq_out, uint8_t th_flags, int32_t err, uint32_t ts, 3324 uint8_t pass, struct rack_sendmap *hintrsm) 3325 { 3326 struct tcp_rack *rack; 3327 struct rack_sendmap *rsm, *nrsm; 3328 register uint32_t snd_max, snd_una; 3329 int32_t idx; 3330 3331 /* 3332 * Add to the RACK log of packets in flight or retransmitted. If 3333 * there is a TS option we will use the TS echoed, if not we will 3334 * grab a TS. 3335 * 3336 * Retransmissions will increment the count and move the ts to its 3337 * proper place. Note that if options do not include TS's then we 3338 * won't be able to effectively use the ACK for an RTT on a retran. 3339 * 3340 * Notes about r_start and r_end. Lets consider a send starting at 3341 * sequence 1 for 10 bytes. In such an example the r_start would be 3342 * 1 (starting sequence) but the r_end would be r_start+len i.e. 11. 3343 * This means that r_end is actually the first sequence for the next 3344 * slot (11). 3345 * 3346 */ 3347 /* 3348 * If err is set what do we do XXXrrs? should we not add the thing? 3349 * -- i.e. return if err != 0 or should we pretend we sent it? -- 3350 * i.e. proceed with add ** do this for now. 3351 */ 3352 INP_WLOCK_ASSERT(tp->t_inpcb); 3353 if (err) 3354 /* 3355 * We don't log errors -- we could but snd_max does not 3356 * advance in this case either. 3357 */ 3358 return; 3359 3360 if (th_flags & TH_RST) { 3361 /* 3362 * We don't log resets and we return immediately from 3363 * sending 3364 */ 3365 return; 3366 } 3367 rack = (struct tcp_rack *)tp->t_fb_ptr; 3368 snd_una = tp->snd_una; 3369 if (SEQ_LEQ((seq_out + len), snd_una)) { 3370 /* Are sending an old segment to induce an ack (keep-alive)? */ 3371 return; 3372 } 3373 if (SEQ_LT(seq_out, snd_una)) { 3374 /* huh? should we panic? */ 3375 uint32_t end; 3376 3377 end = seq_out + len; 3378 seq_out = snd_una; 3379 len = end - seq_out; 3380 } 3381 snd_max = tp->snd_max; 3382 if (th_flags & (TH_SYN | TH_FIN)) { 3383 /* 3384 * The call to rack_log_output is made before bumping 3385 * snd_max. This means we can record one extra byte on a SYN 3386 * or FIN if seq_out is adding more on and a FIN is present 3387 * (and we are not resending). 3388 */ 3389 if (th_flags & TH_SYN) 3390 len++; 3391 if (th_flags & TH_FIN) 3392 len++; 3393 if (SEQ_LT(snd_max, tp->snd_nxt)) { 3394 /* 3395 * The add/update as not been done for the FIN/SYN 3396 * yet. 3397 */ 3398 snd_max = tp->snd_nxt; 3399 } 3400 } 3401 if (len == 0) { 3402 /* We don't log zero window probes */ 3403 return; 3404 } 3405 rack->r_ctl.rc_time_last_sent = ts; 3406 if (IN_RECOVERY(tp->t_flags)) { 3407 rack->r_ctl.rc_prr_out += len; 3408 } 3409 /* First question is it a retransmission? */ 3410 if (seq_out == snd_max) { 3411 again: 3412 rsm = rack_alloc(rack); 3413 if (rsm == NULL) { 3414 /* 3415 * Hmm out of memory and the tcb got destroyed while 3416 * we tried to wait. 3417 */ 3418 #ifdef INVARIANTS 3419 panic("Out of memory when we should not be rack:%p", rack); 3420 #endif 3421 return; 3422 } 3423 if (th_flags & TH_FIN) { 3424 rsm->r_flags = RACK_HAS_FIN; 3425 } else { 3426 rsm->r_flags = 0; 3427 } 3428 rsm->r_tim_lastsent[0] = ts; 3429 rsm->r_rtr_cnt = 1; 3430 rsm->r_rtr_bytes = 0; 3431 if (th_flags & TH_SYN) { 3432 /* The data space is one beyond snd_una */ 3433 rsm->r_start = seq_out + 1; 3434 rsm->r_end = rsm->r_start + (len - 1); 3435 } else { 3436 /* Normal case */ 3437 rsm->r_start = seq_out; 3438 rsm->r_end = rsm->r_start + len; 3439 } 3440 rsm->r_sndcnt = 0; 3441 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_map, rsm, r_next); 3442 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext); 3443 rsm->r_in_tmap = 1; 3444 return; 3445 } 3446 /* 3447 * If we reach here its a retransmission and we need to find it. 3448 */ 3449 more: 3450 if (hintrsm && (hintrsm->r_start == seq_out)) { 3451 rsm = hintrsm; 3452 hintrsm = NULL; 3453 } else if (rack->r_ctl.rc_next) { 3454 /* We have a hint from a previous run */ 3455 rsm = rack->r_ctl.rc_next; 3456 } else { 3457 /* No hints sorry */ 3458 rsm = NULL; 3459 } 3460 if ((rsm) && (rsm->r_start == seq_out)) { 3461 /* 3462 * We used rc_next or hintrsm to retransmit, hopefully the 3463 * likely case. 3464 */ 3465 seq_out = rack_update_entry(tp, rack, rsm, ts, &len); 3466 if (len == 0) { 3467 return; 3468 } else { 3469 goto more; 3470 } 3471 } 3472 /* Ok it was not the last pointer go through it the hard way. */ 3473 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_map, r_next) { 3474 if (rsm->r_start == seq_out) { 3475 seq_out = rack_update_entry(tp, rack, rsm, ts, &len); 3476 rack->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next); 3477 if (len == 0) { 3478 return; 3479 } else { 3480 continue; 3481 } 3482 } 3483 if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) { 3484 /* Transmitted within this piece */ 3485 /* 3486 * Ok we must split off the front and then let the 3487 * update do the rest 3488 */ 3489 nrsm = rack_alloc(rack); 3490 if (nrsm == NULL) { 3491 #ifdef INVARIANTS 3492 panic("Ran out of memory that was preallocated? rack:%p", rack); 3493 #endif 3494 rack_update_rsm(tp, rack, rsm, ts); 3495 return; 3496 } 3497 /* 3498 * copy rsm to nrsm and then trim the front of rsm 3499 * to not include this part. 3500 */ 3501 nrsm->r_start = seq_out; 3502 nrsm->r_end = rsm->r_end; 3503 nrsm->r_rtr_cnt = rsm->r_rtr_cnt; 3504 nrsm->r_flags = rsm->r_flags; 3505 nrsm->r_sndcnt = rsm->r_sndcnt; 3506 nrsm->r_rtr_bytes = 0; 3507 for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { 3508 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; 3509 } 3510 rsm->r_end = nrsm->r_start; 3511 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_map, rsm, nrsm, r_next); 3512 if (rsm->r_in_tmap) { 3513 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 3514 nrsm->r_in_tmap = 1; 3515 } 3516 rsm->r_flags &= (~RACK_HAS_FIN); 3517 seq_out = rack_update_entry(tp, rack, nrsm, ts, &len); 3518 if (len == 0) { 3519 return; 3520 } 3521 } 3522 } 3523 /* 3524 * Hmm not found in map did they retransmit both old and on into the 3525 * new? 3526 */ 3527 if (seq_out == tp->snd_max) { 3528 goto again; 3529 } else if (SEQ_LT(seq_out, tp->snd_max)) { 3530 #ifdef INVARIANTS 3531 printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n", 3532 seq_out, len, tp->snd_una, tp->snd_max); 3533 printf("Starting Dump of all rack entries\n"); 3534 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_map, r_next) { 3535 printf("rsm:%p start:%u end:%u\n", 3536 rsm, rsm->r_start, rsm->r_end); 3537 } 3538 printf("Dump complete\n"); 3539 panic("seq_out not found rack:%p tp:%p", 3540 rack, tp); 3541 #endif 3542 } else { 3543 #ifdef INVARIANTS 3544 /* 3545 * Hmm beyond sndmax? (only if we are using the new rtt-pack 3546 * flag) 3547 */ 3548 panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p", 3549 seq_out, len, tp->snd_max, tp); 3550 #endif 3551 } 3552 } 3553 3554 /* 3555 * Record one of the RTT updates from an ack into 3556 * our sample structure. 3557 */ 3558 static void 3559 tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt) 3560 { 3561 if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) || 3562 (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) { 3563 rack->r_ctl.rack_rs.rs_rtt_lowest = rtt; 3564 } 3565 if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) || 3566 (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) { 3567 rack->r_ctl.rack_rs.rs_rtt_highest = rtt; 3568 } 3569 rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID; 3570 rack->r_ctl.rack_rs.rs_rtt_tot += rtt; 3571 rack->r_ctl.rack_rs.rs_rtt_cnt++; 3572 } 3573 3574 /* 3575 * Collect new round-trip time estimate 3576 * and update averages and current timeout. 3577 */ 3578 static void 3579 tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp) 3580 { 3581 int32_t delta; 3582 uint32_t o_srtt, o_var; 3583 int32_t rtt; 3584 3585 if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) 3586 /* No valid sample */ 3587 return; 3588 if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) { 3589 /* We are to use the lowest RTT seen in a single ack */ 3590 rtt = rack->r_ctl.rack_rs.rs_rtt_lowest; 3591 } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) { 3592 /* We are to use the highest RTT seen in a single ack */ 3593 rtt = rack->r_ctl.rack_rs.rs_rtt_highest; 3594 } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) { 3595 /* We are to use the average RTT seen in a single ack */ 3596 rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot / 3597 (uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt); 3598 } else { 3599 #ifdef INVARIANTS 3600 panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method); 3601 #endif 3602 return; 3603 } 3604 if (rtt == 0) 3605 rtt = 1; 3606 rack_log_rtt_sample(rack, rtt); 3607 o_srtt = tp->t_srtt; 3608 o_var = tp->t_rttvar; 3609 rack = (struct tcp_rack *)tp->t_fb_ptr; 3610 if (tp->t_srtt != 0) { 3611 /* 3612 * srtt is stored as fixed point with 5 bits after the 3613 * binary point (i.e., scaled by 8). The following magic is 3614 * equivalent to the smoothing algorithm in rfc793 with an 3615 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point). 3616 * Adjust rtt to origin 0. 3617 */ 3618 delta = ((rtt - 1) << TCP_DELTA_SHIFT) 3619 - (tp->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT)); 3620 3621 tp->t_srtt += delta; 3622 if (tp->t_srtt <= 0) 3623 tp->t_srtt = 1; 3624 3625 /* 3626 * We accumulate a smoothed rtt variance (actually, a 3627 * smoothed mean difference), then set the retransmit timer 3628 * to smoothed rtt + 4 times the smoothed variance. rttvar 3629 * is stored as fixed point with 4 bits after the binary 3630 * point (scaled by 16). The following is equivalent to 3631 * rfc793 smoothing with an alpha of .75 (rttvar = 3632 * rttvar*3/4 + |delta| / 4). This replaces rfc793's 3633 * wired-in beta. 3634 */ 3635 if (delta < 0) 3636 delta = -delta; 3637 delta -= tp->t_rttvar >> (TCP_RTTVAR_SHIFT - TCP_DELTA_SHIFT); 3638 tp->t_rttvar += delta; 3639 if (tp->t_rttvar <= 0) 3640 tp->t_rttvar = 1; 3641 if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar) 3642 tp->t_rttbest = tp->t_srtt + tp->t_rttvar; 3643 } else { 3644 /* 3645 * No rtt measurement yet - use the unsmoothed rtt. Set the 3646 * variance to half the rtt (so our first retransmit happens 3647 * at 3*rtt). 3648 */ 3649 tp->t_srtt = rtt << TCP_RTT_SHIFT; 3650 tp->t_rttvar = rtt << (TCP_RTTVAR_SHIFT - 1); 3651 tp->t_rttbest = tp->t_srtt + tp->t_rttvar; 3652 } 3653 TCPSTAT_INC(tcps_rttupdated); 3654 rack_log_rtt_upd(tp, rack, rtt, o_srtt, o_var); 3655 tp->t_rttupdated++; 3656 #ifdef NETFLIX_STATS 3657 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt)); 3658 #endif 3659 tp->t_rxtshift = 0; 3660 3661 /* 3662 * the retransmit should happen at rtt + 4 * rttvar. Because of the 3663 * way we do the smoothing, srtt and rttvar will each average +1/2 3664 * tick of bias. When we compute the retransmit timer, we want 1/2 3665 * tick of rounding and 1 extra tick because of +-1/2 tick 3666 * uncertainty in the firing of the timer. The bias will give us 3667 * exactly the 1.5 tick we need. But, because the bias is 3668 * statistical, we have to test that we don't drop below the minimum 3669 * feasible timer (which is 2 ticks). 3670 */ 3671 TCPT_RANGESET(tp->t_rxtcur, TCP_REXMTVAL(tp), 3672 max(MSEC_2_TICKS(rack_rto_min), rtt + 2), MSEC_2_TICKS(rack_rto_max)); 3673 tp->t_softerror = 0; 3674 } 3675 3676 static void 3677 rack_earlier_retran(struct tcpcb *tp, struct rack_sendmap *rsm, 3678 uint32_t t, uint32_t cts) 3679 { 3680 /* 3681 * For this RSM, we acknowledged the data from a previous 3682 * transmission, not the last one we made. This means we did a false 3683 * retransmit. 3684 */ 3685 struct tcp_rack *rack; 3686 3687 if (rsm->r_flags & RACK_HAS_FIN) { 3688 /* 3689 * The sending of the FIN often is multiple sent when we 3690 * have everything outstanding ack'd. We ignore this case 3691 * since its over now. 3692 */ 3693 return; 3694 } 3695 if (rsm->r_flags & RACK_TLP) { 3696 /* 3697 * We expect TLP's to have this occur. 3698 */ 3699 return; 3700 } 3701 rack = (struct tcp_rack *)tp->t_fb_ptr; 3702 /* should we undo cc changes and exit recovery? */ 3703 if (IN_RECOVERY(tp->t_flags)) { 3704 if (rack->r_ctl.rc_rsm_start == rsm->r_start) { 3705 /* 3706 * Undo what we ratched down and exit recovery if 3707 * possible 3708 */ 3709 EXIT_RECOVERY(tp->t_flags); 3710 tp->snd_recover = tp->snd_una; 3711 if (rack->r_ctl.rc_cwnd_at > tp->snd_cwnd) 3712 tp->snd_cwnd = rack->r_ctl.rc_cwnd_at; 3713 if (rack->r_ctl.rc_ssthresh_at > tp->snd_ssthresh) 3714 tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at; 3715 } 3716 } 3717 if (rsm->r_flags & RACK_WAS_SACKPASS) { 3718 /* 3719 * We retransmitted based on a sack and the earlier 3720 * retransmission ack'd it - re-ordering is occuring. 3721 */ 3722 counter_u64_add(rack_reorder_seen, 1); 3723 rack->r_ctl.rc_reorder_ts = cts; 3724 } 3725 counter_u64_add(rack_badfr, 1); 3726 counter_u64_add(rack_badfr_bytes, (rsm->r_end - rsm->r_start)); 3727 } 3728 3729 3730 static int 3731 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack, 3732 struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type) 3733 { 3734 int32_t i; 3735 uint32_t t; 3736 3737 if (rsm->r_flags & RACK_ACKED) 3738 /* Already done */ 3739 return (0); 3740 3741 3742 if ((rsm->r_rtr_cnt == 1) || 3743 ((ack_type == CUM_ACKED) && 3744 (to->to_flags & TOF_TS) && 3745 (to->to_tsecr) && 3746 (rsm->r_tim_lastsent[rsm->r_rtr_cnt - 1] == to->to_tsecr)) 3747 ) { 3748 /* 3749 * We will only find a matching timestamp if its cum-acked. 3750 * But if its only one retransmission its for-sure matching 3751 * :-) 3752 */ 3753 t = cts - rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]; 3754 if ((int)t <= 0) 3755 t = 1; 3756 if (!tp->t_rttlow || tp->t_rttlow > t) 3757 tp->t_rttlow = t; 3758 if (!rack->r_ctl.rc_rack_min_rtt || 3759 SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) { 3760 rack->r_ctl.rc_rack_min_rtt = t; 3761 if (rack->r_ctl.rc_rack_min_rtt == 0) { 3762 rack->r_ctl.rc_rack_min_rtt = 1; 3763 } 3764 } 3765 tcp_rack_xmit_timer(rack, TCP_TS_TO_TICKS(t) + 1); 3766 if ((rsm->r_flags & RACK_TLP) && 3767 (!IN_RECOVERY(tp->t_flags))) { 3768 /* Segment was a TLP and our retrans matched */ 3769 if (rack->r_ctl.rc_tlp_cwnd_reduce) { 3770 rack->r_ctl.rc_rsm_start = tp->snd_max; 3771 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd; 3772 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh; 3773 rack_cong_signal(tp, NULL, CC_NDUPACK); 3774 /* 3775 * When we enter recovery we need to assure 3776 * we send one packet. 3777 */ 3778 rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; 3779 } else 3780 rack->r_ctl.rc_tlp_rtx_out = 0; 3781 } 3782 if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) { 3783 /* New more recent rack_tmit_time */ 3784 rack->r_ctl.rc_rack_tmit_time = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]; 3785 rack->rc_rack_rtt = t; 3786 } 3787 return (1); 3788 } 3789 /* 3790 * We clear the soft/rxtshift since we got an ack. 3791 * There is no assurance we will call the commit() function 3792 * so we need to clear these to avoid incorrect handling. 3793 */ 3794 tp->t_rxtshift = 0; 3795 tp->t_softerror = 0; 3796 if ((to->to_flags & TOF_TS) && 3797 (ack_type == CUM_ACKED) && 3798 (to->to_tsecr) && 3799 ((rsm->r_flags & (RACK_DEFERRED | RACK_OVERMAX)) == 0)) { 3800 /* 3801 * Now which timestamp does it match? In this block the ACK 3802 * must be coming from a previous transmission. 3803 */ 3804 for (i = 0; i < rsm->r_rtr_cnt; i++) { 3805 if (rsm->r_tim_lastsent[i] == to->to_tsecr) { 3806 t = cts - rsm->r_tim_lastsent[i]; 3807 if ((int)t <= 0) 3808 t = 1; 3809 if ((i + 1) < rsm->r_rtr_cnt) { 3810 /* Likely */ 3811 rack_earlier_retran(tp, rsm, t, cts); 3812 } 3813 if (!tp->t_rttlow || tp->t_rttlow > t) 3814 tp->t_rttlow = t; 3815 if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) { 3816 rack->r_ctl.rc_rack_min_rtt = t; 3817 if (rack->r_ctl.rc_rack_min_rtt == 0) { 3818 rack->r_ctl.rc_rack_min_rtt = 1; 3819 } 3820 } 3821 /* 3822 * Note the following calls to 3823 * tcp_rack_xmit_timer() are being commented 3824 * out for now. They give us no more accuracy 3825 * and often lead to a wrong choice. We have 3826 * enough samples that have not been 3827 * retransmitted. I leave the commented out 3828 * code in here in case in the future we 3829 * decide to add it back (though I can't forsee 3830 * doing that). That way we will easily see 3831 * where they need to be placed. 3832 */ 3833 if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, 3834 rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) { 3835 /* New more recent rack_tmit_time */ 3836 rack->r_ctl.rc_rack_tmit_time = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]; 3837 rack->rc_rack_rtt = t; 3838 } 3839 return (1); 3840 } 3841 } 3842 goto ts_not_found; 3843 } else { 3844 /* 3845 * Ok its a SACK block that we retransmitted. or a windows 3846 * machine without timestamps. We can tell nothing from the 3847 * time-stamp since its not there or the time the peer last 3848 * recieved a segment that moved forward its cum-ack point. 3849 */ 3850 ts_not_found: 3851 i = rsm->r_rtr_cnt - 1; 3852 t = cts - rsm->r_tim_lastsent[i]; 3853 if ((int)t <= 0) 3854 t = 1; 3855 if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) { 3856 /* 3857 * We retransmitted and the ack came back in less 3858 * than the smallest rtt we have observed. We most 3859 * likey did an improper retransmit as outlined in 3860 * 4.2 Step 3 point 2 in the rack-draft. 3861 */ 3862 i = rsm->r_rtr_cnt - 2; 3863 t = cts - rsm->r_tim_lastsent[i]; 3864 rack_earlier_retran(tp, rsm, t, cts); 3865 } else if (rack->r_ctl.rc_rack_min_rtt) { 3866 /* 3867 * We retransmitted it and the retransmit did the 3868 * job. 3869 */ 3870 if (!rack->r_ctl.rc_rack_min_rtt || 3871 SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) { 3872 rack->r_ctl.rc_rack_min_rtt = t; 3873 if (rack->r_ctl.rc_rack_min_rtt == 0) { 3874 rack->r_ctl.rc_rack_min_rtt = 1; 3875 } 3876 } 3877 if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, rsm->r_tim_lastsent[i])) { 3878 /* New more recent rack_tmit_time */ 3879 rack->r_ctl.rc_rack_tmit_time = rsm->r_tim_lastsent[i]; 3880 rack->rc_rack_rtt = t; 3881 } 3882 return (1); 3883 } 3884 } 3885 return (0); 3886 } 3887 3888 /* 3889 * Mark the SACK_PASSED flag on all entries prior to rsm send wise. 3890 */ 3891 static void 3892 rack_log_sack_passed(struct tcpcb *tp, 3893 struct tcp_rack *rack, struct rack_sendmap *rsm) 3894 { 3895 struct rack_sendmap *nrsm; 3896 uint32_t ts; 3897 int32_t idx; 3898 3899 idx = rsm->r_rtr_cnt - 1; 3900 ts = rsm->r_tim_lastsent[idx]; 3901 nrsm = rsm; 3902 TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap, 3903 rack_head, r_tnext) { 3904 if (nrsm == rsm) { 3905 /* Skip orginal segment he is acked */ 3906 continue; 3907 } 3908 if (nrsm->r_flags & RACK_ACKED) { 3909 /* Skip ack'd segments */ 3910 continue; 3911 } 3912 idx = nrsm->r_rtr_cnt - 1; 3913 if (ts == nrsm->r_tim_lastsent[idx]) { 3914 /* 3915 * For this case lets use seq no, if we sent in a 3916 * big block (TSO) we would have a bunch of segments 3917 * sent at the same time. 3918 * 3919 * We would only get a report if its SEQ is earlier. 3920 * If we have done multiple retransmits the times 3921 * would not be equal. 3922 */ 3923 if (SEQ_LT(nrsm->r_start, rsm->r_start)) { 3924 nrsm->r_flags |= RACK_SACK_PASSED; 3925 nrsm->r_flags &= ~RACK_WAS_SACKPASS; 3926 } 3927 } else { 3928 /* 3929 * Here they were sent at different times, not a big 3930 * block. Since we transmitted this one later and 3931 * see it sack'd then this must also be missing (or 3932 * we would have gotten a sack block for it) 3933 */ 3934 nrsm->r_flags |= RACK_SACK_PASSED; 3935 nrsm->r_flags &= ~RACK_WAS_SACKPASS; 3936 } 3937 } 3938 } 3939 3940 static uint32_t 3941 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack, 3942 struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts) 3943 { 3944 int32_t idx; 3945 int32_t times = 0; 3946 uint32_t start, end, changed = 0; 3947 struct rack_sendmap *rsm, *nrsm; 3948 int32_t used_ref = 1; 3949 3950 start = sack->start; 3951 end = sack->end; 3952 rsm = *prsm; 3953 if (rsm && SEQ_LT(start, rsm->r_start)) { 3954 TAILQ_FOREACH_REVERSE_FROM(rsm, &rack->r_ctl.rc_map, rack_head, r_next) { 3955 if (SEQ_GEQ(start, rsm->r_start) && 3956 SEQ_LT(start, rsm->r_end)) { 3957 goto do_rest_ofb; 3958 } 3959 } 3960 } 3961 if (rsm == NULL) { 3962 start_at_beginning: 3963 rsm = NULL; 3964 used_ref = 0; 3965 } 3966 /* First lets locate the block where this guy is */ 3967 TAILQ_FOREACH_FROM(rsm, &rack->r_ctl.rc_map, r_next) { 3968 if (SEQ_GEQ(start, rsm->r_start) && 3969 SEQ_LT(start, rsm->r_end)) { 3970 break; 3971 } 3972 } 3973 do_rest_ofb: 3974 if (rsm == NULL) { 3975 /* 3976 * This happens when we get duplicate sack blocks with the 3977 * same end. For example SACK 4: 100 SACK 3: 100 The sort 3978 * will not change there location so we would just start at 3979 * the end of the first one and get lost. 3980 */ 3981 if (tp->t_flags & TF_SENTFIN) { 3982 /* 3983 * Check to see if we have not logged the FIN that 3984 * went out. 3985 */ 3986 nrsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_map, rack_sendmap, r_next); 3987 if (nrsm && (nrsm->r_end + 1) == tp->snd_max) { 3988 /* 3989 * Ok we did not get the FIN logged. 3990 */ 3991 nrsm->r_end++; 3992 rsm = nrsm; 3993 goto do_rest_ofb; 3994 } 3995 } 3996 if (times == 1) { 3997 #ifdef INVARIANTS 3998 panic("tp:%p rack:%p sack:%p to:%p prsm:%p", 3999 tp, rack, sack, to, prsm); 4000 #else 4001 goto out; 4002 #endif 4003 } 4004 times++; 4005 counter_u64_add(rack_sack_proc_restart, 1); 4006 goto start_at_beginning; 4007 } 4008 /* Ok we have an ACK for some piece of rsm */ 4009 if (rsm->r_start != start) { 4010 /* 4011 * Need to split this in two pieces the before and after. 