1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * INET An implementation of the TCP/IP protocol suite for the LINUX 4 * operating system. INET is implemented using the BSD Socket 5 * interface as the means of communication with the user level. 6 * 7 * Implementation of the Transmission Control Protocol(TCP). 8 * 9 * Authors: Ross Biro 10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 11 * Mark Evans, <evansmp@uhura.aston.ac.uk> 12 * Corey Minyard <wf-rch!minyard@relay.EU.net> 13 * Florian La Roche, <flla@stud.uni-sb.de> 14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> 15 * Linus Torvalds, <torvalds@cs.helsinki.fi> 16 * Alan Cox, <gw4pts@gw4pts.ampr.org> 17 * Matthew Dillon, <dillon@apollo.west.oic.com> 18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 19 * Jorge Cwik, <jorge@laser.satlink.net> 20 */ 21 22 /* 23 * Changes: 24 * Pedro Roque : Fast Retransmit/Recovery. 25 * Two receive queues. 26 * Retransmit queue handled by TCP. 27 * Better retransmit timer handling. 28 * New congestion avoidance. 29 * Header prediction. 30 * Variable renaming. 31 * 32 * Eric : Fast Retransmit. 33 * Randy Scott : MSS option defines. 34 * Eric Schenk : Fixes to slow start algorithm. 35 * Eric Schenk : Yet another double ACK bug. 36 * Eric Schenk : Delayed ACK bug fixes. 37 * Eric Schenk : Floyd style fast retrans war avoidance. 38 * David S. Miller : Don't allow zero congestion window. 39 * Eric Schenk : Fix retransmitter so that it sends 40 * next packet on ack of previous packet. 41 * Andi Kleen : Moved open_request checking here 42 * and process RSTs for open_requests. 43 * Andi Kleen : Better prune_queue, and other fixes. 44 * Andrey Savochkin: Fix RTT measurements in the presence of 45 * timestamps. 46 * Andrey Savochkin: Check sequence numbers correctly when 47 * removing SACKs due to in sequence incoming 48 * data segments. 49 * Andi Kleen: Make sure we never ack data there is not 50 * enough room for. Also make this condition 51 * a fatal error if it might still happen. 52 * Andi Kleen: Add tcp_measure_rcv_mss to make 53 * connections with MSS<min(MTU,ann. MSS) 54 * work without delayed acks. 55 * Andi Kleen: Process packets with PSH set in the 56 * fast path. 57 * J Hadi Salim: ECN support 58 * Andrei Gurtov, 59 * Pasi Sarolahti, 60 * Panu Kuhlberg: Experimental audit of TCP (re)transmission 61 * engine. Lots of bugs are found. 62 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs 63 */ 64 65 #define pr_fmt(fmt) "TCP: " fmt 66 67 #include <linux/mm.h> 68 #include <linux/slab.h> 69 #include <linux/module.h> 70 #include <linux/sysctl.h> 71 #include <linux/kernel.h> 72 #include <linux/prefetch.h> 73 #include <net/dst.h> 74 #include <net/tcp.h> 75 #include <net/inet_common.h> 76 #include <linux/ipsec.h> 77 #include <asm/unaligned.h> 78 #include <linux/errqueue.h> 79 #include <trace/events/tcp.h> 80 #include <linux/jump_label_ratelimit.h> 81 #include <net/busy_poll.h> 82 83 int sysctl_tcp_max_orphans __read_mostly = NR_FILE; 84 85 #define FLAG_DATA 0x01 /* Incoming frame contained data. */ 86 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */ 87 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */ 88 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */ 89 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */ 90 #define FLAG_DATA_SACKED 0x20 /* New SACK. */ 91 #define FLAG_ECE 0x40 /* ECE in this ACK */ 92 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */ 93 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/ 94 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */ 95 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */ 96 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */ 97 #define FLAG_SET_XMIT_TIMER 0x1000 /* Set TLP or RTO timer */ 98 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */ 99 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */ 100 #define FLAG_NO_CHALLENGE_ACK 0x8000 /* do not call tcp_send_challenge_ack() */ 101 #define FLAG_ACK_MAYBE_DELAYED 0x10000 /* Likely a delayed ACK */ 102 103 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) 104 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) 105 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK) 106 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) 107 108 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) 109 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH)) 110 111 #define REXMIT_NONE 0 /* no loss recovery to do */ 112 #define REXMIT_LOST 1 /* retransmit packets marked lost */ 113 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */ 114 115 #if IS_ENABLED(CONFIG_TLS_DEVICE) 116 static DEFINE_STATIC_KEY_DEFERRED_FALSE(clean_acked_data_enabled, HZ); 117 118 void clean_acked_data_enable(struct inet_connection_sock *icsk, 119 void (*cad)(struct sock *sk, u32 ack_seq)) 120 { 121 icsk->icsk_clean_acked = cad; 122 static_branch_inc(&clean_acked_data_enabled.key); 123 } 124 EXPORT_SYMBOL_GPL(clean_acked_data_enable); 125 126 void clean_acked_data_disable(struct inet_connection_sock *icsk) 127 { 128 static_branch_slow_dec_deferred(&clean_acked_data_enabled); 129 icsk->icsk_clean_acked = NULL; 130 } 131 EXPORT_SYMBOL_GPL(clean_acked_data_disable); 132 133 void clean_acked_data_flush(void) 134 { 135 static_key_deferred_flush(&clean_acked_data_enabled); 136 } 137 EXPORT_SYMBOL_GPL(clean_acked_data_flush); 138 #endif 139 140 static void tcp_gro_dev_warn(struct sock *sk, const struct sk_buff *skb, 141 unsigned int len) 142 { 143 static bool __once __read_mostly; 144 145 if (!__once) { 146 struct net_device *dev; 147 148 __once = true; 149 150 rcu_read_lock(); 151 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif); 152 if (!dev || len >= dev->mtu) 153 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n", 154 dev ? dev->name : "Unknown driver"); 155 rcu_read_unlock(); 156 } 157 } 158 159 /* Adapt the MSS value used to make delayed ack decision to the 160 * real world. 161 */ 162 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb) 163 { 164 struct inet_connection_sock *icsk = inet_csk(sk); 165 const unsigned int lss = icsk->icsk_ack.last_seg_size; 166 unsigned int len; 167 168 icsk->icsk_ack.last_seg_size = 0; 169 170 /* skb->len may jitter because of SACKs, even if peer 171 * sends good full-sized frames. 172 */ 173 len = skb_shinfo(skb)->gso_size ? : skb->len; 174 if (len >= icsk->icsk_ack.rcv_mss) { 175 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len, 176 tcp_sk(sk)->advmss); 177 /* Account for possibly-removed options */ 178 if (unlikely(len > icsk->icsk_ack.rcv_mss + 179 MAX_TCP_OPTION_SPACE)) 180 tcp_gro_dev_warn(sk, skb, len); 181 } else { 182 /* Otherwise, we make more careful check taking into account, 183 * that SACKs block is variable. 184 * 185 * "len" is invariant segment length, including TCP header. 186 */ 187 len += skb->data - skb_transport_header(skb); 188 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) || 189 /* If PSH is not set, packet should be 190 * full sized, provided peer TCP is not badly broken. 191 * This observation (if it is correct 8)) allows 192 * to handle super-low mtu links fairly. 193 */ 194 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && 195 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) { 196 /* Subtract also invariant (if peer is RFC compliant), 197 * tcp header plus fixed timestamp option length. 198 * Resulting "len" is MSS free of SACK jitter. 199 */ 200 len -= tcp_sk(sk)->tcp_header_len; 201 icsk->icsk_ack.last_seg_size = len; 202 if (len == lss) { 203 icsk->icsk_ack.rcv_mss = len; 204 return; 205 } 206 } 207 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) 208 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; 209 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 210 } 211 } 212 213 static void tcp_incr_quickack(struct sock *sk, unsigned int max_quickacks) 214 { 215 struct inet_connection_sock *icsk = inet_csk(sk); 216 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); 217 218 if (quickacks == 0) 219 quickacks = 2; 220 quickacks = min(quickacks, max_quickacks); 221 if (quickacks > icsk->icsk_ack.quick) 222 icsk->icsk_ack.quick = quickacks; 223 } 224 225 void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks) 226 { 227 struct inet_connection_sock *icsk = inet_csk(sk); 228 229 tcp_incr_quickack(sk, max_quickacks); 230 inet_csk_exit_pingpong_mode(sk); 231 icsk->icsk_ack.ato = TCP_ATO_MIN; 232 } 233 EXPORT_SYMBOL(tcp_enter_quickack_mode); 234 235 /* Send ACKs quickly, if "quick" count is not exhausted 236 * and the session is not interactive. 237 */ 238 239 static bool tcp_in_quickack_mode(struct sock *sk) 240 { 241 const struct inet_connection_sock *icsk = inet_csk(sk); 242 const struct dst_entry *dst = __sk_dst_get(sk); 243 244 return (dst && dst_metric(dst, RTAX_QUICKACK)) || 245 (icsk->icsk_ack.quick && !inet_csk_in_pingpong_mode(sk)); 246 } 247 248 static void tcp_ecn_queue_cwr(struct tcp_sock *tp) 249 { 250 if (tp->ecn_flags & TCP_ECN_OK) 251 tp->ecn_flags |= TCP_ECN_QUEUE_CWR; 252 } 253 254 static void tcp_ecn_accept_cwr(struct sock *sk, const struct sk_buff *skb) 255 { 256 if (tcp_hdr(skb)->cwr) { 257 tcp_sk(sk)->ecn_flags &= ~TCP_ECN_DEMAND_CWR; 258 259 /* If the sender is telling us it has entered CWR, then its 260 * cwnd may be very low (even just 1 packet), so we should ACK 261 * immediately. 262 */ 263 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW; 264 } 265 } 266 267 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp) 268 { 269 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; 270 } 271 272 static void __tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb) 273 { 274 struct tcp_sock *tp = tcp_sk(sk); 275 276 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) { 277 case INET_ECN_NOT_ECT: 278 /* Funny extension: if ECT is not set on a segment, 279 * and we already seen ECT on a previous segment, 280 * it is probably a retransmit. 281 */ 282 if (tp->ecn_flags & TCP_ECN_SEEN) 283 tcp_enter_quickack_mode(sk, 2); 284 break; 285 case INET_ECN_CE: 286 if (tcp_ca_needs_ecn(sk)) 287 tcp_ca_event(sk, CA_EVENT_ECN_IS_CE); 288 289 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) { 290 /* Better not delay acks, sender can have a very low cwnd */ 291 tcp_enter_quickack_mode(sk, 2); 292 tp->ecn_flags |= TCP_ECN_DEMAND_CWR; 293 } 294 tp->ecn_flags |= TCP_ECN_SEEN; 295 break; 296 default: 297 if (tcp_ca_needs_ecn(sk)) 298 tcp_ca_event(sk, CA_EVENT_ECN_NO_CE); 299 tp->ecn_flags |= TCP_ECN_SEEN; 300 break; 301 } 302 } 303 304 static void tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb) 305 { 306 if (tcp_sk(sk)->ecn_flags & TCP_ECN_OK) 307 __tcp_ecn_check_ce(sk, skb); 308 } 309 310 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th) 311 { 312 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr)) 313 tp->ecn_flags &= ~TCP_ECN_OK; 314 } 315 316 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th) 317 { 318 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr)) 319 tp->ecn_flags &= ~TCP_ECN_OK; 320 } 321 322 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th) 323 { 324 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK)) 325 return true; 326 return false; 327 } 328 329 /* Buffer size and advertised window tuning. 330 * 331 * 1. Tuning sk->sk_sndbuf, when connection enters established state. 332 */ 333 334 static void tcp_sndbuf_expand(struct sock *sk) 335 { 336 const struct tcp_sock *tp = tcp_sk(sk); 337 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 338 int sndmem, per_mss; 339 u32 nr_segs; 340 341 /* Worst case is non GSO/TSO : each frame consumes one skb 342 * and skb->head is kmalloced using power of two area of memory 343 */ 344 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) + 345 MAX_TCP_HEADER + 346 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 347 348 per_mss = roundup_pow_of_two(per_mss) + 349 SKB_DATA_ALIGN(sizeof(struct sk_buff)); 350 351 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd); 352 nr_segs = max_t(u32, nr_segs, tp->reordering + 1); 353 354 /* Fast Recovery (RFC 5681 3.2) : 355 * Cubic needs 1.7 factor, rounded to 2 to include 356 * extra cushion (application might react slowly to EPOLLOUT) 357 */ 358 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2; 359 sndmem *= nr_segs * per_mss; 360 361 if (sk->sk_sndbuf < sndmem) 362 sk->sk_sndbuf = min(sndmem, sock_net(sk)->ipv4.sysctl_tcp_wmem[2]); 363 } 364 365 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh) 366 * 367 * All tcp_full_space() is split to two parts: "network" buffer, allocated 368 * forward and advertised in receiver window (tp->rcv_wnd) and 369 * "application buffer", required to isolate scheduling/application 370 * latencies from network. 371 * window_clamp is maximal advertised window. It can be less than 372 * tcp_full_space(), in this case tcp_full_space() - window_clamp 373 * is reserved for "application" buffer. The less window_clamp is 374 * the smoother our behaviour from viewpoint of network, but the lower 375 * throughput and the higher sensitivity of the connection to losses. 8) 376 * 377 * rcv_ssthresh is more strict window_clamp used at "slow start" 378 * phase to predict further behaviour of this connection. 379 * It is used for two goals: 380 * - to enforce header prediction at sender, even when application 381 * requires some significant "application buffer". It is check #1. 382 * - to prevent pruning of receive queue because of misprediction 383 * of receiver window. Check #2. 384 * 385 * The scheme does not work when sender sends good segments opening 386 * window and then starts to feed us spaghetti. But it should work 387 * in common situations. Otherwise, we have to rely on queue collapsing. 388 */ 389 390 /* Slow part of check#2. */ 391 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb) 392 { 393 struct tcp_sock *tp = tcp_sk(sk); 394 /* Optimize this! */ 395 int truesize = tcp_win_from_space(sk, skb->truesize) >> 1; 396 int window = tcp_win_from_space(sk, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1; 397 398 while (tp->rcv_ssthresh <= window) { 399 if (truesize <= skb->len) 400 return 2 * inet_csk(sk)->icsk_ack.rcv_mss; 401 402 truesize >>= 1; 403 window >>= 1; 404 } 405 return 0; 406 } 407 408 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb) 409 { 410 struct tcp_sock *tp = tcp_sk(sk); 411 int room; 412 413 room = min_t(int, tp->window_clamp, tcp_space(sk)) - tp->rcv_ssthresh; 414 415 /* Check #1 */ 416 if (room > 0 && !tcp_under_memory_pressure(sk)) { 417 int incr; 418 419 /* Check #2. Increase window, if skb with such overhead 420 * will fit to rcvbuf in future. 421 */ 422 if (tcp_win_from_space(sk, skb->truesize) <= skb->len) 423 incr = 2 * tp->advmss; 424 else 425 incr = __tcp_grow_window(sk, skb); 426 427 if (incr) { 428 incr = max_t(int, incr, 2 * skb->len); 429 tp->rcv_ssthresh += min(room, incr); 430 inet_csk(sk)->icsk_ack.quick |= 1; 431 } 432 } 433 } 434 435 /* 3. Try to fixup all. It is made immediately after connection enters 436 * established state. 437 */ 438 void tcp_init_buffer_space(struct sock *sk) 439 { 440 int tcp_app_win = sock_net(sk)->ipv4.sysctl_tcp_app_win; 441 struct tcp_sock *tp = tcp_sk(sk); 442 int maxwin; 443 444 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) 445 tcp_sndbuf_expand(sk); 446 447 tp->rcvq_space.space = min_t(u32, tp->rcv_wnd, TCP_INIT_CWND * tp->advmss); 448 tcp_mstamp_refresh(tp); 449 tp->rcvq_space.time = tp->tcp_mstamp; 450 tp->rcvq_space.seq = tp->copied_seq; 451 452 maxwin = tcp_full_space(sk); 453 454 if (tp->window_clamp >= maxwin) { 455 tp->window_clamp = maxwin; 456 457 if (tcp_app_win && maxwin > 4 * tp->advmss) 458 tp->window_clamp = max(maxwin - 459 (maxwin >> tcp_app_win), 460 4 * tp->advmss); 461 } 462 463 /* Force reservation of one segment. */ 464 if (tcp_app_win && 465 tp->window_clamp > 2 * tp->advmss && 466 tp->window_clamp + tp->advmss > maxwin) 467 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss); 468 469 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); 470 tp->snd_cwnd_stamp = tcp_jiffies32; 471 } 472 473 /* 4. Recalculate window clamp after socket hit its memory bounds. */ 474 static void tcp_clamp_window(struct sock *sk) 475 { 476 struct tcp_sock *tp = tcp_sk(sk); 477 struct inet_connection_sock *icsk = inet_csk(sk); 478 struct net *net = sock_net(sk); 479 480 icsk->icsk_ack.quick = 0; 481 482 if (sk->sk_rcvbuf < net->ipv4.sysctl_tcp_rmem[2] && 483 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && 484 !tcp_under_memory_pressure(sk) && 485 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) { 486 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc), 487 net->ipv4.sysctl_tcp_rmem[2]); 488 } 489 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) 490 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss); 491 } 492 493 /* Initialize RCV_MSS value. 494 * RCV_MSS is an our guess about MSS used by the peer. 495 * We haven't any direct information about the MSS. 496 * It's better to underestimate the RCV_MSS rather than overestimate. 497 * Overestimations make us ACKing less frequently than needed. 498 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). 499 */ 500 void tcp_initialize_rcv_mss(struct sock *sk) 501 { 502 const struct tcp_sock *tp = tcp_sk(sk); 503 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); 504 505 hint = min(hint, tp->rcv_wnd / 2); 506 hint = min(hint, TCP_MSS_DEFAULT); 507 hint = max(hint, TCP_MIN_MSS); 508 509 inet_csk(sk)->icsk_ack.rcv_mss = hint; 510 } 511 EXPORT_SYMBOL(tcp_initialize_rcv_mss); 512 513 /* Receiver "autotuning" code. 514 * 515 * The algorithm for RTT estimation w/o timestamps is based on 516 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. 517 * <http://public.lanl.gov/radiant/pubs.html#DRS> 518 * 519 * More detail on this code can be found at 520 * <http://staff.psc.edu/jheffner/>, 521 * though this reference is out of date. A new paper 522 * is pending. 523 */ 524 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) 525 { 526 u32 new_sample = tp->rcv_rtt_est.rtt_us; 527 long m = sample; 528 529 if (new_sample != 0) { 530 /* If we sample in larger samples in the non-timestamp 531 * case, we could grossly overestimate the RTT especially 532 * with chatty applications or bulk transfer apps which 533 * are stalled on filesystem I/O. 534 * 535 * Also, since we are only going for a minimum in the 536 * non-timestamp case, we do not smooth things out 537 * else with timestamps disabled convergence takes too 538 * long. 539 */ 540 if (!win_dep) { 541 m -= (new_sample >> 3); 542 new_sample += m; 543 } else { 544 m <<= 3; 545 if (m < new_sample) 546 new_sample = m; 547 } 548 } else { 549 /* No previous measure. */ 550 new_sample = m << 3; 551 } 552 553 tp->rcv_rtt_est.rtt_us = new_sample; 554 } 555 556 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) 557 { 558 u32 delta_us; 559 560 if (tp->rcv_rtt_est.time == 0) 561 goto new_measure; 562 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) 563 return; 564 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time); 565 if (!delta_us) 566 delta_us = 1; 567 tcp_rcv_rtt_update(tp, delta_us, 1); 568 569 new_measure: 570 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; 571 tp->rcv_rtt_est.time = tp->tcp_mstamp; 572 } 573 574 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, 575 const struct sk_buff *skb) 576 { 577 struct tcp_sock *tp = tcp_sk(sk); 578 579 if (tp->rx_opt.rcv_tsecr == tp->rcv_rtt_last_tsecr) 580 return; 581 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr; 582 583 if (TCP_SKB_CB(skb)->end_seq - 584 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) { 585 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr; 586 u32 delta_us; 587 588 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) { 589 if (!delta) 590 delta = 1; 591 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ); 592 tcp_rcv_rtt_update(tp, delta_us, 0); 593 } 594 } 595 } 596 597 /* 598 * This function should be called every time data is copied to user space. 599 * It calculates the appropriate TCP receive buffer space. 600 */ 601 void tcp_rcv_space_adjust(struct sock *sk) 602 { 603 struct tcp_sock *tp = tcp_sk(sk); 604 u32 copied; 605 int time; 606 607 trace_tcp_rcv_space_adjust(sk); 608 609 tcp_mstamp_refresh(tp); 610 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time); 611 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0) 612 return; 613 614 /* Number of bytes copied to user in last RTT */ 615 copied = tp->copied_seq - tp->rcvq_space.seq; 616 if (copied <= tp->rcvq_space.space) 617 goto new_measure; 618 619 /* A bit of theory : 620 * copied = bytes received in previous RTT, our base window 621 * To cope with packet losses, we need a 2x factor 622 * To cope with slow start, and sender growing its cwin by 100 % 623 * every RTT, we need a 4x factor, because the ACK we are sending 624 * now is for the next RTT, not the current one : 625 * <prev RTT . ><current RTT .. ><next RTT .... > 626 */ 627 628 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf && 629 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { 630 int rcvmem, rcvbuf; 631 u64 rcvwin, grow; 632 633 /* minimal window to cope with packet losses, assuming 634 * steady state. Add some cushion because of small variations. 635 */ 636 rcvwin = ((u64)copied << 1) + 16 * tp->advmss; 637 638 /* Accommodate for sender rate increase (eg. slow start) */ 639 grow = rcvwin * (copied - tp->rcvq_space.space); 640 do_div(grow, tp->rcvq_space.space); 641 rcvwin += (grow << 1); 642 643 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER); 644 while (tcp_win_from_space(sk, rcvmem) < tp->advmss) 645 rcvmem += 128; 646 647 do_div(rcvwin, tp->advmss); 648 rcvbuf = min_t(u64, rcvwin * rcvmem, 649 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]); 650 if (rcvbuf > sk->sk_rcvbuf) { 651 sk->sk_rcvbuf = rcvbuf; 652 653 /* Make the window clamp follow along. */ 654 tp->window_clamp = tcp_win_from_space(sk, rcvbuf); 655 } 656 } 657 tp->rcvq_space.space = copied; 658 659 new_measure: 660 tp->rcvq_space.seq = tp->copied_seq; 661 tp->rcvq_space.time = tp->tcp_mstamp; 662 } 663 664 /* There is something which you must keep in mind when you analyze the 665 * behavior of the tp->ato delayed ack timeout interval. When a 666 * connection starts up, we want to ack as quickly as possible. The 667 * problem is that "good" TCP's do slow start at the beginning of data 668 * transmission. The means that until we send the first few ACK's the 669 * sender will sit on his end and only queue most of his data, because 670 * he can only send snd_cwnd unacked packets at any given time. For 671 * each ACK we send, he increments snd_cwnd and transmits more of his 672 * queue. -DaveM 673 */ 674 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb) 675 { 676 struct tcp_sock *tp = tcp_sk(sk); 677 struct inet_connection_sock *icsk = inet_csk(sk); 678 u32 now; 679 680 inet_csk_schedule_ack(sk); 681 682 tcp_measure_rcv_mss(sk, skb); 683 684 tcp_rcv_rtt_measure(tp); 685 686 now = tcp_jiffies32; 687 688 if (!icsk->icsk_ack.ato) { 689 /* The _first_ data packet received, initialize 690 * delayed ACK engine. 691 */ 692 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS); 693 icsk->icsk_ack.ato = TCP_ATO_MIN; 694 } else { 695 int m = now - icsk->icsk_ack.lrcvtime; 696 697 if (m <= TCP_ATO_MIN / 2) { 698 /* The fastest case is the first. */ 699 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; 700 } else if (m < icsk->icsk_ack.ato) { 701 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; 702 if (icsk->icsk_ack.ato > icsk->icsk_rto) 703 icsk->icsk_ack.ato = icsk->icsk_rto; 704 } else if (m > icsk->icsk_rto) { 705 /* Too long gap. Apparently sender failed to 706 * restart window, so that we send ACKs quickly. 707 */ 708 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS); 709 sk_mem_reclaim(sk); 710 } 711 } 712 icsk->icsk_ack.lrcvtime = now; 713 714 tcp_ecn_check_ce(sk, skb); 715 716 if (skb->len >= 128) 717 tcp_grow_window(sk, skb); 718 } 719 720 /* Called to compute a smoothed rtt estimate. The data fed to this 721 * routine either comes from timestamps, or from segments that were 722 * known _not_ to have been retransmitted [see Karn/Partridge 723 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 724 * piece by Van Jacobson. 725 * NOTE: the next three routines used to be one big routine. 726 * To save cycles in the RFC 1323 implementation it was better to break 727 * it up into three procedures. -- erics 728 */ 729 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us) 730 { 731 struct tcp_sock *tp = tcp_sk(sk); 732 long m = mrtt_us; /* RTT */ 733 u32 srtt = tp->srtt_us; 734 735 /* The following amusing code comes from Jacobson's 736 * article in SIGCOMM '88. Note that rtt and mdev 737 * are scaled versions of rtt and mean deviation. 738 * This is designed to be as fast as possible 739 * m stands for "measurement". 740 * 741 * On a 1990 paper the rto value is changed to: 742 * RTO = rtt + 4 * mdev 743 * 744 * Funny. This algorithm seems to be very broken. 745 * These formulae increase RTO, when it should be decreased, increase 746 * too slowly, when it should be increased quickly, decrease too quickly 747 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely 748 * does not matter how to _calculate_ it. Seems, it was trap 749 * that VJ failed to avoid. 8) 750 */ 751 if (srtt != 0) { 752 m -= (srtt >> 3); /* m is now error in rtt est */ 753 srtt += m; /* rtt = 7/8 rtt + 1/8 new */ 754 if (m < 0) { 755 m = -m; /* m is now abs(error) */ 756 m -= (tp->mdev_us >> 2); /* similar update on mdev */ 757 /* This is similar to one of Eifel findings. 758 * Eifel blocks mdev updates when rtt decreases. 759 * This solution is a bit different: we use finer gain 760 * for mdev in this case (alpha*beta). 761 * Like Eifel it also prevents growth of rto, 762 * but also it limits too fast rto decreases, 763 * happening in pure Eifel. 764 */ 765 if (m > 0) 766 m >>= 3; 767 } else { 768 m -= (tp->mdev_us >> 2); /* similar update on mdev */ 769 } 770 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */ 771 if (tp->mdev_us > tp->mdev_max_us) { 772 tp->mdev_max_us = tp->mdev_us; 773 if (tp->mdev_max_us > tp->rttvar_us) 774 tp->rttvar_us = tp->mdev_max_us; 775 } 776 if (after(tp->snd_una, tp->rtt_seq)) { 777 if (tp->mdev_max_us < tp->rttvar_us) 778 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2; 779 tp->rtt_seq = tp->snd_nxt; 780 tp->mdev_max_us = tcp_rto_min_us(sk); 781 } 782 } else { 783 /* no previous measure. */ 784 srtt = m << 3; /* take the measured time to be rtt */ 785 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */ 786 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk)); 787 tp->mdev_max_us = tp->rttvar_us; 788 tp->rtt_seq = tp->snd_nxt; 789 } 790 tp->srtt_us = max(1U, srtt); 791 } 792 793 static void tcp_update_pacing_rate(struct sock *sk) 794 { 795 const struct tcp_sock *tp = tcp_sk(sk); 796 u64 rate; 797 798 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */ 799 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3); 800 801 /* current rate is (cwnd * mss) / srtt 802 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate. 803 * In Congestion Avoidance phase, set it to 120 % the current rate. 804 * 805 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh) 806 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching 807 * end of slow start and should slow down. 808 */ 809 if (tp->snd_cwnd < tp->snd_ssthresh / 2) 810 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio; 811 else 812 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio; 813 814 rate *= max(tp->snd_cwnd, tp->packets_out); 815 816 if (likely(tp->srtt_us)) 817 do_div(rate, tp->srtt_us); 818 819 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate 820 * without any lock. We want to make sure compiler wont store 821 * intermediate values in this location. 822 */ 823 WRITE_ONCE(sk->sk_pacing_rate, min_t(u64, rate, 824 sk->sk_max_pacing_rate)); 825 } 826 827 /* Calculate rto without backoff. This is the second half of Van Jacobson's 828 * routine referred to above. 829 */ 830 static void tcp_set_rto(struct sock *sk) 831 { 832 const struct tcp_sock *tp = tcp_sk(sk); 833 /* Old crap is replaced with new one. 8) 834 * 835 * More seriously: 836 * 1. If rtt variance happened to be less 50msec, it is hallucination. 837 * It cannot be less due to utterly erratic ACK generation made 838 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ 839 * to do with delayed acks, because at cwnd>2 true delack timeout 840 * is invisible. Actually, Linux-2.4 also generates erratic 841 * ACKs in some circumstances. 842 */ 843 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp); 844 845 /* 2. Fixups made earlier cannot be right. 846 * If we do not estimate RTO correctly without them, 847 * all the algo is pure shit and should be replaced 848 * with correct one. It is exactly, which we pretend to do. 849 */ 850 851 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo 852 * guarantees that rto is higher. 853 */ 854 tcp_bound_rto(sk); 855 } 856 857 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst) 858 { 859 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); 860 861 if (!cwnd) 862 cwnd = TCP_INIT_CWND; 863 return min_t(__u32, cwnd, tp->snd_cwnd_clamp); 864 } 865 866 /* Take a notice that peer is sending D-SACKs */ 867 static void tcp_dsack_seen(struct tcp_sock *tp) 868 { 869 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN; 870 tp->rack.dsack_seen = 1; 871 tp->dsack_dups++; 872 } 873 874 /* It's reordering when higher sequence was delivered (i.e. sacked) before 875 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering 876 * distance is approximated in full-mss packet distance ("reordering"). 877 */ 878 static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq, 879 const int ts) 880 { 881 struct tcp_sock *tp = tcp_sk(sk); 882 const u32 mss = tp->mss_cache; 883 u32 fack, metric; 884 885 fack = tcp_highest_sack_seq(tp); 886 if (!before(low_seq, fack)) 887 return; 888 889 metric = fack - low_seq; 890 if ((metric > tp->reordering * mss) && mss) { 891 #if FASTRETRANS_DEBUG > 1 892 pr_debug("Disorder%d %d %u f%u s%u rr%d\n", 893 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, 894 tp->reordering, 895 0, 896 tp->sacked_out, 897 tp->undo_marker ? tp->undo_retrans : 0); 898 #endif 899 tp->reordering = min_t(u32, (metric + mss - 1) / mss, 900 sock_net(sk)->ipv4.sysctl_tcp_max_reordering); 901 } 902 903 /* This exciting event is worth to be remembered. 8) */ 904 tp->reord_seen++; 905 NET_INC_STATS(sock_net(sk), 906 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER); 907 } 908 909 /* This must be called before lost_out is incremented */ 910 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb) 911 { 912 if (!tp->retransmit_skb_hint || 913 before(TCP_SKB_CB(skb)->seq, 914 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) 915 tp->retransmit_skb_hint = skb; 916 } 917 918 /* Sum the number of packets on the wire we have marked as lost. 919 * There are two cases we care about here: 920 * a) Packet hasn't been marked lost (nor retransmitted), 921 * and this is the first loss. 922 * b) Packet has been marked both lost and retransmitted, 923 * and this means we think it was lost again. 924 */ 925 static void tcp_sum_lost(struct tcp_sock *tp, struct sk_buff *skb) 926 { 927 __u8 sacked = TCP_SKB_CB(skb)->sacked; 928 929 if (!(sacked & TCPCB_LOST) || 930 ((sacked & TCPCB_LOST) && (sacked & TCPCB_SACKED_RETRANS))) 931 tp->lost += tcp_skb_pcount(skb); 932 } 933 934 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb) 935 { 936 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { 937 tcp_verify_retransmit_hint(tp, skb); 938 939 tp->lost_out += tcp_skb_pcount(skb); 940 tcp_sum_lost(tp, skb); 941 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 942 } 943 } 944 945 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb) 946 { 947 tcp_verify_retransmit_hint(tp, skb); 948 949 tcp_sum_lost(tp, skb); 950 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { 951 tp->lost_out += tcp_skb_pcount(skb); 952 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 953 } 954 } 955 956 /* This procedure tags the retransmission queue when SACKs arrive. 957 * 958 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). 959 * Packets in queue with these bits set are counted in variables 960 * sacked_out, retrans_out and lost_out, correspondingly. 961 * 962 * Valid combinations are: 963 * Tag InFlight Description 964 * 0 1 - orig segment is in flight. 965 * S 0 - nothing flies, orig reached receiver. 966 * L 0 - nothing flies, orig lost by net. 967 * R 2 - both orig and retransmit are in flight. 968 * L|R 1 - orig is lost, retransmit is in flight. 969 * S|R 1 - orig reached receiver, retrans is still in flight. 970 * (L|S|R is logically valid, it could occur when L|R is sacked, 971 * but it is equivalent to plain S and code short-curcuits it to S. 972 * L|S is logically invalid, it would mean -1 packet in flight 8)) 973 * 974 * These 6 states form finite state machine, controlled by the following events: 975 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) 976 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) 977 * 3. Loss detection event of two flavors: 978 * A. Scoreboard estimator decided the packet is lost. 979 * A'. Reno "three dupacks" marks head of queue lost. 980 * B. SACK arrives sacking SND.NXT at the moment, when the 981 * segment was retransmitted. 982 * 4. D-SACK added new rule: D-SACK changes any tag to S. 983 * 984 * It is pleasant to note, that state diagram turns out to be commutative, 985 * so that we are allowed not to be bothered by order of our actions, 986 * when multiple events arrive simultaneously. (see the function below). 987 * 988 * Reordering detection. 989 * -------------------- 990 * Reordering metric is maximal distance, which a packet can be displaced 991 * in packet stream. With SACKs we can estimate it: 992 * 993 * 1. SACK fills old hole and the corresponding segment was not 994 * ever retransmitted -> reordering. Alas, we cannot use it 995 * when segment was retransmitted. 996 * 2. The last flaw is solved with D-SACK. D-SACK arrives 997 * for retransmitted and already SACKed segment -> reordering.. 998 * Both of these heuristics are not used in Loss state, when we cannot 999 * account for retransmits accurately. 1000 * 1001 * SACK block validation. 1002 * ---------------------- 1003 * 1004 * SACK block range validation checks that the received SACK block fits to 1005 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT. 1006 * Note that SND.UNA is not included to the range though being valid because 1007 * it means that the receiver is rather inconsistent with itself reporting 1008 * SACK reneging when it should advance SND.UNA. Such SACK block this is 1009 * perfectly valid, however, in light of RFC2018 which explicitly states 1010 * that "SACK block MUST reflect the newest segment. Even if the newest 1011 * segment is going to be discarded ...", not that it looks very clever 1012 * in case of head skb. Due to potentional receiver driven attacks, we 1013 * choose to avoid immediate execution of a walk in write queue due to 1014 * reneging and defer head skb's loss recovery to standard loss recovery 1015 * procedure that will eventually trigger (nothing forbids us doing this). 1016 * 1017 * Implements also blockage to start_seq wrap-around. Problem lies in the 1018 * fact that though start_seq (s) is before end_seq (i.e., not reversed), 1019 * there's no guarantee that it will be before snd_nxt (n). The problem 1020 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt 1021 * wrap (s_w): 1022 * 1023 * <- outs wnd -> <- wrapzone -> 1024 * u e n u_w e_w s n_w 1025 * | | | | | | | 1026 * |<------------+------+----- TCP seqno space --------------+---------->| 1027 * ...-- <2^31 ->| |<--------... 1028 * ...---- >2^31 ------>| |<--------... 1029 * 1030 * Current code wouldn't be vulnerable but it's better still to discard such 1031 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat 1032 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in 1033 * snd_nxt wrap -> snd_una region will then become "well defined", i.e., 1034 * equal to the ideal case (infinite seqno space without wrap caused issues). 1035 * 1036 * With D-SACK the lower bound is extended to cover sequence space below 1037 * SND.UNA down to undo_marker, which is the last point of interest. Yet 1038 * again, D-SACK block must not to go across snd_una (for the same reason as 1039 * for the normal SACK blocks, explained above). But there all simplicity 1040 * ends, TCP might receive valid D-SACKs below that. As long as they reside 1041 * fully below undo_marker they do not affect behavior in anyway and can 1042 * therefore be safely ignored. In rare cases (which are more or less 1043 * theoretical ones), the D-SACK will nicely cross that boundary due to skb 1044 * fragmentation and packet reordering past skb's retransmission. To consider 1045 * them correctly, the acceptable range must be extended even more though 1046 * the exact amount is rather hard to quantify. However, tp->max_window can 1047 * be used as an exaggerated estimate. 1048 */ 1049 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack, 1050 u32 start_seq, u32 end_seq) 1051 { 1052 /* Too far in future, or reversed (interpretation is ambiguous) */ 1053 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq)) 1054 return false; 1055 1056 /* Nasty start_seq wrap-around check (see comments above) */ 1057 if (!before(start_seq, tp->snd_nxt)) 1058 return false; 1059 1060 /* In outstanding window? ...This is valid exit for D-SACKs too. 1061 * start_seq == snd_una is non-sensical (see comments above) 1062 */ 1063 if (after(start_seq, tp->snd_una)) 1064 return true; 1065 1066 if (!is_dsack || !tp->undo_marker) 1067 return false; 1068 1069 /* ...Then it's D-SACK, and must reside below snd_una completely */ 1070 if (after(end_seq, tp->snd_una)) 1071 return false; 1072 1073 if (!before(start_seq, tp->undo_marker)) 1074 return true; 1075 1076 /* Too old */ 1077 if (!after(end_seq, tp->undo_marker)) 1078 return false; 1079 1080 /* Undo_marker boundary crossing (overestimates a lot). Known already: 1081 * start_seq < undo_marker and end_seq >= undo_marker. 1082 */ 1083 return !before(start_seq, end_seq - tp->max_window); 1084 } 1085 1086 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb, 1087 struct tcp_sack_block_wire *sp, int num_sacks, 1088 u32 prior_snd_una) 1089 { 1090 struct tcp_sock *tp = tcp_sk(sk); 1091 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq); 1092 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq); 1093 bool dup_sack = false; 1094 1095 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) { 1096 dup_sack = true; 1097 tcp_dsack_seen(tp); 1098 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV); 1099 } else if (num_sacks > 1) { 1100 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq); 1101 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq); 1102 1103 if (!after(end_seq_0, end_seq_1) && 1104 !before(start_seq_0, start_seq_1)) { 1105 dup_sack = true; 1106 tcp_dsack_seen(tp); 1107 NET_INC_STATS(sock_net(sk), 1108 LINUX_MIB_TCPDSACKOFORECV); 1109 } 1110 } 1111 1112 /* D-SACK for already forgotten data... Do dumb counting. */ 1113 if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 && 1114 !after(end_seq_0, prior_snd_una) && 1115 after(end_seq_0, tp->undo_marker)) 1116 tp->undo_retrans--; 1117 1118 return dup_sack; 1119 } 1120 1121 struct tcp_sacktag_state { 1122 u32 reord; 1123 /* Timestamps for earliest and latest never-retransmitted segment 1124 * that was SACKed. RTO needs the earliest RTT to stay conservative, 1125 * but congestion control should still get an accurate delay signal. 1126 */ 1127 u64 first_sackt; 1128 u64 last_sackt; 1129 struct rate_sample *rate; 1130 int flag; 1131 unsigned int mss_now; 1132 }; 1133 1134 /* Check if skb is fully within the SACK block. In presence of GSO skbs, 1135 * the incoming SACK may not exactly match but we can find smaller MSS 1136 * aligned portion of it that matches. Therefore we might need to fragment 1137 * which may fail and creates some hassle (caller must handle error case 1138 * returns). 1139 * 1140 * FIXME: this could be merged to shift decision code 1141 */ 1142 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb, 1143 u32 start_seq, u32 end_seq) 1144 { 1145 int err; 1146 bool in_sack; 1147 unsigned int pkt_len; 1148 unsigned int mss; 1149 1150 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1151 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1152 1153 if (tcp_skb_pcount(skb) > 1 && !in_sack && 1154 after(TCP_SKB_CB(skb)->end_seq, start_seq)) { 1155 mss = tcp_skb_mss(skb); 1156 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); 1157 1158 if (!in_sack) { 1159 pkt_len = start_seq - TCP_SKB_CB(skb)->seq; 1160 if (pkt_len < mss) 1161 pkt_len = mss; 1162 } else { 1163 pkt_len = end_seq - TCP_SKB_CB(skb)->seq; 1164 if (pkt_len < mss) 1165 return -EINVAL; 1166 } 1167 1168 /* Round if necessary so that SACKs cover only full MSSes 1169 * and/or the remaining small portion (if present) 1170 */ 1171 if (pkt_len > mss) { 1172 unsigned int new_len = (pkt_len / mss) * mss; 1173 if (!in_sack && new_len < pkt_len) 1174 new_len += mss; 1175 pkt_len = new_len; 1176 } 1177 1178 if (pkt_len >= skb->len && !in_sack) 1179 return 0; 1180 1181 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, 1182 pkt_len, mss, GFP_ATOMIC); 1183 if (err < 0) 1184 return err; 1185 } 1186 1187 return in_sack; 1188 } 1189 1190 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */ 1191 static u8 tcp_sacktag_one(struct sock *sk, 1192 struct tcp_sacktag_state *state, u8 sacked, 1193 u32 start_seq, u32 end_seq, 1194 int dup_sack, int pcount, 1195 u64 xmit_time) 1196 { 1197 struct tcp_sock *tp = tcp_sk(sk); 1198 1199 /* Account D-SACK for retransmitted packet. */ 1200 if (dup_sack && (sacked & TCPCB_RETRANS)) { 1201 if (tp->undo_marker && tp->undo_retrans > 0 && 1202 after(end_seq, tp->undo_marker)) 1203 tp->undo_retrans--; 1204 if ((sacked & TCPCB_SACKED_ACKED) && 1205 before(start_seq, state->reord)) 1206 state->reord = start_seq; 1207 } 1208 1209 /* Nothing to do; acked frame is about to be dropped (was ACKed). */ 1210 if (!after(end_seq, tp->snd_una)) 1211 return sacked; 1212 1213 if (!(sacked & TCPCB_SACKED_ACKED)) { 1214 tcp_rack_advance(tp, sacked, end_seq, xmit_time); 1215 1216 if (sacked & TCPCB_SACKED_RETRANS) { 1217 /* If the segment is not tagged as lost, 1218 * we do not clear RETRANS, believing 1219 * that retransmission is still in flight. 1220 */ 1221 if (sacked & TCPCB_LOST) { 1222 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); 1223 tp->lost_out -= pcount; 1224 tp->retrans_out -= pcount; 1225 } 1226 } else { 1227 if (!(sacked & TCPCB_RETRANS)) { 1228 /* New sack for not retransmitted frame, 1229 * which was in hole. It is reordering. 1230 */ 1231 if (before(start_seq, 1232 tcp_highest_sack_seq(tp)) && 1233 before(start_seq, state->reord)) 1234 state->reord = start_seq; 1235 1236 if (!after(end_seq, tp->high_seq)) 1237 state->flag |= FLAG_ORIG_SACK_ACKED; 1238 if (state->first_sackt == 0) 1239 state->first_sackt = xmit_time; 1240 state->last_sackt = xmit_time; 1241 } 1242 1243 if (sacked & TCPCB_LOST) { 1244 sacked &= ~TCPCB_LOST; 1245 tp->lost_out -= pcount; 1246 } 1247 } 1248 1249 sacked |= TCPCB_SACKED_ACKED; 1250 state->flag |= FLAG_DATA_SACKED; 1251 tp->sacked_out += pcount; 1252 tp->delivered += pcount; /* Out-of-order packets delivered */ 1253 1254 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */ 1255 if (tp->lost_skb_hint && 1256 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq)) 1257 tp->lost_cnt_hint += pcount; 1258 } 1259 1260 /* D-SACK. We can detect redundant retransmission in S|R and plain R 1261 * frames and clear it. undo_retrans is decreased above, L|R frames 1262 * are accounted above as well. 1263 */ 1264 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) { 1265 sacked &= ~TCPCB_SACKED_RETRANS; 1266 tp->retrans_out -= pcount; 1267 } 1268 1269 return sacked; 1270 } 1271 1272 /* Shift newly-SACKed bytes from this skb to the immediately previous 1273 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such. 1274 */ 1275 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev, 1276 struct sk_buff *skb, 1277 struct tcp_sacktag_state *state, 1278 unsigned int pcount, int shifted, int mss, 1279 bool dup_sack) 1280 { 1281 struct tcp_sock *tp = tcp_sk(sk); 1282 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */ 1283 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */ 1284 1285 BUG_ON(!pcount); 1286 1287 /* Adjust counters and hints for the newly sacked sequence 1288 * range but discard the return value since prev is already 1289 * marked. We must tag the range first because the seq 1290 * advancement below implicitly advances 1291 * tcp_highest_sack_seq() when skb is highest_sack. 1292 */ 1293 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked, 1294 start_seq, end_seq, dup_sack, pcount, 1295 tcp_skb_timestamp_us(skb)); 1296 tcp_rate_skb_delivered(sk, skb, state->rate); 1297 1298 if (skb == tp->lost_skb_hint) 1299 tp->lost_cnt_hint += pcount; 1300 1301 TCP_SKB_CB(prev)->end_seq += shifted; 1302 TCP_SKB_CB(skb)->seq += shifted; 1303 1304 tcp_skb_pcount_add(prev, pcount); 1305 WARN_ON_ONCE(tcp_skb_pcount(skb) < pcount); 1306 tcp_skb_pcount_add(skb, -pcount); 1307 1308 /* When we're adding to gso_segs == 1, gso_size will be zero, 1309 * in theory this shouldn't be necessary but as long as DSACK 1310 * code can come after this skb later on it's better to keep 1311 * setting gso_size to something. 