4012 */ 4013 nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT); 4014 if (nrsm == NULL) { 4015 /* 4016 * failed XXXrrs what can we do but loose the sack 4017 * info? 4018 */ 4019 goto out; 4020 } 4021 nrsm->r_start = start; 4022 nrsm->r_rtr_bytes = 0; 4023 nrsm->r_end = rsm->r_end; 4024 nrsm->r_rtr_cnt = rsm->r_rtr_cnt; 4025 nrsm->r_flags = rsm->r_flags; 4026 nrsm->r_sndcnt = rsm->r_sndcnt; 4027 for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { 4028 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; 4029 } 4030 rsm->r_end = nrsm->r_start; 4031 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_map, rsm, nrsm, r_next); 4032 if (rsm->r_in_tmap) { 4033 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 4034 nrsm->r_in_tmap = 1; 4035 } 4036 rsm->r_flags &= (~RACK_HAS_FIN); 4037 rsm = nrsm; 4038 } 4039 if (SEQ_GEQ(end, rsm->r_end)) { 4040 /* 4041 * The end of this block is either beyond this guy or right 4042 * at this guy. 4043 */ 4044 4045 if ((rsm->r_flags & RACK_ACKED) == 0) { 4046 rack_update_rtt(tp, rack, rsm, to, cts, SACKED); 4047 changed += (rsm->r_end - rsm->r_start); 4048 rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start); 4049 rack_log_sack_passed(tp, rack, rsm); 4050 /* Is Reordering occuring? */ 4051 if (rsm->r_flags & RACK_SACK_PASSED) { 4052 counter_u64_add(rack_reorder_seen, 1); 4053 rack->r_ctl.rc_reorder_ts = cts; 4054 } 4055 rsm->r_flags |= RACK_ACKED; 4056 rsm->r_flags &= ~RACK_TLP; 4057 if (rsm->r_in_tmap) { 4058 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext); 4059 rsm->r_in_tmap = 0; 4060 } 4061 } 4062 if (end == rsm->r_end) { 4063 /* This block only - done */ 4064 goto out; 4065 } 4066 /* There is more not coverend by this rsm move on */ 4067 start = rsm->r_end; 4068 nrsm = TAILQ_NEXT(rsm, r_next); 4069 rsm = nrsm; 4070 times = 0; 4071 goto do_rest_ofb; 4072 } 4073 /* Ok we need to split off this one at the tail */ 4074 nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT); 4075 if (nrsm == NULL) { 4076 /* failed rrs what can we do but loose the sack info? */ 4077 goto out; 4078 } 4079 /* Clone it */ 4080 nrsm->r_start = end; 4081 nrsm->r_end = rsm->r_end; 4082 nrsm->r_rtr_bytes = 0; 4083 nrsm->r_rtr_cnt = rsm->r_rtr_cnt; 4084 nrsm->r_flags = rsm->r_flags; 4085 nrsm->r_sndcnt = rsm->r_sndcnt; 4086 for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { 4087 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; 4088 } 4089 /* The sack block does not cover this guy fully */ 4090 rsm->r_flags &= (~RACK_HAS_FIN); 4091 rsm->r_end = end; 4092 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_map, rsm, nrsm, r_next); 4093 if (rsm->r_in_tmap) { 4094 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 4095 nrsm->r_in_tmap = 1; 4096 } 4097 if (rsm->r_flags & RACK_ACKED) { 4098 /* Been here done that */ 4099 goto out; 4100 } 4101 rack_update_rtt(tp, rack, rsm, to, cts, SACKED); 4102 changed += (rsm->r_end - rsm->r_start); 4103 rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start); 4104 rack_log_sack_passed(tp, rack, rsm); 4105 /* Is Reordering occuring? */ 4106 if (rsm->r_flags & RACK_SACK_PASSED) { 4107 counter_u64_add(rack_reorder_seen, 1); 4108 rack->r_ctl.rc_reorder_ts = cts; 4109 } 4110 rsm->r_flags |= RACK_ACKED; 4111 rsm->r_flags &= ~RACK_TLP; 4112 if (rsm->r_in_tmap) { 4113 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext); 4114 rsm->r_in_tmap = 0; 4115 } 4116 out: 4117 if (used_ref == 0) { 4118 counter_u64_add(rack_sack_proc_all, 1); 4119 } else { 4120 counter_u64_add(rack_sack_proc_short, 1); 4121 } 4122 /* Save off where we last were */ 4123 if (rsm) 4124 rack->r_ctl.rc_sacklast = TAILQ_NEXT(rsm, r_next); 4125 else 4126 rack->r_ctl.rc_sacklast = NULL; 4127 *prsm = rsm; 4128 return (changed); 4129 } 4130 4131 static void inline 4132 rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack) 4133 { 4134 struct rack_sendmap *tmap; 4135 4136 tmap = NULL; 4137 while (rsm && (rsm->r_flags & RACK_ACKED)) { 4138 /* Its no longer sacked, mark it so */ 4139 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); 4140 #ifdef INVARIANTS 4141 if (rsm->r_in_tmap) { 4142 panic("rack:%p rsm:%p flags:0x%x in tmap?", 4143 rack, rsm, rsm->r_flags); 4144 } 4145 #endif 4146 rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS); 4147 /* Rebuild it into our tmap */ 4148 if (tmap == NULL) { 4149 TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext); 4150 tmap = rsm; 4151 } else { 4152 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext); 4153 tmap = rsm; 4154 } 4155 tmap->r_in_tmap = 1; 4156 rsm = TAILQ_NEXT(rsm, r_next); 4157 } 4158 /* 4159 * Now lets possibly clear the sack filter so we start 4160 * recognizing sacks that cover this area. 4161 */ 4162 if (rack_use_sack_filter) 4163 sack_filter_clear(&rack->r_ctl.rack_sf, th_ack); 4164 4165 } 4166 4167 static void 4168 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th) 4169 { 4170 uint32_t changed, last_seq, entered_recovery = 0; 4171 struct tcp_rack *rack; 4172 struct rack_sendmap *rsm; 4173 struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1]; 4174 register uint32_t th_ack; 4175 int32_t i, j, k, num_sack_blks = 0; 4176 uint32_t cts, acked, ack_point, sack_changed = 0; 4177 4178 INP_WLOCK_ASSERT(tp->t_inpcb); 4179 if (th->th_flags & TH_RST) { 4180 /* We don't log resets */ 4181 return; 4182 } 4183 rack = (struct tcp_rack *)tp->t_fb_ptr; 4184 cts = tcp_ts_getticks(); 4185 rsm = TAILQ_FIRST(&rack->r_ctl.rc_map); 4186 changed = 0; 4187 th_ack = th->th_ack; 4188 4189 if (SEQ_GT(th_ack, tp->snd_una)) { 4190 rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__); 4191 tp->t_acktime = ticks; 4192 } 4193 if (rsm && SEQ_GT(th_ack, rsm->r_start)) 4194 changed = th_ack - rsm->r_start; 4195 if (changed) { 4196 /* 4197 * The ACK point is advancing to th_ack, we must drop off 4198 * the packets in the rack log and calculate any eligble 4199 * RTT's. 4200 */ 4201 rack->r_wanted_output++; 4202 more: 4203 rsm = TAILQ_FIRST(&rack->r_ctl.rc_map); 4204 if (rsm == NULL) { 4205 if ((th_ack - 1) == tp->iss) { 4206 /* 4207 * For the SYN incoming case we will not 4208 * have called tcp_output for the sending of 4209 * the SYN, so there will be no map. All 4210 * other cases should probably be a panic. 4211 */ 4212 goto proc_sack; 4213 } 4214 if (tp->t_flags & TF_SENTFIN) { 4215 /* if we send a FIN we will not hav a map */ 4216 goto proc_sack; 4217 } 4218 #ifdef INVARIANTS 4219 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", 4220 tp, 4221 th, tp->t_state, rack, 4222 tp->snd_una, tp->snd_max, tp->snd_nxt, changed); 4223 #endif 4224 goto proc_sack; 4225 } 4226 if (SEQ_LT(th_ack, rsm->r_start)) { 4227 /* Huh map is missing this */ 4228 #ifdef INVARIANTS 4229 printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n", 4230 rsm->r_start, 4231 th_ack, tp->t_state, rack->r_state); 4232 #endif 4233 goto proc_sack; 4234 } 4235 rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED); 4236 /* Now do we consume the whole thing? */ 4237 if (SEQ_GEQ(th_ack, rsm->r_end)) { 4238 /* Its all consumed. */ 4239 uint32_t left; 4240 4241 rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; 4242 rsm->r_rtr_bytes = 0; 4243 TAILQ_REMOVE(&rack->r_ctl.rc_map, rsm, r_next); 4244 if (rsm->r_in_tmap) { 4245 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext); 4246 rsm->r_in_tmap = 0; 4247 } 4248 if (rack->r_ctl.rc_next == rsm) { 4249 /* scoot along the marker */ 4250 rack->r_ctl.rc_next = TAILQ_FIRST(&rack->r_ctl.rc_map); 4251 } 4252 if (rsm->r_flags & RACK_ACKED) { 4253 /* 4254 * It was acked on the scoreboard -- remove 4255 * it from total 4256 */ 4257 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); 4258 } else if (rsm->r_flags & RACK_SACK_PASSED) { 4259 /* 4260 * There are acked segments ACKED on the 4261 * scoreboard further up. We are seeing 4262 * reordering. 4263 */ 4264 counter_u64_add(rack_reorder_seen, 1); 4265 rsm->r_flags |= RACK_ACKED; 4266 rack->r_ctl.rc_reorder_ts = cts; 4267 } 4268 left = th_ack - rsm->r_end; 4269 if (rsm->r_rtr_cnt > 1) { 4270 /* 4271 * Technically we should make r_rtr_cnt be 4272 * monotonicly increasing and just mod it to 4273 * the timestamp it is replacing.. that way 4274 * we would have the last 3 retransmits. Now 4275 * rc_loss_count will be wrong if we 4276 * retransmit something more than 2 times in 4277 * recovery :( 4278 */ 4279 rack->r_ctl.rc_loss_count += (rsm->r_rtr_cnt - 1); 4280 } 4281 /* Free back to zone */ 4282 rack_free(rack, rsm); 4283 if (left) { 4284 goto more; 4285 } 4286 goto proc_sack; 4287 } 4288 if (rsm->r_flags & RACK_ACKED) { 4289 /* 4290 * It was acked on the scoreboard -- remove it from 4291 * total for the part being cum-acked. 4292 */ 4293 rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start); 4294 } 4295 rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; 4296 rsm->r_rtr_bytes = 0; 4297 rsm->r_start = th_ack; 4298 } 4299 proc_sack: 4300 /* Check for reneging */ 4301 rsm = TAILQ_FIRST(&rack->r_ctl.rc_map); 4302 if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) { 4303 /* 4304 * The peer has moved snd_una up to 4305 * the edge of this send, i.e. one 4306 * that it had previously acked. The only 4307 * way that can be true if the peer threw 4308 * away data (space issues) that it had 4309 * previously sacked (else it would have 4310 * given us snd_una up to (rsm->r_end). 4311 * We need to undo the acked markings here. 4312 * 4313 * Note we have to look to make sure th_ack is 4314 * our rsm->r_start in case we get an old ack 4315 * where th_ack is behind snd_una. 4316 */ 4317 rack_peer_reneges(rack, rsm, th->th_ack); 4318 } 4319 if ((to->to_flags & TOF_SACK) == 0) { 4320 /* We are done nothing left to log */ 4321 goto out; 4322 } 4323 rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_map, rack_sendmap, r_next); 4324 if (rsm) { 4325 last_seq = rsm->r_end; 4326 } else { 4327 last_seq = tp->snd_max; 4328 } 4329 /* Sack block processing */ 4330 if (SEQ_GT(th_ack, tp->snd_una)) 4331 ack_point = th_ack; 4332 else 4333 ack_point = tp->snd_una; 4334 for (i = 0; i < to->to_nsacks; i++) { 4335 bcopy((to->to_sacks + i * TCPOLEN_SACK), 4336 &sack, sizeof(sack)); 4337 sack.start = ntohl(sack.start); 4338 sack.end = ntohl(sack.end); 4339 if (SEQ_GT(sack.end, sack.start) && 4340 SEQ_GT(sack.start, ack_point) && 4341 SEQ_LT(sack.start, tp->snd_max) && 4342 SEQ_GT(sack.end, ack_point) && 4343 SEQ_LEQ(sack.end, tp->snd_max)) { 4344 if ((rack->r_ctl.rc_num_maps_alloced > rack_sack_block_limit) && 4345 (SEQ_LT(sack.end, last_seq)) && 4346 ((sack.end - sack.start) < (tp->t_maxseg / 8))) { 4347 /* 4348 * Not the last piece and its smaller than 4349 * 1/8th of a MSS. We ignore this. 4350 */ 4351 counter_u64_add(rack_runt_sacks, 1); 4352 continue; 4353 } 4354 sack_blocks[num_sack_blks] = sack; 4355 num_sack_blks++; 4356 #ifdef NETFLIX_STATS 4357 } else if (SEQ_LEQ(sack.start, th_ack) && 4358 SEQ_LEQ(sack.end, th_ack)) { 4359 /* 4360 * Its a D-SACK block. 4361 */ 4362 tcp_record_dsack(sack.start, sack.end); 4363 #endif 4364 } 4365 4366 } 4367 if (num_sack_blks == 0) 4368 goto out; 4369 /* 4370 * Sort the SACK blocks so we can update the rack scoreboard with 4371 * just one pass. 4372 */ 4373 if (rack_use_sack_filter) { 4374 num_sack_blks = sack_filter_blks(&rack->r_ctl.rack_sf, sack_blocks, num_sack_blks, th->th_ack); 4375 } 4376 if (num_sack_blks < 2) { 4377 goto do_sack_work; 4378 } 4379 /* Sort the sacks */ 4380 for (i = 0; i < num_sack_blks; i++) { 4381 for (j = i + 1; j < num_sack_blks; j++) { 4382 if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) { 4383 sack = sack_blocks[i]; 4384 sack_blocks[i] = sack_blocks[j]; 4385 sack_blocks[j] = sack; 4386 } 4387 } 4388 } 4389 /* 4390 * Now are any of the sack block ends the same (yes some 4391 * implememtations send these)? 4392 */ 4393 again: 4394 if (num_sack_blks > 1) { 4395 for (i = 0; i < num_sack_blks; i++) { 4396 for (j = i + 1; j < num_sack_blks; j++) { 4397 if (sack_blocks[i].end == sack_blocks[j].end) { 4398 /* 4399 * Ok these two have the same end we 4400 * want the smallest end and then 4401 * throw away the larger and start 4402 * again. 4403 */ 4404 if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) { 4405 /* 4406 * The second block covers 4407 * more area use that 4408 */ 4409 sack_blocks[i].start = sack_blocks[j].start; 4410 } 4411 /* 4412 * Now collapse out the dup-sack and 4413 * lower the count 4414 */ 4415 for (k = (j + 1); k < num_sack_blks; k++) { 4416 sack_blocks[j].start = sack_blocks[k].start; 4417 sack_blocks[j].end = sack_blocks[k].end; 4418 j++; 4419 } 4420 num_sack_blks--; 4421 goto again; 4422 } 4423 } 4424 } 4425 } 4426 do_sack_work: 4427 rsm = rack->r_ctl.rc_sacklast; 4428 for (i = 0; i < num_sack_blks; i++) { 4429 acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts); 4430 if (acked) { 4431 rack->r_wanted_output++; 4432 changed += acked; 4433 sack_changed += acked; 4434 } 4435 } 4436 out: 4437 if (changed) { 4438 /* Something changed cancel the rack timer */ 4439 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 4440 } 4441 if ((sack_changed) && (!IN_RECOVERY(tp->t_flags))) { 4442 /* 4443 * Ok we have a high probability that we need to go in to 4444 * recovery since we have data sack'd 4445 */ 4446 struct rack_sendmap *rsm; 4447 uint32_t tsused; 4448 4449 tsused = tcp_ts_getticks(); 4450 rsm = tcp_rack_output(tp, rack, tsused); 4451 if (rsm) { 4452 /* Enter recovery */ 4453 rack->r_ctl.rc_rsm_start = rsm->r_start; 4454 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd; 4455 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh; 4456 entered_recovery = 1; 4457 rack_cong_signal(tp, NULL, CC_NDUPACK); 4458 /* 4459 * When we enter recovery we need to assure we send 4460 * one packet. 4461 */ 4462 rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; 4463 rack->r_timer_override = 1; 4464 } 4465 } 4466 if (IN_RECOVERY(tp->t_flags) && (entered_recovery == 0)) { 4467 /* Deal with changed an PRR here (in recovery only) */ 4468 uint32_t pipe, snd_una; 4469 4470 rack->r_ctl.rc_prr_delivered += changed; 4471 /* Compute prr_sndcnt */ 4472 if (SEQ_GT(tp->snd_una, th_ack)) { 4473 snd_una = tp->snd_una; 4474 } else { 4475 snd_una = th_ack; 4476 } 4477 pipe = ((tp->snd_max - snd_una) - rack->r_ctl.rc_sacked) + rack->r_ctl.rc_holes_rxt; 4478 if (pipe > tp->snd_ssthresh) { 4479 long sndcnt; 4480 4481 sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh; 4482 if (rack->r_ctl.rc_prr_recovery_fs > 0) 4483 sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs; 4484 else { 4485 rack->r_ctl.rc_prr_sndcnt = 0; 4486 sndcnt = 0; 4487 } 4488 sndcnt++; 4489 if (sndcnt > (long)rack->r_ctl.rc_prr_out) 4490 sndcnt -= rack->r_ctl.rc_prr_out; 4491 else 4492 sndcnt = 0; 4493 rack->r_ctl.rc_prr_sndcnt = sndcnt; 4494 } else { 4495 uint32_t limit; 4496 4497 if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out) 4498 limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out); 4499 else 4500 limit = 0; 4501 if (changed > limit) 4502 limit = changed; 4503 limit += tp->t_maxseg; 4504 if (tp->snd_ssthresh > pipe) { 4505 rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit); 4506 } else { 4507 rack->r_ctl.rc_prr_sndcnt = min(0, limit); 4508 } 4509 } 4510 if (rack->r_ctl.rc_prr_sndcnt >= tp->t_maxseg) { 4511 rack->r_timer_override = 1; 4512 } 4513 } 4514 } 4515 4516 /* 4517 * Return value of 1, we do not need to call rack_process_data(). 4518 * return value of 0, rack_process_data can be called. 4519 * For ret_val if its 0 the TCP is locked, if its non-zero 4520 * its unlocked and probably unsafe to touch the TCB. 4521 */ 4522 static int 4523 rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so, 4524 struct tcpcb *tp, struct tcpopt *to, 4525 uint32_t tiwin, int32_t tlen, 4526 int32_t * ofia, int32_t thflags, int32_t * ret_val) 4527 { 4528 int32_t ourfinisacked = 0; 4529 int32_t nsegs, acked_amount; 4530 int32_t acked; 4531 struct mbuf *mfree; 4532 struct tcp_rack *rack; 4533 int32_t recovery = 0; 4534 4535 rack = (struct tcp_rack *)tp->t_fb_ptr; 4536 if (SEQ_GT(th->th_ack, tp->snd_max)) { 4537 rack_do_dropafterack(m, tp, th, thflags, tlen, ret_val); 4538 return (1); 4539 } 4540 if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) { 4541 rack_log_ack(tp, to, th); 4542 } 4543 if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) { 4544 /* 4545 * Old ack, behind (or duplicate to) the last one rcv'd 4546 * Note: Should mark reordering is occuring! We should also 4547 * look for sack blocks arriving e.g. ack 1, 4-4 then ack 1, 4548 * 3-3, 4-4 would be reording. As well as ack 1, 3-3 <no 4549 * retran and> ack 3 4550 */ 4551 return (0); 4552 } 4553 /* 4554 * If we reach this point, ACK is not a duplicate, i.e., it ACKs 4555 * something we sent. 4556 */ 4557 if (tp->t_flags & TF_NEEDSYN) { 4558 /* 4559 * T/TCP: Connection was half-synchronized, and our SYN has 4560 * been ACK'd (so connection is now fully synchronized). Go 4561 * to non-starred state, increment snd_una for ACK of SYN, 4562 * and check if we can do window scaling. 4563 */ 4564 tp->t_flags &= ~TF_NEEDSYN; 4565 tp->snd_una++; 4566 /* Do window scaling? */ 4567 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == 4568 (TF_RCVD_SCALE | TF_REQ_SCALE)) { 4569 tp->rcv_scale = tp->request_r_scale; 4570 /* Send window already scaled. */ 4571 } 4572 } 4573 nsegs = max(1, m->m_pkthdr.lro_nsegs); 4574 INP_WLOCK_ASSERT(tp->t_inpcb); 4575 4576 acked = BYTES_THIS_ACK(tp, th); 4577 TCPSTAT_ADD(tcps_rcvackpack, nsegs); 4578 TCPSTAT_ADD(tcps_rcvackbyte, acked); 4579 4580 /* 4581 * If we just performed our first retransmit, and the ACK arrives 4582 * within our recovery window, then it was a mistake to do the 4583 * retransmit in the first place. Recover our original cwnd and 4584 * ssthresh, and proceed to transmit where we left off. 4585 */ 4586 if (tp->t_flags & TF_PREVVALID) { 4587 tp->t_flags &= ~TF_PREVVALID; 4588 if (tp->t_rxtshift == 1 && 4589 (int)(ticks - tp->t_badrxtwin) < 0) 4590 rack_cong_signal(tp, th, CC_RTO_ERR); 4591 } 4592 /* 4593 * If we have a timestamp reply, update smoothed round trip time. If 4594 * no timestamp is present but transmit timer is running and timed 4595 * sequence number was acked, update smoothed round trip time. Since 4596 * we now have an rtt measurement, cancel the timer backoff (cf., 4597 * Phil Karn's retransmit alg.). Recompute the initial retransmit 4598 * timer. 4599 * 4600 * Some boxes send broken timestamp replies during the SYN+ACK 4601 * phase, ignore timestamps of 0 or we could calculate a huge RTT 4602 * and blow up the retransmit timer. 4603 */ 4604 /* 4605 * If all outstanding data is acked, stop retransmit timer and 4606 * remember to restart (more output or persist). If there is more 4607 * data to be acked, restart retransmit timer, using current 4608 * (possibly backed-off) value. 4609 */ 4610 if (th->th_ack == tp->snd_max) { 4611 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 4612 rack->r_wanted_output++; 4613 } 4614 /* 4615 * If no data (only SYN) was ACK'd, skip rest of ACK processing. 4616 */ 4617 if (acked == 0) { 4618 if (ofia) 4619 *ofia = ourfinisacked; 4620 return (0); 4621 } 4622 if (rack->r_ctl.rc_early_recovery) { 4623 if (IN_FASTRECOVERY(tp->t_flags)) { 4624 if (SEQ_LT(th->th_ack, tp->snd_recover)) { 4625 tcp_rack_partialack(tp, th); 4626 } else { 4627 rack_post_recovery(tp, th); 4628 recovery = 1; 4629 } 4630 } 4631 } 4632 /* 4633 * Let the congestion control algorithm update congestion control 4634 * related information. This typically means increasing the 4635 * congestion window. 4636 */ 4637 rack_ack_received(tp, rack, th, nsegs, CC_ACK, recovery); 4638 SOCKBUF_LOCK(&so->so_snd); 4639 acked_amount = min(acked, (int)sbavail(&so->so_snd)); 4640 tp->snd_wnd -= acked_amount; 4641 mfree = sbcut_locked(&so->so_snd, acked_amount); 4642 if ((sbused(&so->so_snd) == 0) && 4643 (acked > acked_amount) && 4644 (tp->t_state >= TCPS_FIN_WAIT_1)) { 4645 ourfinisacked = 1; 4646 } 4647 /* NB: sowwakeup_locked() does an implicit unlock. */ 4648 sowwakeup_locked(so); 4649 m_freem(mfree); 4650 if (rack->r_ctl.rc_early_recovery == 0) { 4651 if (IN_FASTRECOVERY(tp->t_flags)) { 4652 if (SEQ_LT(th->th_ack, tp->snd_recover)) { 4653 tcp_rack_partialack(tp, th); 4654 } else { 4655 rack_post_recovery(tp, th); 4656 } 4657 } 4658 } 4659 tp->snd_una = th->th_ack; 4660 if (SEQ_GT(tp->snd_una, tp->snd_recover)) 4661 tp->snd_recover = tp->snd_una; 4662 4663 if (SEQ_LT(tp->snd_nxt, tp->snd_una)) { 4664 tp->snd_nxt = tp->snd_una; 4665 } 4666 if (tp->snd_una == tp->snd_max) { 4667 /* Nothing left outstanding */ 4668 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__); 4669 tp->t_acktime = 0; 4670 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 4671 /* Set need output so persist might get set */ 4672 rack->r_wanted_output++; 4673 if (rack_use_sack_filter) 4674 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una); 4675 if ((tp->t_state >= TCPS_FIN_WAIT_1) && 4676 (sbavail(&so->so_snd) == 0) && 4677 (tp->t_flags2 & TF2_DROP_AF_DATA)) { 4678 /* 4679 * The socket was gone and the 4680 * peer sent data, time to 4681 * reset him. 4682 */ 4683 *ret_val = 1; 4684 tp = tcp_close(tp); 4685 rack_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen); 4686 return (1); 4687 } 4688 } 4689 if (ofia) 4690 *ofia = ourfinisacked; 4691 return (0); 4692 } 4693 4694 4695 /* 4696 * Return value of 1, the TCB is unlocked and most 4697 * likely gone, return value of 0, the TCP is still 4698 * locked. 4699 */ 4700 static int 4701 rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so, 4702 struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, 4703 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 4704 { 4705 /* 4706 * Update window information. Don't look at window if no ACK: TAC's 4707 * send garbage on first SYN. 4708 */ 4709 int32_t nsegs; 4710 int32_t tfo_syn; 4711 struct tcp_rack *rack; 4712 4713 rack = (struct tcp_rack *)tp->t_fb_ptr; 4714 INP_WLOCK_ASSERT(tp->t_inpcb); 4715 nsegs = max(1, m->m_pkthdr.lro_nsegs); 4716 if ((thflags & TH_ACK) && 4717 (SEQ_LT(tp->snd_wl1, th->th_seq) || 4718 (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) || 4719 (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) { 4720 /* keep track of pure window updates */ 4721 if (tlen == 0 && 4722 tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd) 4723 TCPSTAT_INC(tcps_rcvwinupd); 4724 tp->snd_wnd = tiwin; 4725 tp->snd_wl1 = th->th_seq; 4726 tp->snd_wl2 = th->th_ack; 4727 if (tp->snd_wnd > tp->max_sndwnd) 4728 tp->max_sndwnd = tp->snd_wnd; 4729 rack->r_wanted_output++; 4730 } else if (thflags & TH_ACK) { 4731 if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) { 4732 tp->snd_wnd = tiwin; 4733 tp->snd_wl1 = th->th_seq; 4734 tp->snd_wl2 = th->th_ack; 4735 } 4736 } 4737 /* Was persist timer active and now we have window space? */ 4738 if ((rack->rc_in_persist != 0) && tp->snd_wnd) { 4739 rack_exit_persist(tp, rack); 4740 tp->snd_nxt = tp->snd_max; 4741 /* Make sure we output to start the timer */ 4742 rack->r_wanted_output++; 4743 } 4744 if (tp->t_flags2 & TF2_DROP_AF_DATA) { 4745 m_freem(m); 4746 return (0); 4747 } 4748 /* 4749 * Process segments with URG. 4750 */ 4751 if ((thflags & TH_URG) && th->th_urp && 4752 TCPS_HAVERCVDFIN(tp->t_state) == 0) { 4753 /* 4754 * This is a kludge, but if we receive and accept random 4755 * urgent pointers, we'll crash in soreceive. It's hard to 4756 * imagine someone actually wanting to send this much urgent 4757 * data. 4758 */ 4759 SOCKBUF_LOCK(&so->so_rcv); 4760 if (th->th_urp + sbavail(&so->so_rcv) > sb_max) { 4761 th->th_urp = 0; /* XXX */ 4762 thflags &= ~TH_URG; /* XXX */ 4763 SOCKBUF_UNLOCK(&so->so_rcv); /* XXX */ 4764 goto dodata; /* XXX */ 4765 } 4766 /* 4767 * If this segment advances the known urgent pointer, then 4768 * mark the data stream. This should not happen in 4769 * CLOSE_WAIT, CLOSING, LAST_ACK or TIME_WAIT STATES since a 4770 * FIN has been received from the remote side. In these 4771 * states we ignore the URG. 4772 * 4773 * According to RFC961 (Assigned Protocols), the urgent 4774 * pointer points to the last octet of urgent data. We 4775 * continue, however, to consider it to indicate the first 4776 * octet of data past the urgent section as the original 4777 * spec states (in one of two places). 4778 */ 4779 if (SEQ_GT(th->th_seq + th->th_urp, tp->rcv_up)) { 4780 tp->rcv_up = th->th_seq + th->th_urp; 4781 so->so_oobmark = sbavail(&so->so_rcv) + 4782 (tp->rcv_up - tp->rcv_nxt) - 1; 4783 if (so->so_oobmark == 0) 4784 so->so_rcv.sb_state |= SBS_RCVATMARK; 4785 sohasoutofband(so); 4786 tp->t_oobflags &= ~(TCPOOB_HAVEDATA | TCPOOB_HADDATA); 4787 } 4788 SOCKBUF_UNLOCK(&so->so_rcv); 4789 /* 4790 * Remove out of band data so doesn't get presented to user. 4791 * This can happen independent of advancing the URG pointer, 4792 * but if two URG's are pending at once, some out-of-band 4793 * data may creep in... ick. 4794 */ 4795 if (th->th_urp <= (uint32_t) tlen && 4796 !(so->so_options & SO_OOBINLINE)) { 4797 /* hdr drop is delayed */ 4798 tcp_pulloutofband(so, th, m, drop_hdrlen); 4799 } 4800 } else { 4801 /* 4802 * If no out of band data is expected, pull receive urgent 4803 * pointer along with the receive window. 4804 */ 4805 if (SEQ_GT(tp->rcv_nxt, tp->rcv_up)) 4806 tp->rcv_up = tp->rcv_nxt; 4807 } 4808 dodata: /* XXX */ 4809 INP_WLOCK_ASSERT(tp->t_inpcb); 4810 4811 /* 4812 * Process the segment text, merging it into the TCP sequencing 4813 * queue, and arranging for acknowledgment of receipt if necessary. 