1312 */ 1313 if (!TCP_SKB_CB(prev)->tcp_gso_size) 1314 TCP_SKB_CB(prev)->tcp_gso_size = mss; 1315 1316 /* CHECKME: To clear or not to clear? Mimics normal skb currently */ 1317 if (tcp_skb_pcount(skb) <= 1) 1318 TCP_SKB_CB(skb)->tcp_gso_size = 0; 1319 1320 /* Difference in this won't matter, both ACKed by the same cumul. ACK */ 1321 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS); 1322 1323 if (skb->len > 0) { 1324 BUG_ON(!tcp_skb_pcount(skb)); 1325 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED); 1326 return false; 1327 } 1328 1329 /* Whole SKB was eaten :-) */ 1330 1331 if (skb == tp->retransmit_skb_hint) 1332 tp->retransmit_skb_hint = prev; 1333 if (skb == tp->lost_skb_hint) { 1334 tp->lost_skb_hint = prev; 1335 tp->lost_cnt_hint -= tcp_skb_pcount(prev); 1336 } 1337 1338 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 1339 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor; 1340 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1341 TCP_SKB_CB(prev)->end_seq++; 1342 1343 if (skb == tcp_highest_sack(sk)) 1344 tcp_advance_highest_sack(sk, skb); 1345 1346 tcp_skb_collapse_tstamp(prev, skb); 1347 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp)) 1348 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0; 1349 1350 tcp_rtx_queue_unlink_and_free(skb, sk); 1351 1352 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED); 1353 1354 return true; 1355 } 1356 1357 /* I wish gso_size would have a bit more sane initialization than 1358 * something-or-zero which complicates things 1359 */ 1360 static int tcp_skb_seglen(const struct sk_buff *skb) 1361 { 1362 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb); 1363 } 1364 1365 /* Shifting pages past head area doesn't work */ 1366 static int skb_can_shift(const struct sk_buff *skb) 1367 { 1368 return !skb_headlen(skb) && skb_is_nonlinear(skb); 1369 } 1370 1371 int tcp_skb_shift(struct sk_buff *to, struct sk_buff *from, 1372 int pcount, int shiftlen) 1373 { 1374 /* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE) 1375 * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need 1376 * to make sure not storing more than 65535 * 8 bytes per skb, 1377 * even if current MSS is bigger. 1378 */ 1379 if (unlikely(to->len + shiftlen >= 65535 * TCP_MIN_GSO_SIZE)) 1380 return 0; 1381 if (unlikely(tcp_skb_pcount(to) + pcount > 65535)) 1382 return 0; 1383 return skb_shift(to, from, shiftlen); 1384 } 1385 1386 /* Try collapsing SACK blocks spanning across multiple skbs to a single 1387 * skb. 1388 */ 1389 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb, 1390 struct tcp_sacktag_state *state, 1391 u32 start_seq, u32 end_seq, 1392 bool dup_sack) 1393 { 1394 struct tcp_sock *tp = tcp_sk(sk); 1395 struct sk_buff *prev; 1396 int mss; 1397 int pcount = 0; 1398 int len; 1399 int in_sack; 1400 1401 /* Normally R but no L won't result in plain S */ 1402 if (!dup_sack && 1403 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS) 1404 goto fallback; 1405 if (!skb_can_shift(skb)) 1406 goto fallback; 1407 /* This frame is about to be dropped (was ACKed). */ 1408 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1409 goto fallback; 1410 1411 /* Can only happen with delayed DSACK + discard craziness */ 1412 prev = skb_rb_prev(skb); 1413 if (!prev) 1414 goto fallback; 1415 1416 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) 1417 goto fallback; 1418 1419 if (!tcp_skb_can_collapse_to(prev)) 1420 goto fallback; 1421 1422 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1423 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1424 1425 if (in_sack) { 1426 len = skb->len; 1427 pcount = tcp_skb_pcount(skb); 1428 mss = tcp_skb_seglen(skb); 1429 1430 /* TODO: Fix DSACKs to not fragment already SACKed and we can 1431 * drop this restriction as unnecessary 1432 */ 1433 if (mss != tcp_skb_seglen(prev)) 1434 goto fallback; 1435 } else { 1436 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq)) 1437 goto noop; 1438 /* CHECKME: This is non-MSS split case only?, this will 1439 * cause skipped skbs due to advancing loop btw, original 1440 * has that feature too 1441 */ 1442 if (tcp_skb_pcount(skb) <= 1) 1443 goto noop; 1444 1445 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); 1446 if (!in_sack) { 1447 /* TODO: head merge to next could be attempted here 1448 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)), 1449 * though it might not be worth of the additional hassle 1450 * 1451 * ...we can probably just fallback to what was done 1452 * previously. We could try merging non-SACKed ones 1453 * as well but it probably isn't going to buy off 1454 * because later SACKs might again split them, and 1455 * it would make skb timestamp tracking considerably 1456 * harder problem. 1457 */ 1458 goto fallback; 1459 } 1460 1461 len = end_seq - TCP_SKB_CB(skb)->seq; 1462 BUG_ON(len < 0); 1463 BUG_ON(len > skb->len); 1464 1465 /* MSS boundaries should be honoured or else pcount will 1466 * severely break even though it makes things bit trickier. 1467 * Optimize common case to avoid most of the divides 1468 */ 1469 mss = tcp_skb_mss(skb); 1470 1471 /* TODO: Fix DSACKs to not fragment already SACKed and we can 1472 * drop this restriction as unnecessary 1473 */ 1474 if (mss != tcp_skb_seglen(prev)) 1475 goto fallback; 1476 1477 if (len == mss) { 1478 pcount = 1; 1479 } else if (len < mss) { 1480 goto noop; 1481 } else { 1482 pcount = len / mss; 1483 len = pcount * mss; 1484 } 1485 } 1486 1487 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */ 1488 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una)) 1489 goto fallback; 1490 1491 if (!tcp_skb_shift(prev, skb, pcount, len)) 1492 goto fallback; 1493 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack)) 1494 goto out; 1495 1496 /* Hole filled allows collapsing with the next as well, this is very 1497 * useful when hole on every nth skb pattern happens 1498 */ 1499 skb = skb_rb_next(prev); 1500 if (!skb) 1501 goto out; 1502 1503 if (!skb_can_shift(skb) || 1504 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) || 1505 (mss != tcp_skb_seglen(skb))) 1506 goto out; 1507 1508 len = skb->len; 1509 pcount = tcp_skb_pcount(skb); 1510 if (tcp_skb_shift(prev, skb, pcount, len)) 1511 tcp_shifted_skb(sk, prev, skb, state, pcount, 1512 len, mss, 0); 1513 1514 out: 1515 return prev; 1516 1517 noop: 1518 return skb; 1519 1520 fallback: 1521 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK); 1522 return NULL; 1523 } 1524 1525 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk, 1526 struct tcp_sack_block *next_dup, 1527 struct tcp_sacktag_state *state, 1528 u32 start_seq, u32 end_seq, 1529 bool dup_sack_in) 1530 { 1531 struct tcp_sock *tp = tcp_sk(sk); 1532 struct sk_buff *tmp; 1533 1534 skb_rbtree_walk_from(skb) { 1535 int in_sack = 0; 1536 bool dup_sack = dup_sack_in; 1537 1538 /* queue is in-order => we can short-circuit the walk early */ 1539 if (!before(TCP_SKB_CB(skb)->seq, end_seq)) 1540 break; 1541 1542 if (next_dup && 1543 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) { 1544 in_sack = tcp_match_skb_to_sack(sk, skb, 1545 next_dup->start_seq, 1546 next_dup->end_seq); 1547 if (in_sack > 0) 1548 dup_sack = true; 1549 } 1550 1551 /* skb reference here is a bit tricky to get right, since 1552 * shifting can eat and free both this skb and the next, 1553 * so not even _safe variant of the loop is enough. 1554 */ 1555 if (in_sack <= 0) { 1556 tmp = tcp_shift_skb_data(sk, skb, state, 1557 start_seq, end_seq, dup_sack); 1558 if (tmp) { 1559 if (tmp != skb) { 1560 skb = tmp; 1561 continue; 1562 } 1563 1564 in_sack = 0; 1565 } else { 1566 in_sack = tcp_match_skb_to_sack(sk, skb, 1567 start_seq, 1568 end_seq); 1569 } 1570 } 1571 1572 if (unlikely(in_sack < 0)) 1573 break; 1574 1575 if (in_sack) { 1576 TCP_SKB_CB(skb)->sacked = 1577 tcp_sacktag_one(sk, 1578 state, 1579 TCP_SKB_CB(skb)->sacked, 1580 TCP_SKB_CB(skb)->seq, 1581 TCP_SKB_CB(skb)->end_seq, 1582 dup_sack, 1583 tcp_skb_pcount(skb), 1584 tcp_skb_timestamp_us(skb)); 1585 tcp_rate_skb_delivered(sk, skb, state->rate); 1586 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1587 list_del_init(&skb->tcp_tsorted_anchor); 1588 1589 if (!before(TCP_SKB_CB(skb)->seq, 1590 tcp_highest_sack_seq(tp))) 1591 tcp_advance_highest_sack(sk, skb); 1592 } 1593 } 1594 return skb; 1595 } 1596 1597 static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk, u32 seq) 1598 { 1599 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node; 1600 struct sk_buff *skb; 1601 1602 while (*p) { 1603 parent = *p; 1604 skb = rb_to_skb(parent); 1605 if (before(seq, TCP_SKB_CB(skb)->seq)) { 1606 p = &parent->rb_left; 1607 continue; 1608 } 1609 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) { 1610 p = &parent->rb_right; 1611 continue; 1612 } 1613 return skb; 1614 } 1615 return NULL; 1616 } 1617 1618 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk, 1619 u32 skip_to_seq) 1620 { 1621 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq)) 1622 return skb; 1623 1624 return tcp_sacktag_bsearch(sk, skip_to_seq); 1625 } 1626 1627 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb, 1628 struct sock *sk, 1629 struct tcp_sack_block *next_dup, 1630 struct tcp_sacktag_state *state, 1631 u32 skip_to_seq) 1632 { 1633 if (!next_dup) 1634 return skb; 1635 1636 if (before(next_dup->start_seq, skip_to_seq)) { 1637 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq); 1638 skb = tcp_sacktag_walk(skb, sk, NULL, state, 1639 next_dup->start_seq, next_dup->end_seq, 1640 1); 1641 } 1642 1643 return skb; 1644 } 1645 1646 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache) 1647 { 1648 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); 1649 } 1650 1651 static int 1652 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb, 1653 u32 prior_snd_una, struct tcp_sacktag_state *state) 1654 { 1655 struct tcp_sock *tp = tcp_sk(sk); 1656 const unsigned char *ptr = (skb_transport_header(ack_skb) + 1657 TCP_SKB_CB(ack_skb)->sacked); 1658 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2); 1659 struct tcp_sack_block sp[TCP_NUM_SACKS]; 1660 struct tcp_sack_block *cache; 1661 struct sk_buff *skb; 1662 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3); 1663 int used_sacks; 1664 bool found_dup_sack = false; 1665 int i, j; 1666 int first_sack_index; 1667 1668 state->flag = 0; 1669 state->reord = tp->snd_nxt; 1670 1671 if (!tp->sacked_out) 1672 tcp_highest_sack_reset(sk); 1673 1674 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire, 1675 num_sacks, prior_snd_una); 1676 if (found_dup_sack) { 1677 state->flag |= FLAG_DSACKING_ACK; 1678 tp->delivered++; /* A spurious retransmission is delivered */ 1679 } 1680 1681 /* Eliminate too old ACKs, but take into 1682 * account more or less fresh ones, they can 1683 * contain valid SACK info. 1684 */ 1685 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) 1686 return 0; 1687 1688 if (!tp->packets_out) 1689 goto out; 1690 1691 used_sacks = 0; 1692 first_sack_index = 0; 1693 for (i = 0; i < num_sacks; i++) { 1694 bool dup_sack = !i && found_dup_sack; 1695 1696 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq); 1697 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq); 1698 1699 if (!tcp_is_sackblock_valid(tp, dup_sack, 1700 sp[used_sacks].start_seq, 1701 sp[used_sacks].end_seq)) { 1702 int mib_idx; 1703 1704 if (dup_sack) { 1705 if (!tp->undo_marker) 1706 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO; 1707 else 1708 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD; 1709 } else { 1710 /* Don't count olds caused by ACK reordering */ 1711 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) && 1712 !after(sp[used_sacks].end_seq, tp->snd_una)) 1713 continue; 1714 mib_idx = LINUX_MIB_TCPSACKDISCARD; 1715 } 1716 1717 NET_INC_STATS(sock_net(sk), mib_idx); 1718 if (i == 0) 1719 first_sack_index = -1; 1720 continue; 1721 } 1722 1723 /* Ignore very old stuff early */ 1724 if (!after(sp[used_sacks].end_seq, prior_snd_una)) 1725 continue; 1726 1727 used_sacks++; 1728 } 1729 1730 /* order SACK blocks to allow in order walk of the retrans queue */ 1731 for (i = used_sacks - 1; i > 0; i--) { 1732 for (j = 0; j < i; j++) { 1733 if (after(sp[j].start_seq, sp[j + 1].start_seq)) { 1734 swap(sp[j], sp[j + 1]); 1735 1736 /* Track where the first SACK block goes to */ 1737 if (j == first_sack_index) 1738 first_sack_index = j + 1; 1739 } 1740 } 1741 } 1742 1743 state->mss_now = tcp_current_mss(sk); 1744 skb = NULL; 1745 i = 0; 1746 1747 if (!tp->sacked_out) { 1748 /* It's already past, so skip checking against it */ 1749 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); 1750 } else { 1751 cache = tp->recv_sack_cache; 1752 /* Skip empty blocks in at head of the cache */ 1753 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq && 1754 !cache->end_seq) 1755 cache++; 1756 } 1757 1758 while (i < used_sacks) { 1759 u32 start_seq = sp[i].start_seq; 1760 u32 end_seq = sp[i].end_seq; 1761 bool dup_sack = (found_dup_sack && (i == first_sack_index)); 1762 struct tcp_sack_block *next_dup = NULL; 1763 1764 if (found_dup_sack && ((i + 1) == first_sack_index)) 1765 next_dup = &sp[i + 1]; 1766 1767 /* Skip too early cached blocks */ 1768 while (tcp_sack_cache_ok(tp, cache) && 1769 !before(start_seq, cache->end_seq)) 1770 cache++; 1771 1772 /* Can skip some work by looking recv_sack_cache? */ 1773 if (tcp_sack_cache_ok(tp, cache) && !dup_sack && 1774 after(end_seq, cache->start_seq)) { 1775 1776 /* Head todo? */ 1777 if (before(start_seq, cache->start_seq)) { 1778 skb = tcp_sacktag_skip(skb, sk, start_seq); 1779 skb = tcp_sacktag_walk(skb, sk, next_dup, 1780 state, 1781 start_seq, 1782 cache->start_seq, 1783 dup_sack); 1784 } 1785 1786 /* Rest of the block already fully processed? */ 1787 if (!after(end_seq, cache->end_seq)) 1788 goto advance_sp; 1789 1790 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup, 1791 state, 1792 cache->end_seq); 1793 1794 /* ...tail remains todo... */ 1795 if (tcp_highest_sack_seq(tp) == cache->end_seq) { 1796 /* ...but better entrypoint exists! */ 1797 skb = tcp_highest_sack(sk); 1798 if (!skb) 1799 break; 1800 cache++; 1801 goto walk; 1802 } 1803 1804 skb = tcp_sacktag_skip(skb, sk, cache->end_seq); 1805 /* Check overlap against next cached too (past this one already) */ 1806 cache++; 1807 continue; 1808 } 1809 1810 if (!before(start_seq, tcp_highest_sack_seq(tp))) { 1811 skb = tcp_highest_sack(sk); 1812 if (!skb) 1813 break; 1814 } 1815 skb = tcp_sacktag_skip(skb, sk, start_seq); 1816 1817 walk: 1818 skb = tcp_sacktag_walk(skb, sk, next_dup, state, 1819 start_seq, end_seq, dup_sack); 1820 1821 advance_sp: 1822 i++; 1823 } 1824 1825 /* Clear the head of the cache sack blocks so we can skip it next time */ 1826 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) { 1827 tp->recv_sack_cache[i].start_seq = 0; 1828 tp->recv_sack_cache[i].end_seq = 0; 1829 } 1830 for (j = 0; j < used_sacks; j++) 1831 tp->recv_sack_cache[i++] = sp[j]; 1832 1833 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker) 1834 tcp_check_sack_reordering(sk, state->reord, 0); 1835 1836 tcp_verify_left_out(tp); 1837 out: 1838 1839 #if FASTRETRANS_DEBUG > 0 1840 WARN_ON((int)tp->sacked_out < 0); 1841 WARN_ON((int)tp->lost_out < 0); 1842 WARN_ON((int)tp->retrans_out < 0); 1843 WARN_ON((int)tcp_packets_in_flight(tp) < 0); 1844 #endif 1845 return state->flag; 1846 } 1847 1848 /* Limits sacked_out so that sum with lost_out isn't ever larger than 1849 * packets_out. Returns false if sacked_out adjustement wasn't necessary. 1850 */ 1851 static bool tcp_limit_reno_sacked(struct tcp_sock *tp) 1852 { 1853 u32 holes; 1854 1855 holes = max(tp->lost_out, 1U); 1856 holes = min(holes, tp->packets_out); 1857 1858 if ((tp->sacked_out + holes) > tp->packets_out) { 1859 tp->sacked_out = tp->packets_out - holes; 1860 return true; 1861 } 1862 return false; 1863 } 1864 1865 /* If we receive more dupacks than we expected counting segments 1866 * in assumption of absent reordering, interpret this as reordering. 1867 * The only another reason could be bug in receiver TCP. 1868 */ 1869 static void tcp_check_reno_reordering(struct sock *sk, const int addend) 1870 { 1871 struct tcp_sock *tp = tcp_sk(sk); 1872 1873 if (!tcp_limit_reno_sacked(tp)) 1874 return; 1875 1876 tp->reordering = min_t(u32, tp->packets_out + addend, 1877 sock_net(sk)->ipv4.sysctl_tcp_max_reordering); 1878 tp->reord_seen++; 1879 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER); 1880 } 1881 1882 /* Emulate SACKs for SACKless connection: account for a new dupack. */ 1883 1884 static void tcp_add_reno_sack(struct sock *sk, int num_dupack) 1885 { 1886 if (num_dupack) { 1887 struct tcp_sock *tp = tcp_sk(sk); 1888 u32 prior_sacked = tp->sacked_out; 1889 s32 delivered; 1890 1891 tp->sacked_out += num_dupack; 1892 tcp_check_reno_reordering(sk, 0); 1893 delivered = tp->sacked_out - prior_sacked; 1894 if (delivered > 0) 1895 tp->delivered += delivered; 1896 tcp_verify_left_out(tp); 1897 } 1898 } 1899 1900 /* Account for ACK, ACKing some data in Reno Recovery phase. */ 1901 1902 static void tcp_remove_reno_sacks(struct sock *sk, int acked) 1903 { 1904 struct tcp_sock *tp = tcp_sk(sk); 1905 1906 if (acked > 0) { 1907 /* One ACK acked hole. The rest eat duplicate ACKs. */ 1908 tp->delivered += max_t(int, acked - tp->sacked_out, 1); 1909 if (acked - 1 >= tp->sacked_out) 1910 tp->sacked_out = 0; 1911 else 1912 tp->sacked_out -= acked - 1; 1913 } 1914 tcp_check_reno_reordering(sk, acked); 1915 tcp_verify_left_out(tp); 1916 } 1917 1918 static inline void tcp_reset_reno_sack(struct tcp_sock *tp) 1919 { 1920 tp->sacked_out = 0; 1921 } 1922 1923 void tcp_clear_retrans(struct tcp_sock *tp) 1924 { 1925 tp->retrans_out = 0; 1926 tp->lost_out = 0; 1927 tp->undo_marker = 0; 1928 tp->undo_retrans = -1; 1929 tp->sacked_out = 0; 1930 } 1931 1932 static inline void tcp_init_undo(struct tcp_sock *tp) 1933 { 1934 tp->undo_marker = tp->snd_una; 1935 /* Retransmission still in flight may cause DSACKs later. */ 1936 tp->undo_retrans = tp->retrans_out ? : -1; 1937 } 1938 1939 static bool tcp_is_rack(const struct sock *sk) 1940 { 1941 return sock_net(sk)->ipv4.sysctl_tcp_recovery & TCP_RACK_LOSS_DETECTION; 1942 } 1943 1944 /* If we detect SACK reneging, forget all SACK information 1945 * and reset tags completely, otherwise preserve SACKs. If receiver 1946 * dropped its ofo queue, we will know this due to reneging detection. 1947 */ 1948 static void tcp_timeout_mark_lost(struct sock *sk) 1949 { 1950 struct tcp_sock *tp = tcp_sk(sk); 1951 struct sk_buff *skb, *head; 1952 bool is_reneg; /* is receiver reneging on SACKs? */ 1953 1954 head = tcp_rtx_queue_head(sk); 1955 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED); 1956 if (is_reneg) { 1957 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING); 1958 tp->sacked_out = 0; 1959 /* Mark SACK reneging until we recover from this loss event. */ 1960 tp->is_sack_reneg = 1; 1961 } else if (tcp_is_reno(tp)) { 1962 tcp_reset_reno_sack(tp); 1963 } 1964 1965 skb = head; 1966 skb_rbtree_walk_from(skb) { 1967 if (is_reneg) 1968 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED; 1969 else if (tcp_is_rack(sk) && skb != head && 1970 tcp_rack_skb_timeout(tp, skb, 0) > 0) 1971 continue; /* Don't mark recently sent ones lost yet */ 1972 tcp_mark_skb_lost(sk, skb); 1973 } 1974 tcp_verify_left_out(tp); 1975 tcp_clear_all_retrans_hints(tp); 1976 } 1977 1978 /* Enter Loss state. */ 1979 void tcp_enter_loss(struct sock *sk) 1980 { 1981 const struct inet_connection_sock *icsk = inet_csk(sk); 1982 struct tcp_sock *tp = tcp_sk(sk); 1983 struct net *net = sock_net(sk); 1984 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery; 1985 1986 tcp_timeout_mark_lost(sk); 1987 1988 /* Reduce ssthresh if it has not yet been made inside this window. */ 1989 if (icsk->icsk_ca_state <= TCP_CA_Disorder || 1990 !after(tp->high_seq, tp->snd_una) || 1991 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) { 1992 tp->prior_ssthresh = tcp_current_ssthresh(sk); 1993 tp->prior_cwnd = tp->snd_cwnd; 1994 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 1995 tcp_ca_event(sk, CA_EVENT_LOSS); 1996 tcp_init_undo(tp); 1997 } 1998 tp->snd_cwnd = tcp_packets_in_flight(tp) + 1; 1999 tp->snd_cwnd_cnt = 0; 2000 tp->snd_cwnd_stamp = tcp_jiffies32; 2001 2002 /* Timeout in disordered state after receiving substantial DUPACKs 2003 * suggests that the degree of reordering is over-estimated. 2004 */ 2005 if (icsk->icsk_ca_state <= TCP_CA_Disorder && 2006 tp->sacked_out >= net->ipv4.sysctl_tcp_reordering) 2007 tp->reordering = min_t(unsigned int, tp->reordering, 2008 net->ipv4.sysctl_tcp_reordering); 2009 tcp_set_ca_state(sk, TCP_CA_Loss); 2010 tp->high_seq = tp->snd_nxt; 2011 tcp_ecn_queue_cwr(tp); 2012 2013 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous 2014 * loss recovery is underway except recurring timeout(s) on 2015 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing 2016 */ 2017 tp->frto = net->ipv4.sysctl_tcp_frto && 2018 (new_recovery || icsk->icsk_retransmits) && 2019 !inet_csk(sk)->icsk_mtup.probe_size; 2020 } 2021 2022 /* If ACK arrived pointing to a remembered SACK, it means that our 2023 * remembered SACKs do not reflect real state of receiver i.e. 2024 * receiver _host_ is heavily congested (or buggy). 2025 * 2026 * To avoid big spurious retransmission bursts due to transient SACK 2027 * scoreboard oddities that look like reneging, we give the receiver a 2028 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will 2029 * restore sanity to the SACK scoreboard. If the apparent reneging 2030 * persists until this RTO then we'll clear the SACK scoreboard. 2031 */ 2032 static bool tcp_check_sack_reneging(struct sock *sk, int flag) 2033 { 2034 if (flag & FLAG_SACK_RENEGING) { 2035 struct tcp_sock *tp = tcp_sk(sk); 2036 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4), 2037 msecs_to_jiffies(10)); 2038 2039 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 2040 delay, TCP_RTO_MAX); 2041 return true; 2042 } 2043 return false; 2044 } 2045 2046 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs 2047 * counter when SACK is enabled (without SACK, sacked_out is used for 2048 * that purpose). 2049 * 2050 * With reordering, holes may still be in flight, so RFC3517 recovery 2051 * uses pure sacked_out (total number of SACKed segments) even though 2052 * it violates the RFC that uses duplicate ACKs, often these are equal 2053 * but when e.g. out-of-window ACKs or packet duplication occurs, 2054 * they differ. Since neither occurs due to loss, TCP should really 2055 * ignore them. 2056 */ 2057 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp) 2058 { 2059 return tp->sacked_out + 1; 2060 } 2061 2062 /* Linux NewReno/SACK/ECN state machine. 2063 * -------------------------------------- 2064 * 2065 * "Open" Normal state, no dubious events, fast path. 2066 * "Disorder" In all the respects it is "Open", 2067 * but requires a bit more attention. It is entered when 2068 * we see some SACKs or dupacks. It is split of "Open" 2069 * mainly to move some processing from fast path to slow one. 2070 * "CWR" CWND was reduced due to some Congestion Notification event. 2071 * It can be ECN, ICMP source quench, local device congestion. 2072 * "Recovery" CWND was reduced, we are fast-retransmitting. 2073 * "Loss" CWND was reduced due to RTO timeout or SACK reneging. 2074 * 2075 * tcp_fastretrans_alert() is entered: 2076 * - each incoming ACK, if state is not "Open" 2077 * - when arrived ACK is unusual, namely: 2078 * * SACK 2079 * * Duplicate ACK. 2080 * * ECN ECE. 2081 * 2082 * Counting packets in flight is pretty simple. 2083 * 2084 * in_flight = packets_out - left_out + retrans_out 2085 * 2086 * packets_out is SND.NXT-SND.UNA counted in packets. 2087 * 2088 * retrans_out is number of retransmitted segments. 2089 * 2090 * left_out is number of segments left network, but not ACKed yet. 2091 * 2092 * left_out = sacked_out + lost_out 2093 * 2094 * sacked_out: Packets, which arrived to receiver out of order 2095 * and hence not ACKed. With SACKs this number is simply 2096 * amount of SACKed data. Even without SACKs 2097 * it is easy to give pretty reliable estimate of this number, 2098 * counting duplicate ACKs. 2099 * 2100 * lost_out: Packets lost by network. TCP has no explicit 2101 * "loss notification" feedback from network (for now). 2102 * It means that this number can be only _guessed_. 2103 * Actually, it is the heuristics to predict lossage that 2104 * distinguishes different algorithms. 2105 * 2106 * F.e. after RTO, when all the queue is considered as lost, 2107 * lost_out = packets_out and in_flight = retrans_out. 2108 * 2109 * Essentially, we have now a few algorithms detecting 2110 * lost packets. 2111 * 2112 * If the receiver supports SACK: 2113 * 2114 * RFC6675/3517: It is the conventional algorithm. A packet is 2115 * considered lost if the number of higher sequence packets 2116 * SACKed is greater than or equal the DUPACK thoreshold 2117 * (reordering). This is implemented in tcp_mark_head_lost and 2118 * tcp_update_scoreboard. 2119 * 2120 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm 2121 * (2017-) that checks timing instead of counting DUPACKs. 2122 * Essentially a packet is considered lost if it's not S/ACKed 2123 * after RTT + reordering_window, where both metrics are 2124 * dynamically measured and adjusted. This is implemented in 2125 * tcp_rack_mark_lost. 2126 * 2127 * If the receiver does not support SACK: 2128 * 2129 * NewReno (RFC6582): in Recovery we assume that one segment 2130 * is lost (classic Reno). While we are in Recovery and 2131 * a partial ACK arrives, we assume that one more packet 2132 * is lost (NewReno). This heuristics are the same in NewReno 2133 * and SACK. 2134 * 2135 * Really tricky (and requiring careful tuning) part of algorithm 2136 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue(). 2137 * The first determines the moment _when_ we should reduce CWND and, 2138 * hence, slow down forward transmission. In fact, it determines the moment 2139 * when we decide that hole is caused by loss, rather than by a reorder. 2140 * 2141 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill 2142 * holes, caused by lost packets. 2143 * 2144 * And the most logically complicated part of algorithm is undo 2145 * heuristics. We detect false retransmits due to both too early 2146 * fast retransmit (reordering) and underestimated RTO, analyzing 2147 * timestamps and D-SACKs. When we detect that some segments were 2148 * retransmitted by mistake and CWND reduction was wrong, we undo 2149 * window reduction and abort recovery phase. This logic is hidden 2150 * inside several functions named tcp_try_undo_<something>. 2151 */ 2152 2153 /* This function decides, when we should leave Disordered state 2154 * and enter Recovery phase, reducing congestion window. 2155 * 2156 * Main question: may we further continue forward transmission 2157 * with the same cwnd? 2158 */ 2159 static bool tcp_time_to_recover(struct sock *sk, int flag) 2160 { 2161 struct tcp_sock *tp = tcp_sk(sk); 2162 2163 /* Trick#1: The loss is proven. */ 2164 if (tp->lost_out) 2165 return true; 2166 2167 /* Not-A-Trick#2 : Classic rule... */ 2168 if (!tcp_is_rack(sk) && tcp_dupack_heuristics(tp) > tp->reordering) 2169 return true; 2170 2171 return false; 2172 } 2173 2174 /* Detect loss in event "A" above by marking head of queue up as lost. 2175 * For non-SACK(Reno) senders, the first "packets" number of segments 2176 * are considered lost. For RFC3517 SACK, a segment is considered lost if it 2177 * has at least tp->reordering SACKed seqments above it; "packets" refers to 2178 * the maximum SACKed segments to pass before reaching this limit. 2179 */ 2180 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head) 2181 { 2182 struct tcp_sock *tp = tcp_sk(sk); 2183 struct sk_buff *skb; 2184 int cnt, oldcnt, lost; 2185 unsigned int mss; 2186 /* Use SACK to deduce losses of new sequences sent during recovery */ 2187 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq; 2188 2189 WARN_ON(packets > tp->packets_out); 2190 skb = tp->lost_skb_hint; 2191 if (skb) { 2192 /* Head already handled? */ 2193 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una)) 2194 return; 2195 cnt = tp->lost_cnt_hint; 2196 } else { 2197 skb = tcp_rtx_queue_head(sk); 2198 cnt = 0; 2199 } 2200 2201 skb_rbtree_walk_from(skb) { 2202 /* TODO: do this better */ 2203 /* this is not the most efficient way to do this... */ 2204 tp->lost_skb_hint = skb; 2205 tp->lost_cnt_hint = cnt; 2206 2207 if (after(TCP_SKB_CB(skb)->end_seq, loss_high)) 2208 break; 2209 2210 oldcnt = cnt; 2211 if (tcp_is_reno(tp) || 2212 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 2213 cnt += tcp_skb_pcount(skb); 2214 2215 if (cnt > packets) { 2216 if (tcp_is_sack(tp) || 2217 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) || 2218 (oldcnt >= packets)) 2219 break; 2220 2221 mss = tcp_skb_mss(skb); 2222 /* If needed, chop off the prefix to mark as lost. */ 2223 lost = (packets - oldcnt) * mss; 2224 if (lost < skb->len && 2225 tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, 2226 lost, mss, GFP_ATOMIC) < 0) 2227 break; 2228 cnt = packets; 2229 } 2230 2231 tcp_skb_mark_lost(tp, skb); 2232 2233 if (mark_head) 2234 break; 2235 } 2236 tcp_verify_left_out(tp); 2237 } 2238 2239 /* Account newly detected lost packet(s) */ 2240 2241 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit) 2242 { 2243 struct tcp_sock *tp = tcp_sk(sk); 2244 2245 if (tcp_is_sack(tp)) { 2246 int sacked_upto = tp->sacked_out - tp->reordering; 2247 if (sacked_upto >= 0) 2248 tcp_mark_head_lost(sk, sacked_upto, 0); 2249 else if (fast_rexmit) 2250 tcp_mark_head_lost(sk, 1, 1); 2251 } 2252 } 2253 2254 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when) 2255 { 2256 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 2257 before(tp->rx_opt.rcv_tsecr, when); 2258 } 2259 2260 /* skb is spurious retransmitted if the returned timestamp echo 2261 * reply is prior to the skb transmission time 2262 */ 2263 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp, 2264 const struct sk_buff *skb) 2265 { 2266 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) && 2267 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb)); 2268 } 2269 2270 /* Nothing was retransmitted or returned timestamp is less 2271 * than timestamp of the first retransmission. 2272 */ 2273 static inline bool tcp_packet_delayed(const struct tcp_sock *tp) 2274 { 2275 return tp->retrans_stamp && 2276 tcp_tsopt_ecr_before(tp, tp->retrans_stamp); 2277 } 2278 2279 /* Undo procedures. */ 2280 2281 /* We can clear retrans_stamp when there are no retransmissions in the 2282 * window. It would seem that it is trivially available for us in 2283 * tp->retrans_out, however, that kind of assumptions doesn't consider 2284 * what will happen if errors occur when sending retransmission for the 2285 * second time. ...It could the that such segment has only 2286 * TCPCB_EVER_RETRANS set at the present time. It seems that checking 2287 * the head skb is enough except for some reneging corner cases that 2288 * are not worth the effort. 2289 * 2290 * Main reason for all this complexity is the fact that connection dying 2291 * time now depends on the validity of the retrans_stamp, in particular, 2292 * that successive retransmissions of a segment must not advance 2293 * retrans_stamp under any conditions. 2294 */ 2295 static bool tcp_any_retrans_done(const struct sock *sk) 2296 { 2297 const struct tcp_sock *tp = tcp_sk(sk); 2298 struct sk_buff *skb; 2299 2300 if (tp->retrans_out) 2301 return true; 2302 2303 skb = tcp_rtx_queue_head(sk); 2304 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS)) 2305 return true; 2306 2307 return false; 2308 } 2309 2310 static void DBGUNDO(struct sock *sk, const char *msg) 2311 { 2312 #if FASTRETRANS_DEBUG > 1 2313 struct tcp_sock *tp = tcp_sk(sk); 2314 struct inet_sock *inet = inet_sk(sk); 2315 2316 if (sk->sk_family == AF_INET) { 2317 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n", 2318 msg, 2319 &inet->inet_daddr, ntohs(inet->inet_dport), 2320 tp->snd_cwnd, tcp_left_out(tp), 2321 tp->snd_ssthresh, tp->prior_ssthresh, 2322 tp->packets_out); 2323 } 2324 #if IS_ENABLED(CONFIG_IPV6) 2325 else if (sk->sk_family == AF_INET6) { 2326 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n", 2327 msg, 2328 &sk->sk_v6_daddr, ntohs(inet->inet_dport), 2329 tp->snd_cwnd, tcp_left_out(tp), 2330 tp->snd_ssthresh, tp->prior_ssthresh, 2331 tp->packets_out); 2332 } 2333 #endif 2334 #endif 2335 } 2336 2337 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss) 2338 { 2339 struct tcp_sock *tp = tcp_sk(sk); 2340 2341 if (unmark_loss) { 2342 struct sk_buff *skb; 2343 2344 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) { 2345 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 2346 } 2347 tp->lost_out = 0; 2348 tcp_clear_all_retrans_hints(tp); 2349 } 2350 2351 if (tp->prior_ssthresh) { 2352 const struct inet_connection_sock *icsk = inet_csk(sk); 2353 2354 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk); 2355 2356 if (tp->prior_ssthresh > tp->snd_ssthresh) { 2357 tp->snd_ssthresh = tp->prior_ssthresh; 2358 tcp_ecn_withdraw_cwr(tp); 2359 } 2360 } 2361 tp->snd_cwnd_stamp = tcp_jiffies32; 2362 tp->undo_marker = 0; 2363 tp->rack.advanced = 1; /* Force RACK to re-exam losses */ 2364 } 2365 2366 static inline bool tcp_may_undo(const struct tcp_sock *tp) 2367 { 2368 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp)); 2369 } 2370 2371 /* People celebrate: "We love our President!" */ 2372 static bool tcp_try_undo_recovery(struct sock *sk) 2373 { 2374 struct tcp_sock *tp = tcp_sk(sk); 2375 2376 if (tcp_may_undo(tp)) { 2377 int mib_idx; 2378 2379 /* Happy end! We did not retransmit anything 2380 * or our original transmission succeeded. 2381 */ 2382 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans"); 2383 tcp_undo_cwnd_reduction(sk, false); 2384 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss) 2385 mib_idx = LINUX_MIB_TCPLOSSUNDO; 2386 else 2387 mib_idx = LINUX_MIB_TCPFULLUNDO; 2388 2389 NET_INC_STATS(sock_net(sk), mib_idx); 2390 } else if (tp->rack.reo_wnd_persist) { 2391 tp->rack.reo_wnd_persist--; 2392 } 2393 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) { 2394 /* Hold old state until something *above* high_seq 2395 * is ACKed. For Reno it is MUST to prevent false 2396 * fast retransmits (RFC2582). SACK TCP is safe. */ 2397 if (!tcp_any_retrans_done(sk)) 2398 tp->retrans_stamp = 0; 2399 return true; 2400 } 2401 tcp_set_ca_state(sk, TCP_CA_Open); 2402 tp->is_sack_reneg = 0; 2403 return false; 2404 } 2405 2406 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */ 2407 static bool tcp_try_undo_dsack(struct sock *sk) 2408 { 2409 struct tcp_sock *tp = tcp_sk(sk); 2410 2411 if (tp->undo_marker && !tp->undo_retrans) { 2412 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH, 2413 tp->rack.reo_wnd_persist + 1); 2414 DBGUNDO(sk, "D-SACK"); 2415 tcp_undo_cwnd_reduction(sk, false); 2416 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO); 2417 return true; 2418 } 2419 return false; 2420 } 2421 2422 /* Undo during loss recovery after partial ACK or using F-RTO. */ 2423 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo) 2424 { 2425 struct tcp_sock *tp = tcp_sk(sk); 2426 2427 if (frto_undo || tcp_may_undo(tp)) { 2428 tcp_undo_cwnd_reduction(sk, true); 2429 2430 DBGUNDO(sk, "partial loss"); 2431 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO); 2432 if (frto_undo) 2433 NET_INC_STATS(sock_net(sk), 2434 LINUX_MIB_TCPSPURIOUSRTOS); 2435 inet_csk(sk)->icsk_retransmits = 0; 2436 if (frto_undo || tcp_is_sack(tp)) { 2437 tcp_set_ca_state(sk, TCP_CA_Open); 2438 tp->is_sack_reneg = 0; 2439 } 2440 return true; 2441 } 2442 return false; 2443 } 2444 2445 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937. 2446 * It computes the number of packets to send (sndcnt) based on packets newly 2447 * delivered: 2448 * 1) If the packets in flight is larger than ssthresh, PRR spreads the 2449 * cwnd reductions across a full RTT. 2450 * 2) Otherwise PRR uses packet conservation to send as much as delivered. 2451 * But when the retransmits are acked without further losses, PRR 2452 * slow starts cwnd up to ssthresh to speed up the recovery. 2453 */ 2454 static void tcp_init_cwnd_reduction(struct sock *sk) 2455 { 2456 struct tcp_sock *tp = tcp_sk(sk); 2457 2458 tp->high_seq = tp->snd_nxt; 2459 tp->tlp_high_seq = 0; 2460 tp->snd_cwnd_cnt = 0; 2461 tp->prior_cwnd = tp->snd_cwnd; 2462 tp->prr_delivered = 0; 2463 tp->prr_out = 0; 2464 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk); 2465 tcp_ecn_queue_cwr(tp); 2466 } 2467 2468 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int flag) 2469 { 2470 struct tcp_sock *tp = tcp_sk(sk); 2471 int sndcnt = 0; 2472 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp); 2473 2474 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd)) 2475 return; 2476 2477 tp->prr_delivered += newly_acked_sacked; 2478 if (delta < 0) { 2479 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered + 2480 tp->prior_cwnd - 1; 2481 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out; 2482 } else if ((flag & (FLAG_RETRANS_DATA_ACKED | FLAG_LOST_RETRANS)) == 2483 FLAG_RETRANS_DATA_ACKED) { 2484 sndcnt = min_t(int, delta, 2485 max_t(int, tp->prr_delivered - tp->prr_out, 2486 newly_acked_sacked) + 1); 2487 } else { 2488 sndcnt = min(delta, newly_acked_sacked); 2489 } 2490 /* Force a fast retransmit upon entering fast recovery */ 2491 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1)); 2492 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt; 2493 } 2494 2495 static inline void tcp_end_cwnd_reduction(struct sock *sk) 2496 { 2497 struct tcp_sock *tp = tcp_sk(sk); 2498 2499 if (inet_csk(sk)->icsk_ca_ops->cong_control) 2500 return; 2501 2502 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */ 2503 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH && 2504 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) { 2505 tp->snd_cwnd = tp->snd_ssthresh; 2506 tp->snd_cwnd_stamp = tcp_jiffies32; 2507 } 2508 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR); 2509 } 2510 2511 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */ 2512 void tcp_enter_cwr(struct sock *sk) 2513 { 2514 struct tcp_sock *tp = tcp_sk(sk); 2515 2516 tp->prior_ssthresh = 0; 2517 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) { 2518 tp->undo_marker = 0; 2519 tcp_init_cwnd_reduction(sk); 2520 tcp_set_ca_state(sk, TCP_CA_CWR); 2521 } 2522 } 2523 EXPORT_SYMBOL(tcp_enter_cwr); 2524 2525 static void tcp_try_keep_open(struct sock *sk) 2526 { 2527 struct tcp_sock *tp = tcp_sk(sk); 2528 int state = TCP_CA_Open; 2529 2530 if (tcp_left_out(tp) || tcp_any_retrans_done(sk)) 2531 state = TCP_CA_Disorder; 2532 2533 if (inet_csk(sk)->icsk_ca_state != state) { 2534 tcp_set_ca_state(sk, state); 2535 tp->high_seq = tp->snd_nxt; 2536 } 2537 } 2538 2539 static void tcp_try_to_open(struct sock *sk, int flag) 2540 { 2541 struct tcp_sock *tp = tcp_sk(sk); 2542 2543 tcp_verify_left_out(tp); 2544 2545 if (!tcp_any_retrans_done(sk)) 2546 tp->retrans_stamp = 0; 2547 2548 if (flag & FLAG_ECE) 2549 tcp_enter_cwr(sk); 2550 2551 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) { 2552 tcp_try_keep_open(sk); 2553 } 2554 } 2555 2556 static void tcp_mtup_probe_failed(struct sock *sk) 2557 { 2558 struct inet_connection_sock *icsk = inet_csk(sk); 2559 2560 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1; 2561 icsk->icsk_mtup.probe_size = 0; 2562 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL); 2563 } 2564 2565 static void tcp_mtup_probe_success(struct sock *sk) 2566 { 2567 struct tcp_sock *tp = tcp_sk(sk); 2568 struct inet_connection_sock *icsk = inet_csk(sk); 2569 2570 /* FIXME: breaks with very large cwnd */ 2571 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2572 tp->snd_cwnd = tp->snd_cwnd * 2573 tcp_mss_to_mtu(sk, tp->mss_cache) / 2574 icsk->icsk_mtup.probe_size; 2575 tp->snd_cwnd_cnt = 0; 2576 tp->snd_cwnd_stamp = tcp_jiffies32; 2577 tp->snd_ssthresh = tcp_current_ssthresh(sk); 2578 2579 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size; 2580 icsk->icsk_mtup.probe_size = 0; 2581 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 2582 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS); 2583 } 2584 2585 /* Do a simple retransmit without using the backoff mechanisms in 2586 * tcp_timer. This is used for path mtu discovery. 2587 * The socket is already locked here. 2588 */ 2589 void tcp_simple_retransmit(struct sock *sk) 2590 { 2591 const struct inet_connection_sock *icsk = inet_csk(sk); 2592 struct tcp_sock *tp = tcp_sk(sk); 2593 struct sk_buff *skb; 2594 unsigned int mss = tcp_current_mss(sk); 2595 2596 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) { 2597 if (tcp_skb_seglen(skb) > mss && 2598 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { 2599 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2600 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 2601 tp->retrans_out -= tcp_skb_pcount(skb); 2602 } 2603 tcp_skb_mark_lost_uncond_verify(tp, skb); 2604 } 2605 } 2606 2607 tcp_clear_retrans_hints_partial(tp); 2608 2609 if (!tp->lost_out) 2610 return; 2611 2612 if (tcp_is_reno(tp)) 2613 tcp_limit_reno_sacked(tp); 2614 2615 tcp_verify_left_out(tp); 2616 2617 /* Don't muck with the congestion window here. 2618 * Reason is that we do not increase amount of _data_ 2619 * in network, but units changed and effective 2620 * cwnd/ssthresh really reduced now. 2621 */ 2622 if (icsk->icsk_ca_state != TCP_CA_Loss) { 2623 tp->high_seq = tp->snd_nxt; 2624 tp->snd_ssthresh = tcp_current_ssthresh(sk); 2625 tp->prior_ssthresh = 0; 2626 tp->undo_marker = 0; 2627 tcp_set_ca_state(sk, TCP_CA_Loss); 2628 } 2629 tcp_xmit_retransmit_queue(sk); 2630 } 2631 EXPORT_SYMBOL(tcp_simple_retransmit); 2632 2633 void tcp_enter_recovery(struct sock *sk, bool ece_ack) 2634 { 2635 struct tcp_sock *tp = tcp_sk(sk); 2636 int mib_idx; 2637 2638 if (tcp_is_reno(tp)) 2639 mib_idx = LINUX_MIB_TCPRENORECOVERY; 2640 else 2641 mib_idx = LINUX_MIB_TCPSACKRECOVERY; 2642 2643 NET_INC_STATS(sock_net(sk), mib_idx); 2644 2645 tp->prior_ssthresh = 0; 2646 tcp_init_undo(tp); 2647 2648 if (!tcp_in_cwnd_reduction(sk)) { 2649 if (!ece_ack) 2650 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2651 tcp_init_cwnd_reduction(sk); 2652 } 2653 tcp_set_ca_state(sk, TCP_CA_Recovery); 2654 } 2655 2656 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are 2657 * recovered or spurious. Otherwise retransmits more on partial ACKs. 2658 */ 2659 static void tcp_process_loss(struct sock *sk, int flag, int num_dupack, 2660 int *rexmit) 2661 { 2662 struct tcp_sock *tp = tcp_sk(sk); 2663 bool recovered = !before(tp->snd_una, tp->high_seq); 2664 2665 if ((flag & FLAG_SND_UNA_ADVANCED || tp->fastopen_rsk) && 2666 tcp_try_undo_loss(sk, false)) 2667 return; 2668 2669 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */ 2670 /* Step 3.b. A timeout is spurious if not all data are 2671 * lost, i.e., never-retransmitted data are (s)acked. 2672 */ 2673 if ((flag & FLAG_ORIG_SACK_ACKED) && 2674 tcp_try_undo_loss(sk, true)) 2675 return; 2676 2677 if (after(tp->snd_nxt, tp->high_seq)) { 2678 if (flag & FLAG_DATA_SACKED || num_dupack) 2679 tp->frto = 0; /* Step 3.a. loss was real */ 2680 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) { 2681 tp->high_seq = tp->snd_nxt; 2682 /* Step 2.b. Try send new data (but deferred until cwnd 2683 * is updated in tcp_ack()). Otherwise fall back to 2684 * the conventional recovery. 2685 */ 2686 if (!tcp_write_queue_empty(sk) && 2687 after(tcp_wnd_end(tp), tp->snd_nxt)) { 2688 *rexmit = REXMIT_NEW; 2689 return; 2690 } 2691 tp->frto = 0; 2692 } 2693 } 2694 2695 if (recovered) { 2696 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */ 2697 tcp_try_undo_recovery(sk); 2698 return; 2699 } 2700 if (tcp_is_reno(tp)) { 2701 /* A Reno DUPACK means new data in F-RTO step 2.b above are 2702 * delivered. Lower inflight to clock out (re)tranmissions. 2703 */ 2704 if (after(tp->snd_nxt, tp->high_seq) && num_dupack) 2705 tcp_add_reno_sack(sk, num_dupack); 2706 else if (flag & FLAG_SND_UNA_ADVANCED) 2707 tcp_reset_reno_sack(tp); 2708 } 2709 *rexmit = REXMIT_LOST; 2710 } 2711 2712 /* Undo during fast recovery after partial ACK. */ 2713 static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una) 2714 { 2715 struct tcp_sock *tp = tcp_sk(sk); 2716 2717 if (tp->undo_marker && tcp_packet_delayed(tp)) { 2718 /* Plain luck! Hole if filled with delayed 2719 * packet, rather than with a retransmit. Check reordering. 