4814 * This process logically involves adjusting tp->rcv_wnd as data is 4815 * presented to the user (this happens in tcp_usrreq.c, case 4816 * PRU_RCVD). If a FIN has already been received on this connection 4817 * then we just ignore the text. 4818 */ 4819 tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) && 4820 IS_FASTOPEN(tp->t_flags)); 4821 if ((tlen || (thflags & TH_FIN) || tfo_syn) && 4822 TCPS_HAVERCVDFIN(tp->t_state) == 0) { 4823 tcp_seq save_start = th->th_seq; 4824 tcp_seq save_rnxt = tp->rcv_nxt; 4825 int save_tlen = tlen; 4826 4827 m_adj(m, drop_hdrlen); /* delayed header drop */ 4828 /* 4829 * Insert segment which includes th into TCP reassembly 4830 * queue with control block tp. Set thflags to whether 4831 * reassembly now includes a segment with FIN. This handles 4832 * the common case inline (segment is the next to be 4833 * received on an established connection, and the queue is 4834 * empty), avoiding linkage into and removal from the queue 4835 * and repetition of various conversions. Set DELACK for 4836 * segments received in order, but ack immediately when 4837 * segments are out of order (so fast retransmit can work). 4838 */ 4839 if (th->th_seq == tp->rcv_nxt && 4840 SEGQ_EMPTY(tp) && 4841 (TCPS_HAVEESTABLISHED(tp->t_state) || 4842 tfo_syn)) { 4843 if (DELAY_ACK(tp, tlen) || tfo_syn) { 4844 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 4845 tp->t_flags |= TF_DELACK; 4846 } else { 4847 rack->r_wanted_output++; 4848 tp->t_flags |= TF_ACKNOW; 4849 } 4850 tp->rcv_nxt += tlen; 4851 thflags = th->th_flags & TH_FIN; 4852 TCPSTAT_ADD(tcps_rcvpack, nsegs); 4853 TCPSTAT_ADD(tcps_rcvbyte, tlen); 4854 SOCKBUF_LOCK(&so->so_rcv); 4855 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) 4856 m_freem(m); 4857 else 4858 sbappendstream_locked(&so->so_rcv, m, 0); 4859 /* NB: sorwakeup_locked() does an implicit unlock. */ 4860 sorwakeup_locked(so); 4861 } else { 4862 /* 4863 * XXX: Due to the header drop above "th" is 4864 * theoretically invalid by now. Fortunately 4865 * m_adj() doesn't actually frees any mbufs when 4866 * trimming from the head. 4867 */ 4868 tcp_seq temp = save_start; 4869 thflags = tcp_reass(tp, th, &temp, &tlen, m); 4870 tp->t_flags |= TF_ACKNOW; 4871 } 4872 if (((tlen == 0) && (save_tlen > 0) && 4873 (SEQ_LT(save_start, save_rnxt)))) { 4874 /* 4875 * DSACK actually handled in the fastpath 4876 * above. 4877 */ 4878 tcp_update_sack_list(tp, save_start, save_start + save_tlen); 4879 } else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) { 4880 /* 4881 * Cleaning sackblks by using zero length 4882 * update. 4883 */ 4884 tcp_update_sack_list(tp, save_start, save_start); 4885 } else if ((tlen > 0) && (tlen >= save_tlen)) { 4886 /* Update of sackblks. */ 4887 tcp_update_sack_list(tp, save_start, save_start + save_tlen); 4888 } else if (tlen > 0) { 4889 tcp_update_sack_list(tp, save_start, save_start+tlen); 4890 } 4891 } else { 4892 m_freem(m); 4893 thflags &= ~TH_FIN; 4894 } 4895 4896 /* 4897 * If FIN is received ACK the FIN and let the user know that the 4898 * connection is closing. 4899 */ 4900 if (thflags & TH_FIN) { 4901 if (TCPS_HAVERCVDFIN(tp->t_state) == 0) { 4902 socantrcvmore(so); 4903 /* 4904 * If connection is half-synchronized (ie NEEDSYN 4905 * flag on) then delay ACK, so it may be piggybacked 4906 * when SYN is sent. Otherwise, since we received a 4907 * FIN then no more input can be expected, send ACK 4908 * now. 4909 */ 4910 if (tp->t_flags & TF_NEEDSYN) { 4911 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 4912 tp->t_flags |= TF_DELACK; 4913 } else { 4914 tp->t_flags |= TF_ACKNOW; 4915 } 4916 tp->rcv_nxt++; 4917 } 4918 switch (tp->t_state) { 4919 4920 /* 4921 * In SYN_RECEIVED and ESTABLISHED STATES enter the 4922 * CLOSE_WAIT state. 4923 */ 4924 case TCPS_SYN_RECEIVED: 4925 tp->t_starttime = ticks; 4926 /* FALLTHROUGH */ 4927 case TCPS_ESTABLISHED: 4928 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 4929 tcp_state_change(tp, TCPS_CLOSE_WAIT); 4930 break; 4931 4932 /* 4933 * If still in FIN_WAIT_1 STATE FIN has not been 4934 * acked so enter the CLOSING state. 4935 */ 4936 case TCPS_FIN_WAIT_1: 4937 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 4938 tcp_state_change(tp, TCPS_CLOSING); 4939 break; 4940 4941 /* 4942 * In FIN_WAIT_2 state enter the TIME_WAIT state, 4943 * starting the time-wait timer, turning off the 4944 * other standard timers. 4945 */ 4946 case TCPS_FIN_WAIT_2: 4947 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 4948 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 4949 tcp_twstart(tp); 4950 return (1); 4951 } 4952 } 4953 /* 4954 * Return any desired output. 4955 */ 4956 if ((tp->t_flags & TF_ACKNOW) || (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) { 4957 rack->r_wanted_output++; 4958 } 4959 INP_WLOCK_ASSERT(tp->t_inpcb); 4960 return (0); 4961 } 4962 4963 /* 4964 * Here nothing is really faster, its just that we 4965 * have broken out the fast-data path also just like 4966 * the fast-ack. 4967 */ 4968 static int 4969 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so, 4970 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 4971 uint32_t tiwin, int32_t nxt_pkt) 4972 { 4973 int32_t nsegs; 4974 int32_t newsize = 0; /* automatic sockbuf scaling */ 4975 struct tcp_rack *rack; 4976 #ifdef TCPDEBUG 4977 /* 4978 * The size of tcp_saveipgen must be the size of the max ip header, 4979 * now IPv6. 4980 */ 4981 u_char tcp_saveipgen[IP6_HDR_LEN]; 4982 struct tcphdr tcp_savetcp; 4983 short ostate = 0; 4984 4985 #endif 4986 /* 4987 * If last ACK falls within this segment's sequence numbers, record 4988 * the timestamp. NOTE that the test is modified according to the 4989 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26). 4990 */ 4991 if (__predict_false(th->th_seq != tp->rcv_nxt)) { 4992 return (0); 4993 } 4994 if (__predict_false(tp->snd_nxt != tp->snd_max)) { 4995 return (0); 4996 } 4997 if (tiwin && tiwin != tp->snd_wnd) { 4998 return (0); 4999 } 5000 if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) { 5001 return (0); 5002 } 5003 if (__predict_false((to->to_flags & TOF_TS) && 5004 (TSTMP_LT(to->to_tsval, tp->ts_recent)))) { 5005 return (0); 5006 } 5007 if (__predict_false((th->th_ack != tp->snd_una))) { 5008 return (0); 5009 } 5010 if (__predict_false(tlen > sbspace(&so->so_rcv))) { 5011 return (0); 5012 } 5013 if ((to->to_flags & TOF_TS) != 0 && 5014 SEQ_LEQ(th->th_seq, tp->last_ack_sent)) { 5015 tp->ts_recent_age = tcp_ts_getticks(); 5016 tp->ts_recent = to->to_tsval; 5017 } 5018 rack = (struct tcp_rack *)tp->t_fb_ptr; 5019 /* 5020 * This is a pure, in-sequence data packet with nothing on the 5021 * reassembly queue and we have enough buffer space to take it. 5022 */ 5023 nsegs = max(1, m->m_pkthdr.lro_nsegs); 5024 5025 5026 /* Clean receiver SACK report if present */ 5027 if (tp->rcv_numsacks) 5028 tcp_clean_sackreport(tp); 5029 TCPSTAT_INC(tcps_preddat); 5030 tp->rcv_nxt += tlen; 5031 /* 5032 * Pull snd_wl1 up to prevent seq wrap relative to th_seq. 5033 */ 5034 tp->snd_wl1 = th->th_seq; 5035 /* 5036 * Pull rcv_up up to prevent seq wrap relative to rcv_nxt. 5037 */ 5038 tp->rcv_up = tp->rcv_nxt; 5039 TCPSTAT_ADD(tcps_rcvpack, nsegs); 5040 TCPSTAT_ADD(tcps_rcvbyte, tlen); 5041 #ifdef TCPDEBUG 5042 if (so->so_options & SO_DEBUG) 5043 tcp_trace(TA_INPUT, ostate, tp, 5044 (void *)tcp_saveipgen, &tcp_savetcp, 0); 5045 #endif 5046 newsize = tcp_autorcvbuf(m, th, so, tp, tlen); 5047 5048 /* Add data to socket buffer. */ 5049 SOCKBUF_LOCK(&so->so_rcv); 5050 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 5051 m_freem(m); 5052 } else { 5053 /* 5054 * Set new socket buffer size. Give up when limit is 5055 * reached. 5056 */ 5057 if (newsize) 5058 if (!sbreserve_locked(&so->so_rcv, 5059 newsize, so, NULL)) 5060 so->so_rcv.sb_flags &= ~SB_AUTOSIZE; 5061 m_adj(m, drop_hdrlen); /* delayed header drop */ 5062 sbappendstream_locked(&so->so_rcv, m, 0); 5063 rack_calc_rwin(so, tp); 5064 } 5065 /* NB: sorwakeup_locked() does an implicit unlock. */ 5066 sorwakeup_locked(so); 5067 if (DELAY_ACK(tp, tlen)) { 5068 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 5069 tp->t_flags |= TF_DELACK; 5070 } else { 5071 tp->t_flags |= TF_ACKNOW; 5072 rack->r_wanted_output++; 5073 } 5074 if ((tp->snd_una == tp->snd_max) && rack_use_sack_filter) 5075 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una); 5076 return (1); 5077 } 5078 5079 /* 5080 * This subfunction is used to try to highly optimize the 5081 * fast path. We again allow window updates that are 5082 * in sequence to remain in the fast-path. We also add 5083 * in the __predict's to attempt to help the compiler. 5084 * Note that if we return a 0, then we can *not* process 5085 * it and the caller should push the packet into the 5086 * slow-path. 5087 */ 5088 static int 5089 rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so, 5090 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 5091 uint32_t tiwin, int32_t nxt_pkt, uint32_t cts) 5092 { 5093 int32_t acked; 5094 int32_t nsegs; 5095 5096 #ifdef TCPDEBUG 5097 /* 5098 * The size of tcp_saveipgen must be the size of the max ip header, 5099 * now IPv6. 5100 */ 5101 u_char tcp_saveipgen[IP6_HDR_LEN]; 5102 struct tcphdr tcp_savetcp; 5103 short ostate = 0; 5104 5105 #endif 5106 struct tcp_rack *rack; 5107 5108 if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) { 5109 /* Old ack, behind (or duplicate to) the last one rcv'd */ 5110 return (0); 5111 } 5112 if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) { 5113 /* Above what we have sent? */ 5114 return (0); 5115 } 5116 if (__predict_false(tp->snd_nxt != tp->snd_max)) { 5117 /* We are retransmitting */ 5118 return (0); 5119 } 5120 if (__predict_false(tiwin == 0)) { 5121 /* zero window */ 5122 return (0); 5123 } 5124 if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) { 5125 /* We need a SYN or a FIN, unlikely.. */ 5126 return (0); 5127 } 5128 if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) { 5129 /* Timestamp is behind .. old ack with seq wrap? */ 5130 return (0); 5131 } 5132 if (__predict_false(IN_RECOVERY(tp->t_flags))) { 5133 /* Still recovering */ 5134 return (0); 5135 } 5136 rack = (struct tcp_rack *)tp->t_fb_ptr; 5137 if (rack->r_ctl.rc_sacked) { 5138 /* We have sack holes on our scoreboard */ 5139 return (0); 5140 } 5141 /* Ok if we reach here, we can process a fast-ack */ 5142 nsegs = max(1, m->m_pkthdr.lro_nsegs); 5143 rack_log_ack(tp, to, th); 5144 /* Did the window get updated? */ 5145 if (tiwin != tp->snd_wnd) { 5146 tp->snd_wnd = tiwin; 5147 tp->snd_wl1 = th->th_seq; 5148 if (tp->snd_wnd > tp->max_sndwnd) 5149 tp->max_sndwnd = tp->snd_wnd; 5150 } 5151 if ((rack->rc_in_persist != 0) && (tp->snd_wnd >= tp->t_maxseg)) { 5152 rack_exit_persist(tp, rack); 5153 } 5154 /* 5155 * If last ACK falls within this segment's sequence numbers, record 5156 * the timestamp. NOTE that the test is modified according to the 5157 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26). 5158 */ 5159 if ((to->to_flags & TOF_TS) != 0 && 5160 SEQ_LEQ(th->th_seq, tp->last_ack_sent)) { 5161 tp->ts_recent_age = tcp_ts_getticks(); 5162 tp->ts_recent = to->to_tsval; 5163 } 5164 /* 5165 * This is a pure ack for outstanding data. 5166 */ 5167 TCPSTAT_INC(tcps_predack); 5168 5169 /* 5170 * "bad retransmit" recovery. 5171 */ 5172 if (tp->t_flags & TF_PREVVALID) { 5173 tp->t_flags &= ~TF_PREVVALID; 5174 if (tp->t_rxtshift == 1 && 5175 (int)(ticks - tp->t_badrxtwin) < 0) 5176 rack_cong_signal(tp, th, CC_RTO_ERR); 5177 } 5178 /* 5179 * Recalculate the transmit timer / rtt. 5180 * 5181 * Some boxes send broken timestamp replies during the SYN+ACK 5182 * phase, ignore timestamps of 0 or we could calculate a huge RTT 5183 * and blow up the retransmit timer. 5184 */ 5185 acked = BYTES_THIS_ACK(tp, th); 5186 5187 #ifdef TCP_HHOOK 5188 /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */ 5189 hhook_run_tcp_est_in(tp, th, to); 5190 #endif 5191 5192 TCPSTAT_ADD(tcps_rcvackpack, nsegs); 5193 TCPSTAT_ADD(tcps_rcvackbyte, acked); 5194 sbdrop(&so->so_snd, acked); 5195 /* 5196 * Let the congestion control algorithm update congestion control 5197 * related information. This typically means increasing the 5198 * congestion window. 5199 */ 5200 rack_ack_received(tp, rack, th, nsegs, CC_ACK, 0); 5201 5202 tp->snd_una = th->th_ack; 5203 /* 5204 * Pull snd_wl2 up to prevent seq wrap relative to th_ack. 5205 */ 5206 tp->snd_wl2 = th->th_ack; 5207 tp->t_dupacks = 0; 5208 m_freem(m); 5209 /* ND6_HINT(tp); *//* Some progress has been made. */ 5210 5211 /* 5212 * If all outstanding data are acked, stop retransmit timer, 5213 * otherwise restart timer using current (possibly backed-off) 5214 * value. If process is waiting for space, wakeup/selwakeup/signal. 5215 * If data are ready to send, let tcp_output decide between more 5216 * output or persist. 5217 */ 5218 #ifdef TCPDEBUG 5219 if (so->so_options & SO_DEBUG) 5220 tcp_trace(TA_INPUT, ostate, tp, 5221 (void *)tcp_saveipgen, 5222 &tcp_savetcp, 0); 5223 #endif 5224 if (tp->snd_una == tp->snd_max) { 5225 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__); 5226 tp->t_acktime = 0; 5227 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 5228 } 5229 /* Wake up the socket if we have room to write more */ 5230 sowwakeup(so); 5231 if (sbavail(&so->so_snd)) { 5232 rack->r_wanted_output++; 5233 } 5234 return (1); 5235 } 5236 5237 /* 5238 * Return value of 1, the TCB is unlocked and most 5239 * likely gone, return value of 0, the TCP is still 5240 * locked. 5241 */ 5242 static int 5243 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so, 5244 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 5245 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 5246 { 5247 int32_t ret_val = 0; 5248 int32_t todrop; 5249 int32_t ourfinisacked = 0; 5250 5251 rack_calc_rwin(so, tp); 5252 /* 5253 * If the state is SYN_SENT: if seg contains an ACK, but not for our 5254 * SYN, drop the input. if seg contains a RST, then drop the 5255 * connection. if seg does not contain SYN, then drop it. Otherwise 5256 * this is an acceptable SYN segment initialize tp->rcv_nxt and 5257 * tp->irs if seg contains ack then advance tp->snd_una if seg 5258 * contains an ECE and ECN support is enabled, the stream is ECN 5259 * capable. if SYN has been acked change to ESTABLISHED else 5260 * SYN_RCVD state arrange for segment to be acked (eventually) 5261 * continue processing rest of data/controls, beginning with URG 5262 */ 5263 if ((thflags & TH_ACK) && 5264 (SEQ_LEQ(th->th_ack, tp->iss) || 5265 SEQ_GT(th->th_ack, tp->snd_max))) { 5266 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 5267 return (1); 5268 } 5269 if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) { 5270 TCP_PROBE5(connect__refused, NULL, tp, 5271 mtod(m, const char *), tp, th); 5272 tp = tcp_drop(tp, ECONNREFUSED); 5273 rack_do_drop(m, tp); 5274 return (1); 5275 } 5276 if (thflags & TH_RST) { 5277 rack_do_drop(m, tp); 5278 return (1); 5279 } 5280 if (!(thflags & TH_SYN)) { 5281 rack_do_drop(m, tp); 5282 return (1); 5283 } 5284 tp->irs = th->th_seq; 5285 tcp_rcvseqinit(tp); 5286 if (thflags & TH_ACK) { 5287 int tfo_partial = 0; 5288 5289 TCPSTAT_INC(tcps_connects); 5290 soisconnected(so); 5291 #ifdef MAC 5292 mac_socketpeer_set_from_mbuf(m, so); 5293 #endif 5294 /* Do window scaling on this connection? */ 5295 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == 5296 (TF_RCVD_SCALE | TF_REQ_SCALE)) { 5297 tp->rcv_scale = tp->request_r_scale; 5298 } 5299 tp->rcv_adv += min(tp->rcv_wnd, 5300 TCP_MAXWIN << tp->rcv_scale); 5301 /* 5302 * If not all the data that was sent in the TFO SYN 5303 * has been acked, resend the remainder right away. 5304 */ 5305 if (IS_FASTOPEN(tp->t_flags) && 5306 (tp->snd_una != tp->snd_max)) { 5307 tp->snd_nxt = th->th_ack; 5308 tfo_partial = 1; 5309 } 5310 /* 5311 * If there's data, delay ACK; if there's also a FIN ACKNOW 5312 * will be turned on later. 5313 */ 5314 if (DELAY_ACK(tp, tlen) && tlen != 0 && (tfo_partial == 0)) { 5315 rack_timer_cancel(tp, (struct tcp_rack *)tp->t_fb_ptr, 5316 ((struct tcp_rack *)tp->t_fb_ptr)->r_ctl.rc_rcvtime, __LINE__); 5317 tp->t_flags |= TF_DELACK; 5318 } else { 5319 ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output++; 5320 tp->t_flags |= TF_ACKNOW; 5321 } 5322 5323 if (((thflags & (TH_CWR | TH_ECE)) == TH_ECE) && 5324 V_tcp_do_ecn) { 5325 tp->t_flags |= TF_ECN_PERMIT; 5326 TCPSTAT_INC(tcps_ecn_shs); 5327 } 5328 if (SEQ_GT(th->th_ack, tp->snd_una)) { 5329 /* 5330 * We advance snd_una for the 5331 * fast open case. If th_ack is 5332 * acknowledging data beyond 5333 * snd_una we can't just call 5334 * ack-processing since the 5335 * data stream in our send-map 5336 * will start at snd_una + 1 (one 5337 * beyond the SYN). If its just 5338 * equal we don't need to do that 5339 * and there is no send_map. 5340 */ 5341 tp->snd_una++; 5342 } 5343 /* 5344 * Received <SYN,ACK> in SYN_SENT[*] state. Transitions: 5345 * SYN_SENT --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1 5346 */ 5347 tp->t_starttime = ticks; 5348 if (tp->t_flags & TF_NEEDFIN) { 5349 tcp_state_change(tp, TCPS_FIN_WAIT_1); 5350 tp->t_flags &= ~TF_NEEDFIN; 5351 thflags &= ~TH_SYN; 5352 } else { 5353 tcp_state_change(tp, TCPS_ESTABLISHED); 5354 TCP_PROBE5(connect__established, NULL, tp, 5355 mtod(m, const char *), tp, th); 5356 cc_conn_init(tp); 5357 } 5358 } else { 5359 /* 5360 * Received initial SYN in SYN-SENT[*] state => simultaneous 5361 * open. If segment contains CC option and there is a 5362 * cached CC, apply TAO test. If it succeeds, connection is * 5363 * half-synchronized. Otherwise, do 3-way handshake: 5364 * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If 5365 * there was no CC option, clear cached CC value. 5366 */ 5367 tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN); 5368 tcp_state_change(tp, TCPS_SYN_RECEIVED); 5369 } 5370 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 5371 INP_WLOCK_ASSERT(tp->t_inpcb); 5372 /* 5373 * Advance th->th_seq to correspond to first data byte. If data, 5374 * trim to stay within window, dropping FIN if necessary. 5375 */ 5376 th->th_seq++; 5377 if (tlen > tp->rcv_wnd) { 5378 todrop = tlen - tp->rcv_wnd; 5379 m_adj(m, -todrop); 5380 tlen = tp->rcv_wnd; 5381 thflags &= ~TH_FIN; 5382 TCPSTAT_INC(tcps_rcvpackafterwin); 5383 TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop); 5384 } 5385 tp->snd_wl1 = th->th_seq - 1; 5386 tp->rcv_up = th->th_seq; 5387 /* 5388 * Client side of transaction: already sent SYN and data. If the 5389 * remote host used T/TCP to validate the SYN, our data will be 5390 * ACK'd; if so, enter normal data segment processing in the middle 5391 * of step 5, ack processing. Otherwise, goto step 6. 5392 */ 5393 if (thflags & TH_ACK) { 5394 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) 5395 return (ret_val); 5396 /* We may have changed to FIN_WAIT_1 above */ 5397 if (tp->t_state == TCPS_FIN_WAIT_1) { 5398 /* 5399 * In FIN_WAIT_1 STATE in addition to the processing 5400 * for the ESTABLISHED state if our FIN is now 5401 * acknowledged then enter FIN_WAIT_2. 5402 */ 5403 if (ourfinisacked) { 5404 /* 5405 * If we can't receive any more data, then 5406 * closing user can proceed. Starting the 5407 * timer is contrary to the specification, 5408 * but if we don't get a FIN we'll hang 5409 * forever. 5410 * 5411 * XXXjl: we should release the tp also, and 5412 * use a compressed state. 5413 */ 5414 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 5415 soisdisconnected(so); 5416 tcp_timer_activate(tp, TT_2MSL, 5417 (tcp_fast_finwait2_recycle ? 5418 tcp_finwait2_timeout : 5419 TP_MAXIDLE(tp))); 5420 } 5421 tcp_state_change(tp, TCPS_FIN_WAIT_2); 5422 } 5423 } 5424 } 5425 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5426 tiwin, thflags, nxt_pkt)); 5427 } 5428 5429 /* 5430 * Return value of 1, the TCB is unlocked and most 5431 * likely gone, return value of 0, the TCP is still 5432 * locked. 5433 */ 5434 static int 5435 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so, 5436 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 5437 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 5438 { 5439 int32_t ret_val = 0; 5440 int32_t ourfinisacked = 0; 5441 5442 rack_calc_rwin(so, tp); 5443 5444 if ((thflags & TH_ACK) && 5445 (SEQ_LEQ(th->th_ack, tp->snd_una) || 5446 SEQ_GT(th->th_ack, tp->snd_max))) { 5447 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 5448 return (1); 5449 } 5450 if (IS_FASTOPEN(tp->t_flags)) { 5451 /* 5452 * When a TFO connection is in SYN_RECEIVED, the 5453 * only valid packets are the initial SYN, a 5454 * retransmit/copy of the initial SYN (possibly with 5455 * a subset of the original data), a valid ACK, a 5456 * FIN, or a RST. 5457 */ 5458 if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) { 5459 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 5460 return (1); 5461 } else if (thflags & TH_SYN) { 5462 /* non-initial SYN is ignored */ 5463 struct tcp_rack *rack; 5464 5465 rack = (struct tcp_rack *)tp->t_fb_ptr; 5466 if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) || 5467 (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) || 5468 (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) { 5469 rack_do_drop(m, NULL); 5470 return (0); 5471 } 5472 } else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) { 5473 rack_do_drop(m, NULL); 5474 return (0); 5475 } 5476 } 5477 if (thflags & TH_RST) 5478 return (rack_process_rst(m, th, so, tp)); 5479 /* 5480 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 5481 * it's less than ts_recent, drop it. 5482 */ 5483 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 5484 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 5485 if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val)) 5486 return (ret_val); 5487 } 5488 /* 5489 * In the SYN-RECEIVED state, validate that the packet belongs to 5490 * this connection before trimming the data to fit the receive 5491 * window. Check the sequence number versus IRS since we know the 5492 * sequence numbers haven't wrapped. This is a partial fix for the 5493 * "LAND" DoS attack. 5494 */ 5495 if (SEQ_LT(th->th_seq, tp->irs)) { 5496 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 5497 return (1); 5498 } 5499 if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 5500 return (ret_val); 5501 } 5502 /* 5503 * If last ACK falls within this segment's sequence numbers, record 5504 * its timestamp. NOTE: 1) That the test incorporates suggestions 5505 * from the latest proposal of the tcplw@cray.com list (Braden 5506 * 1993/04/26). 2) That updating only on newer timestamps interferes 5507 * with our earlier PAWS tests, so this check should be solely 5508 * predicated on the sequence space of this segment. 3) That we 5509 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 5510 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 5511 * SEG.Len, This modified check allows us to overcome RFC1323's 5512 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 5513 * p.869. In such cases, we can still calculate the RTT correctly 5514 * when RCV.NXT == Last.ACK.Sent. 5515 */ 5516 if ((to->to_flags & TOF_TS) != 0 && 5517 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 5518 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 5519 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 5520 tp->ts_recent_age = tcp_ts_getticks(); 5521 tp->ts_recent = to->to_tsval; 5522 } 5523 tp->snd_wnd = tiwin; 5524 /* 5525 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 5526 * is on (half-synchronized state), then queue data for later 5527 * processing; else drop segment and return. 