2720 */ 2721 tcp_check_sack_reordering(sk, prior_snd_una, 1); 2722 2723 /* We are getting evidence that the reordering degree is higher 2724 * than we realized. If there are no retransmits out then we 2725 * can undo. Otherwise we clock out new packets but do not 2726 * mark more packets lost or retransmit more. 2727 */ 2728 if (tp->retrans_out) 2729 return true; 2730 2731 if (!tcp_any_retrans_done(sk)) 2732 tp->retrans_stamp = 0; 2733 2734 DBGUNDO(sk, "partial recovery"); 2735 tcp_undo_cwnd_reduction(sk, true); 2736 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO); 2737 tcp_try_keep_open(sk); 2738 return true; 2739 } 2740 return false; 2741 } 2742 2743 static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag) 2744 { 2745 struct tcp_sock *tp = tcp_sk(sk); 2746 2747 if (tcp_rtx_queue_empty(sk)) 2748 return; 2749 2750 if (unlikely(tcp_is_reno(tp))) { 2751 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED); 2752 } else if (tcp_is_rack(sk)) { 2753 u32 prior_retrans = tp->retrans_out; 2754 2755 tcp_rack_mark_lost(sk); 2756 if (prior_retrans > tp->retrans_out) 2757 *ack_flag |= FLAG_LOST_RETRANS; 2758 } 2759 } 2760 2761 static bool tcp_force_fast_retransmit(struct sock *sk) 2762 { 2763 struct tcp_sock *tp = tcp_sk(sk); 2764 2765 return after(tcp_highest_sack_seq(tp), 2766 tp->snd_una + tp->reordering * tp->mss_cache); 2767 } 2768 2769 /* Process an event, which can update packets-in-flight not trivially. 2770 * Main goal of this function is to calculate new estimate for left_out, 2771 * taking into account both packets sitting in receiver's buffer and 2772 * packets lost by network. 2773 * 2774 * Besides that it updates the congestion state when packet loss or ECN 2775 * is detected. But it does not reduce the cwnd, it is done by the 2776 * congestion control later. 2777 * 2778 * It does _not_ decide what to send, it is made in function 2779 * tcp_xmit_retransmit_queue(). 2780 */ 2781 static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una, 2782 int num_dupack, int *ack_flag, int *rexmit) 2783 { 2784 struct inet_connection_sock *icsk = inet_csk(sk); 2785 struct tcp_sock *tp = tcp_sk(sk); 2786 int fast_rexmit = 0, flag = *ack_flag; 2787 bool do_lost = num_dupack || ((flag & FLAG_DATA_SACKED) && 2788 tcp_force_fast_retransmit(sk)); 2789 2790 if (!tp->packets_out && tp->sacked_out) 2791 tp->sacked_out = 0; 2792 2793 /* Now state machine starts. 2794 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */ 2795 if (flag & FLAG_ECE) 2796 tp->prior_ssthresh = 0; 2797 2798 /* B. In all the states check for reneging SACKs. */ 2799 if (tcp_check_sack_reneging(sk, flag)) 2800 return; 2801 2802 /* C. Check consistency of the current state. */ 2803 tcp_verify_left_out(tp); 2804 2805 /* D. Check state exit conditions. State can be terminated 2806 * when high_seq is ACKed. */ 2807 if (icsk->icsk_ca_state == TCP_CA_Open) { 2808 WARN_ON(tp->retrans_out != 0); 2809 tp->retrans_stamp = 0; 2810 } else if (!before(tp->snd_una, tp->high_seq)) { 2811 switch (icsk->icsk_ca_state) { 2812 case TCP_CA_CWR: 2813 /* CWR is to be held something *above* high_seq 2814 * is ACKed for CWR bit to reach receiver. */ 2815 if (tp->snd_una != tp->high_seq) { 2816 tcp_end_cwnd_reduction(sk); 2817 tcp_set_ca_state(sk, TCP_CA_Open); 2818 } 2819 break; 2820 2821 case TCP_CA_Recovery: 2822 if (tcp_is_reno(tp)) 2823 tcp_reset_reno_sack(tp); 2824 if (tcp_try_undo_recovery(sk)) 2825 return; 2826 tcp_end_cwnd_reduction(sk); 2827 break; 2828 } 2829 } 2830 2831 /* E. Process state. */ 2832 switch (icsk->icsk_ca_state) { 2833 case TCP_CA_Recovery: 2834 if (!(flag & FLAG_SND_UNA_ADVANCED)) { 2835 if (tcp_is_reno(tp)) 2836 tcp_add_reno_sack(sk, num_dupack); 2837 } else { 2838 if (tcp_try_undo_partial(sk, prior_snd_una)) 2839 return; 2840 /* Partial ACK arrived. Force fast retransmit. */ 2841 do_lost = tcp_is_reno(tp) || 2842 tcp_force_fast_retransmit(sk); 2843 } 2844 if (tcp_try_undo_dsack(sk)) { 2845 tcp_try_keep_open(sk); 2846 return; 2847 } 2848 tcp_identify_packet_loss(sk, ack_flag); 2849 break; 2850 case TCP_CA_Loss: 2851 tcp_process_loss(sk, flag, num_dupack, rexmit); 2852 tcp_identify_packet_loss(sk, ack_flag); 2853 if (!(icsk->icsk_ca_state == TCP_CA_Open || 2854 (*ack_flag & FLAG_LOST_RETRANS))) 2855 return; 2856 /* Change state if cwnd is undone or retransmits are lost */ 2857 /* fall through */ 2858 default: 2859 if (tcp_is_reno(tp)) { 2860 if (flag & FLAG_SND_UNA_ADVANCED) 2861 tcp_reset_reno_sack(tp); 2862 tcp_add_reno_sack(sk, num_dupack); 2863 } 2864 2865 if (icsk->icsk_ca_state <= TCP_CA_Disorder) 2866 tcp_try_undo_dsack(sk); 2867 2868 tcp_identify_packet_loss(sk, ack_flag); 2869 if (!tcp_time_to_recover(sk, flag)) { 2870 tcp_try_to_open(sk, flag); 2871 return; 2872 } 2873 2874 /* MTU probe failure: don't reduce cwnd */ 2875 if (icsk->icsk_ca_state < TCP_CA_CWR && 2876 icsk->icsk_mtup.probe_size && 2877 tp->snd_una == tp->mtu_probe.probe_seq_start) { 2878 tcp_mtup_probe_failed(sk); 2879 /* Restores the reduction we did in tcp_mtup_probe() */ 2880 tp->snd_cwnd++; 2881 tcp_simple_retransmit(sk); 2882 return; 2883 } 2884 2885 /* Otherwise enter Recovery state */ 2886 tcp_enter_recovery(sk, (flag & FLAG_ECE)); 2887 fast_rexmit = 1; 2888 } 2889 2890 if (!tcp_is_rack(sk) && do_lost) 2891 tcp_update_scoreboard(sk, fast_rexmit); 2892 *rexmit = REXMIT_LOST; 2893 } 2894 2895 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag) 2896 { 2897 u32 wlen = sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen * HZ; 2898 struct tcp_sock *tp = tcp_sk(sk); 2899 2900 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) { 2901 /* If the remote keeps returning delayed ACKs, eventually 2902 * the min filter would pick it up and overestimate the 2903 * prop. delay when it expires. Skip suspected delayed ACKs. 2904 */ 2905 return; 2906 } 2907 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32, 2908 rtt_us ? : jiffies_to_usecs(1)); 2909 } 2910 2911 static bool tcp_ack_update_rtt(struct sock *sk, const int flag, 2912 long seq_rtt_us, long sack_rtt_us, 2913 long ca_rtt_us, struct rate_sample *rs) 2914 { 2915 const struct tcp_sock *tp = tcp_sk(sk); 2916 2917 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because 2918 * broken middle-boxes or peers may corrupt TS-ECR fields. But 2919 * Karn's algorithm forbids taking RTT if some retransmitted data 2920 * is acked (RFC6298). 2921 */ 2922 if (seq_rtt_us < 0) 2923 seq_rtt_us = sack_rtt_us; 2924 2925 /* RTTM Rule: A TSecr value received in a segment is used to 2926 * update the averaged RTT measurement only if the segment 2927 * acknowledges some new data, i.e., only if it advances the 2928 * left edge of the send window. 2929 * See draft-ietf-tcplw-high-performance-00, section 3.3. 2930 */ 2931 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 2932 flag & FLAG_ACKED) { 2933 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr; 2934 2935 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) { 2936 seq_rtt_us = delta * (USEC_PER_SEC / TCP_TS_HZ); 2937 ca_rtt_us = seq_rtt_us; 2938 } 2939 } 2940 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */ 2941 if (seq_rtt_us < 0) 2942 return false; 2943 2944 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is 2945 * always taken together with ACK, SACK, or TS-opts. Any negative 2946 * values will be skipped with the seq_rtt_us < 0 check above. 2947 */ 2948 tcp_update_rtt_min(sk, ca_rtt_us, flag); 2949 tcp_rtt_estimator(sk, seq_rtt_us); 2950 tcp_set_rto(sk); 2951 2952 /* RFC6298: only reset backoff on valid RTT measurement. */ 2953 inet_csk(sk)->icsk_backoff = 0; 2954 return true; 2955 } 2956 2957 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */ 2958 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req) 2959 { 2960 struct rate_sample rs; 2961 long rtt_us = -1L; 2962 2963 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack) 2964 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack); 2965 2966 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs); 2967 } 2968 2969 2970 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked) 2971 { 2972 const struct inet_connection_sock *icsk = inet_csk(sk); 2973 2974 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked); 2975 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32; 2976 } 2977 2978 /* Restart timer after forward progress on connection. 2979 * RFC2988 recommends to restart timer to now+rto. 2980 */ 2981 void tcp_rearm_rto(struct sock *sk) 2982 { 2983 const struct inet_connection_sock *icsk = inet_csk(sk); 2984 struct tcp_sock *tp = tcp_sk(sk); 2985 2986 /* If the retrans timer is currently being used by Fast Open 2987 * for SYN-ACK retrans purpose, stay put. 2988 */ 2989 if (tp->fastopen_rsk) 2990 return; 2991 2992 if (!tp->packets_out) { 2993 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS); 2994 } else { 2995 u32 rto = inet_csk(sk)->icsk_rto; 2996 /* Offset the time elapsed after installing regular RTO */ 2997 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT || 2998 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) { 2999 s64 delta_us = tcp_rto_delta_us(sk); 3000 /* delta_us may not be positive if the socket is locked 3001 * when the retrans timer fires and is rescheduled. 3002 */ 3003 rto = usecs_to_jiffies(max_t(int, delta_us, 1)); 3004 } 3005 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto, 3006 TCP_RTO_MAX, tcp_rtx_queue_head(sk)); 3007 } 3008 } 3009 3010 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */ 3011 static void tcp_set_xmit_timer(struct sock *sk) 3012 { 3013 if (!tcp_schedule_loss_probe(sk, true)) 3014 tcp_rearm_rto(sk); 3015 } 3016 3017 /* If we get here, the whole TSO packet has not been acked. */ 3018 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb) 3019 { 3020 struct tcp_sock *tp = tcp_sk(sk); 3021 u32 packets_acked; 3022 3023 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)); 3024 3025 packets_acked = tcp_skb_pcount(skb); 3026 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 3027 return 0; 3028 packets_acked -= tcp_skb_pcount(skb); 3029 3030 if (packets_acked) { 3031 BUG_ON(tcp_skb_pcount(skb) == 0); 3032 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)); 3033 } 3034 3035 return packets_acked; 3036 } 3037 3038 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb, 3039 u32 prior_snd_una) 3040 { 3041 const struct skb_shared_info *shinfo; 3042 3043 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */ 3044 if (likely(!TCP_SKB_CB(skb)->txstamp_ack)) 3045 return; 3046 3047 shinfo = skb_shinfo(skb); 3048 if (!before(shinfo->tskey, prior_snd_una) && 3049 before(shinfo->tskey, tcp_sk(sk)->snd_una)) { 3050 tcp_skb_tsorted_save(skb) { 3051 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK); 3052 } tcp_skb_tsorted_restore(skb); 3053 } 3054 } 3055 3056 /* Remove acknowledged frames from the retransmission queue. If our packet 3057 * is before the ack sequence we can discard it as it's confirmed to have 3058 * arrived at the other end. 3059 */ 3060 static int tcp_clean_rtx_queue(struct sock *sk, u32 prior_fack, 3061 u32 prior_snd_una, 3062 struct tcp_sacktag_state *sack) 3063 { 3064 const struct inet_connection_sock *icsk = inet_csk(sk); 3065 u64 first_ackt, last_ackt; 3066 struct tcp_sock *tp = tcp_sk(sk); 3067 u32 prior_sacked = tp->sacked_out; 3068 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */ 3069 struct sk_buff *skb, *next; 3070 bool fully_acked = true; 3071 long sack_rtt_us = -1L; 3072 long seq_rtt_us = -1L; 3073 long ca_rtt_us = -1L; 3074 u32 pkts_acked = 0; 3075 u32 last_in_flight = 0; 3076 bool rtt_update; 3077 int flag = 0; 3078 3079 first_ackt = 0; 3080 3081 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) { 3082 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 3083 const u32 start_seq = scb->seq; 3084 u8 sacked = scb->sacked; 3085 u32 acked_pcount; 3086 3087 tcp_ack_tstamp(sk, skb, prior_snd_una); 3088 3089 /* Determine how many packets and what bytes were acked, tso and else */ 3090 if (after(scb->end_seq, tp->snd_una)) { 3091 if (tcp_skb_pcount(skb) == 1 || 3092 !after(tp->snd_una, scb->seq)) 3093 break; 3094 3095 acked_pcount = tcp_tso_acked(sk, skb); 3096 if (!acked_pcount) 3097 break; 3098 fully_acked = false; 3099 } else { 3100 acked_pcount = tcp_skb_pcount(skb); 3101 } 3102 3103 if (unlikely(sacked & TCPCB_RETRANS)) { 3104 if (sacked & TCPCB_SACKED_RETRANS) 3105 tp->retrans_out -= acked_pcount; 3106 flag |= FLAG_RETRANS_DATA_ACKED; 3107 } else if (!(sacked & TCPCB_SACKED_ACKED)) { 3108 last_ackt = tcp_skb_timestamp_us(skb); 3109 WARN_ON_ONCE(last_ackt == 0); 3110 if (!first_ackt) 3111 first_ackt = last_ackt; 3112 3113 last_in_flight = TCP_SKB_CB(skb)->tx.in_flight; 3114 if (before(start_seq, reord)) 3115 reord = start_seq; 3116 if (!after(scb->end_seq, tp->high_seq)) 3117 flag |= FLAG_ORIG_SACK_ACKED; 3118 } 3119 3120 if (sacked & TCPCB_SACKED_ACKED) { 3121 tp->sacked_out -= acked_pcount; 3122 } else if (tcp_is_sack(tp)) { 3123 tp->delivered += acked_pcount; 3124 if (!tcp_skb_spurious_retrans(tp, skb)) 3125 tcp_rack_advance(tp, sacked, scb->end_seq, 3126 tcp_skb_timestamp_us(skb)); 3127 } 3128 if (sacked & TCPCB_LOST) 3129 tp->lost_out -= acked_pcount; 3130 3131 tp->packets_out -= acked_pcount; 3132 pkts_acked += acked_pcount; 3133 tcp_rate_skb_delivered(sk, skb, sack->rate); 3134 3135 /* Initial outgoing SYN's get put onto the write_queue 3136 * just like anything else we transmit. It is not 3137 * true data, and if we misinform our callers that 3138 * this ACK acks real data, we will erroneously exit 3139 * connection startup slow start one packet too 3140 * quickly. This is severely frowned upon behavior. 3141 */ 3142 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) { 3143 flag |= FLAG_DATA_ACKED; 3144 } else { 3145 flag |= FLAG_SYN_ACKED; 3146 tp->retrans_stamp = 0; 3147 } 3148 3149 if (!fully_acked) 3150 break; 3151 3152 next = skb_rb_next(skb); 3153 if (unlikely(skb == tp->retransmit_skb_hint)) 3154 tp->retransmit_skb_hint = NULL; 3155 if (unlikely(skb == tp->lost_skb_hint)) 3156 tp->lost_skb_hint = NULL; 3157 tcp_rtx_queue_unlink_and_free(skb, sk); 3158 } 3159 3160 if (!skb) 3161 tcp_chrono_stop(sk, TCP_CHRONO_BUSY); 3162 3163 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una))) 3164 tp->snd_up = tp->snd_una; 3165 3166 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 3167 flag |= FLAG_SACK_RENEGING; 3168 3169 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) { 3170 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt); 3171 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt); 3172 3173 if (pkts_acked == 1 && last_in_flight < tp->mss_cache && 3174 last_in_flight && !prior_sacked && fully_acked && 3175 sack->rate->prior_delivered + 1 == tp->delivered && 3176 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) { 3177 /* Conservatively mark a delayed ACK. It's typically 3178 * from a lone runt packet over the round trip to 3179 * a receiver w/o out-of-order or CE events. 3180 */ 3181 flag |= FLAG_ACK_MAYBE_DELAYED; 3182 } 3183 } 3184 if (sack->first_sackt) { 3185 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt); 3186 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt); 3187 } 3188 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us, 3189 ca_rtt_us, sack->rate); 3190 3191 if (flag & FLAG_ACKED) { 3192 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */ 3193 if (unlikely(icsk->icsk_mtup.probe_size && 3194 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) { 3195 tcp_mtup_probe_success(sk); 3196 } 3197 3198 if (tcp_is_reno(tp)) { 3199 tcp_remove_reno_sacks(sk, pkts_acked); 3200 3201 /* If any of the cumulatively ACKed segments was 3202 * retransmitted, non-SACK case cannot confirm that 3203 * progress was due to original transmission due to 3204 * lack of TCPCB_SACKED_ACKED bits even if some of 3205 * the packets may have been never retransmitted. 3206 */ 3207 if (flag & FLAG_RETRANS_DATA_ACKED) 3208 flag &= ~FLAG_ORIG_SACK_ACKED; 3209 } else { 3210 int delta; 3211 3212 /* Non-retransmitted hole got filled? That's reordering */ 3213 if (before(reord, prior_fack)) 3214 tcp_check_sack_reordering(sk, reord, 0); 3215 3216 delta = prior_sacked - tp->sacked_out; 3217 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta); 3218 } 3219 } else if (skb && rtt_update && sack_rtt_us >= 0 && 3220 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp, 3221 tcp_skb_timestamp_us(skb))) { 3222 /* Do not re-arm RTO if the sack RTT is measured from data sent 3223 * after when the head was last (re)transmitted. Otherwise the 3224 * timeout may continue to extend in loss recovery. 3225 */ 3226 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */ 3227 } 3228 3229 if (icsk->icsk_ca_ops->pkts_acked) { 3230 struct ack_sample sample = { .pkts_acked = pkts_acked, 3231 .rtt_us = sack->rate->rtt_us, 3232 .in_flight = last_in_flight }; 3233 3234 icsk->icsk_ca_ops->pkts_acked(sk, &sample); 3235 } 3236 3237 #if FASTRETRANS_DEBUG > 0 3238 WARN_ON((int)tp->sacked_out < 0); 3239 WARN_ON((int)tp->lost_out < 0); 3240 WARN_ON((int)tp->retrans_out < 0); 3241 if (!tp->packets_out && tcp_is_sack(tp)) { 3242 icsk = inet_csk(sk); 3243 if (tp->lost_out) { 3244 pr_debug("Leak l=%u %d\n", 3245 tp->lost_out, icsk->icsk_ca_state); 3246 tp->lost_out = 0; 3247 } 3248 if (tp->sacked_out) { 3249 pr_debug("Leak s=%u %d\n", 3250 tp->sacked_out, icsk->icsk_ca_state); 3251 tp->sacked_out = 0; 3252 } 3253 if (tp->retrans_out) { 3254 pr_debug("Leak r=%u %d\n", 3255 tp->retrans_out, icsk->icsk_ca_state); 3256 tp->retrans_out = 0; 3257 } 3258 } 3259 #endif 3260 return flag; 3261 } 3262 3263 static void tcp_ack_probe(struct sock *sk) 3264 { 3265 struct inet_connection_sock *icsk = inet_csk(sk); 3266 struct sk_buff *head = tcp_send_head(sk); 3267 const struct tcp_sock *tp = tcp_sk(sk); 3268 3269 /* Was it a usable window open? */ 3270 if (!head) 3271 return; 3272 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) { 3273 icsk->icsk_backoff = 0; 3274 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0); 3275 /* Socket must be waked up by subsequent tcp_data_snd_check(). 3276 * This function is not for random using! 3277 */ 3278 } else { 3279 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX); 3280 3281 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 3282 when, TCP_RTO_MAX, NULL); 3283 } 3284 } 3285 3286 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag) 3287 { 3288 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) || 3289 inet_csk(sk)->icsk_ca_state != TCP_CA_Open; 3290 } 3291 3292 /* Decide wheather to run the increase function of congestion control. */ 3293 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag) 3294 { 3295 /* If reordering is high then always grow cwnd whenever data is 3296 * delivered regardless of its ordering. Otherwise stay conservative 3297 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/ 3298 * new SACK or ECE mark may first advance cwnd here and later reduce 3299 * cwnd in tcp_fastretrans_alert() based on more states. 3300 */ 3301 if (tcp_sk(sk)->reordering > sock_net(sk)->ipv4.sysctl_tcp_reordering) 3302 return flag & FLAG_FORWARD_PROGRESS; 3303 3304 return flag & FLAG_DATA_ACKED; 3305 } 3306 3307 /* The "ultimate" congestion control function that aims to replace the rigid 3308 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction). 3309 * It's called toward the end of processing an ACK with precise rate 3310 * information. All transmission or retransmission are delayed afterwards. 3311 */ 3312 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked, 3313 int flag, const struct rate_sample *rs) 3314 { 3315 const struct inet_connection_sock *icsk = inet_csk(sk); 3316 3317 if (icsk->icsk_ca_ops->cong_control) { 3318 icsk->icsk_ca_ops->cong_control(sk, rs); 3319 return; 3320 } 3321 3322 if (tcp_in_cwnd_reduction(sk)) { 3323 /* Reduce cwnd if state mandates */ 3324 tcp_cwnd_reduction(sk, acked_sacked, flag); 3325 } else if (tcp_may_raise_cwnd(sk, flag)) { 3326 /* Advance cwnd if state allows */ 3327 tcp_cong_avoid(sk, ack, acked_sacked); 3328 } 3329 tcp_update_pacing_rate(sk); 3330 } 3331 3332 /* Check that window update is acceptable. 3333 * The function assumes that snd_una<=ack<=snd_next. 3334 */ 3335 static inline bool tcp_may_update_window(const struct tcp_sock *tp, 3336 const u32 ack, const u32 ack_seq, 3337 const u32 nwin) 3338 { 3339 return after(ack, tp->snd_una) || 3340 after(ack_seq, tp->snd_wl1) || 3341 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd); 3342 } 3343 3344 /* If we update tp->snd_una, also update tp->bytes_acked */ 3345 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack) 3346 { 3347 u32 delta = ack - tp->snd_una; 3348 3349 sock_owned_by_me((struct sock *)tp); 3350 tp->bytes_acked += delta; 3351 tp->snd_una = ack; 3352 } 3353 3354 /* If we update tp->rcv_nxt, also update tp->bytes_received */ 3355 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq) 3356 { 3357 u32 delta = seq - tp->rcv_nxt; 3358 3359 sock_owned_by_me((struct sock *)tp); 3360 tp->bytes_received += delta; 3361 tp->rcv_nxt = seq; 3362 } 3363 3364 /* Update our send window. 3365 * 3366 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2 3367 * and in FreeBSD. NetBSD's one is even worse.) is wrong. 3368 */ 3369 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack, 3370 u32 ack_seq) 3371 { 3372 struct tcp_sock *tp = tcp_sk(sk); 3373 int flag = 0; 3374 u32 nwin = ntohs(tcp_hdr(skb)->window); 3375 3376 if (likely(!tcp_hdr(skb)->syn)) 3377 nwin <<= tp->rx_opt.snd_wscale; 3378 3379 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) { 3380 flag |= FLAG_WIN_UPDATE; 3381 tcp_update_wl(tp, ack_seq); 3382 3383 if (tp->snd_wnd != nwin) { 3384 tp->snd_wnd = nwin; 3385 3386 /* Note, it is the only place, where 3387 * fast path is recovered for sending TCP. 3388 */ 3389 tp->pred_flags = 0; 3390 tcp_fast_path_check(sk); 3391 3392 if (!tcp_write_queue_empty(sk)) 3393 tcp_slow_start_after_idle_check(sk); 3394 3395 if (nwin > tp->max_window) { 3396 tp->max_window = nwin; 3397 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie); 3398 } 3399 } 3400 } 3401 3402 tcp_snd_una_update(tp, ack); 3403 3404 return flag; 3405 } 3406 3407 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx, 3408 u32 *last_oow_ack_time) 3409 { 3410 if (*last_oow_ack_time) { 3411 s32 elapsed = (s32)(tcp_jiffies32 - *last_oow_ack_time); 3412 3413 if (0 <= elapsed && elapsed < net->ipv4.sysctl_tcp_invalid_ratelimit) { 3414 NET_INC_STATS(net, mib_idx); 3415 return true; /* rate-limited: don't send yet! */ 3416 } 3417 } 3418 3419 *last_oow_ack_time = tcp_jiffies32; 3420 3421 return false; /* not rate-limited: go ahead, send dupack now! */ 3422 } 3423 3424 /* Return true if we're currently rate-limiting out-of-window ACKs and 3425 * thus shouldn't send a dupack right now. We rate-limit dupacks in 3426 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS 3427 * attacks that send repeated SYNs or ACKs for the same connection. To 3428 * do this, we do not send a duplicate SYNACK or ACK if the remote 3429 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate. 3430 */ 3431 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb, 3432 int mib_idx, u32 *last_oow_ack_time) 3433 { 3434 /* Data packets without SYNs are not likely part of an ACK loop. */ 3435 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) && 3436 !tcp_hdr(skb)->syn) 3437 return false; 3438 3439 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time); 3440 } 3441 3442 /* RFC 5961 7 [ACK Throttling] */ 3443 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb) 3444 { 3445 /* unprotected vars, we dont care of overwrites */ 3446 static u32 challenge_timestamp; 3447 static unsigned int challenge_count; 3448 struct tcp_sock *tp = tcp_sk(sk); 3449 struct net *net = sock_net(sk); 3450 u32 count, now; 3451 3452 /* First check our per-socket dupack rate limit. */ 3453 if (__tcp_oow_rate_limited(net, 3454 LINUX_MIB_TCPACKSKIPPEDCHALLENGE, 3455 &tp->last_oow_ack_time)) 3456 return; 3457 3458 /* Then check host-wide RFC 5961 rate limit. */ 3459 now = jiffies / HZ; 3460 if (now != challenge_timestamp) { 3461 u32 ack_limit = net->ipv4.sysctl_tcp_challenge_ack_limit; 3462 u32 half = (ack_limit + 1) >> 1; 3463 3464 challenge_timestamp = now; 3465 WRITE_ONCE(challenge_count, half + prandom_u32_max(ack_limit)); 3466 } 3467 count = READ_ONCE(challenge_count); 3468 if (count > 0) { 3469 WRITE_ONCE(challenge_count, count - 1); 3470 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK); 3471 tcp_send_ack(sk); 3472 } 3473 } 3474 3475 static void tcp_store_ts_recent(struct tcp_sock *tp) 3476 { 3477 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval; 3478 tp->rx_opt.ts_recent_stamp = ktime_get_seconds(); 3479 } 3480 3481 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq) 3482 { 3483 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) { 3484 /* PAWS bug workaround wrt. ACK frames, the PAWS discard 3485 * extra check below makes sure this can only happen 3486 * for pure ACK frames. -DaveM 3487 * 3488 * Not only, also it occurs for expired timestamps. 3489 */ 3490 3491 if (tcp_paws_check(&tp->rx_opt, 0)) 3492 tcp_store_ts_recent(tp); 3493 } 3494 } 3495 3496 /* This routine deals with acks during a TLP episode. 3497 * We mark the end of a TLP episode on receiving TLP dupack or when 3498 * ack is after tlp_high_seq. 3499 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe. 3500 */ 3501 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag) 3502 { 3503 struct tcp_sock *tp = tcp_sk(sk); 3504 3505 if (before(ack, tp->tlp_high_seq)) 3506 return; 3507 3508 if (flag & FLAG_DSACKING_ACK) { 3509 /* This DSACK means original and TLP probe arrived; no loss */ 3510 tp->tlp_high_seq = 0; 3511 } else if (after(ack, tp->tlp_high_seq)) { 3512 /* ACK advances: there was a loss, so reduce cwnd. Reset 3513 * tlp_high_seq in tcp_init_cwnd_reduction() 3514 */ 3515 tcp_init_cwnd_reduction(sk); 3516 tcp_set_ca_state(sk, TCP_CA_CWR); 3517 tcp_end_cwnd_reduction(sk); 3518 tcp_try_keep_open(sk); 3519 NET_INC_STATS(sock_net(sk), 3520 LINUX_MIB_TCPLOSSPROBERECOVERY); 3521 } else if (!(flag & (FLAG_SND_UNA_ADVANCED | 3522 FLAG_NOT_DUP | FLAG_DATA_SACKED))) { 3523 /* Pure dupack: original and TLP probe arrived; no loss */ 3524 tp->tlp_high_seq = 0; 3525 } 3526 } 3527 3528 static inline void tcp_in_ack_event(struct sock *sk, u32 flags) 3529 { 3530 const struct inet_connection_sock *icsk = inet_csk(sk); 3531 3532 if (icsk->icsk_ca_ops->in_ack_event) 3533 icsk->icsk_ca_ops->in_ack_event(sk, flags); 3534 } 3535 3536 /* Congestion control has updated the cwnd already. So if we're in 3537 * loss recovery then now we do any new sends (for FRTO) or 3538 * retransmits (for CA_Loss or CA_recovery) that make sense. 3539 */ 3540 static void tcp_xmit_recovery(struct sock *sk, int rexmit) 3541 { 3542 struct tcp_sock *tp = tcp_sk(sk); 3543 3544 if (rexmit == REXMIT_NONE || sk->sk_state == TCP_SYN_SENT) 3545 return; 3546 3547 if (unlikely(rexmit == 2)) { 3548 __tcp_push_pending_frames(sk, tcp_current_mss(sk), 3549 TCP_NAGLE_OFF); 3550 if (after(tp->snd_nxt, tp->high_seq)) 3551 return; 3552 tp->frto = 0; 3553 } 3554 tcp_xmit_retransmit_queue(sk); 3555 } 3556 3557 /* Returns the number of packets newly acked or sacked by the current ACK */ 3558 static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, int flag) 3559 { 3560 const struct net *net = sock_net(sk); 3561 struct tcp_sock *tp = tcp_sk(sk); 3562 u32 delivered; 3563 3564 delivered = tp->delivered - prior_delivered; 3565 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered); 3566 if (flag & FLAG_ECE) { 3567 tp->delivered_ce += delivered; 3568 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, delivered); 3569 } 3570 return delivered; 3571 } 3572 3573 /* This routine deals with incoming acks, but not outgoing ones. */ 3574 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag) 3575 { 3576 struct inet_connection_sock *icsk = inet_csk(sk); 3577 struct tcp_sock *tp = tcp_sk(sk); 3578 struct tcp_sacktag_state sack_state; 3579 struct rate_sample rs = { .prior_delivered = 0 }; 3580 u32 prior_snd_una = tp->snd_una; 3581 bool is_sack_reneg = tp->is_sack_reneg; 3582 u32 ack_seq = TCP_SKB_CB(skb)->seq; 3583 u32 ack = TCP_SKB_CB(skb)->ack_seq; 3584 int num_dupack = 0; 3585 int prior_packets = tp->packets_out; 3586 u32 delivered = tp->delivered; 3587 u32 lost = tp->lost; 3588 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */ 3589 u32 prior_fack; 3590 3591 sack_state.first_sackt = 0; 3592 sack_state.rate = &rs; 3593 3594 /* We very likely will need to access rtx queue. */ 3595 prefetch(sk->tcp_rtx_queue.rb_node); 3596 3597 /* If the ack is older than previous acks 3598 * then we can probably ignore it. 3599 */ 3600 if (before(ack, prior_snd_una)) { 3601 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */ 3602 if (before(ack, prior_snd_una - tp->max_window)) { 3603 if (!(flag & FLAG_NO_CHALLENGE_ACK)) 3604 tcp_send_challenge_ack(sk, skb); 3605 return -1; 3606 } 3607 goto old_ack; 3608 } 3609 3610 /* If the ack includes data we haven't sent yet, discard 3611 * this segment (RFC793 Section 3.9). 3612 */ 3613 if (after(ack, tp->snd_nxt)) 3614 return -1; 3615 3616 if (after(ack, prior_snd_una)) { 3617 flag |= FLAG_SND_UNA_ADVANCED; 3618 icsk->icsk_retransmits = 0; 3619 3620 #if IS_ENABLED(CONFIG_TLS_DEVICE) 3621 if (static_branch_unlikely(&clean_acked_data_enabled.key)) 3622 if (icsk->icsk_clean_acked) 3623 icsk->icsk_clean_acked(sk, ack); 3624 #endif 3625 } 3626 3627 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una; 3628 rs.prior_in_flight = tcp_packets_in_flight(tp); 3629 3630 /* ts_recent update must be made after we are sure that the packet 3631 * is in window. 3632 */ 3633 if (flag & FLAG_UPDATE_TS_RECENT) 3634 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); 3635 3636 if ((flag & (FLAG_SLOWPATH | FLAG_SND_UNA_ADVANCED)) == 3637 FLAG_SND_UNA_ADVANCED) { 3638 /* Window is constant, pure forward advance. 3639 * No more checks are required. 3640 * Note, we use the fact that SND.UNA>=SND.WL2. 3641 */ 3642 tcp_update_wl(tp, ack_seq); 3643 tcp_snd_una_update(tp, ack); 3644 flag |= FLAG_WIN_UPDATE; 3645 3646 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE); 3647 3648 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS); 3649 } else { 3650 u32 ack_ev_flags = CA_ACK_SLOWPATH; 3651 3652 if (ack_seq != TCP_SKB_CB(skb)->end_seq) 3653 flag |= FLAG_DATA; 3654 else 3655 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS); 3656 3657 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq); 3658 3659 if (TCP_SKB_CB(skb)->sacked) 3660 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una, 3661 &sack_state); 3662 3663 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) { 3664 flag |= FLAG_ECE; 3665 ack_ev_flags |= CA_ACK_ECE; 3666 } 3667 3668 if (flag & FLAG_WIN_UPDATE) 3669 ack_ev_flags |= CA_ACK_WIN_UPDATE; 3670 3671 tcp_in_ack_event(sk, ack_ev_flags); 3672 } 3673 3674 /* We passed data and got it acked, remove any soft error 3675 * log. Something worked... 3676 */ 3677 sk->sk_err_soft = 0; 3678 icsk->icsk_probes_out = 0; 3679 tp->rcv_tstamp = tcp_jiffies32; 3680 if (!prior_packets) 3681 goto no_queue; 3682 3683 /* See if we can take anything off of the retransmit queue. */ 3684 flag |= tcp_clean_rtx_queue(sk, prior_fack, prior_snd_una, &sack_state); 3685 3686 tcp_rack_update_reo_wnd(sk, &rs); 3687 3688 if (tp->tlp_high_seq) 3689 tcp_process_tlp_ack(sk, ack, flag); 3690 /* If needed, reset TLP/RTO timer; RACK may later override this. */ 3691 if (flag & FLAG_SET_XMIT_TIMER) 3692 tcp_set_xmit_timer(sk); 3693 3694 if (tcp_ack_is_dubious(sk, flag)) { 3695 if (!(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP))) { 3696 num_dupack = 1; 3697 /* Consider if pure acks were aggregated in tcp_add_backlog() */ 3698 if (!(flag & FLAG_DATA)) 3699 num_dupack = max_t(u16, 1, skb_shinfo(skb)->gso_segs); 3700 } 3701 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag, 3702 &rexmit); 3703 } 3704 3705 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) 3706 sk_dst_confirm(sk); 3707 3708 delivered = tcp_newly_delivered(sk, delivered, flag); 3709 lost = tp->lost - lost; /* freshly marked lost */ 3710 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED); 3711 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate); 3712 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate); 3713 tcp_xmit_recovery(sk, rexmit); 3714 return 1; 3715 3716 no_queue: 3717 /* If data was DSACKed, see if we can undo a cwnd reduction. */ 3718 if (flag & FLAG_DSACKING_ACK) { 3719 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag, 3720 &rexmit); 3721 tcp_newly_delivered(sk, delivered, flag); 3722 } 3723 /* If this ack opens up a zero window, clear backoff. It was 3724 * being used to time the probes, and is probably far higher than 3725 * it needs to be for normal retransmission. 3726 */ 3727 tcp_ack_probe(sk); 3728 3729 if (tp->tlp_high_seq) 3730 tcp_process_tlp_ack(sk, ack, flag); 3731 return 1; 3732 3733 old_ack: 3734 /* If data was SACKed, tag it and see if we should send more data. 3735 * If data was DSACKed, see if we can undo a cwnd reduction. 3736 */ 3737 if (TCP_SKB_CB(skb)->sacked) { 3738 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una, 3739 &sack_state); 3740 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag, 3741 &rexmit); 3742 tcp_newly_delivered(sk, delivered, flag); 3743 tcp_xmit_recovery(sk, rexmit); 3744 } 3745 3746 return 0; 3747 } 3748 3749 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie, 3750 bool syn, struct tcp_fastopen_cookie *foc, 3751 bool exp_opt) 3752 { 3753 /* Valid only in SYN or SYN-ACK with an even length. */ 3754 if (!foc || !syn || len < 0 || (len & 1)) 3755 return; 3756 3757 if (len >= TCP_FASTOPEN_COOKIE_MIN && 3758 len <= TCP_FASTOPEN_COOKIE_MAX) 3759 memcpy(foc->val, cookie, len); 3760 else if (len != 0) 3761 len = -1; 3762 foc->len = len; 3763 foc->exp = exp_opt; 3764 } 3765 3766 static void smc_parse_options(const struct tcphdr *th, 3767 struct tcp_options_received *opt_rx, 3768 const unsigned char *ptr, 3769 int opsize) 3770 { 3771 #if IS_ENABLED(CONFIG_SMC) 3772 if (static_branch_unlikely(&tcp_have_smc)) { 3773 if (th->syn && !(opsize & 1) && 3774 opsize >= TCPOLEN_EXP_SMC_BASE && 3775 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC) 3776 opt_rx->smc_ok = 1; 3777 } 3778 #endif 3779 } 3780 3781 /* Look for tcp options. Normally only called on SYN and SYNACK packets. 3782 * But, this can also be called on packets in the established flow when 3783 * the fast version below fails. 3784 */ 3785 void tcp_parse_options(const struct net *net, 3786 const struct sk_buff *skb, 3787 struct tcp_options_received *opt_rx, int estab, 3788 struct tcp_fastopen_cookie *foc) 3789 { 3790 const unsigned char *ptr; 3791 const struct tcphdr *th = tcp_hdr(skb); 3792 int length = (th->doff * 4) - sizeof(struct tcphdr); 3793 3794 ptr = (const unsigned char *)(th + 1); 3795 opt_rx->saw_tstamp = 0; 3796 3797 while (length > 0) { 3798 int opcode = *ptr++; 3799 int opsize; 3800 3801 switch (opcode) { 3802 case TCPOPT_EOL: 3803 return; 3804 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ 3805 length--; 3806 continue; 3807 default: 3808 if (length < 2) 3809 return; 3810 opsize = *ptr++; 3811 if (opsize < 2) /* "silly options" */ 3812 return; 3813 if (opsize > length) 3814 return; /* don't parse partial options */ 3815 switch (opcode) { 3816 case TCPOPT_MSS: 3817 if (opsize == TCPOLEN_MSS && th->syn && !estab) { 3818 u16 in_mss = get_unaligned_be16(ptr); 3819 if (in_mss) { 3820 if (opt_rx->user_mss && 3821 opt_rx->user_mss < in_mss) 3822 in_mss = opt_rx->user_mss; 3823 opt_rx->mss_clamp = in_mss; 3824 } 3825 } 3826 break; 3827 case TCPOPT_WINDOW: 3828 if (opsize == TCPOLEN_WINDOW && th->syn && 3829 !estab && net->ipv4.sysctl_tcp_window_scaling) { 3830 __u8 snd_wscale = *(__u8 *)ptr; 3831 opt_rx->wscale_ok = 1; 3832 if (snd_wscale > TCP_MAX_WSCALE) { 3833 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n", 3834 __func__, 3835 snd_wscale, 3836 TCP_MAX_WSCALE); 3837 snd_wscale = TCP_MAX_WSCALE; 3838 } 3839 opt_rx->snd_wscale = snd_wscale; 3840 } 3841 break; 3842 case TCPOPT_TIMESTAMP: 3843 if ((opsize == TCPOLEN_TIMESTAMP) && 3844 ((estab && opt_rx->tstamp_ok) || 3845 (!estab && net->ipv4.sysctl_tcp_timestamps))) { 3846 opt_rx->saw_tstamp = 1; 3847 opt_rx->rcv_tsval = get_unaligned_be32(ptr); 3848 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4); 3849 } 3850 break; 3851 case TCPOPT_SACK_PERM: 3852 if (opsize == TCPOLEN_SACK_PERM && th->syn && 3853 !estab && net->ipv4.sysctl_tcp_sack) { 3854 opt_rx->sack_ok = TCP_SACK_SEEN; 3855 tcp_sack_reset(opt_rx); 3856 } 3857 break; 3858 3859 case TCPOPT_SACK: 3860 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) && 3861 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) && 3862 opt_rx->sack_ok) { 3863 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th; 3864 } 3865 break; 3866 #ifdef CONFIG_TCP_MD5SIG 3867 case TCPOPT_MD5SIG: 3868 /* 3869 * The MD5 Hash has already been 3870 * checked (see tcp_v{4,6}_do_rcv()). 3871 */ 3872 break; 3873 #endif 3874 case TCPOPT_FASTOPEN: 3875 tcp_parse_fastopen_option( 3876 opsize - TCPOLEN_FASTOPEN_BASE, 3877 ptr, th->syn, foc, false); 3878 break; 3879 3880 case TCPOPT_EXP: 3881 /* Fast Open option shares code 254 using a 3882 * 16 bits magic number. 3883 */ 3884 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE && 3885 get_unaligned_be16(ptr) == 3886 TCPOPT_FASTOPEN_MAGIC) 3887 tcp_parse_fastopen_option(opsize - 3888 TCPOLEN_EXP_FASTOPEN_BASE, 3889 ptr + 2, th->syn, foc, true); 3890 else 3891 smc_parse_options(th, opt_rx, ptr, 3892 opsize); 3893 break; 3894 3895 } 3896 ptr += opsize-2; 3897 length -= opsize; 3898 } 3899 } 3900 } 3901 EXPORT_SYMBOL(tcp_parse_options); 3902 3903 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th) 3904 { 3905 const __be32 *ptr = (const __be32 *)(th + 1); 3906 3907 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) 3908 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) { 3909 tp->rx_opt.saw_tstamp = 1; 3910 ++ptr; 3911 tp->rx_opt.rcv_tsval = ntohl(*ptr); 3912 ++ptr; 3913 if (*ptr) 3914 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset; 3915 else 3916 tp->rx_opt.rcv_tsecr = 0; 3917 return true; 3918 } 3919 return false; 3920 } 3921 3922 /* Fast parse options. This hopes to only see timestamps. 3923 * If it is wrong it falls back on tcp_parse_options(). 3924 */ 3925 static bool tcp_fast_parse_options(const struct net *net, 3926 const struct sk_buff *skb, 3927 const struct tcphdr *th, struct tcp_sock *tp) 3928 { 3929 /* In the spirit of fast parsing, compare doff directly to constant 3930 * values. Because equality is used, short doff can be ignored here. 3931 */ 3932 if (th->doff == (sizeof(*th) / 4)) { 3933 tp->rx_opt.saw_tstamp = 0; 3934 return false; 3935 } else if (tp->rx_opt.tstamp_ok && 3936 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) { 3937 if (tcp_parse_aligned_timestamp(tp, th)) 3938 return true; 3939 } 3940 3941 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL); 3942 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) 3943 tp->rx_opt.rcv_tsecr -= tp->tsoffset; 3944 3945 return true; 3946 } 3947 3948 #ifdef CONFIG_TCP_MD5SIG 3949 /* 3950 * Parse MD5 Signature option 3951 */ 3952 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th) 3953 { 3954 int length = (th->doff << 2) - sizeof(*th); 3955 const u8 *ptr = (const u8 *)(th + 1); 3956 3957 /* If not enough data remaining, we can short cut */ 3958 while (length >= TCPOLEN_MD5SIG) { 3959 int opcode = *ptr++; 3960 int opsize; 3961 3962 switch (opcode) { 3963 case TCPOPT_EOL: 3964 return NULL; 3965 case TCPOPT_NOP: 3966 length--; 3967 continue; 3968 default: 3969 opsize = *ptr++; 3970 if (opsize < 2 || opsize > length) 3971 return NULL; 3972 if (opcode == TCPOPT_MD5SIG) 3973 return opsize == TCPOLEN_MD5SIG ? ptr : NULL; 3974 } 3975 ptr += opsize - 2; 3976 length -= opsize; 3977 } 3978 return NULL; 3979 } 3980 EXPORT_SYMBOL(tcp_parse_md5sig_option); 3981 #endif 3982 3983 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM 3984 * 3985 * It is not fatal. If this ACK does _not_ change critical state (seqs, window) 3986 * it can pass through stack. So, the following predicate verifies that 3987 * this segment is not used for anything but congestion avoidance or 3988 * fast retransmit. Moreover, we even are able to eliminate most of such 3989 * second order effects, if we apply some small "replay" window (~RTO) 3990 * to timestamp space. 3991 * 3992 * All these measures still do not guarantee that we reject wrapped ACKs 3993 * on networks with high bandwidth, when sequence space is recycled fastly, 3994 * but it guarantees that such events will be very rare and do not affect 3995 * connection seriously. This doesn't look nice, but alas, PAWS is really 3996 * buggy extension. 3997 * 3998 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC 3999 * states that events when retransmit arrives after original data are rare. 4000 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is 4001 * the biggest problem on large power networks even with minor reordering. 4002 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe 4003 * up to bandwidth of 18Gigabit/sec. 8) ] 4004 */ 4005 4006 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb) 4007 { 4008 const struct tcp_sock *tp = tcp_sk(sk); 4009 const struct tcphdr *th = tcp_hdr(skb); 4010 u32 seq = TCP_SKB_CB(skb)->seq; 4011 u32 ack = TCP_SKB_CB(skb)->ack_seq; 4012 4013 return (/* 1. Pure ACK with correct sequence number. */ 4014 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) && 4015 4016 /* 2. ... and duplicate ACK. */ 4017 ack == tp->snd_una && 4018 4019 /* 3. ... and does not update window. */ 4020 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) && 4021 4022 /* 4. ... and sits in replay window. */ 4023 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ); 4024 } 4025 4026 static inline bool tcp_paws_discard(const struct sock *sk, 4027 const struct sk_buff *skb) 4028 { 4029 const struct tcp_sock *tp = tcp_sk(sk); 4030 4031 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) && 4032 !tcp_disordered_ack(sk, skb); 4033 } 4034 4035 /* Check segment sequence number for validity. 4036 * 4037 * Segment controls are considered valid, if the segment 4038 * fits to the window after truncation to the window. Acceptability 4039 * of data (and SYN, FIN, of course) is checked separately. 4040 * See tcp_data_queue(), for example. 4041 * 4042 * Also, controls (RST is main one) are accepted using RCV.WUP instead 4043 * of RCV.NXT. Peer still did not advance his SND.UNA when we 4044 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP. 4045 * (borrowed from freebsd) 4046 */ 4047 4048 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq) 4049 { 4050 return !before(end_seq, tp->rcv_wup) && 4051 !after(seq, tp->rcv_nxt + tcp_receive_window(tp)); 4052 } 4053 4054 /* When we get a reset we do this. */ 4055 void tcp_reset(struct sock *sk) 4056 { 4057 trace_tcp_receive_reset(sk); 4058 4059 /* We want the right error as BSD sees it (and indeed as we do). */ 4060 switch (sk->sk_state) { 4061 case TCP_SYN_SENT: 4062 sk->sk_err = ECONNREFUSED; 4063 break; 4064 case TCP_CLOSE_WAIT: 4065 sk->sk_err = EPIPE; 4066 break; 4067 case TCP_CLOSE: 4068 return; 4069 default: 4070 sk->sk_err = ECONNRESET; 4071 } 4072 /* This barrier is coupled with smp_rmb() in tcp_poll() */ 4073 smp_wmb(); 4074 4075 tcp_write_queue_purge(sk); 4076 tcp_done(sk); 4077 4078 if (!sock_flag(sk, SOCK_DEAD)) 4079 sk->sk_error_report(sk); 4080 } 4081 4082 /* 4083 * Process the FIN bit. This now behaves as it is supposed to work 4084 * and the FIN takes effect when it is validly part of sequence 4085 * space. Not before when we get holes. 4086 * 4087 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT 4088 * (and thence onto LAST-ACK and finally, CLOSE, we never enter 4089 * TIME-WAIT) 4090 * 4091 * If we are in FINWAIT-1, a received FIN indicates simultaneous 4092 * close and we go into CLOSING (and later onto TIME-WAIT) 4093 * 4094 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT. 4095 */ 4096 void tcp_fin(struct sock *sk) 4097 { 4098 struct tcp_sock *tp = tcp_sk(sk); 4099 4100 inet_csk_schedule_ack(sk); 4101 4102 sk->sk_shutdown |= RCV_SHUTDOWN; 4103 sock_set_flag(sk, SOCK_DONE); 4104 4105 switch (sk->sk_state) { 4106 case TCP_SYN_RECV: 4107 case TCP_ESTABLISHED: 4108 /* Move to CLOSE_WAIT */ 4109 tcp_set_state(sk, TCP_CLOSE_WAIT); 4110 inet_csk_enter_pingpong_mode(sk); 4111 break; 4112 4113 case TCP_CLOSE_WAIT: 4114 case TCP_CLOSING: 4115 /* Received a retransmission of the FIN, do 4116 * nothing. 4117 */ 4118 break; 4119 case TCP_LAST_ACK: 4120 /* RFC793: Remain in the LAST-ACK state. */ 4121 break; 4122 4123 case TCP_FIN_WAIT1: 4124 /* This case occurs when a simultaneous close 4125 * happens, we must ack the received FIN and 4126 * enter the CLOSING state. 