5528 */ 5529 if ((thflags & TH_ACK) == 0) { 5530 if (IS_FASTOPEN(tp->t_flags)) { 5531 cc_conn_init(tp); 5532 } 5533 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5534 tiwin, thflags, nxt_pkt)); 5535 } 5536 TCPSTAT_INC(tcps_connects); 5537 soisconnected(so); 5538 /* Do window scaling? */ 5539 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == 5540 (TF_RCVD_SCALE | TF_REQ_SCALE)) { 5541 tp->rcv_scale = tp->request_r_scale; 5542 } 5543 /* 5544 * Make transitions: SYN-RECEIVED -> ESTABLISHED SYN-RECEIVED* -> 5545 * FIN-WAIT-1 5546 */ 5547 tp->t_starttime = ticks; 5548 if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) { 5549 tcp_fastopen_decrement_counter(tp->t_tfo_pending); 5550 tp->t_tfo_pending = NULL; 5551 5552 /* 5553 * Account for the ACK of our SYN prior to 5554 * regular ACK processing below. 5555 */ 5556 tp->snd_una++; 5557 } 5558 if (tp->t_flags & TF_NEEDFIN) { 5559 tcp_state_change(tp, TCPS_FIN_WAIT_1); 5560 tp->t_flags &= ~TF_NEEDFIN; 5561 } else { 5562 tcp_state_change(tp, TCPS_ESTABLISHED); 5563 TCP_PROBE5(accept__established, NULL, tp, 5564 mtod(m, const char *), tp, th); 5565 /* 5566 * TFO connections call cc_conn_init() during SYN 5567 * processing. Calling it again here for such connections 5568 * is not harmless as it would undo the snd_cwnd reduction 5569 * that occurs when a TFO SYN|ACK is retransmitted. 5570 */ 5571 if (!IS_FASTOPEN(tp->t_flags)) 5572 cc_conn_init(tp); 5573 } 5574 /* 5575 * If segment contains data or ACK, will call tcp_reass() later; if 5576 * not, do so now to pass queued data to user. 5577 */ 5578 if (tlen == 0 && (thflags & TH_FIN) == 0) 5579 (void) tcp_reass(tp, (struct tcphdr *)0, NULL, 0, 5580 (struct mbuf *)0); 5581 tp->snd_wl1 = th->th_seq - 1; 5582 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 5583 return (ret_val); 5584 } 5585 if (tp->t_state == TCPS_FIN_WAIT_1) { 5586 /* We could have went to FIN_WAIT_1 (or EST) above */ 5587 /* 5588 * In FIN_WAIT_1 STATE in addition to the processing for the 5589 * ESTABLISHED state if our FIN is now acknowledged then 5590 * enter FIN_WAIT_2. 5591 */ 5592 if (ourfinisacked) { 5593 /* 5594 * If we can't receive any more data, then closing 5595 * user can proceed. Starting the timer is contrary 5596 * to the specification, but if we don't get a FIN 5597 * we'll hang forever. 5598 * 5599 * XXXjl: we should release the tp also, and use a 5600 * compressed state. 5601 */ 5602 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 5603 soisdisconnected(so); 5604 tcp_timer_activate(tp, TT_2MSL, 5605 (tcp_fast_finwait2_recycle ? 5606 tcp_finwait2_timeout : 5607 TP_MAXIDLE(tp))); 5608 } 5609 tcp_state_change(tp, TCPS_FIN_WAIT_2); 5610 } 5611 } 5612 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5613 tiwin, thflags, nxt_pkt)); 5614 } 5615 5616 /* 5617 * Return value of 1, the TCB is unlocked and most 5618 * likely gone, return value of 0, the TCP is still 5619 * locked. 5620 */ 5621 static int 5622 rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so, 5623 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 5624 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 5625 { 5626 int32_t ret_val = 0; 5627 5628 /* 5629 * Header prediction: check for the two common cases of a 5630 * uni-directional data xfer. If the packet has no control flags, 5631 * is in-sequence, the window didn't change and we're not 5632 * retransmitting, it's a candidate. If the length is zero and the 5633 * ack moved forward, we're the sender side of the xfer. Just free 5634 * the data acked & wake any higher level process that was blocked 5635 * waiting for space. If the length is non-zero and the ack didn't 5636 * move, we're the receiver side. If we're getting packets in-order 5637 * (the reassembly queue is empty), add the data toc The socket 5638 * buffer and note that we need a delayed ack. Make sure that the 5639 * hidden state-flags are also off. Since we check for 5640 * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN. 5641 */ 5642 if (__predict_true(((to->to_flags & TOF_SACK) == 0)) && 5643 __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_URG | TH_ACK)) == TH_ACK) && 5644 __predict_true(SEGQ_EMPTY(tp)) && 5645 __predict_true(th->th_seq == tp->rcv_nxt)) { 5646 struct tcp_rack *rack; 5647 5648 rack = (struct tcp_rack *)tp->t_fb_ptr; 5649 if (tlen == 0) { 5650 if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen, 5651 tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) { 5652 return (0); 5653 } 5654 } else { 5655 if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen, 5656 tiwin, nxt_pkt)) { 5657 return (0); 5658 } 5659 } 5660 } 5661 rack_calc_rwin(so, tp); 5662 5663 if (thflags & TH_RST) 5664 return (rack_process_rst(m, th, so, tp)); 5665 5666 /* 5667 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 5668 * synchronized state. 5669 */ 5670 if (thflags & TH_SYN) { 5671 rack_challenge_ack(m, th, tp, &ret_val); 5672 return (ret_val); 5673 } 5674 /* 5675 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 5676 * it's less than ts_recent, drop it. 5677 */ 5678 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 5679 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 5680 if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val)) 5681 return (ret_val); 5682 } 5683 if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 5684 return (ret_val); 5685 } 5686 /* 5687 * If last ACK falls within this segment's sequence numbers, record 5688 * its timestamp. NOTE: 1) That the test incorporates suggestions 5689 * from the latest proposal of the tcplw@cray.com list (Braden 5690 * 1993/04/26). 2) That updating only on newer timestamps interferes 5691 * with our earlier PAWS tests, so this check should be solely 5692 * predicated on the sequence space of this segment. 3) That we 5693 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 5694 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 5695 * SEG.Len, This modified check allows us to overcome RFC1323's 5696 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 5697 * p.869. In such cases, we can still calculate the RTT correctly 5698 * when RCV.NXT == Last.ACK.Sent. 5699 */ 5700 if ((to->to_flags & TOF_TS) != 0 && 5701 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 5702 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 5703 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 5704 tp->ts_recent_age = tcp_ts_getticks(); 5705 tp->ts_recent = to->to_tsval; 5706 } 5707 /* 5708 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 5709 * is on (half-synchronized state), then queue data for later 5710 * processing; else drop segment and return. 5711 */ 5712 if ((thflags & TH_ACK) == 0) { 5713 if (tp->t_flags & TF_NEEDSYN) { 5714 5715 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5716 tiwin, thflags, nxt_pkt)); 5717 5718 } else if (tp->t_flags & TF_ACKNOW) { 5719 rack_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 5720 return (ret_val); 5721 } else { 5722 rack_do_drop(m, NULL); 5723 return (0); 5724 } 5725 } 5726 /* 5727 * Ack processing. 5728 */ 5729 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) { 5730 return (ret_val); 5731 } 5732 if (sbavail(&so->so_snd)) { 5733 if (rack_progress_timeout_check(tp)) { 5734 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); 5735 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 5736 return (1); 5737 } 5738 } 5739 /* State changes only happen in rack_process_data() */ 5740 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5741 tiwin, thflags, nxt_pkt)); 5742 } 5743 5744 /* 5745 * Return value of 1, the TCB is unlocked and most 5746 * likely gone, return value of 0, the TCP is still 5747 * locked. 5748 */ 5749 static int 5750 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so, 5751 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 5752 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 5753 { 5754 int32_t ret_val = 0; 5755 5756 rack_calc_rwin(so, tp); 5757 if (thflags & TH_RST) 5758 return (rack_process_rst(m, th, so, tp)); 5759 /* 5760 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 5761 * synchronized state. 5762 */ 5763 if (thflags & TH_SYN) { 5764 rack_challenge_ack(m, th, tp, &ret_val); 5765 return (ret_val); 5766 } 5767 /* 5768 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 5769 * it's less than ts_recent, drop it. 5770 */ 5771 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 5772 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 5773 if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val)) 5774 return (ret_val); 5775 } 5776 if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 5777 return (ret_val); 5778 } 5779 /* 5780 * If last ACK falls within this segment's sequence numbers, record 5781 * its timestamp. NOTE: 1) That the test incorporates suggestions 5782 * from the latest proposal of the tcplw@cray.com list (Braden 5783 * 1993/04/26). 2) That updating only on newer timestamps interferes 5784 * with our earlier PAWS tests, so this check should be solely 5785 * predicated on the sequence space of this segment. 3) That we 5786 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 5787 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 5788 * SEG.Len, This modified check allows us to overcome RFC1323's 5789 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 5790 * p.869. In such cases, we can still calculate the RTT correctly 5791 * when RCV.NXT == Last.ACK.Sent. 5792 */ 5793 if ((to->to_flags & TOF_TS) != 0 && 5794 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 5795 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 5796 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 5797 tp->ts_recent_age = tcp_ts_getticks(); 5798 tp->ts_recent = to->to_tsval; 5799 } 5800 /* 5801 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 5802 * is on (half-synchronized state), then queue data for later 5803 * processing; else drop segment and return. 5804 */ 5805 if ((thflags & TH_ACK) == 0) { 5806 if (tp->t_flags & TF_NEEDSYN) { 5807 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5808 tiwin, thflags, nxt_pkt)); 5809 5810 } else if (tp->t_flags & TF_ACKNOW) { 5811 rack_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 5812 return (ret_val); 5813 } else { 5814 rack_do_drop(m, NULL); 5815 return (0); 5816 } 5817 } 5818 /* 5819 * Ack processing. 5820 */ 5821 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) { 5822 return (ret_val); 5823 } 5824 if (sbavail(&so->so_snd)) { 5825 if (rack_progress_timeout_check(tp)) { 5826 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); 5827 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 5828 return (1); 5829 } 5830 } 5831 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5832 tiwin, thflags, nxt_pkt)); 5833 } 5834 5835 static int 5836 rack_check_data_after_close(struct mbuf *m, 5837 struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so) 5838 { 5839 struct tcp_rack *rack; 5840 5841 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 5842 rack = (struct tcp_rack *)tp->t_fb_ptr; 5843 if (rack->rc_allow_data_af_clo == 0) { 5844 close_now: 5845 tp = tcp_close(tp); 5846 TCPSTAT_INC(tcps_rcvafterclose); 5847 rack_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen)); 5848 return (1); 5849 } 5850 if (sbavail(&so->so_snd) == 0) 5851 goto close_now; 5852 /* Ok we allow data that is ignored and a followup reset */ 5853 tp->rcv_nxt = th->th_seq + *tlen; 5854 tp->t_flags2 |= TF2_DROP_AF_DATA; 5855 rack->r_wanted_output = 1; 5856 *tlen = 0; 5857 return (0); 5858 } 5859 5860 /* 5861 * Return value of 1, the TCB is unlocked and most 5862 * likely gone, return value of 0, the TCP is still 5863 * locked. 5864 */ 5865 static int 5866 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so, 5867 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 5868 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 5869 { 5870 int32_t ret_val = 0; 5871 int32_t ourfinisacked = 0; 5872 5873 rack_calc_rwin(so, tp); 5874 5875 if (thflags & TH_RST) 5876 return (rack_process_rst(m, th, so, tp)); 5877 /* 5878 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 5879 * synchronized state. 5880 */ 5881 if (thflags & TH_SYN) { 5882 rack_challenge_ack(m, th, tp, &ret_val); 5883 return (ret_val); 5884 } 5885 /* 5886 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 5887 * it's less than ts_recent, drop it. 5888 */ 5889 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 5890 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 5891 if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val)) 5892 return (ret_val); 5893 } 5894 if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 5895 return (ret_val); 5896 } 5897 /* 5898 * If new data are received on a connection after the user processes 5899 * are gone, then RST the other end. 5900 */ 5901 if ((so->so_state & SS_NOFDREF) && tlen) { 5902 if (rack_check_data_after_close(m, tp, &tlen, th, so)) 5903 return (1); 5904 } 5905 /* 5906 * If last ACK falls within this segment's sequence numbers, record 5907 * its timestamp. NOTE: 1) That the test incorporates suggestions 5908 * from the latest proposal of the tcplw@cray.com list (Braden 5909 * 1993/04/26). 2) That updating only on newer timestamps interferes 5910 * with our earlier PAWS tests, so this check should be solely 5911 * predicated on the sequence space of this segment. 3) That we 5912 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 5913 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 5914 * SEG.Len, This modified check allows us to overcome RFC1323's 5915 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 5916 * p.869. In such cases, we can still calculate the RTT correctly 5917 * when RCV.NXT == Last.ACK.Sent. 5918 */ 5919 if ((to->to_flags & TOF_TS) != 0 && 5920 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 5921 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 5922 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 5923 tp->ts_recent_age = tcp_ts_getticks(); 5924 tp->ts_recent = to->to_tsval; 5925 } 5926 /* 5927 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 5928 * is on (half-synchronized state), then queue data for later 5929 * processing; else drop segment and return. 5930 */ 5931 if ((thflags & TH_ACK) == 0) { 5932 if (tp->t_flags & TF_NEEDSYN) { 5933 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5934 tiwin, thflags, nxt_pkt)); 5935 } else if (tp->t_flags & TF_ACKNOW) { 5936 rack_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 5937 return (ret_val); 5938 } else { 5939 rack_do_drop(m, NULL); 5940 return (0); 5941 } 5942 } 5943 /* 5944 * Ack processing. 5945 */ 5946 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 5947 return (ret_val); 5948 } 5949 if (ourfinisacked) { 5950 /* 5951 * If we can't receive any more data, then closing user can 5952 * proceed. Starting the timer is contrary to the 5953 * specification, but if we don't get a FIN we'll hang 5954 * forever. 5955 * 5956 * XXXjl: we should release the tp also, and use a 5957 * compressed state. 5958 */ 5959 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 5960 soisdisconnected(so); 5961 tcp_timer_activate(tp, TT_2MSL, 5962 (tcp_fast_finwait2_recycle ? 5963 tcp_finwait2_timeout : 5964 TP_MAXIDLE(tp))); 5965 } 5966 tcp_state_change(tp, TCPS_FIN_WAIT_2); 5967 } 5968 if (sbavail(&so->so_snd)) { 5969 if (rack_progress_timeout_check(tp)) { 5970 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); 5971 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 5972 return (1); 5973 } 5974 } 5975 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5976 tiwin, thflags, nxt_pkt)); 5977 } 5978 5979 /* 5980 * Return value of 1, the TCB is unlocked and most 5981 * likely gone, return value of 0, the TCP is still 5982 * locked. 5983 */ 5984 static int 5985 rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so, 5986 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 5987 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 5988 { 5989 int32_t ret_val = 0; 5990 int32_t ourfinisacked = 0; 5991 5992 rack_calc_rwin(so, tp); 5993 5994 if (thflags & TH_RST) 5995 return (rack_process_rst(m, th, so, tp)); 5996 /* 5997 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 5998 * synchronized state. 5999 */ 6000 if (thflags & TH_SYN) { 6001 rack_challenge_ack(m, th, tp, &ret_val); 6002 return (ret_val); 6003 } 6004 /* 6005 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 6006 * it's less than ts_recent, drop it. 6007 */ 6008 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 6009 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 6010 if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val)) 6011 return (ret_val); 6012 } 6013 if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 6014 return (ret_val); 6015 } 6016 /* 6017 * If new data are received on a connection after the user processes 6018 * are gone, then RST the other end. 6019 */ 6020 if ((so->so_state & SS_NOFDREF) && tlen) { 6021 if (rack_check_data_after_close(m, tp, &tlen, th, so)) 6022 return (1); 6023 } 6024 /* 6025 * If last ACK falls within this segment's sequence numbers, record 6026 * its timestamp. NOTE: 1) That the test incorporates suggestions 6027 * from the latest proposal of the tcplw@cray.com list (Braden 6028 * 1993/04/26). 2) That updating only on newer timestamps interferes 6029 * with our earlier PAWS tests, so this check should be solely 6030 * predicated on the sequence space of this segment. 3) That we 6031 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 6032 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 6033 * SEG.Len, This modified check allows us to overcome RFC1323's 6034 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 6035 * p.869. In such cases, we can still calculate the RTT correctly 6036 * when RCV.NXT == Last.ACK.Sent. 6037 */ 6038 if ((to->to_flags & TOF_TS) != 0 && 6039 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 6040 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 6041 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 6042 tp->ts_recent_age = tcp_ts_getticks(); 6043 tp->ts_recent = to->to_tsval; 6044 } 6045 /* 6046 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 6047 * is on (half-synchronized state), then queue data for later 6048 * processing; else drop segment and return. 6049 */ 6050 if ((thflags & TH_ACK) == 0) { 6051 if (tp->t_flags & TF_NEEDSYN) { 6052 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 6053 tiwin, thflags, nxt_pkt)); 6054 } else if (tp->t_flags & TF_ACKNOW) { 6055 rack_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 6056 return (ret_val); 6057 } else { 6058 rack_do_drop(m, NULL); 6059 return (0); 6060 } 6061 } 6062 /* 6063 * Ack processing. 6064 */ 6065 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 6066 return (ret_val); 6067 } 6068 if (ourfinisacked) { 6069 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 6070 tcp_twstart(tp); 6071 m_freem(m); 6072 return (1); 6073 } 6074 if (sbavail(&so->so_snd)) { 6075 if (rack_progress_timeout_check(tp)) { 6076 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); 6077 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 6078 return (1); 6079 } 6080 } 6081 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 6082 tiwin, thflags, nxt_pkt)); 6083 } 6084 6085 /* 6086 * Return value of 1, the TCB is unlocked and most 6087 * likely gone, return value of 0, the TCP is still 6088 * locked. 6089 */ 6090 static int 6091 rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so, 6092 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 6093 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 6094 { 6095 int32_t ret_val = 0; 6096 int32_t ourfinisacked = 0; 6097 6098 rack_calc_rwin(so, tp); 6099 6100 if (thflags & TH_RST) 6101 return (rack_process_rst(m, th, so, tp)); 6102 /* 6103 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 6104 * synchronized state. 6105 */ 6106 if (thflags & TH_SYN) { 6107 rack_challenge_ack(m, th, tp, &ret_val); 6108 return (ret_val); 6109 } 6110 /* 6111 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 6112 * it's less than ts_recent, drop it. 6113 */ 6114 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 6115 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 6116 if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val)) 6117 return (ret_val); 6118 } 6119 if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 6120 return (ret_val); 6121 } 6122 /* 6123 * If new data are received on a connection after the user processes 6124 * are gone, then RST the other end. 6125 */ 6126 if ((so->so_state & SS_NOFDREF) && tlen) { 6127 if (rack_check_data_after_close(m, tp, &tlen, th, so)) 6128 return (1); 6129 } 6130 /* 6131 * If last ACK falls within this segment's sequence numbers, record 6132 * its timestamp. NOTE: 1) That the test incorporates suggestions 6133 * from the latest proposal of the tcplw@cray.com list (Braden 6134 * 1993/04/26). 2) That updating only on newer timestamps interferes 6135 * with our earlier PAWS tests, so this check should be solely 6136 * predicated on the sequence space of this segment. 3) That we 6137 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 6138 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 6139 * SEG.Len, This modified check allows us to overcome RFC1323's 6140 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 6141 * p.869. In such cases, we can still calculate the RTT correctly 6142 * when RCV.NXT == Last.ACK.Sent. 6143 */ 6144 if ((to->to_flags & TOF_TS) != 0 && 6145 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 6146 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 6147 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 6148 tp->ts_recent_age = tcp_ts_getticks(); 6149 tp->ts_recent = to->to_tsval; 6150 } 6151 /* 6152 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 6153 * is on (half-synchronized state), then queue data for later 6154 * processing; else drop segment and return. 6155 */ 6156 if ((thflags & TH_ACK) == 0) { 6157 if (tp->t_flags & TF_NEEDSYN) { 6158 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 6159 tiwin, thflags, nxt_pkt)); 6160 } else if (tp->t_flags & TF_ACKNOW) { 6161 rack_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 6162 return (ret_val); 6163 } else { 6164 rack_do_drop(m, NULL); 6165 return (0); 6166 } 6167 } 6168 /* 6169 * case TCPS_LAST_ACK: Ack processing. 6170 */ 6171 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 6172 return (ret_val); 6173 } 6174 if (ourfinisacked) { 6175 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 6176 tp = tcp_close(tp); 6177 rack_do_drop(m, tp); 6178 return (1); 6179 } 6180 if (sbavail(&so->so_snd)) { 6181 if (rack_progress_timeout_check(tp)) { 6182 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); 6183 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 6184 return (1); 6185 } 6186 } 6187 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 6188 tiwin, thflags, nxt_pkt)); 6189 } 6190 6191 6192 /* 6193 * Return value of 1, the TCB is unlocked and most 6194 * likely gone, return value of 0, the TCP is still 6195 * locked. 6196 */ 6197 static int 6198 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so, 6199 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 6200 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 6201 { 6202 int32_t ret_val = 0; 6203 int32_t ourfinisacked = 0; 6204 6205 rack_calc_rwin(so, tp); 6206 6207 /* Reset receive buffer auto scaling when not in bulk receive mode. */ 6208 if (thflags & TH_RST) 6209 return (rack_process_rst(m, th, so, tp)); 6210 /* 6211 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 6212 * synchronized state. 6213 */ 6214 if (thflags & TH_SYN) { 6215 rack_challenge_ack(m, th, tp, &ret_val); 6216 return (ret_val); 6217 } 6218 /* 6219 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 6220 * it's less than ts_recent, drop it. 6221 */ 6222 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 6223 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 6224 if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val)) 6225 return (ret_val); 6226 } 6227 if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 6228 return (ret_val); 6229 } 6230 /* 6231 * If new data are received on a connection after the user processes 6232 * are gone, then RST the other end. 6233 */ 6234 if ((so->so_state & SS_NOFDREF) && 6235 tlen) { 6236 if (rack_check_data_after_close(m, tp, &tlen, th, so)) 6237 return (1); 6238 } 6239 /* 6240 * If last ACK falls within this segment's sequence numbers, record 6241 * its timestamp. NOTE: 1) That the test incorporates suggestions 6242 * from the latest proposal of the tcplw@cray.com list (Braden 6243 * 1993/04/26). 2) That updating only on newer timestamps interferes 6244 * with our earlier PAWS tests, so this check should be solely 6245 * predicated on the sequence space of this segment. 3) That we 6246 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 6247 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 6248 * SEG.Len, This modified check allows us to overcome RFC1323's 6249 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 6250 * p.869. In such cases, we can still calculate the RTT correctly 6251 * when RCV.NXT == Last.ACK.Sent. 6252 */ 6253 if ((to->to_flags & TOF_TS) != 0 && 6254 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 6255 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 6256 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 6257 tp->ts_recent_age = tcp_ts_getticks(); 6258 tp->ts_recent = to->to_tsval; 6259 } 6260 /* 6261 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 6262 * is on (half-synchronized state), then queue data for later 6263 * processing; else drop segment and return. 6264 */ 6265 if ((thflags & TH_ACK) == 0) { 6266 if (tp->t_flags & TF_NEEDSYN) { 6267 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 6268 tiwin, thflags, nxt_pkt)); 6269 } else if (tp->t_flags & TF_ACKNOW) { 6270 rack_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 6271 return (ret_val); 6272 } else { 6273 rack_do_drop(m, NULL); 6274 return (0); 6275 } 6276 } 6277 /* 6278 * Ack processing. 6279 */ 6280 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 6281 return (ret_val); 6282 } 6283 if (sbavail(&so->so_snd)) { 6284 if (rack_progress_timeout_check(tp)) { 6285 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); 6286 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 6287 return (1); 6288 } 6289 } 6290 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 6291 tiwin, thflags, nxt_pkt)); 6292 } 6293 6294 6295 static void inline 6296 rack_clear_rate_sample(struct tcp_rack *rack) 6297 { 6298 rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY; 6299 rack->r_ctl.rack_rs.rs_rtt_cnt = 0; 6300 rack->r_ctl.rack_rs.rs_rtt_tot = 0; 6301 } 6302 6303 static int 6304 rack_init(struct tcpcb *tp) 6305 { 6306 struct tcp_rack *rack = NULL; 6307 6308 tp->t_fb_ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT); 6309 if (tp->t_fb_ptr == NULL) { 6310 /* 6311 * We need to allocate memory but cant. The INP and INP_INFO 6312 * locks and they are recusive (happens during setup. So a 6313 * scheme to drop the locks fails :( 6314 * 6315 */ 6316 return (ENOMEM); 6317 } 6318 memset(tp->t_fb_ptr, 0, sizeof(struct tcp_rack)); 6319 6320 rack = (struct tcp_rack *)tp->t_fb_ptr; 6321 TAILQ_INIT(&rack->r_ctl.rc_map); 6322 TAILQ_INIT(&rack->r_ctl.rc_free); 6323 TAILQ_INIT(&rack->r_ctl.rc_tmap); 6324 rack->rc_tp = tp; 6325 if (tp->t_inpcb) { 6326 rack->rc_inp = tp->t_inpcb; 6327 } 6328 /* Probably not needed but lets be sure */ 6329 rack_clear_rate_sample(rack); 6330 rack->r_cpu = 0; 6331 rack->r_ctl.rc_reorder_fade = rack_reorder_fade; 6332 rack->rc_allow_data_af_clo = rack_ignore_data_after_close; 6333 rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh; 6334 rack->rc_pace_reduce = rack_slot_reduction; 6335 if (V_tcp_delack_enabled) 6336 tp->t_delayed_ack = 1; 6337 else 6338 tp->t_delayed_ack = 0; 6339 rack->rc_pace_max_segs = rack_hptsi_segments; 6340 rack->r_ctl.rc_early_recovery_segs = rack_early_recovery_max_seg; 6341 rack->r_ctl.rc_reorder_shift = rack_reorder_thresh; 6342 rack->r_ctl.rc_pkt_delay = rack_pkt_delay; 6343 rack->r_ctl.rc_prop_reduce = rack_use_proportional_reduce; 6344 rack->r_idle_reduce_largest = rack_reduce_largest_on_idle; 6345 rack->r_enforce_min_pace = rack_min_pace_time; 6346 rack->r_min_pace_seg_thresh = rack_min_pace_time_seg_req; 6347 rack->r_ctl.rc_prop_rate = rack_proportional_rate; 6348 rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp; 6349 rack->r_ctl.rc_early_recovery = rack_early_recovery; 6350 rack->rc_always_pace = rack_pace_every_seg; 6351 rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method; 6352 rack->rack_tlp_threshold_use = rack_tlp_threshold_use; 6353 rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr; 6354 rack->r_ctl.rc_min_to = rack_min_to; 6355 rack->r_ctl.rc_prr_inc_var = rack_inc_var; 6356 rack_start_hpts_timer(rack, tp, tcp_ts_getticks(), __LINE__, 0, 0, 0); 6357 if (tp->snd_una != tp->snd_max) { 6358 /* Create a send map for the current outstanding data */ 6359 struct rack_sendmap *rsm; 6360 6361 rsm = rack_alloc(rack); 6362 if (rsm == NULL) { 6363 uma_zfree(rack_pcb_zone, tp->t_fb_ptr); 6364 tp->t_fb_ptr = NULL; 6365 return (ENOMEM); 6366 } 6367 rsm->r_flags = RACK_OVERMAX; 6368 rsm->r_tim_lastsent[0] = tcp_ts_getticks(); 6369 rsm->r_rtr_cnt = 1; 6370 rsm->r_rtr_bytes = 0; 6371 rsm->r_start = tp->snd_una; 6372 rsm->r_end = tp->snd_max; 6373 rsm->r_sndcnt = 0; 6374 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_map, rsm, r_next); 6375 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext); 6376 rsm->r_in_tmap = 1; 6377 } 6378 return (0); 6379 } 6380 6381 static int 6382 rack_handoff_ok(struct tcpcb *tp) 6383 { 6384 if ((tp->t_state == TCPS_CLOSED) || 6385 (tp->t_state == TCPS_LISTEN)) { 6386 /* Sure no problem though it may not stick */ 6387 return (0); 6388 } 6389 if ((tp->t_state == TCPS_SYN_SENT) || 6390 (tp->t_state == TCPS_SYN_RECEIVED)) { 6391 /* 6392 * We really don't know you have to get to ESTAB or beyond 6393 * to tell. 6394 */ 6395 return (EAGAIN); 6396 } 6397 if (tp->t_flags & TF_SACK_PERMIT) { 6398 return (0); 6399 } 6400 /* 6401 * If we reach here we don't do SACK on this connection so we can 6402 * never do rack. 6403 */ 6404 return (EINVAL); 6405 } 6406 6407 static void 6408 rack_fini(struct tcpcb *tp, int32_t tcb_is_purged) 6409 { 6410 if (tp->t_fb_ptr) { 6411 struct tcp_rack *rack; 6412 struct rack_sendmap *rsm; 6413 6414 rack = (struct tcp_rack *)tp->t_fb_ptr; 6415 #ifdef TCP_BLACKBOX 6416 tcp_log_flowend(tp); 6417 #endif 6418 rsm = TAILQ_FIRST(&rack->r_ctl.rc_map); 6419 while (rsm) { 6420 TAILQ_REMOVE(&rack->r_ctl.rc_map, rsm, r_next); 6421 uma_zfree(rack_zone, rsm); 6422 rsm = TAILQ_FIRST(&rack->r_ctl.rc_map); 6423 } 6424 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free); 6425 while (rsm) { 6426 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_next); 6427 uma_zfree(rack_zone, rsm); 6428 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free); 6429 } 6430 rack->rc_free_cnt = 0; 6431 uma_zfree(rack_pcb_zone, tp->t_fb_ptr); 6432 tp->t_fb_ptr = NULL; 6433 } 6434 } 6435 6436 static void 6437 rack_set_state(struct tcpcb *tp, struct tcp_rack *rack) 6438 { 6439 switch (tp->t_state) { 6440 case TCPS_SYN_SENT: 6441 rack->r_state = TCPS_SYN_SENT; 6442 rack->r_substate = rack_do_syn_sent; 6443 break; 6444 case TCPS_SYN_RECEIVED: 6445 rack->r_state = TCPS_SYN_RECEIVED; 6446 rack->r_substate = rack_do_syn_recv; 6447 break; 6448 case TCPS_ESTABLISHED: 6449 rack->r_state = TCPS_ESTABLISHED; 6450 rack->r_substate = rack_do_established; 6451 break; 6452 case TCPS_CLOSE_WAIT: 6453 rack->r_state = TCPS_CLOSE_WAIT; 6454 rack->r_substate = rack_do_close_wait; 6455 break; 6456 case TCPS_FIN_WAIT_1: 6457 rack->r_state = TCPS_FIN_WAIT_1; 6458 rack->r_substate = rack_do_fin_wait_1; 6459 break; 6460 case TCPS_CLOSING: 6461 rack->r_state = TCPS_CLOSING; 6462 rack->r_substate = rack_do_closing; 6463 break; 6464 case TCPS_LAST_ACK: 6465 rack->r_state = TCPS_LAST_ACK; 6466 rack->r_substate = rack_do_lastack; 6467 break; 6468 case TCPS_FIN_WAIT_2: 6469 rack->r_state = TCPS_FIN_WAIT_2; 6470 rack->r_substate = rack_do_fin_wait_2; 6471 break; 6472 case TCPS_LISTEN: 6473 case TCPS_CLOSED: 6474 case TCPS_TIME_WAIT: 6475 default: 6476 #ifdef INVARIANTS 6477 panic("tcp tp:%p state:%d sees impossible state?", tp, tp->t_state); 6478 #endif 6479 break; 6480 }; 6481 } 6482 6483 6484 static void 6485 rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb) 6486 { 6487 /* 6488 * We received an ack, and then did not 6489 * call send or were bounced out due to the 6490 * hpts was running. Now a timer is up as well, is 6491 * it the right timer? 6492 */ 6493 struct rack_sendmap *rsm; 6494 int tmr_up; 6495 6496 tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK; 6497 if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT)) 6498 return; 6499 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); 6500 if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) && 6501 (tmr_up == PACE_TMR_RXT)) { 6502 /* Should be an RXT */ 6503 return; 6504 } 6505 if (rsm == NULL) { 6506 /* Nothing outstanding? */ 6507 if (tp->t_flags & TF_DELACK) { 6508 if (tmr_up == PACE_TMR_DELACK) 6509 /* We are supposed to have delayed ack up and we do */ 6510 return; 6511 } else if (sbavail(&tp->t_inpcb->inp_socket->so_snd) && (tmr_up == PACE_TMR_RXT)) { 6512 /* 6513 * if we hit enobufs then we would expect the possiblity 6514 * of nothing outstanding and the RXT up (and the hptsi timer). 6515 */ 6516 return; 6517 } else if (((tcp_always_keepalive || 6518 rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) && 6519 (tp->t_state <= TCPS_CLOSING)) && 6520 (tmr_up == PACE_TMR_KEEP) && 6521 (tp->snd_max == tp->snd_una)) { 6522 /* We should have keep alive up and we do */ 6523 return; 6524 } 6525 } 6526 if (rsm && (rsm->r_flags & RACK_SACK_PASSED)) { 6527 if ((tp->t_flags & TF_SENTFIN) && 6528 ((tp->snd_max - tp->snd_una) == 1) && 6529 (rsm->r_flags & RACK_HAS_FIN)) { 6530 /* needs to be a RXT */ 6531 if (tmr_up == PACE_TMR_RXT) 6532 return; 6533 } else if (tmr_up == PACE_TMR_RACK) 6534 return; 6535 } else if (SEQ_GT(tp->snd_max,tp->snd_una) && 6536 ((tmr_up == PACE_TMR_TLP) || 6537 (tmr_up == PACE_TMR_RXT))) { 6538 /* 6539 * Either a TLP or RXT is fine if no sack-passed 6540 * is in place and data is outstanding. 6541 */ 6542 return; 6543 } else if (tmr_up == PACE_TMR_DELACK) { 6544 /* 6545 * If the delayed ack was going to go off 6546 * before the rtx/tlp/rack timer were going to 6547 * expire, then that would be the timer in control. 6548 * Note we don't check the time here trusting the 6549 * code is correct. 6550 */ 6551 return; 6552 } 6553 /* 6554 * Ok the timer originally started is not what we want now. 6555 * We will force the hpts to be stopped if any, and restart 6556 * with the slot set to what was in the saved slot. 6557 */ 6558 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 6559 rack_start_hpts_timer(rack, tp, tcp_ts_getticks(), __LINE__, 0, 0, 0); 6560 } 6561 6562 static void 6563 rack_hpts_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so, 6564 struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos, 6565 int32_t nxt_pkt, struct timeval *tv) 6566 { 6567 int32_t thflags, retval, did_out = 0; 6568 int32_t way_out = 0; 6569 uint32_t cts; 6570 uint32_t tiwin; 6571 struct tcpopt to; 6572 struct tcp_rack *rack; 6573 struct rack_sendmap *rsm; 6574 int32_t prev_state = 0; 6575 6576 cts = tcp_tv_to_mssectick(tv); 6577 rack = (struct tcp_rack *)tp->t_fb_ptr; 6578 6579 kern_prefetch(rack, &prev_state); 6580 prev_state = 0; 6581 thflags = th->th_flags; 6582 /* 6583 * If this is either a state-changing packet or current state isn't 6584 * established, we require a read lock on tcbinfo. Otherwise, we 6585 * allow the tcbinfo to be in either locked or unlocked, as the 6586 * caller may have unnecessarily acquired a lock due to a race. 6587 */ 6588 if ((thflags & (TH_SYN | TH_FIN | TH_RST)) != 0 || 6589 tp->t_state != TCPS_ESTABLISHED) { 6590 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 6591 } 6592 INP_WLOCK_ASSERT(tp->t_inpcb); 6593 KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN", 6594 __func__)); 6595 KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT", 6596 __func__)); 6597 { 6598 union tcp_log_stackspecific log; 6599 6600 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 6601 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; 6602 log.u_bbr.ininput = rack->rc_inp->inp_in_input; 6603 TCP_LOG_EVENT(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0, 6604 tlen, &log, true); 6605 } 6606 if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) { 6607 way_out = 4; 6608 goto done_with_input; 6609 } 6610 /* 6611 * If a segment with the ACK-bit set arrives in the SYN-SENT state 6612 * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9. 6613 */ 6614 if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) && 6615 (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) { 6616 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 6617 return; 6618 } 6619 /* 6620 * Segment received on connection. Reset idle time and keep-alive 6621 * timer. XXX: This should be done after segment validation to 6622 * ignore broken/spoofed segs. 6623 */ 6624 if (tp->t_idle_reduce && (tp->snd_max == tp->snd_una)) { 6625 #ifdef NETFLIX_CWV 6626 if ((tp->cwv_enabled) && 6627 ((tp->cwv_cwnd_valid == 0) && 6628 TCPS_HAVEESTABLISHED(tp->t_state) && 6629 (tp->snd_cwnd > tp->snd_cwv.init_cwnd))) { 6630 tcp_newcwv_nvp_closedown(tp); 6631 } else 6632 #endif 6633 if ((ticks - tp->t_rcvtime) >= tp->t_rxtcur) { 6634 counter_u64_add(rack_input_idle_reduces, 1); 6635 rack_cc_after_idle(tp, 6636 (rack->r_idle_reduce_largest ? 1 :0)); 6637 } 6638 } 6639 rack->r_ctl.rc_rcvtime = cts; 6640 tp->t_rcvtime = ticks; 6641 6642 #ifdef NETFLIX_CWV 6643 if (tp->cwv_enabled) { 6644 if ((tp->cwv_cwnd_valid == 0) && 6645 TCPS_HAVEESTABLISHED(tp->t_state) && 6646 (tp->snd_cwnd > tp->snd_cwv.init_cwnd)) 6647 tcp_newcwv_nvp_closedown(tp); 6648 } 6649 #endif 6650 /* 6651 * Unscale the window into a 32-bit value. For the SYN_SENT state 6652 * the scale is zero. 6653 */ 6654 tiwin = th->th_win << tp->snd_scale; 6655 #ifdef NETFLIX_STATS 6656 stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin); 6657 #endif 6658 /* 6659 * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move 6660 * this to occur after we've validated the segment. 6661 */ 6662 if (tp->t_flags & TF_ECN_PERMIT) { 6663 if (thflags & TH_CWR) 6664 tp->t_flags &= ~TF_ECN_SND_ECE; 6665 switch (iptos & IPTOS_ECN_MASK) { 6666 case IPTOS_ECN_CE: 6667 tp->t_flags |= TF_ECN_SND_ECE; 6668 TCPSTAT_INC(tcps_ecn_ce); 6669 break; 6670 case IPTOS_ECN_ECT0: 6671 TCPSTAT_INC(tcps_ecn_ect0); 6672 break; 6673 case IPTOS_ECN_ECT1: 6674 TCPSTAT_INC(tcps_ecn_ect1); 6675 break; 6676 } 6677 /* Congestion experienced. */ 6678 if (thflags & TH_ECE) { 6679 rack_cong_signal(tp, th, CC_ECN); 6680 } 6681 } 6682 /* 6683 * Parse options on any incoming segment. 6684 */ 6685 tcp_dooptions(&to, (u_char *)(th + 1), 6686 (th->th_off << 2) - sizeof(struct tcphdr), 6687 (thflags & TH_SYN) ? TO_SYN : 0); 6688 6689 /* 6690 * If echoed timestamp is later than the current time, fall back to 6691 * non RFC1323 RTT calculation. Normalize timestamp if syncookies 6692 * were used when this connection was established. 6693 */ 6694 if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) { 6695 to.to_tsecr -= tp->ts_offset; 6696 if (TSTMP_GT(to.to_tsecr, cts)) 6697 to.to_tsecr = 0; 6698 } 6699 /* 6700 * If its the first time in we need to take care of options and 6701 * verify we can do SACK for rack! 6702 */ 6703 if (rack->r_state == 0) { 6704 /* Should be init'd by rack_init() */ 6705 KASSERT(rack->rc_inp != NULL, 6706 ("%s: rack->rc_inp unexpectedly NULL", __func__)); 6707 if (rack->rc_inp == NULL) { 6708 rack->rc_inp = tp->t_inpcb; 6709 } 6710 6711 /* 6712 * Process options only when we get SYN/ACK back. The SYN 6713 * case for incoming connections is handled in tcp_syncache. 6714 * According to RFC1323 the window field in a SYN (i.e., a 6715 * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX 6716 * this is traditional behavior, may need to be cleaned up. 6717 */ 6718 rack->r_cpu = inp_to_cpuid(tp->t_inpcb); 6719 if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) { 6720 if ((to.to_flags & TOF_SCALE) && 6721 (tp->t_flags & TF_REQ_SCALE)) { 6722 tp->t_flags |= TF_RCVD_SCALE; 6723 tp->snd_scale = to.to_wscale; 6724 } 6725 /* 6726 * Initial send window. It will be updated with the 6727 * next incoming segment to the scaled value. 6728 */ 6729 tp->snd_wnd = th->th_win; 6730 if (to.to_flags & TOF_TS) { 6731 tp->t_flags |= TF_RCVD_TSTMP; 6732 tp->ts_recent = to.to_tsval; 6733 tp->ts_recent_age = cts; 6734 } 6735 if (to.to_flags & TOF_MSS) 6736 tcp_mss(tp, to.to_mss); 6737 if ((tp->t_flags & TF_SACK_PERMIT) && 6738 (to.to_flags & TOF_SACKPERM) == 0) 6739 tp->t_flags &= ~TF_SACK_PERMIT; 6740 if (IS_FASTOPEN(tp->t_flags)) { 6741 if (to.to_flags & TOF_FASTOPEN) { 6742 uint16_t mss; 6743 6744 if (to.to_flags & TOF_MSS) 6745 mss = to.to_mss; 6746 else 6747 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0) 6748 mss = TCP6_MSS; 6749 else 6750 mss = TCP_MSS; 6751 tcp_fastopen_update_cache(tp, mss, 6752 to.to_tfo_len, to.to_tfo_cookie); 6753 } else 6754 tcp_fastopen_disable_path(tp); 6755 } 6756 } 6757 /* 6758 * At this point we are at the initial call. Here we decide 6759 * if we are doing RACK or not. We do this by seeing if 6760 * TF_SACK_PERMIT is set, if not rack is *not* possible and 6761 * we switch to the default code. 6762 */ 6763 if ((tp->t_flags & TF_SACK_PERMIT) == 0) { 6764 tcp_switch_back_to_default(tp); 6765 (*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen, 6766 tlen, iptos); 6767 return; 6768 } 6769 /* Set the flag */ 6770 rack->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0; 6771 tcp_set_hpts(tp->t_inpcb); 6772 rack_stop_all_timers(tp); 6773 sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack); 6774 } 6775 /* 6776 * This is the one exception case where we set the rack state 6777 * always. All other times (timers etc) we must have a rack-state 6778 * set (so we assure we have done the checks above for SACK). 6779 */ 6780 if (rack->r_state != tp->t_state) 6781 rack_set_state(tp, rack); 6782 if (SEQ_GT(th->th_ack, tp->snd_una) && (rsm = TAILQ_FIRST(&rack->r_ctl.rc_map)) != NULL) 6783 kern_prefetch(rsm, &prev_state); 6784 prev_state = rack->r_state; 6785 rack->r_ctl.rc_tlp_send_cnt = 0; 6786 rack_clear_rate_sample(rack); 6787 retval = (*rack->r_substate) (m, th, so, 6788 tp, &to, drop_hdrlen, 6789 tlen, tiwin, thflags, nxt_pkt); 6790 #ifdef INVARIANTS 6791 if ((retval == 0) && 6792 (tp->t_inpcb == NULL)) { 6793 panic("retval:%d tp:%p t_inpcb:NULL state:%d", 6794 retval, tp, prev_state); 6795 } 6796 #endif 6797 if (retval == 0) { 6798 /* 6799 * If retval is 1 the tcb is unlocked and most likely the tp 6800 * is gone. 6801 */ 6802 INP_WLOCK_ASSERT(tp->t_inpcb); 6803 tcp_rack_xmit_timer_commit(rack, tp); 6804 if (((tp->snd_max - tp->snd_una) > tp->snd_wnd) && 6805 (rack->rc_in_persist == 0)){ 6806 /* 6807 * The peer shrunk its window on us to the point 6808 * where we have sent too much. The only thing 6809 * we can do here is stop any timers and 6810 * enter persist. We most likely lost the last 6811 * bytes we sent but oh well, we will have to 6812 * retransmit them after the peer is caught up. 6813 */ 6814 if (rack->rc_inp->inp_in_hpts) 6815 tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT); 6816 rack_timer_cancel(tp, rack, cts, __LINE__); 6817 rack_enter_persist(tp, rack, cts); 6818 rack_start_hpts_timer(rack, tp, tcp_ts_getticks(), __LINE__, 0, 0, 0); 6819 way_out = 3; 6820 goto done_with_input; 6821 } 6822 if (nxt_pkt == 0) { 6823 if (rack->r_wanted_output != 0) { 6824 did_out = 1; 6825 (void)tp->t_fb->tfb_tcp_output(tp); 6826 } 6827 rack_start_hpts_timer(rack, tp, cts, __LINE__, 0, 0, 0); 6828 } 6829 if (((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) && 6830 (SEQ_GT(tp->snd_max, tp->snd_una) || 6831 (tp->t_flags & TF_DELACK) || 6832 ((tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) && 6833 (tp->t_state <= TCPS_CLOSING)))) { 6834 /* We could not send (probably in the hpts but stopped the timer earlier)? */ 6835 if ((tp->snd_max == tp->snd_una) && 6836 ((tp->t_flags & TF_DELACK) == 0) && 6837 (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) { 6838 /* keep alive not needed if we are hptsi output yet */ 6839 ; 6840 } else { 6841 if (rack->rc_inp->inp_in_hpts) 6842 tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT); 6843 rack_start_hpts_timer(rack, tp, tcp_ts_getticks(), __LINE__, 0, 0, 0); 6844 } 6845 way_out = 1; 6846 } else { 6847 /* Do we have the correct timer running? */ 6848 rack_timer_audit(tp, rack, &so->so_snd); 6849 way_out = 2; 6850 } 6851 done_with_input: 6852 rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out); 6853 if (did_out) 6854 rack->r_wanted_output = 0; 6855 #ifdef INVARIANTS 6856 if (tp->t_inpcb == NULL) { 6857 panic("OP:%d retval:%d tp:%p t_inpcb:NULL state:%d", 6858 did_out, 6859 retval, tp, prev_state); 6860 } 6861 #endif 6862 INP_WUNLOCK(tp->t_inpcb); 6863 } 6864 } 6865 6866 void 6867 rack_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so, 6868 struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos) 6869 { 6870 struct timeval tv; 6871 #ifdef RSS 6872 struct tcp_function_block *tfb; 6873 struct tcp_rack *rack; 6874 struct epoch_tracker et; 6875 6876 rack = (struct tcp_rack *)tp->t_fb_ptr; 6877 if (rack->r_state == 0) { 6878 /* 6879 * Initial input (ACK to SYN-ACK etc)lets go ahead and get 6880 * it processed 6881 */ 6882 INP_INFO_RLOCK_ET(&V_tcbinfo, et); 6883 tcp_get_usecs(&tv); 6884 rack_hpts_do_segment(m, th, so, tp, drop_hdrlen, 6885 tlen, iptos, 0, &tv); 6886 INP_INFO_RUNLOCK_ET(&V_tcbinfo, et); 6887 return; 6888 } 6889 tcp_queue_to_input(tp, m, th, tlen, drop_hdrlen, iptos); 6890 INP_WUNLOCK(tp->t_inpcb); 6891 #else 6892 tcp_get_usecs(&tv); 6893 rack_hpts_do_segment(m, th, so, tp, drop_hdrlen, 6894 tlen, iptos, 0, &tv); 6895 #endif 6896 } 6897 6898 struct rack_sendmap * 6899 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused) 6900 { 6901 struct rack_sendmap *rsm = NULL; 6902 int32_t idx; 6903 uint32_t srtt_cur, srtt = 0, thresh = 0, ts_low = 0; 6904 6905 /* Return the next guy to be re-transmitted */ 6906 if (TAILQ_EMPTY(&rack->r_ctl.rc_map)) { 6907 return (NULL); 6908 } 6909 if (tp->t_flags & TF_SENTFIN) { 6910 /* retran the end FIN? */ 6911 return (NULL); 6912 } 6913 /* ok lets look at this one */ 6914 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); 6915 if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) { 6916 goto check_it; 6917 } 6918 rsm = rack_find_lowest_rsm(rack); 6919 if (rsm == NULL) { 6920 return (NULL); 6921 } 6922 check_it: 6923 srtt_cur = tp->t_srtt >> TCP_RTT_SHIFT; 6924 srtt = TICKS_2_MSEC(srtt_cur); 6925 if (rack->rc_rack_rtt && (srtt > rack->rc_rack_rtt)) 6926 srtt = rack->rc_rack_rtt; 6927 if (rsm->r_flags & RACK_ACKED) { 6928 return (NULL); 6929 } 6930 if ((rsm->r_flags & RACK_SACK_PASSED) == 0) { 6931 /* Its not yet ready */ 6932 return (NULL); 6933 } 6934 idx = rsm->r_rtr_cnt - 1; 6935 ts_low = rsm->r_tim_lastsent[idx]; 6936 thresh = rack_calc_thresh_rack(rack, srtt, tsused); 6937 if (tsused <= ts_low) { 6938 return (NULL); 6939 } 6940 if ((tsused - ts_low) >= thresh) { 6941 return (rsm); 6942 } 6943 return (NULL); 6944 } 6945 6946 static int 6947 rack_output(struct tcpcb *tp) 6948 { 6949 struct socket *so; 6950 uint32_t recwin, sendwin; 6951 uint32_t sb_offset; 6952 int32_t len, flags, error = 0; 6953 struct mbuf *m; 6954 struct mbuf *mb; 6955 uint32_t if_hw_tsomaxsegcount = 0; 6956 uint32_t if_hw_tsomaxsegsize; 6957 long tot_len_this_send = 0; 6958 struct ip *ip = NULL; 6959 #ifdef TCPDEBUG 6960 struct ipovly *ipov = NULL; 6961 #endif 6962 struct udphdr *udp = NULL; 6963 struct tcp_rack *rack; 6964 struct tcphdr *th; 6965 uint8_t pass = 0; 6966 uint8_t wanted_cookie = 0; 6967 u_char opt[TCP_MAXOLEN]; 6968 unsigned ipoptlen, optlen, hdrlen, ulen=0; 6969 uint32_t rack_seq; 6970 6971 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 6972 unsigned ipsec_optlen = 0; 6973 6974 #endif 6975 int32_t idle, sendalot; 6976 int32_t sub_from_prr = 0; 6977 volatile int32_t sack_rxmit; 6978 struct rack_sendmap *rsm = NULL; 6979 int32_t tso, mtu, would_have_fin = 0; 6980 struct tcpopt to; 6981 int32_t slot = 0; 6982 uint32_t cts; 6983 uint8_t hpts_calling, doing_tlp = 0; 6984 int32_t do_a_prefetch; 6985 int32_t prefetch_rsm = 0; 6986 int32_t prefetch_so_done = 0; 6987 struct tcp_log_buffer *lgb = NULL; 6988 struct inpcb *inp; 6989 struct sockbuf *sb; 6990 #ifdef INET6 6991 struct ip6_hdr *ip6 = NULL; 6992 int32_t isipv6; 6993 #endif 6994 /* setup and take the cache hits here */ 6995 rack = (struct tcp_rack *)tp->t_fb_ptr; 6996 inp = rack->rc_inp; 6997 so = inp->inp_socket; 6998 sb = &so->so_snd; 6999 kern_prefetch(sb, &do_a_prefetch); 7000 do_a_prefetch = 1; 7001 7002 INP_WLOCK_ASSERT(inp); 7003 #ifdef TCP_OFFLOAD 7004 if (tp->t_flags & TF_TOE) 7005 return (tcp_offload_output(tp)); 7006 #endif 7007 #ifdef INET6 7008 if (rack->r_state) { 7009 /* Use the cache line loaded if possible */ 7010 isipv6 = rack->r_is_v6; 7011 } else { 7012 isipv6 = (inp->inp_vflag & INP_IPV6) != 0; 7013 } 7014 #endif 7015 cts = tcp_ts_getticks(); 7016 if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) && 7017 inp->inp_in_hpts) { 7018 /* 7019 * We are on the hpts for some timer but not hptsi output. 