4127 */ 4128 tcp_send_ack(sk); 4129 tcp_set_state(sk, TCP_CLOSING); 4130 break; 4131 case TCP_FIN_WAIT2: 4132 /* Received a FIN -- send ACK and enter TIME_WAIT. */ 4133 tcp_send_ack(sk); 4134 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 4135 break; 4136 default: 4137 /* Only TCP_LISTEN and TCP_CLOSE are left, in these 4138 * cases we should never reach this piece of code. 4139 */ 4140 pr_err("%s: Impossible, sk->sk_state=%d\n", 4141 __func__, sk->sk_state); 4142 break; 4143 } 4144 4145 /* It _is_ possible, that we have something out-of-order _after_ FIN. 4146 * Probably, we should reset in this case. For now drop them. 4147 */ 4148 skb_rbtree_purge(&tp->out_of_order_queue); 4149 if (tcp_is_sack(tp)) 4150 tcp_sack_reset(&tp->rx_opt); 4151 sk_mem_reclaim(sk); 4152 4153 if (!sock_flag(sk, SOCK_DEAD)) { 4154 sk->sk_state_change(sk); 4155 4156 /* Do not send POLL_HUP for half duplex close. */ 4157 if (sk->sk_shutdown == SHUTDOWN_MASK || 4158 sk->sk_state == TCP_CLOSE) 4159 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP); 4160 else 4161 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN); 4162 } 4163 } 4164 4165 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, 4166 u32 end_seq) 4167 { 4168 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) { 4169 if (before(seq, sp->start_seq)) 4170 sp->start_seq = seq; 4171 if (after(end_seq, sp->end_seq)) 4172 sp->end_seq = end_seq; 4173 return true; 4174 } 4175 return false; 4176 } 4177 4178 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq) 4179 { 4180 struct tcp_sock *tp = tcp_sk(sk); 4181 4182 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) { 4183 int mib_idx; 4184 4185 if (before(seq, tp->rcv_nxt)) 4186 mib_idx = LINUX_MIB_TCPDSACKOLDSENT; 4187 else 4188 mib_idx = LINUX_MIB_TCPDSACKOFOSENT; 4189 4190 NET_INC_STATS(sock_net(sk), mib_idx); 4191 4192 tp->rx_opt.dsack = 1; 4193 tp->duplicate_sack[0].start_seq = seq; 4194 tp->duplicate_sack[0].end_seq = end_seq; 4195 } 4196 } 4197 4198 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq) 4199 { 4200 struct tcp_sock *tp = tcp_sk(sk); 4201 4202 if (!tp->rx_opt.dsack) 4203 tcp_dsack_set(sk, seq, end_seq); 4204 else 4205 tcp_sack_extend(tp->duplicate_sack, seq, end_seq); 4206 } 4207 4208 static void tcp_rcv_spurious_retrans(struct sock *sk, const struct sk_buff *skb) 4209 { 4210 /* When the ACK path fails or drops most ACKs, the sender would 4211 * timeout and spuriously retransmit the same segment repeatedly. 4212 * The receiver remembers and reflects via DSACKs. Leverage the 4213 * DSACK state and change the txhash to re-route speculatively. 4214 */ 4215 if (TCP_SKB_CB(skb)->seq == tcp_sk(sk)->duplicate_sack[0].start_seq) 4216 sk_rethink_txhash(sk); 4217 } 4218 4219 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb) 4220 { 4221 struct tcp_sock *tp = tcp_sk(sk); 4222 4223 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 4224 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4225 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); 4226 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS); 4227 4228 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) { 4229 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 4230 4231 tcp_rcv_spurious_retrans(sk, skb); 4232 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) 4233 end_seq = tp->rcv_nxt; 4234 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq); 4235 } 4236 } 4237 4238 tcp_send_ack(sk); 4239 } 4240 4241 /* These routines update the SACK block as out-of-order packets arrive or 4242 * in-order packets close up the sequence space. 4243 */ 4244 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp) 4245 { 4246 int this_sack; 4247 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4248 struct tcp_sack_block *swalk = sp + 1; 4249 4250 /* See if the recent change to the first SACK eats into 4251 * or hits the sequence space of other SACK blocks, if so coalesce. 4252 */ 4253 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) { 4254 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) { 4255 int i; 4256 4257 /* Zap SWALK, by moving every further SACK up by one slot. 4258 * Decrease num_sacks. 4259 */ 4260 tp->rx_opt.num_sacks--; 4261 for (i = this_sack; i < tp->rx_opt.num_sacks; i++) 4262 sp[i] = sp[i + 1]; 4263 continue; 4264 } 4265 this_sack++, swalk++; 4266 } 4267 } 4268 4269 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq) 4270 { 4271 struct tcp_sock *tp = tcp_sk(sk); 4272 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4273 int cur_sacks = tp->rx_opt.num_sacks; 4274 int this_sack; 4275 4276 if (!cur_sacks) 4277 goto new_sack; 4278 4279 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) { 4280 if (tcp_sack_extend(sp, seq, end_seq)) { 4281 /* Rotate this_sack to the first one. */ 4282 for (; this_sack > 0; this_sack--, sp--) 4283 swap(*sp, *(sp - 1)); 4284 if (cur_sacks > 1) 4285 tcp_sack_maybe_coalesce(tp); 4286 return; 4287 } 4288 } 4289 4290 /* Could not find an adjacent existing SACK, build a new one, 4291 * put it at the front, and shift everyone else down. We 4292 * always know there is at least one SACK present already here. 4293 * 4294 * If the sack array is full, forget about the last one. 4295 */ 4296 if (this_sack >= TCP_NUM_SACKS) { 4297 if (tp->compressed_ack > TCP_FASTRETRANS_THRESH) 4298 tcp_send_ack(sk); 4299 this_sack--; 4300 tp->rx_opt.num_sacks--; 4301 sp--; 4302 } 4303 for (; this_sack > 0; this_sack--, sp--) 4304 *sp = *(sp - 1); 4305 4306 new_sack: 4307 /* Build the new head SACK, and we're done. */ 4308 sp->start_seq = seq; 4309 sp->end_seq = end_seq; 4310 tp->rx_opt.num_sacks++; 4311 } 4312 4313 /* RCV.NXT advances, some SACKs should be eaten. */ 4314 4315 static void tcp_sack_remove(struct tcp_sock *tp) 4316 { 4317 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4318 int num_sacks = tp->rx_opt.num_sacks; 4319 int this_sack; 4320 4321 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */ 4322 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) { 4323 tp->rx_opt.num_sacks = 0; 4324 return; 4325 } 4326 4327 for (this_sack = 0; this_sack < num_sacks;) { 4328 /* Check if the start of the sack is covered by RCV.NXT. */ 4329 if (!before(tp->rcv_nxt, sp->start_seq)) { 4330 int i; 4331 4332 /* RCV.NXT must cover all the block! */ 4333 WARN_ON(before(tp->rcv_nxt, sp->end_seq)); 4334 4335 /* Zap this SACK, by moving forward any other SACKS. */ 4336 for (i = this_sack+1; i < num_sacks; i++) 4337 tp->selective_acks[i-1] = tp->selective_acks[i]; 4338 num_sacks--; 4339 continue; 4340 } 4341 this_sack++; 4342 sp++; 4343 } 4344 tp->rx_opt.num_sacks = num_sacks; 4345 } 4346 4347 /** 4348 * tcp_try_coalesce - try to merge skb to prior one 4349 * @sk: socket 4350 * @dest: destination queue 4351 * @to: prior buffer 4352 * @from: buffer to add in queue 4353 * @fragstolen: pointer to boolean 4354 * 4355 * Before queueing skb @from after @to, try to merge them 4356 * to reduce overall memory use and queue lengths, if cost is small. 4357 * Packets in ofo or receive queues can stay a long time. 4358 * Better try to coalesce them right now to avoid future collapses. 4359 * Returns true if caller should free @from instead of queueing it 4360 */ 4361 static bool tcp_try_coalesce(struct sock *sk, 4362 struct sk_buff *to, 4363 struct sk_buff *from, 4364 bool *fragstolen) 4365 { 4366 int delta; 4367 4368 *fragstolen = false; 4369 4370 /* Its possible this segment overlaps with prior segment in queue */ 4371 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq) 4372 return false; 4373 4374 #ifdef CONFIG_TLS_DEVICE 4375 if (from->decrypted != to->decrypted) 4376 return false; 4377 #endif 4378 4379 if (!skb_try_coalesce(to, from, fragstolen, &delta)) 4380 return false; 4381 4382 atomic_add(delta, &sk->sk_rmem_alloc); 4383 sk_mem_charge(sk, delta); 4384 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE); 4385 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq; 4386 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq; 4387 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags; 4388 4389 if (TCP_SKB_CB(from)->has_rxtstamp) { 4390 TCP_SKB_CB(to)->has_rxtstamp = true; 4391 to->tstamp = from->tstamp; 4392 skb_hwtstamps(to)->hwtstamp = skb_hwtstamps(from)->hwtstamp; 4393 } 4394 4395 return true; 4396 } 4397 4398 static bool tcp_ooo_try_coalesce(struct sock *sk, 4399 struct sk_buff *to, 4400 struct sk_buff *from, 4401 bool *fragstolen) 4402 { 4403 bool res = tcp_try_coalesce(sk, to, from, fragstolen); 4404 4405 /* In case tcp_drop() is called later, update to->gso_segs */ 4406 if (res) { 4407 u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) + 4408 max_t(u16, 1, skb_shinfo(from)->gso_segs); 4409 4410 skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF); 4411 } 4412 return res; 4413 } 4414 4415 static void tcp_drop(struct sock *sk, struct sk_buff *skb) 4416 { 4417 sk_drops_add(sk, skb); 4418 __kfree_skb(skb); 4419 } 4420 4421 /* This one checks to see if we can put data from the 4422 * out_of_order queue into the receive_queue. 4423 */ 4424 static void tcp_ofo_queue(struct sock *sk) 4425 { 4426 struct tcp_sock *tp = tcp_sk(sk); 4427 __u32 dsack_high = tp->rcv_nxt; 4428 bool fin, fragstolen, eaten; 4429 struct sk_buff *skb, *tail; 4430 struct rb_node *p; 4431 4432 p = rb_first(&tp->out_of_order_queue); 4433 while (p) { 4434 skb = rb_to_skb(p); 4435 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 4436 break; 4437 4438 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) { 4439 __u32 dsack = dsack_high; 4440 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high)) 4441 dsack_high = TCP_SKB_CB(skb)->end_seq; 4442 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack); 4443 } 4444 p = rb_next(p); 4445 rb_erase(&skb->rbnode, &tp->out_of_order_queue); 4446 4447 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) { 4448 tcp_drop(sk, skb); 4449 continue; 4450 } 4451 4452 tail = skb_peek_tail(&sk->sk_receive_queue); 4453 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen); 4454 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq); 4455 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN; 4456 if (!eaten) 4457 __skb_queue_tail(&sk->sk_receive_queue, skb); 4458 else 4459 kfree_skb_partial(skb, fragstolen); 4460 4461 if (unlikely(fin)) { 4462 tcp_fin(sk); 4463 /* tcp_fin() purges tp->out_of_order_queue, 4464 * so we must end this loop right now. 4465 */ 4466 break; 4467 } 4468 } 4469 } 4470 4471 static bool tcp_prune_ofo_queue(struct sock *sk); 4472 static int tcp_prune_queue(struct sock *sk); 4473 4474 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb, 4475 unsigned int size) 4476 { 4477 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || 4478 !sk_rmem_schedule(sk, skb, size)) { 4479 4480 if (tcp_prune_queue(sk) < 0) 4481 return -1; 4482 4483 while (!sk_rmem_schedule(sk, skb, size)) { 4484 if (!tcp_prune_ofo_queue(sk)) 4485 return -1; 4486 } 4487 } 4488 return 0; 4489 } 4490 4491 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb) 4492 { 4493 struct tcp_sock *tp = tcp_sk(sk); 4494 struct rb_node **p, *parent; 4495 struct sk_buff *skb1; 4496 u32 seq, end_seq; 4497 bool fragstolen; 4498 4499 tcp_ecn_check_ce(sk, skb); 4500 4501 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) { 4502 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP); 4503 tcp_drop(sk, skb); 4504 return; 4505 } 4506 4507 /* Disable header prediction. */ 4508 tp->pred_flags = 0; 4509 inet_csk_schedule_ack(sk); 4510 4511 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE); 4512 seq = TCP_SKB_CB(skb)->seq; 4513 end_seq = TCP_SKB_CB(skb)->end_seq; 4514 4515 p = &tp->out_of_order_queue.rb_node; 4516 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) { 4517 /* Initial out of order segment, build 1 SACK. */ 4518 if (tcp_is_sack(tp)) { 4519 tp->rx_opt.num_sacks = 1; 4520 tp->selective_acks[0].start_seq = seq; 4521 tp->selective_acks[0].end_seq = end_seq; 4522 } 4523 rb_link_node(&skb->rbnode, NULL, p); 4524 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue); 4525 tp->ooo_last_skb = skb; 4526 goto end; 4527 } 4528 4529 /* In the typical case, we are adding an skb to the end of the list. 4530 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup. 4531 */ 4532 if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb, 4533 skb, &fragstolen)) { 4534 coalesce_done: 4535 tcp_grow_window(sk, skb); 4536 kfree_skb_partial(skb, fragstolen); 4537 skb = NULL; 4538 goto add_sack; 4539 } 4540 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */ 4541 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) { 4542 parent = &tp->ooo_last_skb->rbnode; 4543 p = &parent->rb_right; 4544 goto insert; 4545 } 4546 4547 /* Find place to insert this segment. Handle overlaps on the way. */ 4548 parent = NULL; 4549 while (*p) { 4550 parent = *p; 4551 skb1 = rb_to_skb(parent); 4552 if (before(seq, TCP_SKB_CB(skb1)->seq)) { 4553 p = &parent->rb_left; 4554 continue; 4555 } 4556 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) { 4557 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 4558 /* All the bits are present. Drop. */ 4559 NET_INC_STATS(sock_net(sk), 4560 LINUX_MIB_TCPOFOMERGE); 4561 tcp_drop(sk, skb); 4562 skb = NULL; 4563 tcp_dsack_set(sk, seq, end_seq); 4564 goto add_sack; 4565 } 4566 if (after(seq, TCP_SKB_CB(skb1)->seq)) { 4567 /* Partial overlap. */ 4568 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq); 4569 } else { 4570 /* skb's seq == skb1's seq and skb covers skb1. 4571 * Replace skb1 with skb. 4572 */ 4573 rb_replace_node(&skb1->rbnode, &skb->rbnode, 4574 &tp->out_of_order_queue); 4575 tcp_dsack_extend(sk, 4576 TCP_SKB_CB(skb1)->seq, 4577 TCP_SKB_CB(skb1)->end_seq); 4578 NET_INC_STATS(sock_net(sk), 4579 LINUX_MIB_TCPOFOMERGE); 4580 tcp_drop(sk, skb1); 4581 goto merge_right; 4582 } 4583 } else if (tcp_ooo_try_coalesce(sk, skb1, 4584 skb, &fragstolen)) { 4585 goto coalesce_done; 4586 } 4587 p = &parent->rb_right; 4588 } 4589 insert: 4590 /* Insert segment into RB tree. */ 4591 rb_link_node(&skb->rbnode, parent, p); 4592 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue); 4593 4594 merge_right: 4595 /* Remove other segments covered by skb. */ 4596 while ((skb1 = skb_rb_next(skb)) != NULL) { 4597 if (!after(end_seq, TCP_SKB_CB(skb1)->seq)) 4598 break; 4599 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 4600 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, 4601 end_seq); 4602 break; 4603 } 4604 rb_erase(&skb1->rbnode, &tp->out_of_order_queue); 4605 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, 4606 TCP_SKB_CB(skb1)->end_seq); 4607 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE); 4608 tcp_drop(sk, skb1); 4609 } 4610 /* If there is no skb after us, we are the last_skb ! */ 4611 if (!skb1) 4612 tp->ooo_last_skb = skb; 4613 4614 add_sack: 4615 if (tcp_is_sack(tp)) 4616 tcp_sack_new_ofo_skb(sk, seq, end_seq); 4617 end: 4618 if (skb) { 4619 tcp_grow_window(sk, skb); 4620 skb_condense(skb); 4621 skb_set_owner_r(skb, sk); 4622 } 4623 } 4624 4625 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, 4626 bool *fragstolen) 4627 { 4628 int eaten; 4629 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue); 4630 4631 eaten = (tail && 4632 tcp_try_coalesce(sk, tail, 4633 skb, fragstolen)) ? 1 : 0; 4634 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq); 4635 if (!eaten) { 4636 __skb_queue_tail(&sk->sk_receive_queue, skb); 4637 skb_set_owner_r(skb, sk); 4638 } 4639 return eaten; 4640 } 4641 4642 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size) 4643 { 4644 struct sk_buff *skb; 4645 int err = -ENOMEM; 4646 int data_len = 0; 4647 bool fragstolen; 4648 4649 if (size == 0) 4650 return 0; 4651 4652 if (size > PAGE_SIZE) { 4653 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS); 4654 4655 data_len = npages << PAGE_SHIFT; 4656 size = data_len + (size & ~PAGE_MASK); 4657 } 4658 skb = alloc_skb_with_frags(size - data_len, data_len, 4659 PAGE_ALLOC_COSTLY_ORDER, 4660 &err, sk->sk_allocation); 4661 if (!skb) 4662 goto err; 4663 4664 skb_put(skb, size - data_len); 4665 skb->data_len = data_len; 4666 skb->len = size; 4667 4668 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) { 4669 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP); 4670 goto err_free; 4671 } 4672 4673 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size); 4674 if (err) 4675 goto err_free; 4676 4677 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt; 4678 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size; 4679 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1; 4680 4681 if (tcp_queue_rcv(sk, skb, &fragstolen)) { 4682 WARN_ON_ONCE(fragstolen); /* should not happen */ 4683 __kfree_skb(skb); 4684 } 4685 return size; 4686 4687 err_free: 4688 kfree_skb(skb); 4689 err: 4690 return err; 4691 4692 } 4693 4694 void tcp_data_ready(struct sock *sk) 4695 { 4696 const struct tcp_sock *tp = tcp_sk(sk); 4697 int avail = tp->rcv_nxt - tp->copied_seq; 4698 4699 if (avail < sk->sk_rcvlowat && !sock_flag(sk, SOCK_DONE)) 4700 return; 4701 4702 sk->sk_data_ready(sk); 4703 } 4704 4705 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb) 4706 { 4707 struct tcp_sock *tp = tcp_sk(sk); 4708 bool fragstolen; 4709 int eaten; 4710 4711 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) { 4712 __kfree_skb(skb); 4713 return; 4714 } 4715 skb_dst_drop(skb); 4716 __skb_pull(skb, tcp_hdr(skb)->doff * 4); 4717 4718 tcp_ecn_accept_cwr(sk, skb); 4719 4720 tp->rx_opt.dsack = 0; 4721 4722 /* Queue data for delivery to the user. 4723 * Packets in sequence go to the receive queue. 4724 * Out of sequence packets to the out_of_order_queue. 4725 */ 4726 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { 4727 if (tcp_receive_window(tp) == 0) { 4728 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP); 4729 goto out_of_window; 4730 } 4731 4732 /* Ok. In sequence. In window. */ 4733 queue_and_out: 4734 if (skb_queue_len(&sk->sk_receive_queue) == 0) 4735 sk_forced_mem_schedule(sk, skb->truesize); 4736 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) { 4737 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP); 4738 goto drop; 4739 } 4740 4741 eaten = tcp_queue_rcv(sk, skb, &fragstolen); 4742 if (skb->len) 4743 tcp_event_data_recv(sk, skb); 4744 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 4745 tcp_fin(sk); 4746 4747 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) { 4748 tcp_ofo_queue(sk); 4749 4750 /* RFC5681. 4.2. SHOULD send immediate ACK, when 4751 * gap in queue is filled. 4752 */ 4753 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) 4754 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW; 4755 } 4756 4757 if (tp->rx_opt.num_sacks) 4758 tcp_sack_remove(tp); 4759 4760 tcp_fast_path_check(sk); 4761 4762 if (eaten > 0) 4763 kfree_skb_partial(skb, fragstolen); 4764 if (!sock_flag(sk, SOCK_DEAD)) 4765 tcp_data_ready(sk); 4766 return; 4767 } 4768 4769 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 4770 tcp_rcv_spurious_retrans(sk, skb); 4771 /* A retransmit, 2nd most common case. Force an immediate ack. */ 4772 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); 4773 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 4774 4775 out_of_window: 4776 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS); 4777 inet_csk_schedule_ack(sk); 4778 drop: 4779 tcp_drop(sk, skb); 4780 return; 4781 } 4782 4783 /* Out of window. F.e. zero window probe. */ 4784 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp))) 4785 goto out_of_window; 4786 4787 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4788 /* Partial packet, seq < rcv_next < end_seq */ 4789 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt); 4790 4791 /* If window is closed, drop tail of packet. But after 4792 * remembering D-SACK for its head made in previous line. 4793 */ 4794 if (!tcp_receive_window(tp)) { 4795 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP); 4796 goto out_of_window; 4797 } 4798 goto queue_and_out; 4799 } 4800 4801 tcp_data_queue_ofo(sk, skb); 4802 } 4803 4804 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list) 4805 { 4806 if (list) 4807 return !skb_queue_is_last(list, skb) ? skb->next : NULL; 4808 4809 return skb_rb_next(skb); 4810 } 4811 4812 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb, 4813 struct sk_buff_head *list, 4814 struct rb_root *root) 4815 { 4816 struct sk_buff *next = tcp_skb_next(skb, list); 4817 4818 if (list) 4819 __skb_unlink(skb, list); 4820 else 4821 rb_erase(&skb->rbnode, root); 4822 4823 __kfree_skb(skb); 4824 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED); 4825 4826 return next; 4827 } 4828 4829 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */ 4830 void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb) 4831 { 4832 struct rb_node **p = &root->rb_node; 4833 struct rb_node *parent = NULL; 4834 struct sk_buff *skb1; 4835 4836 while (*p) { 4837 parent = *p; 4838 skb1 = rb_to_skb(parent); 4839 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq)) 4840 p = &parent->rb_left; 4841 else 4842 p = &parent->rb_right; 4843 } 4844 rb_link_node(&skb->rbnode, parent, p); 4845 rb_insert_color(&skb->rbnode, root); 4846 } 4847 4848 /* Collapse contiguous sequence of skbs head..tail with 4849 * sequence numbers start..end. 4850 * 4851 * If tail is NULL, this means until the end of the queue. 4852 * 4853 * Segments with FIN/SYN are not collapsed (only because this 4854 * simplifies code) 4855 */ 4856 static void 4857 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root, 4858 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end) 4859 { 4860 struct sk_buff *skb = head, *n; 4861 struct sk_buff_head tmp; 4862 bool end_of_skbs; 4863 4864 /* First, check that queue is collapsible and find 4865 * the point where collapsing can be useful. 