7020 * Remove from the hpts unconditionally. 7021 */ 7022 rack_timer_cancel(tp, rack, cts, __LINE__); 7023 } 7024 /* Mark that we have called rack_output(). */ 7025 if ((rack->r_timer_override) || 7026 (tp->t_flags & TF_FORCEDATA) || 7027 (tp->t_state < TCPS_ESTABLISHED)) { 7028 if (tp->t_inpcb->inp_in_hpts) 7029 tcp_hpts_remove(tp->t_inpcb, HPTS_REMOVE_OUTPUT); 7030 } else if (tp->t_inpcb->inp_in_hpts) { 7031 /* 7032 * On the hpts you can't pass even if ACKNOW is on, we will 7033 * when the hpts fires. 7034 */ 7035 counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1); 7036 return (0); 7037 } 7038 hpts_calling = inp->inp_hpts_calls; 7039 inp->inp_hpts_calls = 0; 7040 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) { 7041 if (rack_process_timers(tp, rack, cts, hpts_calling)) { 7042 counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1); 7043 return (0); 7044 } 7045 } 7046 rack->r_wanted_output = 0; 7047 rack->r_timer_override = 0; 7048 /* 7049 * For TFO connections in SYN_SENT or SYN_RECEIVED, 7050 * only allow the initial SYN or SYN|ACK and those sent 7051 * by the retransmit timer. 7052 */ 7053 if (IS_FASTOPEN(tp->t_flags) && 7054 ((tp->t_state == TCPS_SYN_RECEIVED) || 7055 (tp->t_state == TCPS_SYN_SENT)) && 7056 SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */ 7057 (tp->t_rxtshift == 0)) /* not a retransmit */ 7058 return (0); 7059 /* 7060 * Determine length of data that should be transmitted, and flags 7061 * that will be used. If there is some data or critical controls 7062 * (SYN, RST) to send, then transmit; otherwise, investigate 7063 * further. 7064 */ 7065 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una); 7066 #ifdef NETFLIX_CWV 7067 if (tp->cwv_enabled) { 7068 if ((tp->cwv_cwnd_valid == 0) && 7069 TCPS_HAVEESTABLISHED(tp->t_state) && 7070 (tp->snd_cwnd > tp->snd_cwv.init_cwnd)) 7071 tcp_newcwv_nvp_closedown(tp); 7072 } else 7073 #endif 7074 if (tp->t_idle_reduce) { 7075 if (idle && ((ticks - tp->t_rcvtime) >= tp->t_rxtcur)) 7076 rack_cc_after_idle(tp, 7077 (rack->r_idle_reduce_largest ? 1 :0)); 7078 } 7079 tp->t_flags &= ~TF_LASTIDLE; 7080 if (idle) { 7081 if (tp->t_flags & TF_MORETOCOME) { 7082 tp->t_flags |= TF_LASTIDLE; 7083 idle = 0; 7084 } 7085 } 7086 again: 7087 /* 7088 * If we've recently taken a timeout, snd_max will be greater than 7089 * snd_nxt. There may be SACK information that allows us to avoid 7090 * resending already delivered data. Adjust snd_nxt accordingly. 7091 */ 7092 sendalot = 0; 7093 cts = tcp_ts_getticks(); 7094 tso = 0; 7095 mtu = 0; 7096 sb_offset = tp->snd_max - tp->snd_una; 7097 sendwin = min(tp->snd_wnd, tp->snd_cwnd); 7098 7099 flags = tcp_outflags[tp->t_state]; 7100 /* 7101 * Send any SACK-generated retransmissions. If we're explicitly 7102 * trying to send out new data (when sendalot is 1), bypass this 7103 * function. If we retransmit in fast recovery mode, decrement 7104 * snd_cwnd, since we're replacing a (future) new transmission with 7105 * a retransmission now, and we previously incremented snd_cwnd in 7106 * tcp_input(). 7107 */ 7108 /* 7109 * Still in sack recovery , reset rxmit flag to zero. 7110 */ 7111 while (rack->rc_free_cnt < rack_free_cache) { 7112 rsm = rack_alloc(rack); 7113 if (rsm == NULL) { 7114 if (inp->inp_hpts_calls) 7115 /* Retry in a ms */ 7116 slot = 1; 7117 goto just_return_nolock; 7118 } 7119 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_next); 7120 rack->rc_free_cnt++; 7121 rsm = NULL; 7122 } 7123 if (inp->inp_hpts_calls) 7124 inp->inp_hpts_calls = 0; 7125 sack_rxmit = 0; 7126 len = 0; 7127 rsm = NULL; 7128 if (flags & TH_RST) { 7129 SOCKBUF_LOCK(sb); 7130 goto send; 7131 } 7132 if (rack->r_ctl.rc_tlpsend) { 7133 /* Tail loss probe */ 7134 long cwin; 7135 long tlen; 7136 7137 doing_tlp = 1; 7138 rsm = rack->r_ctl.rc_tlpsend; 7139 rack->r_ctl.rc_tlpsend = NULL; 7140 sack_rxmit = 1; 7141 tlen = rsm->r_end - rsm->r_start; 7142 if (tlen > tp->t_maxseg) 7143 tlen = tp->t_maxseg; 7144 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start), 7145 ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p", 7146 __func__, __LINE__, 7147 rsm->r_start, tp->snd_una, tp, rack, rsm)); 7148 sb_offset = rsm->r_start - tp->snd_una; 7149 cwin = min(tp->snd_wnd, tlen); 7150 len = cwin; 7151 } else if (rack->r_ctl.rc_resend) { 7152 /* Retransmit timer */ 7153 rsm = rack->r_ctl.rc_resend; 7154 rack->r_ctl.rc_resend = NULL; 7155 len = rsm->r_end - rsm->r_start; 7156 sack_rxmit = 1; 7157 sendalot = 0; 7158 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start), 7159 ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p", 7160 __func__, __LINE__, 7161 rsm->r_start, tp->snd_una, tp, rack, rsm)); 7162 sb_offset = rsm->r_start - tp->snd_una; 7163 if (len >= tp->t_maxseg) { 7164 len = tp->t_maxseg; 7165 } 7166 } else if ((rack->rc_in_persist == 0) && 7167 ((rsm = tcp_rack_output(tp, rack, cts)) != NULL)) { 7168 long tlen; 7169 7170 if ((!IN_RECOVERY(tp->t_flags)) && 7171 ((tp->t_flags & (TF_WASFRECOVERY | TF_WASCRECOVERY)) == 0)) { 7172 /* Enter recovery if not induced by a time-out */ 7173 rack->r_ctl.rc_rsm_start = rsm->r_start; 7174 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd; 7175 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh; 7176 rack_cong_signal(tp, NULL, CC_NDUPACK); 7177 /* 7178 * When we enter recovery we need to assure we send 7179 * one packet. 7180 */ 7181 rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; 7182 } 7183 #ifdef INVARIANTS 7184 if (SEQ_LT(rsm->r_start, tp->snd_una)) { 7185 panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n", 7186 tp, rack, rsm, rsm->r_start, tp->snd_una); 7187 } 7188 #endif 7189 tlen = rsm->r_end - rsm->r_start; 7190 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start), 7191 ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p", 7192 __func__, __LINE__, 7193 rsm->r_start, tp->snd_una, tp, rack, rsm)); 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 if (len > 0) { 7216 sub_from_prr = 1; 7217 sack_rxmit = 1; 7218 TCPSTAT_INC(tcps_sack_rexmits); 7219 TCPSTAT_ADD(tcps_sack_rexmit_bytes, 7220 min(len, tp->t_maxseg)); 7221 counter_u64_add(rack_rtm_prr_retran, 1); 7222 } 7223 } 7224 if (rsm && (rsm->r_flags & RACK_HAS_FIN)) { 7225 /* we are retransmitting the fin */ 7226 len--; 7227 if (len) { 7228 /* 7229 * When retransmitting data do *not* include the 7230 * FIN. This could happen from a TLP probe. 7231 */ 7232 flags &= ~TH_FIN; 7233 } 7234 } 7235 #ifdef INVARIANTS 7236 /* For debugging */ 7237 rack->r_ctl.rc_rsm_at_retran = rsm; 7238 #endif 7239 /* 7240 * Get standard flags, and add SYN or FIN if requested by 'hidden' 7241 * state flags. 7242 */ 7243 if (tp->t_flags & TF_NEEDFIN) 7244 flags |= TH_FIN; 7245 if (tp->t_flags & TF_NEEDSYN) 7246 flags |= TH_SYN; 7247 if ((sack_rxmit == 0) && (prefetch_rsm == 0)) { 7248 void *end_rsm; 7249 end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext); 7250 if (end_rsm) 7251 kern_prefetch(end_rsm, &prefetch_rsm); 7252 prefetch_rsm = 1; 7253 } 7254 SOCKBUF_LOCK(sb); 7255 /* 7256 * If in persist timeout with window of 0, send 1 byte. Otherwise, 7257 * if window is small but nonzero and time TF_SENTFIN expired, we 7258 * will send what we can and go to transmit state. 7259 */ 7260 if (tp->t_flags & TF_FORCEDATA) { 7261 if (sendwin == 0) { 7262 /* 7263 * If we still have some data to send, then clear 7264 * the FIN bit. Usually this would happen below 7265 * when it realizes that we aren't sending all the 7266 * data. However, if we have exactly 1 byte of 7267 * unsent data, then it won't clear the FIN bit 7268 * below, and if we are in persist state, we wind up 7269 * sending the packet without recording that we sent 7270 * the FIN bit. 7271 * 7272 * We can't just blindly clear the FIN bit, because 7273 * if we don't have any more data to send then the 7274 * probe will be the FIN itself. 7275 */ 7276 if (sb_offset < sbused(sb)) 7277 flags &= ~TH_FIN; 7278 sendwin = 1; 7279 } else { 7280 if (rack->rc_in_persist) 7281 rack_exit_persist(tp, rack); 7282 /* 7283 * If we are dropping persist mode then we need to 7284 * correct snd_nxt/snd_max and off. 7285 */ 7286 tp->snd_nxt = tp->snd_max; 7287 sb_offset = tp->snd_nxt - tp->snd_una; 7288 } 7289 } 7290 /* 7291 * If snd_nxt == snd_max and we have transmitted a FIN, the 7292 * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a 7293 * negative length. This can also occur when TCP opens up its 7294 * congestion window while receiving additional duplicate acks after 7295 * fast-retransmit because TCP will reset snd_nxt to snd_max after 7296 * the fast-retransmit. 7297 * 7298 * In the normal retransmit-FIN-only case, however, snd_nxt will be 7299 * set to snd_una, the sb_offset will be 0, and the length may wind 7300 * up 0. 7301 * 7302 * If sack_rxmit is true we are retransmitting from the scoreboard 7303 * in which case len is already set. 7304 */ 7305 if (sack_rxmit == 0) { 7306 uint32_t avail; 7307 7308 avail = sbavail(sb); 7309 if (SEQ_GT(tp->snd_nxt, tp->snd_una) && avail) 7310 sb_offset = tp->snd_nxt - tp->snd_una; 7311 else 7312 sb_offset = 0; 7313 if (IN_RECOVERY(tp->t_flags) == 0) { 7314 if (rack->r_ctl.rc_tlp_new_data) { 7315 /* TLP is forcing out new data */ 7316 if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) { 7317 rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset); 7318 } 7319 if (rack->r_ctl.rc_tlp_new_data > tp->snd_wnd) 7320 len = tp->snd_wnd; 7321 else 7322 len = rack->r_ctl.rc_tlp_new_data; 7323 rack->r_ctl.rc_tlp_new_data = 0; 7324 doing_tlp = 1; 7325 } else { 7326 if (sendwin > avail) { 7327 /* use the available */ 7328 if (avail > sb_offset) { 7329 len = (int32_t)(avail - sb_offset); 7330 } else { 7331 len = 0; 7332 } 7333 } else { 7334 if (sendwin > sb_offset) { 7335 len = (int32_t)(sendwin - sb_offset); 7336 } else { 7337 len = 0; 7338 } 7339 } 7340 } 7341 } else { 7342 uint32_t outstanding; 7343 7344 /* 7345 * We are inside of a SACK recovery episode and are 7346 * sending new data, having retransmitted all the 7347 * data possible so far in the scoreboard. 7348 */ 7349 outstanding = tp->snd_max - tp->snd_una; 7350 if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) 7351 len = 0; 7352 else if (avail > sb_offset) 7353 len = avail - sb_offset; 7354 else 7355 len = 0; 7356 if (len > 0) { 7357 if (len > rack->r_ctl.rc_prr_sndcnt) 7358 len = rack->r_ctl.rc_prr_sndcnt; 7359 7360 if (len > 0) { 7361 sub_from_prr = 1; 7362 counter_u64_add(rack_rtm_prr_newdata, 1); 7363 } 7364 } 7365 if (len > tp->t_maxseg) { 7366 /* 7367 * We should never send more than a MSS when 7368 * retransmitting or sending new data in prr 7369 * mode unless the override flag is on. Most 7370 * likely the PRR algorithm is not going to 7371 * let us send a lot as well :-) 7372 */ 7373 if (rack->r_ctl.rc_prr_sendalot == 0) 7374 len = tp->t_maxseg; 7375 } else if (len < tp->t_maxseg) { 7376 /* 7377 * Do we send any? The idea here is if the 7378 * send empty's the socket buffer we want to 7379 * do it. However if not then lets just wait 7380 * for our prr_sndcnt to get bigger. 7381 */ 7382 long leftinsb; 7383 7384 leftinsb = sbavail(sb) - sb_offset; 7385 if (leftinsb > len) { 7386 /* This send does not empty the sb */ 7387 len = 0; 7388 } 7389 } 7390 } 7391 } 7392 if (prefetch_so_done == 0) { 7393 kern_prefetch(so, &prefetch_so_done); 7394 prefetch_so_done = 1; 7395 } 7396 /* 7397 * Lop off SYN bit if it has already been sent. However, if this is 7398 * SYN-SENT state and if segment contains data and if we don't know 7399 * that foreign host supports TAO, suppress sending segment. 7400 */ 7401 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una) && 7402 ((sack_rxmit == 0) && (tp->t_rxtshift == 0))) { 7403 if (tp->t_state != TCPS_SYN_RECEIVED) 7404 flags &= ~TH_SYN; 7405 /* 7406 * When sending additional segments following a TFO SYN|ACK, 7407 * do not include the SYN bit. 7408 */ 7409 if (IS_FASTOPEN(tp->t_flags) && 7410 (tp->t_state == TCPS_SYN_RECEIVED)) 7411 flags &= ~TH_SYN; 7412 sb_offset--, len++; 7413 } 7414 /* 7415 * Be careful not to send data and/or FIN on SYN segments. This 7416 * measure is needed to prevent interoperability problems with not 7417 * fully conformant TCP implementations. 7418 */ 7419 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) { 7420 len = 0; 7421 flags &= ~TH_FIN; 7422 } 7423 /* 7424 * On TFO sockets, ensure no data is sent in the following cases: 7425 * 7426 * - When retransmitting SYN|ACK on a passively-created socket 7427 * 7428 * - When retransmitting SYN on an actively created socket 7429 * 7430 * - When sending a zero-length cookie (cookie request) on an 7431 * actively created socket 7432 * 7433 * - When the socket is in the CLOSED state (RST is being sent) 7434 */ 7435 if (IS_FASTOPEN(tp->t_flags) && 7436 (((flags & TH_SYN) && (tp->t_rxtshift > 0)) || 7437 ((tp->t_state == TCPS_SYN_SENT) && 7438 (tp->t_tfo_client_cookie_len == 0)) || 7439 (flags & TH_RST))) { 7440 sack_rxmit = 0; 7441 len = 0; 7442 } 7443 /* Without fast-open there should never be data sent on a SYN */ 7444 if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags))) 7445 len = 0; 7446 if (len <= 0) { 7447 /* 7448 * If FIN has been sent but not acked, but we haven't been 7449 * called to retransmit, len will be < 0. Otherwise, window 7450 * shrank after we sent into it. If window shrank to 0, 7451 * cancel pending retransmit, pull snd_nxt back to (closed) 7452 * window, and set the persist timer if it isn't already 7453 * going. If the window didn't close completely, just wait 7454 * for an ACK. 7455 * 7456 * We also do a general check here to ensure that we will 7457 * set the persist timer when we have data to send, but a 7458 * 0-byte window. This makes sure the persist timer is set 7459 * even if the packet hits one of the "goto send" lines 7460 * below. 7461 */ 7462 len = 0; 7463 if ((tp->snd_wnd == 0) && 7464 (TCPS_HAVEESTABLISHED(tp->t_state)) && 7465 (sb_offset < (int)sbavail(sb))) { 7466 tp->snd_nxt = tp->snd_una; 7467 rack_enter_persist(tp, rack, cts); 7468 } 7469 } 7470 /* len will be >= 0 after this point. */ 7471 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 7472 tcp_sndbuf_autoscale(tp, so, sendwin); 7473 /* 7474 * Decide if we can use TCP Segmentation Offloading (if supported by 7475 * hardware). 7476 * 7477 * TSO may only be used if we are in a pure bulk sending state. The 7478 * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP 7479 * options prevent using TSO. With TSO the TCP header is the same 7480 * (except for the sequence number) for all generated packets. This 7481 * makes it impossible to transmit any options which vary per 7482 * generated segment or packet. 7483 * 7484 * IPv4 handling has a clear separation of ip options and ip header 7485 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does 7486 * the right thing below to provide length of just ip options and thus 7487 * checking for ipoptlen is enough to decide if ip options are present. 7488 */ 7489 7490 #ifdef INET6 7491 if (isipv6) 7492 ipoptlen = ip6_optlen(tp->t_inpcb); 7493 else 7494 #endif 7495 if (tp->t_inpcb->inp_options) 7496 ipoptlen = tp->t_inpcb->inp_options->m_len - 7497 offsetof(struct ipoption, ipopt_list); 7498 else 7499 ipoptlen = 0; 7500 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 7501 /* 7502 * Pre-calculate here as we save another lookup into the darknesses 7503 * of IPsec that way and can actually decide if TSO is ok. 7504 */ 7505 #ifdef INET6 7506 if (isipv6 && IPSEC_ENABLED(ipv6)) 7507 ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb); 7508 #ifdef INET 7509 else 7510 #endif 7511 #endif /* INET6 */ 7512 #ifdef INET 7513 if (IPSEC_ENABLED(ipv4)) 7514 ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb); 7515 #endif /* INET */ 7516 #endif 7517 7518 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 7519 ipoptlen += ipsec_optlen; 7520 #endif 7521 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg && 7522 (tp->t_port == 0) && 7523 ((tp->t_flags & TF_SIGNATURE) == 0) && 7524 tp->rcv_numsacks == 0 && sack_rxmit == 0 && 7525 ipoptlen == 0) 7526 tso = 1; 7527 { 7528 uint32_t outstanding; 7529 7530 outstanding = tp->snd_max - tp->snd_una; 7531 if (tp->t_flags & TF_SENTFIN) { 7532 /* 7533 * If we sent a fin, snd_max is 1 higher than 7534 * snd_una 7535 */ 7536 outstanding--; 7537 } 7538 if (outstanding > 0) { 7539 /* 7540 * This is sub-optimal. We only send a stand alone 7541 * FIN on its own segment. 7542 */ 7543 if (flags & TH_FIN) { 7544 flags &= ~TH_FIN; 7545 would_have_fin = 1; 7546 } 7547 } else if (sack_rxmit) { 7548 if ((rsm->r_flags & RACK_HAS_FIN) == 0) 7549 flags &= ~TH_FIN; 7550 } else { 7551 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + 7552 sbused(sb))) 7553 flags &= ~TH_FIN; 7554 } 7555 } 7556 recwin = sbspace(&so->so_rcv); 7557 7558 /* 7559 * Sender silly window avoidance. We transmit under the following 7560 * conditions when len is non-zero: 7561 * 7562 * - We have a full segment (or more with TSO) - This is the last 7563 * buffer in a write()/send() and we are either idle or running 7564 * NODELAY - we've timed out (e.g. persist timer) - we have more 7565 * then 1/2 the maximum send window's worth of data (receiver may be 7566 * limited the window size) - we need to retransmit 7567 */ 7568 if (len) { 7569 if (len >= tp->t_maxseg) { 7570 pass = 1; 7571 goto send; 7572 } 7573 /* 7574 * NOTE! on localhost connections an 'ack' from the remote 7575 * end may occur synchronously with the output and cause us 7576 * to flush a buffer queued with moretocome. XXX 7577 * 7578 */ 7579 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */ 7580 (idle || (tp->t_flags & TF_NODELAY)) && 7581 ((uint32_t)len + (uint32_t)sb_offset >= sbavail(&so->so_snd)) && 7582 (tp->t_flags & TF_NOPUSH) == 0) { 7583 pass = 2; 7584 goto send; 7585 } 7586 if (tp->t_flags & TF_FORCEDATA) { /* typ. timeout case */ 7587 pass = 3; 7588 goto send; 7589 } 7590 if ((tp->snd_una == tp->snd_max) && len) { /* Nothing outstanding */ 7591 goto send; 7592 } 7593 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) { 7594 pass = 4; 7595 goto send; 7596 } 7597 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { /* retransmit case */ 7598 pass = 5; 7599 goto send; 7600 } 7601 if (sack_rxmit) { 7602 pass = 6; 7603 goto send; 7604 } 7605 } 7606 /* 7607 * Sending of standalone window updates. 7608 * 7609 * Window updates are important when we close our window due to a 7610 * full socket buffer and are opening it again after the application 7611 * reads data from it. Once the window has opened again and the 7612 * remote end starts to send again the ACK clock takes over and 7613 * provides the most current window information. 7614 * 7615 * We must avoid the silly window syndrome whereas every read from 7616 * the receive buffer, no matter how small, causes a window update 7617 * to be sent. We also should avoid sending a flurry of window 7618 * updates when the socket buffer had queued a lot of data and the 7619 * application is doing small reads. 7620 * 7621 * Prevent a flurry of pointless window updates by only sending an 7622 * update when we can increase the advertized window by more than 7623 * 1/4th of the socket buffer capacity. When the buffer is getting 7624 * full or is very small be more aggressive and send an update 7625 * whenever we can increase by two mss sized segments. In all other 7626 * situations the ACK's to new incoming data will carry further 7627 * window increases. 7628 * 7629 * Don't send an independent window update if a delayed ACK is 7630 * pending (it will get piggy-backed on it) or the remote side 7631 * already has done a half-close and won't send more data. Skip 7632 * this if the connection is in T/TCP half-open state. 7633 */ 7634 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) && 7635 !(tp->t_flags & TF_DELACK) && 7636 !TCPS_HAVERCVDFIN(tp->t_state)) { 7637 /* 7638 * "adv" is the amount we could increase the window, taking 7639 * into account that we are limited by TCP_MAXWIN << 7640 * tp->rcv_scale. 7641 */ 7642 int32_t adv; 7643 int oldwin; 7644 7645 adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale); 7646 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) { 7647 oldwin = (tp->rcv_adv - tp->rcv_nxt); 7648 adv -= oldwin; 7649 } else 7650 oldwin = 0; 7651 7652 /* 7653 * If the new window size ends up being the same as the old 7654 * size when it is scaled, then don't force a window update. 7655 */ 7656 if (oldwin >> tp->rcv_scale == (adv + oldwin) >> tp->rcv_scale) 7657 goto dontupdate; 7658 7659 if (adv >= (int32_t)(2 * tp->t_maxseg) && 7660 (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) || 7661 recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) || 7662 so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg)) { 7663 pass = 7; 7664 goto send; 7665 } 7666 if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) 7667 goto send; 7668 } 7669 dontupdate: 7670 7671 /* 7672 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW 7673 * is also a catch-all for the retransmit timer timeout case. 7674 */ 7675 if (tp->t_flags & TF_ACKNOW) { 7676 pass = 8; 7677 goto send; 7678 } 7679 if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) { 7680 pass = 9; 7681 goto send; 7682 } 7683 if (SEQ_GT(tp->snd_up, tp->snd_una)) { 7684 pass = 10; 7685 goto send; 7686 } 7687 /* 7688 * If our state indicates that FIN should be sent and we have not 7689 * yet done so, then we need to send. 7690 */ 7691 if ((flags & TH_FIN) && 7692 (tp->snd_nxt == tp->snd_una)) { 7693 pass = 11; 7694 goto send; 7695 } 7696 /* 7697 * No reason to send a segment, just return. 7698 */ 7699 just_return: 7700 SOCKBUF_UNLOCK(sb); 7701 just_return_nolock: 7702 if (tot_len_this_send == 0) 7703 counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1); 7704 rack_start_hpts_timer(rack, tp, cts, __LINE__, slot, tot_len_this_send, 1); 7705 rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling); 7706 tp->t_flags &= ~TF_FORCEDATA; 7707 return (0); 7708 7709 send: 7710 if (doing_tlp == 0) { 7711 /* 7712 * Data not a TLP, and its not the rxt firing. If it is the 7713 * rxt firing, we want to leave the tlp_in_progress flag on 7714 * so we don't send another TLP. It has to be a rack timer 7715 * or normal send (response to acked data) to clear the tlp 7716 * in progress flag. 7717 */ 7718 rack->rc_tlp_in_progress = 0; 7719 } 7720 SOCKBUF_LOCK_ASSERT(sb); 7721 if (len > 0) { 7722 if (len >= tp->t_maxseg) 7723 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT; 7724 else 7725 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT; 7726 } 7727 /* 7728 * Before ESTABLISHED, force sending of initial options unless TCP 7729 * set not to do any options. NOTE: we assume that the IP/TCP header 7730 * plus TCP options always fit in a single mbuf, leaving room for a 7731 * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr) 7732 * + optlen <= MCLBYTES 7733 */ 7734 optlen = 0; 7735 #ifdef INET6 7736 if (isipv6) 7737 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 7738 else 7739 #endif 7740 hdrlen = sizeof(struct tcpiphdr); 7741 7742 /* 7743 * Compute options for segment. We only have to care about SYN and 7744 * established connection segments. Options for SYN-ACK segments 7745 * are handled in TCP syncache. 7746 */ 7747 to.to_flags = 0; 7748 if ((tp->t_flags & TF_NOOPT) == 0) { 7749 /* Maximum segment size. */ 7750 if (flags & TH_SYN) { 7751 tp->snd_nxt = tp->iss; 7752 to.to_mss = tcp_mssopt(&inp->inp_inc); 7753 #ifdef NETFLIX_TCPOUDP 7754 if (tp->t_port) 7755 to.to_mss -= V_tcp_udp_tunneling_overhead; 7756 #endif 7757 to.to_flags |= TOF_MSS; 7758 7759 /* 7760 * On SYN or SYN|ACK transmits on TFO connections, 7761 * only include the TFO option if it is not a 7762 * retransmit, as the presence of the TFO option may 7763 * have caused the original SYN or SYN|ACK to have 7764 * been dropped by a middlebox. 7765 */ 7766 if (IS_FASTOPEN(tp->t_flags) && 7767 (tp->t_rxtshift == 0)) { 7768 if (tp->t_state == TCPS_SYN_RECEIVED) { 7769 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; 7770 to.to_tfo_cookie = 7771 (u_int8_t *)&tp->t_tfo_cookie.server; 7772 to.to_flags |= TOF_FASTOPEN; 7773 wanted_cookie = 1; 7774 } else if (tp->t_state == TCPS_SYN_SENT) { 7775 to.to_tfo_len = 7776 tp->t_tfo_client_cookie_len; 7777 to.to_tfo_cookie = 7778 tp->t_tfo_cookie.client; 7779 to.to_flags |= TOF_FASTOPEN; 7780 wanted_cookie = 1; 7781 /* 7782 * If we wind up having more data to 7783 * send with the SYN than can fit in 7784 * one segment, don't send any more 7785 * until the SYN|ACK comes back from 7786 * the other end. 7787 */ 7788 sendalot = 0; 7789 } 7790 } 7791 } 7792 /* Window scaling. */ 7793 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) { 7794 to.to_wscale = tp->request_r_scale; 7795 to.to_flags |= TOF_SCALE; 7796 } 7797 /* Timestamps. */ 7798 if ((tp->t_flags & TF_RCVD_TSTMP) || 7799 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) { 7800 to.to_tsval = cts + tp->ts_offset; 7801 to.to_tsecr = tp->ts_recent; 7802 to.to_flags |= TOF_TS; 7803 } 7804 /* Set receive buffer autosizing timestamp. */ 7805 if (tp->rfbuf_ts == 0 && 7806 (so->so_rcv.sb_flags & SB_AUTOSIZE)) 7807 tp->rfbuf_ts = tcp_ts_getticks(); 7808 /* Selective ACK's. */ 7809 if (flags & TH_SYN) 7810 to.to_flags |= TOF_SACKPERM; 7811 else if (TCPS_HAVEESTABLISHED(tp->t_state) && 7812 tp->rcv_numsacks > 0) { 7813 to.to_flags |= TOF_SACK; 7814 to.to_nsacks = tp->rcv_numsacks; 7815 to.to_sacks = (u_char *)tp->sackblks; 7816 } 7817 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 7818 /* TCP-MD5 (RFC2385). */ 7819 if (tp->t_flags & TF_SIGNATURE) 7820 to.to_flags |= TOF_SIGNATURE; 7821 #endif /* TCP_SIGNATURE */ 7822 7823 /* Processing the options. */ 7824 hdrlen += optlen = tcp_addoptions(&to, opt); 7825 /* 7826 * If we wanted a TFO option to be added, but it was unable 7827 * to fit, ensure no data is sent. 7828 */ 7829 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie && 7830 !(to.to_flags & TOF_FASTOPEN)) 7831 len = 0; 7832 } 7833 #ifdef NETFLIX_TCPOUDP 7834 if (tp->t_port) { 7835 if (V_tcp_udp_tunneling_port == 0) { 7836 /* The port was removed?? */ 7837 SOCKBUF_UNLOCK(&so->so_snd); 7838 return (EHOSTUNREACH); 7839 } 7840 hdrlen += sizeof(struct udphdr); 7841 } 7842 #endif 7843 ipoptlen = 0; 7844 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 7845 ipoptlen += ipsec_optlen; 7846 #endif 7847 7848 /* 7849 * Adjust data length if insertion of options will bump the packet 7850 * length beyond the t_maxseg length. Clear the FIN bit because we 7851 * cut off the tail of the segment. 7852 */ 7853 if (len + optlen + ipoptlen > tp->t_maxseg) { 7854 if (flags & TH_FIN) { 7855 would_have_fin = 1; 7856 flags &= ~TH_FIN; 7857 } 7858 if (tso) { 7859 uint32_t if_hw_tsomax; 7860 uint32_t moff; 7861 int32_t max_len; 7862 7863 /* extract TSO information */ 7864 if_hw_tsomax = tp->t_tsomax; 7865 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount; 7866 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize; 7867 KASSERT(ipoptlen == 0, 7868 ("%s: TSO can't do IP options", __func__)); 7869 7870 /* 7871 * Check if we should limit by maximum payload 7872 * length: 7873 */ 7874 if (if_hw_tsomax != 0) { 7875 /* compute maximum TSO length */ 7876 max_len = (if_hw_tsomax - hdrlen - 7877 max_linkhdr); 7878 if (max_len <= 0) { 7879 len = 0; 7880 } else if (len > max_len) { 7881 sendalot = 1; 7882 len = max_len; 7883 } 7884 } 7885 /* 7886 * Prevent the last segment from being fractional 7887 * unless the send sockbuf can be emptied: 7888 */ 7889 max_len = (tp->t_maxseg - optlen); 7890 if ((sb_offset + len) < sbavail(sb)) { 7891 moff = len % (u_int)max_len; 7892 if (moff != 0) { 7893 len -= moff; 7894 sendalot = 1; 7895 } 7896 } 7897 /* 7898 * In case there are too many small fragments don't 7899 * use TSO: 7900 */ 7901 if (len <= max_len) { 7902 len = max_len; 7903 sendalot = 1; 7904 tso = 0; 7905 } 7906 /* 7907 * Send the FIN in a separate segment after the bulk 7908 * sending is done. We don't trust the TSO 7909 * implementations to clear the FIN flag on all but 7910 * the last segment. 7911 */ 7912 if (tp->t_flags & TF_NEEDFIN) 7913 sendalot = 1; 7914 7915 } else { 7916 len = tp->t_maxseg - optlen - ipoptlen; 7917 sendalot = 1; 7918 } 7919 } else 7920 tso = 0; 7921 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET, 7922 ("%s: len > IP_MAXPACKET", __func__)); 7923 #ifdef DIAGNOSTIC 7924 #ifdef INET6 7925 if (max_linkhdr + hdrlen > MCLBYTES) 7926 #else 7927 if (max_linkhdr + hdrlen > MHLEN) 7928 #endif 7929 panic("tcphdr too big"); 7930 #endif 7931 7932 /* 7933 * This KASSERT is here to catch edge cases at a well defined place. 7934 * Before, those had triggered (random) panic conditions further 7935 * down. 7936 */ 7937 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 7938 if ((len == 0) && 7939 (flags & TH_FIN) && 7940 (sbused(sb))) { 7941 /* 7942 * We have outstanding data, don't send a fin by itself!. 7943 */ 7944 goto just_return; 7945 } 7946 /* 7947 * Grab a header mbuf, attaching a copy of data to be transmitted, 7948 * and initialize the header from the template for sends on this 7949 * connection. 7950 */ 7951 if (len) { 7952 uint32_t max_val; 7953 uint32_t moff; 7954 7955 if (rack->rc_pace_max_segs) 7956 max_val = rack->rc_pace_max_segs * tp->t_maxseg; 7957 else 7958 max_val = len; 7959 /* 7960 * We allow a limit on sending with hptsi. 7961 */ 7962 if (len > max_val) { 7963 len = max_val; 7964 } 7965 #ifdef INET6 7966 if (MHLEN < hdrlen + max_linkhdr) 7967 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); 7968 else 7969 #endif 7970 m = m_gethdr(M_NOWAIT, MT_DATA); 7971 7972 if (m == NULL) { 7973 SOCKBUF_UNLOCK(sb); 7974 error = ENOBUFS; 7975 sack_rxmit = 0; 7976 goto out; 7977 } 7978 m->m_data += max_linkhdr; 7979 m->m_len = hdrlen; 7980 7981 /* 7982 * Start the m_copy functions from the closest mbuf to the 7983 * sb_offset in the socket buffer chain. 7984 */ 7985 mb = sbsndptr_noadv(sb, sb_offset, &moff); 7986 if (len <= MHLEN - hdrlen - max_linkhdr) { 7987 m_copydata(mb, moff, (int)len, 7988 mtod(m, caddr_t)+hdrlen); 7989 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 7990 sbsndptr_adv(sb, mb, len); 7991 m->m_len += len; 7992 } else { 7993 struct sockbuf *msb; 7994 7995 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 7996 msb = NULL; 7997 else 7998 msb = sb; 7999 m->m_next = tcp_m_copym(mb, moff, &len, 8000 if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb); 8001 if (len <= (tp->t_maxseg - optlen)) { 8002 /* 8003 * Must have ran out of mbufs for the copy 8004 * shorten it to no longer need tso. Lets 8005 * not put on sendalot since we are low on 8006 * mbufs. 8007 */ 8008 tso = 0; 8009 } 8010 if (m->m_next == NULL) { 8011 SOCKBUF_UNLOCK(sb); 8012 (void)m_free(m); 8013 error = ENOBUFS; 8014 sack_rxmit = 0; 8015 goto out; 8016 } 8017 } 8018 if ((tp->t_flags & TF_FORCEDATA) && len == 1) { 8019 TCPSTAT_INC(tcps_sndprobe); 8020 #ifdef NETFLIX_STATS 8021 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 8022 stats_voi_update_abs_u32(tp->t_stats, 8023 VOI_TCP_RETXPB, len); 8024 else 8025 stats_voi_update_abs_u64(tp->t_stats, 8026 VOI_TCP_TXPB, len); 8027 #endif 8028 } else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) { 8029 if (rsm && (rsm->r_flags & RACK_TLP)) { 8030 /* 8031 * TLP should not count in retran count, but 8032 * in its own bin 8033 */ 8034 counter_u64_add(rack_tlp_retran, 1); 8035 counter_u64_add(rack_tlp_retran_bytes, len); 8036 } else { 8037 tp->t_sndrexmitpack++; 8038 TCPSTAT_INC(tcps_sndrexmitpack); 8039 TCPSTAT_ADD(tcps_sndrexmitbyte, len); 8040 } 8041 #ifdef NETFLIX_STATS 8042 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB, 8043 len); 8044 #endif 8045 } else { 8046 TCPSTAT_INC(tcps_sndpack); 8047 TCPSTAT_ADD(tcps_sndbyte, len); 8048 #ifdef NETFLIX_STATS 8049 stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB, 8050 len); 8051 #endif 8052 } 8053 /* 8054 * If we're sending everything we've got, set PUSH. (This 8055 * will keep happy those implementations which only give 8056 * data to the user when a buffer fills or a PUSH comes in.) 8057 */ 8058 if (sb_offset + len == sbused(sb) && 8059 sbused(sb) && 8060 !(flags & TH_SYN)) 8061 flags |= TH_PUSH; 8062 8063 /* 8064 * Are we doing hptsi, if so we must calculate the slot. We 8065 * only do hptsi in ESTABLISHED and with no RESET being 8066 * sent where we have data to send. 8067 */ 8068 if (((tp->t_state == TCPS_ESTABLISHED) || 8069 (tp->t_state == TCPS_CLOSE_WAIT) || 8070 ((tp->t_state == TCPS_FIN_WAIT_1) && 8071 ((tp->t_flags & TF_SENTFIN) == 0) && 8072 ((flags & TH_FIN) == 0))) && 8073 ((flags & TH_RST) == 0) && 8074 (rack->rc_always_pace)) { 8075 /* 8076 * We use the most optimistic possible cwnd/srtt for 8077 * sending calculations. This will make our 8078 * calculation anticipate getting more through 8079 * quicker then possible. But thats ok we don't want 8080 * the peer to have a gap in data sending. 8081 */ 8082 uint32_t srtt, cwnd, tr_perms = 0; 8083 8084 if (rack->r_ctl.rc_rack_min_rtt) 8085 srtt = rack->r_ctl.rc_rack_min_rtt; 8086 else 8087 srtt = TICKS_2_MSEC((tp->t_srtt >> TCP_RTT_SHIFT)); 8088 if (rack->r_ctl.rc_rack_largest_cwnd) 8089 cwnd = rack->r_ctl.rc_rack_largest_cwnd; 8090 else 8091 cwnd = tp->snd_cwnd; 8092 tr_perms = cwnd / srtt; 8093 if (tr_perms == 0) { 8094 tr_perms = tp->t_maxseg; 8095 } 8096 tot_len_this_send += len; 8097 /* 8098 * Calculate how long this will take to drain, if 8099 * the calculation comes out to zero, thats ok we 8100 * will use send_a_lot to possibly spin around for 8101 * more increasing tot_len_this_send to the point 8102 * that its going to require a pace, or we hit the 8103 * cwnd. Which in that case we are just waiting for 8104 * a ACK. 8105 */ 8106 slot = tot_len_this_send / tr_perms; 8107 /* Now do we reduce the time so we don't run dry? */ 8108 if (slot && rack->rc_pace_reduce) { 8109 int32_t reduce; 8110 8111 reduce = (slot / rack->rc_pace_reduce); 8112 if (reduce < slot) { 8113 slot -= reduce; 8114 } else 8115 slot = 0; 8116 } 8117 if (rack->r_enforce_min_pace && 8118 (slot == 0) && 8119 (tot_len_this_send >= (rack->r_min_pace_seg_thresh * tp->t_maxseg))) { 8120 /* We are enforcing a minimum pace time of 1ms */ 8121 slot = rack->r_enforce_min_pace; 8122 } 8123 } 8124 SOCKBUF_UNLOCK(sb); 8125 } else { 8126 SOCKBUF_UNLOCK(sb); 8127 if (tp->t_flags & TF_ACKNOW) 8128 TCPSTAT_INC(tcps_sndacks); 8129 else if (flags & (TH_SYN | TH_FIN | TH_RST)) 8130 TCPSTAT_INC(tcps_sndctrl); 8131 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 8132 TCPSTAT_INC(tcps_sndurg); 8133 else 8134 TCPSTAT_INC(tcps_sndwinup); 8135 8136 m = m_gethdr(M_NOWAIT, MT_DATA); 8137 if (m == NULL) { 8138 error = ENOBUFS; 8139 sack_rxmit = 0; 8140 goto out; 8141 } 8142 #ifdef INET6 8143 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && 8144 MHLEN >= hdrlen) { 8145 M_ALIGN(m, hdrlen); 8146 } else 8147 #endif 8148 m->m_data += max_linkhdr; 8149 m->m_len = hdrlen; 8150 } 8151 SOCKBUF_UNLOCK_ASSERT(sb); 8152 m->m_pkthdr.rcvif = (struct ifnet *)0; 8153 #ifdef MAC 8154 mac_inpcb_create_mbuf(inp, m); 8155 #endif 8156 #ifdef INET6 8157 if (isipv6) { 8158 ip6 = mtod(m, struct ip6_hdr *); 8159 #ifdef NETFLIX_TCPOUDP 8160 if (tp->t_port) { 8161 udp = (struct udphdr *)((caddr_t)ip6 + ipoptlen + sizeof(struct ip6_hdr)); 8162 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 8163 udp->uh_dport = tp->t_port; 8164 ulen = hdrlen + len - sizeof(struct ip6_hdr); 8165 udp->uh_ulen = htons(ulen); 8166 th = (struct tcphdr *)(udp + 1); 8167 } else 8168 #endif 8169 th = (struct tcphdr *)(ip6 + 1); 8170 tcpip_fillheaders(inp, ip6, th); 8171 } else 8172 #endif /* INET6 */ 8173 { 8174 ip = mtod(m, struct ip *); 8175 #ifdef TCPDEBUG 8176 ipov = (struct ipovly *)ip; 8177 #endif 8178 #ifdef NETFLIX_TCPOUDP 8179 if (tp->t_port) { 8180 udp = (struct udphdr *)((caddr_t)ip + ipoptlen + sizeof(struct ip)); 8181 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 8182 udp->uh_dport = tp->t_port; 8183 ulen = hdrlen + len - sizeof(struct ip); 8184 udp->uh_ulen = htons(ulen); 8185 th = (struct tcphdr *)(udp + 1); 8186 } else 8187 #endif 8188 th = (struct tcphdr *)(ip + 1); 8189 tcpip_fillheaders(inp, ip, th); 8190 } 8191 /* 8192 * Fill in fields, remembering maximum advertised window for use in 8193 * delaying messages about window sizes. If resending a FIN, be sure 8194 * not to use a new sequence number. 8195 */ 8196 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && 8197 tp->snd_nxt == tp->snd_max) 8198 tp->snd_nxt--; 8199 /* 8200 * If we are starting a connection, send ECN setup SYN packet. If we 8201 * are on a retransmit, we may resend those bits a number of times 8202 * as per RFC 3168. 8203 */ 8204 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) { 8205 if (tp->t_rxtshift >= 1) { 8206 if (tp->t_rxtshift <= V_tcp_ecn_maxretries) 8207 flags |= TH_ECE | TH_CWR; 8208 } else 8209 flags |= TH_ECE | TH_CWR; 8210 } 8211 if (tp->t_state == TCPS_ESTABLISHED && 8212 (tp->t_flags & TF_ECN_PERMIT)) { 8213 /* 8214 * If the peer has ECN, mark data packets with ECN capable 8215 * transmission (ECT). Ignore pure ack packets, 8216 * retransmissions and window probes. 8217 */ 8218 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) && 8219 !((tp->t_flags & TF_FORCEDATA) && len == 1)) { 8220 #ifdef INET6 8221 if (isipv6) 8222 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20); 8223 else 8224 #endif 8225 ip->ip_tos |= IPTOS_ECN_ECT0; 8226 TCPSTAT_INC(tcps_ecn_ect0); 8227 } 8228 /* 8229 * Reply with proper ECN notifications. 8230 */ 8231 if (tp->t_flags & TF_ECN_SND_CWR) { 8232 flags |= TH_CWR; 8233 tp->t_flags &= ~TF_ECN_SND_CWR; 8234 } 8235 if (tp->t_flags & TF_ECN_SND_ECE) 8236 flags |= TH_ECE; 8237 } 8238 /* 8239 * If we are doing retransmissions, then snd_nxt will not reflect 8240 * the first unsent octet. For ACK only packets, we do not want the 8241 * sequence number of the retransmitted packet, we want the sequence 8242 * number of the next unsent octet. So, if there is no data (and no 8243 * SYN or FIN), use snd_max instead of snd_nxt when filling in 8244 * ti_seq. But if we are in persist state, snd_max might reflect 8245 * one byte beyond the right edge of the window, so use snd_nxt in 8246 * that case, since we know we aren't doing a retransmission. 8247 * (retransmit and persist are mutually exclusive...) 8248 */ 8249 if (sack_rxmit == 0) { 8250 if (len || (flags & (TH_SYN | TH_FIN)) || 8251 rack->rc_in_persist) { 8252 th->th_seq = htonl(tp->snd_nxt); 8253 rack_seq = tp->snd_nxt; 8254 } else if (flags & TH_RST) { 8255 /* 8256 * For a Reset send the last cum ack in sequence 8257 * (this like any other choice may still generate a 8258 * challenge ack, if a ack-update packet is in 8259 * flight). 8260 */ 8261 th->th_seq = htonl(tp->snd_una); 8262 rack_seq = tp->snd_una; 8263 } else { 8264 th->th_seq = htonl(tp->snd_max); 8265 rack_seq = tp->snd_max; 8266 } 8267 } else { 8268 th->th_seq = htonl(rsm->r_start); 8269 rack_seq = rsm->r_start; 8270 } 8271 th->th_ack = htonl(tp->rcv_nxt); 8272 if (optlen) { 8273 bcopy(opt, th + 1, optlen); 8274 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 8275 } 8276 th->th_flags = flags; 8277 /* 8278 * Calculate receive window. Don't shrink window, but avoid silly 8279 * window syndrome. 8280 * If a RST segment is sent, advertise a window of zero. 8281 */ 8282 if (flags & TH_RST) { 8283 recwin = 0; 8284 } else { 8285 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) && 8286 recwin < (long)tp->t_maxseg) 8287 recwin = 0; 8288 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) && 8289 recwin < (long)(tp->rcv_adv - tp->rcv_nxt)) 8290 recwin = (long)(tp->rcv_adv - tp->rcv_nxt); 8291 if (recwin > (long)TCP_MAXWIN << tp->rcv_scale) 8292 recwin = (long)TCP_MAXWIN << tp->rcv_scale; 8293 } 8294 8295 /* 8296 * According to RFC1323 the window field in a SYN (i.e., a <SYN> or 8297 * <SYN,ACK>) segment itself is never scaled. The <SYN,ACK> case is 8298 * handled in syncache. 8299 */ 8300 if (flags & TH_SYN) 8301 th->th_win = htons((u_short) 8302 (min(sbspace(&so->so_rcv), TCP_MAXWIN))); 8303 else 8304 th->th_win = htons((u_short)(recwin >> tp->rcv_scale)); 8305 /* 8306 * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0 8307 * window. This may cause the remote transmitter to stall. This 8308 * flag tells soreceive() to disable delayed acknowledgements when 8309 * draining the buffer. This can occur if the receiver is 8310 * attempting to read more data than can be buffered prior to 8311 * transmitting on the connection. 8312 */ 8313 if (th->th_win == 0) { 8314 tp->t_sndzerowin++; 8315 tp->t_flags |= TF_RXWIN0SENT; 8316 } else 8317 tp->t_flags &= ~TF_RXWIN0SENT; 8318 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 8319 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); 8320 th->th_flags |= TH_URG; 8321 } else 8322 /* 8323 * If no urgent pointer to send, then we pull the urgent 8324 * pointer to the left edge of the send window so that it 8325 * doesn't drift into the send window on sequence number 8326 * wraparound. 8327 */ 8328 tp->snd_up = tp->snd_una; /* drag it along */ 8329 8330 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 8331 if (to.to_flags & TOF_SIGNATURE) { 8332 /* 8333 * Calculate MD5 signature and put it into the place 8334 * determined before. 8335 * NOTE: since TCP options buffer doesn't point into 8336 * mbuf's data, calculate offset and use it. 8337 */ 8338 if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th, 8339 (u_char *)(th + 1) + (to.to_signature - opt)) != 0) { 8340 /* 8341 * Do not send segment if the calculation of MD5 8342 * digest has failed. 8343 */ 8344 goto out; 8345 } 8346 } 8347 #endif 8348 8349 /* 8350 * Put TCP length in extended header, and then checksum extended 8351 * header and data. 8352 */ 8353 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 8354 #ifdef INET6 8355 if (isipv6) { 8356 /* 8357 * ip6_plen is not need to be filled now, and will be filled 8358 * in ip6_output. 8359 */ 8360 if (tp->t_port) { 8361 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6; 8362 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 8363 udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0); 8364 th->th_sum = htons(0); 8365 } else { 8366 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 8367 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 8368 th->th_sum = in6_cksum_pseudo(ip6, 8369 sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP, 8370 0); 8371 } 8372 } 8373 #endif 8374 #if defined(INET6) && defined(INET) 8375 else 8376 #endif 8377 #ifdef INET 8378 { 8379 if (tp->t_port) { 8380 m->m_pkthdr.csum_flags = CSUM_UDP; 8381 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 8382 udp->uh_sum = in_pseudo(ip->ip_src.s_addr, 8383 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP)); 8384 th->th_sum = htons(0); 8385 } else { 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 } 8392 /* IP version must be set here for ipv4/ipv6 checking later */ 8393 KASSERT(ip->ip_v == IPVERSION, 8394 ("%s: IP version incorrect: %d", __func__, ip->ip_v)); 8395 } 8396 #endif 8397 8398 /* 8399 * Enable TSO and specify the size of the segments. The TCP pseudo 8400 * header checksum is always provided. XXX: Fixme: This is currently 8401 * not the case for IPv6. 8402 */ 8403 if (tso) { 8404 KASSERT(len > tp->t_maxseg - optlen, 8405 ("%s: len <= tso_segsz", __func__)); 8406 m->m_pkthdr.csum_flags |= CSUM_TSO; 8407 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen; 8408 } 8409 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 8410 KASSERT(len + hdrlen + ipoptlen - ipsec_optlen == m_length(m, NULL), 8411 ("%s: mbuf chain shorter than expected: %d + %u + %u - %u != %u", 8412 __func__, len, hdrlen, ipoptlen, ipsec_optlen, m_length(m, NULL))); 8413 #else 8414 KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL), 8415 ("%s: mbuf chain shorter than expected: %d + %u + %u != %u", 8416 __func__, len, hdrlen, ipoptlen, m_length(m, NULL))); 8417 #endif 8418 8419 #ifdef TCP_HHOOK 8420 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */ 8421 hhook_run_tcp_est_out(tp, th, &to, len, tso); 8422 #endif 8423 8424 #ifdef TCPDEBUG 8425 /* 8426 * Trace. 8427 */ 8428 if (so->so_options & SO_DEBUG) { 8429 u_short save = 0; 8430 8431 #ifdef INET6 8432 if (!isipv6) 8433 #endif 8434 { 8435 save = ipov->ih_len; 8436 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + 8437 * (th->th_off << 2) */ ); 8438 } 8439 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0); 8440 #ifdef INET6 8441 if (!isipv6) 8442 #endif 8443 ipov->ih_len = save; 8444 } 8445 #endif /* TCPDEBUG */ 8446 8447 /* We're getting ready to send; log now. */ 8448 if (tp->t_logstate != TCP_LOG_STATE_OFF) { 8449 union tcp_log_stackspecific log; 8450 8451 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 8452 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; 8453 log.u_bbr.ininput = rack->rc_inp->inp_in_input; 8454 log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt; 8455 if (rsm || sack_rxmit) { 8456 log.u_bbr.flex8 = 1; 8457 } else { 8458 log.u_bbr.flex8 = 0; 8459 } 8460 lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK, 8461 len, &log, false, NULL, NULL, 0, NULL); 8462 } else 8463 lgb = NULL; 8464 8465 /* 8466 * Fill in IP length and desired time to live and send to IP level. 8467 * There should be a better way to handle ttl and tos; we could keep 8468 * them in the template, but need a way to checksum without them. 8469 */ 8470 /* 8471 * m->m_pkthdr.len should have been set before cksum calcuration, 8472 * because in6_cksum() need it. 8473 */ 8474 #ifdef INET6 8475 if (isipv6) { 8476 /* 8477 * we separately set hoplimit for every segment, since the 8478 * user might want to change the value via setsockopt. Also, 8479 * desired default hop limit might be changed via Neighbor 8480 * Discovery. 8481 */ 8482 ip6->ip6_hlim = in6_selecthlim(inp, NULL); 8483 8484 /* 8485 * Set the packet size here for the benefit of DTrace 8486 * probes. ip6_output() will set it properly; it's supposed 8487 * to include the option header lengths as well. 8488 */ 8489 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6)); 8490 8491 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) 8492 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 8493 else 8494 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 8495 8496 if (tp->t_state == TCPS_SYN_SENT) 8497 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th); 8498 8499 TCP_PROBE5(send, NULL, tp, ip6, tp, th); 8500 /* TODO: IPv6 IP6TOS_ECT bit on */ 8501 error = ip6_output(m, tp->t_inpcb->in6p_outputopts, 8502 &inp->inp_route6, 8503 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 8504 NULL, NULL, inp); 8505 8506 if (error == EMSGSIZE && inp->inp_route6.ro_rt != NULL) 8507 mtu = inp->inp_route6.ro_rt->rt_mtu; 8508 } 8509 #endif /* INET6 */ 8510 #if defined(INET) && defined(INET6) 8511 else 8512 #endif 8513 #ifdef INET 8514 { 8515 ip->ip_len = htons(m->m_pkthdr.len); 8516 #ifdef INET6 8517 if (inp->inp_vflag & INP_IPV6PROTO) 8518 ip->ip_ttl = in6_selecthlim(inp, NULL); 8519 #endif /* INET6 */ 8520 /* 8521 * If we do path MTU discovery, then we set DF on every 8522 * packet. This might not be the best thing to do according 8523 * to RFC3390 Section 2. However the tcp hostcache migitates 8524 * the problem so it affects only the first tcp connection 8525 * with a host. 8526 * 8527 * NB: Don't set DF on small MTU/MSS to have a safe 8528 * fallback. 8529 */ 8530 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) { 8531 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 8532 if (tp->t_port == 0 || len < V_tcp_minmss) { 8533 ip->ip_off |= htons(IP_DF); 8534 } 8535 } else { 8536 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 8537 } 8538 8539 if (tp->t_state == TCPS_SYN_SENT) 8540 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th); 8541 8542 TCP_PROBE5(send, NULL, tp, ip, tp, th); 8543 8544 error = ip_output(m, tp->t_inpcb->inp_options, &inp->inp_route, 8545 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, 8546 inp); 8547 if (error == EMSGSIZE && inp->inp_route.ro_rt != NULL) 8548 mtu = inp->inp_route.ro_rt->rt_mtu; 8549 } 8550 #endif /* INET */ 8551 8552 out: 8553 if (lgb) { 8554 lgb->tlb_errno = error; 8555 lgb = NULL; 8556 } 8557 /* 8558 * In transmit state, time the transmission and arrange for the 8559 * retransmit. In persist state, just set snd_max. 8560 */ 8561 if (error == 0) { 8562 if (len == 0) 8563 counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1); 8564 else if (len == 1) { 8565 counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1); 8566 } else if (len > 1) { 8567 int idx; 8568 8569 idx = (len / tp->t_maxseg) + 3; 8570 if (idx >= TCP_MSS_ACCT_ATIMER) 8571 counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1); 8572 else 8573 counter_u64_add(rack_out_size[idx], 1); 8574 } 8575 } 8576 if (sub_from_prr && (error == 0)) { 8577 rack->r_ctl.rc_prr_sndcnt -= len; 8578 } 8579 sub_from_prr = 0; 8580 rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error, cts, 8581 pass, rsm); 8582 if ((tp->t_flags & TF_FORCEDATA) == 0 || 8583 (rack->rc_in_persist == 0)) { 8584 tcp_seq startseq = tp->snd_nxt; 8585 8586 /* 8587 * Advance snd_nxt over sequence space of this segment. 8588 */ 8589 if (error) 8590 /* We don't log or do anything with errors */ 8591 goto timer; 8592 8593 if (flags & (TH_SYN | TH_FIN)) { 8594 if (flags & TH_SYN) 8595 tp->snd_nxt++; 8596 if (flags & TH_FIN) { 8597 tp->snd_nxt++; 8598 tp->t_flags |= TF_SENTFIN; 8599 } 8600 } 8601 /* In the ENOBUFS case we do *not* update snd_max */ 8602 if (sack_rxmit) 8603 goto timer; 8604 8605 tp->snd_nxt += len; 8606 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 8607 if (tp->snd_una == tp->snd_max) { 8608 /* 8609 * Update the time we just added data since 8610 * none was outstanding. 8611 */ 8612 rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__); 8613 tp->t_acktime = ticks; 8614 } 8615 tp->snd_max = tp->snd_nxt; 8616 /* 8617 * Time this transmission if not a retransmission and 8618 * not currently timing anything. 8619 * This is only relevant in case of switching back to 8620 * the base stack. 8621 */ 8622 if (tp->t_rtttime == 0) { 8623 tp->t_rtttime = ticks; 8624 tp->t_rtseq = startseq; 8625 TCPSTAT_INC(tcps_segstimed); 8626 } 8627 #ifdef NETFLIX_STATS 8628 if (!(tp->t_flags & TF_GPUTINPROG) && len) { 8629 tp->t_flags |= TF_GPUTINPROG; 8630 tp->gput_seq = startseq; 8631 tp->gput_ack = startseq + 8632 ulmin(sbavail(sb) - sb_offset, sendwin); 8633 tp->gput_ts = tcp_ts_getticks(); 8634 } 8635 #endif 8636 } 8637 /* 8638 * Set retransmit timer if not currently set, and not doing 8639 * a pure ack or a keep-alive probe. Initial value for 8640 * retransmit timer is smoothed round-trip time + 2 * 8641 * round-trip time variance. Initialize shift counter which 8642 * is used for backoff of retransmit time. 8643 */ 8644 timer: 8645 if ((tp->snd_wnd == 0) && 8646 TCPS_HAVEESTABLISHED(tp->t_state)) { 8647 /* 8648 * If the persists timer was set above (right before 8649 * the goto send), and still needs to be on. Lets 8650 * make sure all is canceled. If the persist timer 8651 * is not running, we want to get it up. 8652 */ 8653 if (rack->rc_in_persist == 0) { 8654 rack_enter_persist(tp, rack, cts); 8655 } 8656 } 8657 } else { 8658 /* 8659 * Persist case, update snd_max but since we are in persist 8660 * mode (no window) we do not update snd_nxt. 8661 */ 8662 int32_t xlen = len; 8663 8664 if (error) 8665 goto nomore; 8666 8667 if (flags & TH_SYN) 8668 ++xlen; 8669 if (flags & TH_FIN) { 8670 ++xlen; 8671 tp->t_flags |= TF_SENTFIN; 8672 } 8673 /* In the ENOBUFS case we do *not* update snd_max */ 8674 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) { 8675 if (tp->snd_una == tp->snd_max) { 8676 /* 8677 * Update the time we just added data since 8678 * none was outstanding. 8679 */ 8680 rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__); 8681 tp->t_acktime = ticks; 8682 } 8683 tp->snd_max = tp->snd_nxt + len; 8684 } 8685 } 8686 nomore: 8687 if (error) { 8688 SOCKBUF_UNLOCK_ASSERT(sb); /* Check gotos. */ 8689 /* 8690 * Failures do not advance the seq counter above. For the 8691 * case of ENOBUFS we will fall out and retry in 1ms with 8692 * the hpts. Everything else will just have to retransmit 8693 * with the timer. 8694 * 8695 * In any case, we do not want to loop around for another 8696 * send without a good reason. 8697 */ 8698 sendalot = 0; 8699 switch (error) { 8700 case EPERM: 8701 tp->t_flags &= ~TF_FORCEDATA; 8702 tp->t_softerror = error; 8703 return (error); 8704 case ENOBUFS: 8705 if (slot == 0) { 8706 /* 8707 * Pace us right away to retry in a some 8708 * time 8709 */ 8710 slot = 1 + rack->rc_enobuf; 8711 if (rack->rc_enobuf < 255) 8712 rack->rc_enobuf++; 8713 if (slot > (rack->rc_rack_rtt / 2)) { 8714 slot = rack->rc_rack_rtt / 2; 8715 } 8716 if (slot < 10) 8717 slot = 10; 8718 } 8719 counter_u64_add(rack_saw_enobuf, 1); 8720 error = 0; 8721 goto enobufs; 8722 case EMSGSIZE: 8723 /* 8724 * For some reason the interface we used initially 8725 * to send segments changed to another or lowered 8726 * its MTU. If TSO was active we either got an 8727 * interface without TSO capabilits or TSO was 8728 * turned off. If we obtained mtu from ip_output() 8729 * then update it and try again. 8730 */ 8731 if (tso) 8732 tp->t_flags &= ~TF_TSO; 8733 if (mtu != 0) { 8734 tcp_mss_update(tp, -1, mtu, NULL, NULL); 8735 goto again; 8736 } 8737 slot = 10; 8738 rack_start_hpts_timer(rack, tp, cts, __LINE__, slot, 0, 1); 8739 tp->t_flags &= ~TF_FORCEDATA; 8740 return (error); 8741 case ENETUNREACH: 8742 counter_u64_add(rack_saw_enetunreach, 1); 8743 case EHOSTDOWN: 8744 case EHOSTUNREACH: 8745 case ENETDOWN: 8746 if (TCPS_HAVERCVDSYN(tp->t_state)) { 8747 tp->t_softerror = error; 8748 } 8749 /* FALLTHROUGH */ 8750 default: 8751 slot = 10; 8752 rack_start_hpts_timer(rack, tp, cts, __LINE__, slot, 0, 1); 8753 tp->t_flags &= ~TF_FORCEDATA; 8754 return (error); 8755 } 8756 } else { 8757 rack->rc_enobuf = 0; 8758 } 8759 TCPSTAT_INC(tcps_sndtotal); 8760 8761 /* 8762 * Data sent (as far as we can tell). If this advertises a larger 8763 * window than any other segment, then remember the size of the 8764 * advertised window. Any pending ACK has now been sent. 8765 */ 8766 if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) 8767 tp->rcv_adv = tp->rcv_nxt + recwin; 8768 tp->last_ack_sent = tp->rcv_nxt; 8769 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); 8770 enobufs: 8771 rack->r_tlp_running = 0; 8772 if ((flags & TH_RST) || (would_have_fin == 1)) { 8773 /* 8774 * We don't send again after a RST. We also do *not* send 8775 * again if we would have had a find, but now have 8776 * outstanding data. 8777 */ 8778 slot = 0; 8779 sendalot = 0; 8780 } 8781 if (slot) { 8782 /* set the rack tcb into the slot N */ 8783 counter_u64_add(rack_paced_segments, 1); 8784 } else if (sendalot) { 8785 if (len) 8786 counter_u64_add(rack_unpaced_segments, 1); 8787 sack_rxmit = 0; 8788 tp->t_flags &= ~TF_FORCEDATA; 8789 goto again; 8790 } else if (len) { 8791 counter_u64_add(rack_unpaced_segments, 1); 8792 } 8793 tp->t_flags &= ~TF_FORCEDATA; 8794 rack_start_hpts_timer(rack, tp, cts, __LINE__, slot, tot_len_this_send, 1); 8795 return (error); 8796 } 8797 8798 /* 8799 * rack_ctloutput() must drop the inpcb lock before performing copyin on 8800 * socket option arguments. When it re-acquires the lock after the copy, it 8801 * has to revalidate that the connection is still valid for the socket 8802 * option. 8803 */ 8804 static int 8805 rack_set_sockopt(struct socket *so, struct sockopt *sopt, 8806 struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack) 8807 { 8808 int32_t error = 0, optval; 8809 8810 switch (sopt->sopt_name) { 8811 case TCP_RACK_PROP_RATE: 8812 case TCP_RACK_PROP: 8813 case TCP_RACK_TLP_REDUCE: 8814 case TCP_RACK_EARLY_RECOV: 8815 case TCP_RACK_PACE_ALWAYS: 8816 case TCP_DELACK: 8817 case TCP_RACK_PACE_REDUCE: 8818 case TCP_RACK_PACE_MAX_SEG: 8819 case TCP_RACK_PRR_SENDALOT: 8820 case TCP_RACK_MIN_TO: 8821 case TCP_RACK_EARLY_SEG: 8822 case TCP_RACK_REORD_THRESH: 8823 case TCP_RACK_REORD_FADE: 8824 case TCP_RACK_TLP_THRESH: 8825 case TCP_RACK_PKT_DELAY: 8826 case TCP_RACK_TLP_USE: 8827 case TCP_RACK_TLP_INC_VAR: 8828 case TCP_RACK_IDLE_REDUCE_HIGH: 8829 case TCP_RACK_MIN_PACE: 8830 case TCP_RACK_MIN_PACE_SEG: 8831 case TCP_BBR_RACK_RTT_USE: 8832 case TCP_DATA_AFTER_CLOSE: 8833 break; 8834 default: 8835 return (tcp_default_ctloutput(so, sopt, inp, tp)); 8836 break; 8837 } 8838 INP_WUNLOCK(inp); 8839 error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval)); 8840 if (error) 8841 return (error); 8842 INP_WLOCK(inp); 8843 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { 8844 INP_WUNLOCK(inp); 8845 return (ECONNRESET); 8846 } 8847 tp = intotcpcb(inp); 8848 rack = (struct tcp_rack *)tp->t_fb_ptr; 8849 switch (sopt->sopt_name) { 8850 case TCP_RACK_PROP_RATE: 8851 if ((optval <= 0) || (optval >= 100)) { 8852 error = EINVAL; 8853 break; 8854 } 8855 RACK_OPTS_INC(tcp_rack_prop_rate); 8856 rack->r_ctl.rc_prop_rate = optval; 8857 break; 8858 case TCP_RACK_TLP_USE: 8859 if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) { 8860 error = EINVAL; 8861 break; 8862 } 8863 RACK_OPTS_INC(tcp_tlp_use); 8864 rack->rack_tlp_threshold_use = optval; 8865 break; 8866 case TCP_RACK_PROP: 8867 /* RACK proportional rate reduction (bool) */ 8868 RACK_OPTS_INC(tcp_rack_prop); 8869 rack->r_ctl.rc_prop_reduce = optval; 8870 break; 8871 case TCP_RACK_TLP_REDUCE: 8872 /* RACK TLP cwnd reduction (bool) */ 8873 RACK_OPTS_INC(tcp_rack_tlp_reduce); 8874 rack->r_ctl.rc_tlp_cwnd_reduce = optval; 8875 break; 8876 case TCP_RACK_EARLY_RECOV: 8877 /* Should recovery happen early (bool) */ 8878 RACK_OPTS_INC(tcp_rack_early_recov); 8879 rack->r_ctl.rc_early_recovery = optval; 8880 break; 8881 case TCP_RACK_PACE_ALWAYS: 8882 /* Use the always pace method (bool) */ 8883 RACK_OPTS_INC(tcp_rack_pace_always); 8884 if (optval > 0) 8885 rack->rc_always_pace = 1; 8886 else 8887 rack->rc_always_pace = 0; 8888 break; 8889 case TCP_RACK_PACE_REDUCE: 8890 /* RACK Hptsi reduction factor (divisor) */ 8891 RACK_OPTS_INC(tcp_rack_pace_reduce); 8892 if (optval) 8893 /* Must be non-zero */ 8894 rack->rc_pace_reduce = optval; 8895 else 8896 error = EINVAL; 8897 break; 8898 case TCP_RACK_PACE_MAX_SEG: 8899 /* Max segments in a pace */ 8900 RACK_OPTS_INC(tcp_rack_max_seg); 8901 rack->rc_pace_max_segs = optval; 8902 break; 8903 case TCP_RACK_PRR_SENDALOT: 8904 /* Allow PRR to send more than one seg */ 8905 RACK_OPTS_INC(tcp_rack_prr_sendalot); 8906 rack->r_ctl.rc_prr_sendalot = optval; 8907 break; 8908 case TCP_RACK_MIN_TO: 8909 /* Minimum time between rack t-o's in ms */ 8910 RACK_OPTS_INC(tcp_rack_min_to); 8911 rack->r_ctl.rc_min_to = optval; 8912 break; 8913 case TCP_RACK_EARLY_SEG: 8914 /* If early recovery max segments */ 8915 RACK_OPTS_INC(tcp_rack_early_seg); 8916 rack->r_ctl.rc_early_recovery_segs = optval; 8917 break; 8918 case TCP_RACK_REORD_THRESH: 8919 /* RACK reorder threshold (shift amount) */ 8920 RACK_OPTS_INC(tcp_rack_reord_thresh); 8921 if ((optval > 0) && (optval < 31)) 8922 rack->r_ctl.rc_reorder_shift = optval; 8923 else 8924 error = EINVAL; 8925 break; 8926 case TCP_RACK_REORD_FADE: 8927 /* Does reordering fade after ms time */ 8928 RACK_OPTS_INC(tcp_rack_reord_fade); 8929 rack->r_ctl.rc_reorder_fade = optval; 8930 break; 8931 case TCP_RACK_TLP_THRESH: 8932 /* RACK TLP theshold i.e. srtt+(srtt/N) */ 8933 RACK_OPTS_INC(tcp_rack_tlp_thresh); 8934 if (optval) 8935 rack->r_ctl.rc_tlp_threshold = optval; 8936 else 8937 error = EINVAL; 8938 break; 8939 case TCP_RACK_PKT_DELAY: 8940 /* RACK added ms i.e. rack-rtt + reord + N */ 8941 RACK_OPTS_INC(tcp_rack_pkt_delay); 8942 rack->r_ctl.rc_pkt_delay = optval; 8943 break; 8944 case TCP_RACK_TLP_INC_VAR: 8945 /* Does TLP include rtt variance in t-o */ 8946 RACK_OPTS_INC(tcp_rack_tlp_inc_var); 8947 rack->r_ctl.rc_prr_inc_var = optval; 8948 break; 8949 case TCP_RACK_IDLE_REDUCE_HIGH: 8950 RACK_OPTS_INC(tcp_rack_idle_reduce_high); 8951 if (optval) 8952 rack->r_idle_reduce_largest = 1; 8953 else 8954 rack->r_idle_reduce_largest = 0; 8955 break; 8956 case TCP_DELACK: 8957 if (optval == 0) 8958 tp->t_delayed_ack = 0; 8959 else 8960 tp->t_delayed_ack = 1; 8961 if (tp->t_flags & TF_DELACK) { 8962 tp->t_flags &= ~TF_DELACK; 8963 tp->t_flags |= TF_ACKNOW; 8964 rack_output(tp); 8965 } 8966 break; 8967 case TCP_RACK_MIN_PACE: 8968 RACK_OPTS_INC(tcp_rack_min_pace); 8969 if (optval > 3) 8970 rack->r_enforce_min_pace = 3; 8971 else 8972 rack->r_enforce_min_pace = optval; 8973 break; 8974 case TCP_RACK_MIN_PACE_SEG: 8975 RACK_OPTS_INC(tcp_rack_min_pace_seg); 8976 if (optval >= 16) 8977 rack->r_min_pace_seg_thresh = 15; 8978 else 8979 rack->r_min_pace_seg_thresh = optval; 8980 break; 8981 case TCP_BBR_RACK_RTT_USE: 8982 if ((optval != USE_RTT_HIGH) && 8983 (optval != USE_RTT_LOW) && 8984 (optval != USE_RTT_AVG)) 8985 error = EINVAL; 8986 else 8987 rack->r_ctl.rc_rate_sample_method = optval; 8988 break; 8989 case TCP_DATA_AFTER_CLOSE: 8990 if (optval) 8991 rack->rc_allow_data_af_clo = 1; 8992 else 8993 rack->rc_allow_data_af_clo = 0; 8994 break; 8995 default: 8996 return (tcp_default_ctloutput(so, sopt, inp, tp)); 8997 break; 8998 } 8999 #ifdef NETFLIX_STATS 9000 tcp_log_socket_option(tp, sopt->sopt_name, optval, error); 9001 #endif 9002 INP_WUNLOCK(inp); 9003 return (error); 9004 } 9005 9006 static int 9007 rack_get_sockopt(struct socket *so, struct sockopt *sopt, 9008 struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack) 9009 { 9010 int32_t error, optval; 9011 9012 /* 9013 * Because all our options are either boolean or an int, we can just 9014 * pull everything into optval and then unlock and copy. If we ever 9015 * add a option that is not a int, then this will have quite an 9016 * impact to this routine. 9017 */ 9018 switch (sopt->sopt_name) { 9019 case TCP_RACK_PROP_RATE: 9020 optval = rack->r_ctl.rc_prop_rate; 9021 break; 9022 case TCP_RACK_PROP: 9023 /* RACK proportional rate reduction (bool) */ 9024 optval = rack->r_ctl.rc_prop_reduce; 9025 break; 9026 case TCP_RACK_TLP_REDUCE: 9027 /* RACK TLP cwnd reduction (bool) */ 9028 optval = rack->r_ctl.rc_tlp_cwnd_reduce; 9029 break; 9030 case TCP_RACK_EARLY_RECOV: 9031 /* Should recovery happen early (bool) */ 9032 optval = rack->r_ctl.rc_early_recovery; 9033 break; 9034 case TCP_RACK_PACE_REDUCE: 9035 /* RACK Hptsi reduction factor (divisor) */ 9036 optval = rack->rc_pace_reduce; 9037 break; 9038 case TCP_RACK_PACE_MAX_SEG: 9039 /* Max segments in a pace */ 9040 optval = rack->rc_pace_max_segs; 9041 break; 9042 case TCP_RACK_PACE_ALWAYS: 9043 /* Use the always pace method */ 9044 optval = rack->rc_always_pace; 9045 break; 9046 case TCP_RACK_PRR_SENDALOT: 9047 /* Allow PRR to send more than one seg */ 9048 optval = rack->r_ctl.rc_prr_sendalot; 9049 break; 9050 case TCP_RACK_MIN_TO: 9051 /* Minimum time between rack t-o's in ms */ 9052 optval = rack->r_ctl.rc_min_to; 9053 break; 9054 case TCP_RACK_EARLY_SEG: 9055 /* If early recovery max segments */ 9056 optval = rack->r_ctl.rc_early_recovery_segs; 9057 break; 9058 case TCP_RACK_REORD_THRESH: 9059 /* RACK reorder threshold (shift amount) */ 9060 optval = rack->r_ctl.rc_reorder_shift; 9061 break; 9062 case TCP_RACK_REORD_FADE: 9063 /* Does reordering fade after ms time */ 9064 optval = rack->r_ctl.rc_reorder_fade; 9065 break; 9066 case TCP_RACK_TLP_THRESH: 9067 /* RACK TLP theshold i.e. srtt+(srtt/N) */ 9068 optval = rack->r_ctl.rc_tlp_threshold; 9069 break; 9070 case TCP_RACK_PKT_DELAY: 9071 /* RACK added ms i.e. rack-rtt + reord + N */ 9072 optval = rack->r_ctl.rc_pkt_delay; 9073 break; 9074 case TCP_RACK_TLP_USE: 9075 optval = rack->rack_tlp_threshold_use; 9076 break; 9077 case TCP_RACK_TLP_INC_VAR: 9078 /* Does TLP include rtt variance in t-o */ 9079 optval = rack->r_ctl.rc_prr_inc_var; 9080 break; 9081 case TCP_RACK_IDLE_REDUCE_HIGH: 9082 optval = rack->r_idle_reduce_largest; 9083 break; 9084 case TCP_RACK_MIN_PACE: 9085 optval = rack->r_enforce_min_pace; 9086 break; 9087 case TCP_RACK_MIN_PACE_SEG: 9088 optval = rack->r_min_pace_seg_thresh; 9089 break; 9090 case TCP_BBR_RACK_RTT_USE: 9091 optval = rack->r_ctl.rc_rate_sample_method; 9092 break; 9093 case TCP_DELACK: 9094 optval = tp->t_delayed_ack; 9095 break; 9096 case TCP_DATA_AFTER_CLOSE: 9097 optval = rack->rc_allow_data_af_clo; 9098 break; 9099 default: 9100 return (tcp_default_ctloutput(so, sopt, inp, tp)); 9101 break; 9102 } 9103 INP_WUNLOCK(inp); 9104 error = sooptcopyout(sopt, &optval, sizeof optval); 9105 return (error); 9106 } 9107 9108 static int 9109 rack_ctloutput(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp) 9110 { 9111 int32_t error = EINVAL; 9112 struct tcp_rack *rack; 9113 9114 rack = (struct tcp_rack *)tp->t_fb_ptr; 9115 if (rack == NULL) { 9116 /* Huh? */ 9117 goto out; 9118 } 9119 if (sopt->sopt_dir == SOPT_SET) { 9120 return (rack_set_sockopt(so, sopt, inp, tp, rack)); 9121 } else if (sopt->sopt_dir == SOPT_GET) { 9122 return (rack_get_sockopt(so, sopt, inp, tp, rack)); 9123 } 9124 out: 9125 INP_WUNLOCK(inp); 9126 return (error); 9127 } 9128 9129 9130 struct tcp_function_block __tcp_rack = { 9131 .tfb_tcp_block_name = __XSTRING(STACKNAME), 9132 .tfb_tcp_output = rack_output, 9133 .tfb_tcp_do_segment = rack_do_segment, 9134 .tfb_tcp_hpts_do_segment = rack_hpts_do_segment, 9135 .tfb_tcp_ctloutput = rack_ctloutput, 9136 .tfb_tcp_fb_init = rack_init, 9137 .tfb_tcp_fb_fini = rack_fini, 9138 .tfb_tcp_timer_stop_all = rack_stopall, 9139 .tfb_tcp_timer_activate = rack_timer_activate, 9140 .tfb_tcp_timer_active = rack_timer_active, 9141 .tfb_tcp_timer_stop = rack_timer_stop, 9142 .tfb_tcp_rexmit_tmr = rack_remxt_tmr, 9143 .tfb_tcp_handoff_ok = rack_handoff_ok 9144 }; 9145 9146 static const char *rack_stack_names[] = { 9147 __XSTRING(STACKNAME), 9148 #ifdef STACKALIAS 9149 __XSTRING(STACKALIAS), 9150 #endif 9151 }; 9152 9153 static int 9154 rack_ctor(void *mem, int32_t size, void *arg, int32_t how) 9155 { 9156 memset(mem, 0, size); 9157 return (0); 9158 } 9159 9160 static void 9161 rack_dtor(void *mem, int32_t size, void *arg) 9162 { 9163 9164 } 9165 9166 static bool rack_mod_inited = false; 9167 9168 static int 9169 tcp_addrack(module_t mod, int32_t type, void *data) 9170 { 9171 int32_t err = 0; 9172 int num_stacks; 9173 9174 switch (type) { 9175 case MOD_LOAD: 9176 rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map", 9177 sizeof(struct rack_sendmap), 9178 rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0); 9179 9180 rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb", 9181 sizeof(struct tcp_rack), 9182 rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0); 9183 9184 sysctl_ctx_init(&rack_sysctl_ctx); 9185 rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx, 9186 SYSCTL_STATIC_CHILDREN(_net_inet_tcp), 9187 OID_AUTO, 9188 __XSTRING(STACKNAME), 9189 CTLFLAG_RW, 0, 9190 ""); 9191 if (rack_sysctl_root == NULL) { 9192 printf("Failed to add sysctl node\n"); 9193 err = EFAULT; 9194 goto free_uma; 9195 } 9196 rack_init_sysctls(); 9197 num_stacks = nitems(rack_stack_names); 9198 err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK, 9199 rack_stack_names, &num_stacks); 9200 if (err) { 9201 printf("Failed to register %s stack name for " 9202 "%s module\n", rack_stack_names[num_stacks], 9203 __XSTRING(MODNAME)); 9204 sysctl_ctx_free(&rack_sysctl_ctx); 9205 free_uma: 9206 uma_zdestroy(rack_zone); 9207 uma_zdestroy(rack_pcb_zone); 9208 rack_counter_destroy(); 9209 printf("Failed to register rack module -- err:%d\n", err); 9210 return (err); 9211 } 9212 rack_mod_inited = true; 9213 break; 9214 case MOD_QUIESCE: 9215 err = deregister_tcp_functions(&__tcp_rack, true, false); 9216 break; 9217 case MOD_UNLOAD: 9218 err = deregister_tcp_functions(&__tcp_rack, false, true); 9219 if (err == EBUSY) 9220 break; 9221 if (rack_mod_inited) { 9222 uma_zdestroy(rack_zone); 9223 uma_zdestroy(rack_pcb_zone); 9224 sysctl_ctx_free(&rack_sysctl_ctx); 9225 rack_counter_destroy(); 9226 rack_mod_inited = false; 9227 } 9228 err = 0; 9229 break; 9230 default: 9231 return (EOPNOTSUPP); 9232 } 9233 return (err); 9234 } 9235 9236 static moduledata_t tcp_rack = { 9237 .name = __XSTRING(MODNAME), 9238 .evhand = tcp_addrack, 9239 .priv = 0 9240 }; 9241 9242 MODULE_VERSION(MODNAME, 1); 9243 DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY); 9244 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1); 9245