4866 */ 4867 restart: 4868 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) { 4869 n = tcp_skb_next(skb, list); 4870 4871 /* No new bits? It is possible on ofo queue. */ 4872 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 4873 skb = tcp_collapse_one(sk, skb, list, root); 4874 if (!skb) 4875 break; 4876 goto restart; 4877 } 4878 4879 /* The first skb to collapse is: 4880 * - not SYN/FIN and 4881 * - bloated or contains data before "start" or 4882 * overlaps to the next one. 4883 */ 4884 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) && 4885 (tcp_win_from_space(sk, skb->truesize) > skb->len || 4886 before(TCP_SKB_CB(skb)->seq, start))) { 4887 end_of_skbs = false; 4888 break; 4889 } 4890 4891 if (n && n != tail && 4892 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) { 4893 end_of_skbs = false; 4894 break; 4895 } 4896 4897 /* Decided to skip this, advance start seq. */ 4898 start = TCP_SKB_CB(skb)->end_seq; 4899 } 4900 if (end_of_skbs || 4901 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN))) 4902 return; 4903 4904 __skb_queue_head_init(&tmp); 4905 4906 while (before(start, end)) { 4907 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start); 4908 struct sk_buff *nskb; 4909 4910 nskb = alloc_skb(copy, GFP_ATOMIC); 4911 if (!nskb) 4912 break; 4913 4914 memcpy(nskb->cb, skb->cb, sizeof(skb->cb)); 4915 #ifdef CONFIG_TLS_DEVICE 4916 nskb->decrypted = skb->decrypted; 4917 #endif 4918 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start; 4919 if (list) 4920 __skb_queue_before(list, skb, nskb); 4921 else 4922 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */ 4923 skb_set_owner_r(nskb, sk); 4924 4925 /* Copy data, releasing collapsed skbs. */ 4926 while (copy > 0) { 4927 int offset = start - TCP_SKB_CB(skb)->seq; 4928 int size = TCP_SKB_CB(skb)->end_seq - start; 4929 4930 BUG_ON(offset < 0); 4931 if (size > 0) { 4932 size = min(copy, size); 4933 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size)) 4934 BUG(); 4935 TCP_SKB_CB(nskb)->end_seq += size; 4936 copy -= size; 4937 start += size; 4938 } 4939 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 4940 skb = tcp_collapse_one(sk, skb, list, root); 4941 if (!skb || 4942 skb == tail || 4943 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN))) 4944 goto end; 4945 #ifdef CONFIG_TLS_DEVICE 4946 if (skb->decrypted != nskb->decrypted) 4947 goto end; 4948 #endif 4949 } 4950 } 4951 } 4952 end: 4953 skb_queue_walk_safe(&tmp, skb, n) 4954 tcp_rbtree_insert(root, skb); 4955 } 4956 4957 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs 4958 * and tcp_collapse() them until all the queue is collapsed. 4959 */ 4960 static void tcp_collapse_ofo_queue(struct sock *sk) 4961 { 4962 struct tcp_sock *tp = tcp_sk(sk); 4963 u32 range_truesize, sum_tiny = 0; 4964 struct sk_buff *skb, *head; 4965 u32 start, end; 4966 4967 skb = skb_rb_first(&tp->out_of_order_queue); 4968 new_range: 4969 if (!skb) { 4970 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue); 4971 return; 4972 } 4973 start = TCP_SKB_CB(skb)->seq; 4974 end = TCP_SKB_CB(skb)->end_seq; 4975 range_truesize = skb->truesize; 4976 4977 for (head = skb;;) { 4978 skb = skb_rb_next(skb); 4979 4980 /* Range is terminated when we see a gap or when 4981 * we are at the queue end. 4982 */ 4983 if (!skb || 4984 after(TCP_SKB_CB(skb)->seq, end) || 4985 before(TCP_SKB_CB(skb)->end_seq, start)) { 4986 /* Do not attempt collapsing tiny skbs */ 4987 if (range_truesize != head->truesize || 4988 end - start >= SKB_WITH_OVERHEAD(SK_MEM_QUANTUM)) { 4989 tcp_collapse(sk, NULL, &tp->out_of_order_queue, 4990 head, skb, start, end); 4991 } else { 4992 sum_tiny += range_truesize; 4993 if (sum_tiny > sk->sk_rcvbuf >> 3) 4994 return; 4995 } 4996 goto new_range; 4997 } 4998 4999 range_truesize += skb->truesize; 5000 if (unlikely(before(TCP_SKB_CB(skb)->seq, start))) 5001 start = TCP_SKB_CB(skb)->seq; 5002 if (after(TCP_SKB_CB(skb)->end_seq, end)) 5003 end = TCP_SKB_CB(skb)->end_seq; 5004 } 5005 } 5006 5007 /* 5008 * Clean the out-of-order queue to make room. 5009 * We drop high sequences packets to : 5010 * 1) Let a chance for holes to be filled. 5011 * 2) not add too big latencies if thousands of packets sit there. 5012 * (But if application shrinks SO_RCVBUF, we could still end up 5013 * freeing whole queue here) 5014 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks. 5015 * 5016 * Return true if queue has shrunk. 5017 */ 5018 static bool tcp_prune_ofo_queue(struct sock *sk) 5019 { 5020 struct tcp_sock *tp = tcp_sk(sk); 5021 struct rb_node *node, *prev; 5022 int goal; 5023 5024 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) 5025 return false; 5026 5027 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED); 5028 goal = sk->sk_rcvbuf >> 3; 5029 node = &tp->ooo_last_skb->rbnode; 5030 do { 5031 prev = rb_prev(node); 5032 rb_erase(node, &tp->out_of_order_queue); 5033 goal -= rb_to_skb(node)->truesize; 5034 tcp_drop(sk, rb_to_skb(node)); 5035 if (!prev || goal <= 0) { 5036 sk_mem_reclaim(sk); 5037 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf && 5038 !tcp_under_memory_pressure(sk)) 5039 break; 5040 goal = sk->sk_rcvbuf >> 3; 5041 } 5042 node = prev; 5043 } while (node); 5044 tp->ooo_last_skb = rb_to_skb(prev); 5045 5046 /* Reset SACK state. A conforming SACK implementation will 5047 * do the same at a timeout based retransmit. When a connection 5048 * is in a sad state like this, we care only about integrity 5049 * of the connection not performance. 5050 */ 5051 if (tp->rx_opt.sack_ok) 5052 tcp_sack_reset(&tp->rx_opt); 5053 return true; 5054 } 5055 5056 /* Reduce allocated memory if we can, trying to get 5057 * the socket within its memory limits again. 5058 * 5059 * Return less than zero if we should start dropping frames 5060 * until the socket owning process reads some of the data 5061 * to stabilize the situation. 5062 */ 5063 static int tcp_prune_queue(struct sock *sk) 5064 { 5065 struct tcp_sock *tp = tcp_sk(sk); 5066 5067 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED); 5068 5069 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf) 5070 tcp_clamp_window(sk); 5071 else if (tcp_under_memory_pressure(sk)) 5072 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss); 5073 5074 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 5075 return 0; 5076 5077 tcp_collapse_ofo_queue(sk); 5078 if (!skb_queue_empty(&sk->sk_receive_queue)) 5079 tcp_collapse(sk, &sk->sk_receive_queue, NULL, 5080 skb_peek(&sk->sk_receive_queue), 5081 NULL, 5082 tp->copied_seq, tp->rcv_nxt); 5083 sk_mem_reclaim(sk); 5084 5085 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 5086 return 0; 5087 5088 /* Collapsing did not help, destructive actions follow. 5089 * This must not ever occur. */ 5090 5091 tcp_prune_ofo_queue(sk); 5092 5093 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 5094 return 0; 5095 5096 /* If we are really being abused, tell the caller to silently 5097 * drop receive data on the floor. It will get retransmitted 5098 * and hopefully then we'll have sufficient space. 5099 */ 5100 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED); 5101 5102 /* Massive buffer overcommit. */ 5103 tp->pred_flags = 0; 5104 return -1; 5105 } 5106 5107 static bool tcp_should_expand_sndbuf(const struct sock *sk) 5108 { 5109 const struct tcp_sock *tp = tcp_sk(sk); 5110 5111 /* If the user specified a specific send buffer setting, do 5112 * not modify it. 5113 */ 5114 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK) 5115 return false; 5116 5117 /* If we are under global TCP memory pressure, do not expand. */ 5118 if (tcp_under_memory_pressure(sk)) 5119 return false; 5120 5121 /* If we are under soft global TCP memory pressure, do not expand. */ 5122 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0)) 5123 return false; 5124 5125 /* If we filled the congestion window, do not expand. */ 5126 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 5127 return false; 5128 5129 return true; 5130 } 5131 5132 /* When incoming ACK allowed to free some skb from write_queue, 5133 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket 5134 * on the exit from tcp input handler. 5135 * 5136 * PROBLEM: sndbuf expansion does not work well with largesend. 5137 */ 5138 static void tcp_new_space(struct sock *sk) 5139 { 5140 struct tcp_sock *tp = tcp_sk(sk); 5141 5142 if (tcp_should_expand_sndbuf(sk)) { 5143 tcp_sndbuf_expand(sk); 5144 tp->snd_cwnd_stamp = tcp_jiffies32; 5145 } 5146 5147 sk->sk_write_space(sk); 5148 } 5149 5150 static void tcp_check_space(struct sock *sk) 5151 { 5152 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) { 5153 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK); 5154 /* pairs with tcp_poll() */ 5155 smp_mb(); 5156 if (sk->sk_socket && 5157 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 5158 tcp_new_space(sk); 5159 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) 5160 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED); 5161 } 5162 } 5163 } 5164 5165 static inline void tcp_data_snd_check(struct sock *sk) 5166 { 5167 tcp_push_pending_frames(sk); 5168 tcp_check_space(sk); 5169 } 5170 5171 /* 5172 * Check if sending an ack is needed. 5173 */ 5174 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible) 5175 { 5176 struct tcp_sock *tp = tcp_sk(sk); 5177 unsigned long rtt, delay; 5178 5179 /* More than one full frame received... */ 5180 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss && 5181 /* ... and right edge of window advances far enough. 5182 * (tcp_recvmsg() will send ACK otherwise). 5183 * If application uses SO_RCVLOWAT, we want send ack now if 5184 * we have not received enough bytes to satisfy the condition. 5185 */ 5186 (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat || 5187 __tcp_select_window(sk) >= tp->rcv_wnd)) || 5188 /* We ACK each frame or... */ 5189 tcp_in_quickack_mode(sk) || 5190 /* Protocol state mandates a one-time immediate ACK */ 5191 inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) { 5192 send_now: 5193 tcp_send_ack(sk); 5194 return; 5195 } 5196 5197 if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) { 5198 tcp_send_delayed_ack(sk); 5199 return; 5200 } 5201 5202 if (!tcp_is_sack(tp) || 5203 tp->compressed_ack >= sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr) 5204 goto send_now; 5205 5206 if (tp->compressed_ack_rcv_nxt != tp->rcv_nxt) { 5207 tp->compressed_ack_rcv_nxt = tp->rcv_nxt; 5208 if (tp->compressed_ack > TCP_FASTRETRANS_THRESH) 5209 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED, 5210 tp->compressed_ack - TCP_FASTRETRANS_THRESH); 5211 tp->compressed_ack = 0; 5212 } 5213 5214 if (++tp->compressed_ack <= TCP_FASTRETRANS_THRESH) 5215 goto send_now; 5216 5217 if (hrtimer_is_queued(&tp->compressed_ack_timer)) 5218 return; 5219 5220 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */ 5221 5222 rtt = tp->rcv_rtt_est.rtt_us; 5223 if (tp->srtt_us && tp->srtt_us < rtt) 5224 rtt = tp->srtt_us; 5225 5226 delay = min_t(unsigned long, sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns, 5227 rtt * (NSEC_PER_USEC >> 3)/20); 5228 sock_hold(sk); 5229 hrtimer_start(&tp->compressed_ack_timer, ns_to_ktime(delay), 5230 HRTIMER_MODE_REL_PINNED_SOFT); 5231 } 5232 5233 static inline void tcp_ack_snd_check(struct sock *sk) 5234 { 5235 if (!inet_csk_ack_scheduled(sk)) { 5236 /* We sent a data segment already. */ 5237 return; 5238 } 5239 __tcp_ack_snd_check(sk, 1); 5240 } 5241 5242 /* 5243 * This routine is only called when we have urgent data 5244 * signaled. Its the 'slow' part of tcp_urg. It could be 5245 * moved inline now as tcp_urg is only called from one 5246 * place. We handle URGent data wrong. We have to - as 5247 * BSD still doesn't use the correction from RFC961. 5248 * For 1003.1g we should support a new option TCP_STDURG to permit 5249 * either form (or just set the sysctl tcp_stdurg). 5250 */ 5251 5252 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th) 5253 { 5254 struct tcp_sock *tp = tcp_sk(sk); 5255 u32 ptr = ntohs(th->urg_ptr); 5256 5257 if (ptr && !sock_net(sk)->ipv4.sysctl_tcp_stdurg) 5258 ptr--; 5259 ptr += ntohl(th->seq); 5260 5261 /* Ignore urgent data that we've already seen and read. */ 5262 if (after(tp->copied_seq, ptr)) 5263 return; 5264 5265 /* Do not replay urg ptr. 5266 * 5267 * NOTE: interesting situation not covered by specs. 5268 * Misbehaving sender may send urg ptr, pointing to segment, 5269 * which we already have in ofo queue. We are not able to fetch 5270 * such data and will stay in TCP_URG_NOTYET until will be eaten 5271 * by recvmsg(). Seems, we are not obliged to handle such wicked 5272 * situations. But it is worth to think about possibility of some 5273 * DoSes using some hypothetical application level deadlock. 5274 */ 5275 if (before(ptr, tp->rcv_nxt)) 5276 return; 5277 5278 /* Do we already have a newer (or duplicate) urgent pointer? */ 5279 if (tp->urg_data && !after(ptr, tp->urg_seq)) 5280 return; 5281 5282 /* Tell the world about our new urgent pointer. */ 5283 sk_send_sigurg(sk); 5284 5285 /* We may be adding urgent data when the last byte read was 5286 * urgent. To do this requires some care. We cannot just ignore 5287 * tp->copied_seq since we would read the last urgent byte again 5288 * as data, nor can we alter copied_seq until this data arrives 5289 * or we break the semantics of SIOCATMARK (and thus sockatmark()) 5290 * 5291 * NOTE. Double Dutch. Rendering to plain English: author of comment 5292 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB); 5293 * and expect that both A and B disappear from stream. This is _wrong_. 5294 * Though this happens in BSD with high probability, this is occasional. 5295 * Any application relying on this is buggy. Note also, that fix "works" 5296 * only in this artificial test. Insert some normal data between A and B and we will 5297 * decline of BSD again. Verdict: it is better to remove to trap 5298 * buggy users. 5299 */ 5300 if (tp->urg_seq == tp->copied_seq && tp->urg_data && 5301 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) { 5302 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 5303 tp->copied_seq++; 5304 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) { 5305 __skb_unlink(skb, &sk->sk_receive_queue); 5306 __kfree_skb(skb); 5307 } 5308 } 5309 5310 tp->urg_data = TCP_URG_NOTYET; 5311 tp->urg_seq = ptr; 5312 5313 /* Disable header prediction. */ 5314 tp->pred_flags = 0; 5315 } 5316 5317 /* This is the 'fast' part of urgent handling. */ 5318 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th) 5319 { 5320 struct tcp_sock *tp = tcp_sk(sk); 5321 5322 /* Check if we get a new urgent pointer - normally not. */ 5323 if (th->urg) 5324 tcp_check_urg(sk, th); 5325 5326 /* Do we wait for any urgent data? - normally not... */ 5327 if (tp->urg_data == TCP_URG_NOTYET) { 5328 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) - 5329 th->syn; 5330 5331 /* Is the urgent pointer pointing into this packet? */ 5332 if (ptr < skb->len) { 5333 u8 tmp; 5334 if (skb_copy_bits(skb, ptr, &tmp, 1)) 5335 BUG(); 5336 tp->urg_data = TCP_URG_VALID | tmp; 5337 if (!sock_flag(sk, SOCK_DEAD)) 5338 sk->sk_data_ready(sk); 5339 } 5340 } 5341 } 5342 5343 /* Accept RST for rcv_nxt - 1 after a FIN. 5344 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a 5345 * FIN is sent followed by a RST packet. The RST is sent with the same 5346 * sequence number as the FIN, and thus according to RFC 5961 a challenge 5347 * ACK should be sent. However, Mac OSX rate limits replies to challenge 5348 * ACKs on the closed socket. In addition middleboxes can drop either the 5349 * challenge ACK or a subsequent RST. 5350 */ 5351 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb) 5352 { 5353 struct tcp_sock *tp = tcp_sk(sk); 5354 5355 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) && 5356 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK | 5357 TCPF_CLOSING)); 5358 } 5359 5360 /* Does PAWS and seqno based validation of an incoming segment, flags will 5361 * play significant role here. 5362 */ 5363 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb, 5364 const struct tcphdr *th, int syn_inerr) 5365 { 5366 struct tcp_sock *tp = tcp_sk(sk); 5367 bool rst_seq_match = false; 5368 5369 /* RFC1323: H1. Apply PAWS check first. */ 5370 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) && 5371 tp->rx_opt.saw_tstamp && 5372 tcp_paws_discard(sk, skb)) { 5373 if (!th->rst) { 5374 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED); 5375 if (!tcp_oow_rate_limited(sock_net(sk), skb, 5376 LINUX_MIB_TCPACKSKIPPEDPAWS, 5377 &tp->last_oow_ack_time)) 5378 tcp_send_dupack(sk, skb); 5379 goto discard; 5380 } 5381 /* Reset is accepted even if it did not pass PAWS. */ 5382 } 5383 5384 /* Step 1: check sequence number */ 5385 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { 5386 /* RFC793, page 37: "In all states except SYN-SENT, all reset 5387 * (RST) segments are validated by checking their SEQ-fields." 5388 * And page 69: "If an incoming segment is not acceptable, 5389 * an acknowledgment should be sent in reply (unless the RST 5390 * bit is set, if so drop the segment and return)". 5391 */ 5392 if (!th->rst) { 5393 if (th->syn) 5394 goto syn_challenge; 5395 if (!tcp_oow_rate_limited(sock_net(sk), skb, 5396 LINUX_MIB_TCPACKSKIPPEDSEQ, 5397 &tp->last_oow_ack_time)) 5398 tcp_send_dupack(sk, skb); 5399 } else if (tcp_reset_check(sk, skb)) { 5400 tcp_reset(sk); 5401 } 5402 goto discard; 5403 } 5404 5405 /* Step 2: check RST bit */ 5406 if (th->rst) { 5407 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a 5408 * FIN and SACK too if available): 5409 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or 5410 * the right-most SACK block, 5411 * then 5412 * RESET the connection 5413 * else 5414 * Send a challenge ACK 5415 */ 5416 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt || 5417 tcp_reset_check(sk, skb)) { 5418 rst_seq_match = true; 5419 } else if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) { 5420 struct tcp_sack_block *sp = &tp->selective_acks[0]; 5421 int max_sack = sp[0].end_seq; 5422 int this_sack; 5423 5424 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; 5425 ++this_sack) { 5426 max_sack = after(sp[this_sack].end_seq, 5427 max_sack) ? 5428 sp[this_sack].end_seq : max_sack; 5429 } 5430 5431 if (TCP_SKB_CB(skb)->seq == max_sack) 5432 rst_seq_match = true; 5433 } 5434 5435 if (rst_seq_match) 5436 tcp_reset(sk); 5437 else { 5438 /* Disable TFO if RST is out-of-order 5439 * and no data has been received 5440 * for current active TFO socket 5441 */ 5442 if (tp->syn_fastopen && !tp->data_segs_in && 5443 sk->sk_state == TCP_ESTABLISHED) 5444 tcp_fastopen_active_disable(sk); 5445 tcp_send_challenge_ack(sk, skb); 5446 } 5447 goto discard; 5448 } 5449 5450 /* step 3: check security and precedence [ignored] */ 5451 5452 /* step 4: Check for a SYN 5453 * RFC 5961 4.2 : Send a challenge ack 5454 */ 5455 if (th->syn) { 5456 syn_challenge: 5457 if (syn_inerr) 5458 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS); 5459 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE); 5460 tcp_send_challenge_ack(sk, skb); 5461 goto discard; 5462 } 5463 5464 return true; 5465 5466 discard: 5467 tcp_drop(sk, skb); 5468 return false; 5469 } 5470 5471 /* 5472 * TCP receive function for the ESTABLISHED state. 5473 * 5474 * It is split into a fast path and a slow path. The fast path is 5475 * disabled when: 5476 * - A zero window was announced from us - zero window probing 5477 * is only handled properly in the slow path. 5478 * - Out of order segments arrived. 5479 * - Urgent data is expected. 5480 * - There is no buffer space left 5481 * - Unexpected TCP flags/window values/header lengths are received 5482 * (detected by checking the TCP header against pred_flags) 5483 * - Data is sent in both directions. Fast path only supports pure senders 5484 * or pure receivers (this means either the sequence number or the ack 5485 * value must stay constant) 5486 * - Unexpected TCP option. 5487 * 5488 * When these conditions are not satisfied it drops into a standard 5489 * receive procedure patterned after RFC793 to handle all cases. 5490 * The first three cases are guaranteed by proper pred_flags setting, 5491 * the rest is checked inline. Fast processing is turned on in 5492 * tcp_data_queue when everything is OK. 5493 */ 5494 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb) 5495 { 5496 const struct tcphdr *th = (const struct tcphdr *)skb->data; 5497 struct tcp_sock *tp = tcp_sk(sk); 5498 unsigned int len = skb->len; 5499 5500 /* TCP congestion window tracking */ 5501 trace_tcp_probe(sk, skb); 5502 5503 tcp_mstamp_refresh(tp); 5504 if (unlikely(!sk->sk_rx_dst)) 5505 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb); 5506 /* 5507 * Header prediction. 5508 * The code loosely follows the one in the famous 5509 * "30 instruction TCP receive" Van Jacobson mail. 5510 * 5511 * Van's trick is to deposit buffers into socket queue 5512 * on a device interrupt, to call tcp_recv function 5513 * on the receive process context and checksum and copy 5514 * the buffer to user space. smart... 5515 * 5516 * Our current scheme is not silly either but we take the 5517 * extra cost of the net_bh soft interrupt processing... 5518 * We do checksum and copy also but from device to kernel. 5519 */ 5520 5521 tp->rx_opt.saw_tstamp = 0; 5522 5523 /* pred_flags is 0xS?10 << 16 + snd_wnd 5524 * if header_prediction is to be made 5525 * 'S' will always be tp->tcp_header_len >> 2 5526 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to 5527 * turn it off (when there are holes in the receive 5528 * space for instance) 5529 * PSH flag is ignored. 5530 */ 5531 5532 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags && 5533 TCP_SKB_CB(skb)->seq == tp->rcv_nxt && 5534 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) { 5535 int tcp_header_len = tp->tcp_header_len; 5536 5537 /* Timestamp header prediction: tcp_header_len 5538 * is automatically equal to th->doff*4 due to pred_flags 5539 * match. 5540 */ 5541 5542 /* Check timestamp */ 5543 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) { 5544 /* No? Slow path! */ 5545 if (!tcp_parse_aligned_timestamp(tp, th)) 5546 goto slow_path; 5547 5548 /* If PAWS failed, check it more carefully in slow path */ 5549 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0) 5550 goto slow_path; 5551 5552 /* DO NOT update ts_recent here, if checksum fails 5553 * and timestamp was corrupted part, it will result 5554 * in a hung connection since we will drop all 5555 * future packets due to the PAWS test. 5556 */ 5557 } 5558 5559 if (len <= tcp_header_len) { 5560 /* Bulk data transfer: sender */ 5561 if (len == tcp_header_len) { 5562 /* Predicted packet is in window by definition. 5563 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 5564 * Hence, check seq<=rcv_wup reduces to: 5565 */ 5566 if (tcp_header_len == 5567 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 5568 tp->rcv_nxt == tp->rcv_wup) 5569 tcp_store_ts_recent(tp); 5570 5571 /* We know that such packets are checksummed 5572 * on entry. 5573 */ 5574 tcp_ack(sk, skb, 0); 5575 __kfree_skb(skb); 5576 tcp_data_snd_check(sk); 5577 /* When receiving pure ack in fast path, update 5578 * last ts ecr directly instead of calling 5579 * tcp_rcv_rtt_measure_ts() 5580 */ 5581 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr; 5582 return; 5583 } else { /* Header too small */ 5584 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS); 5585 goto discard; 5586 } 5587 } else { 5588 int eaten = 0; 5589 bool fragstolen = false; 5590 5591 if (tcp_checksum_complete(skb)) 5592 goto csum_error; 5593 5594 if ((int)skb->truesize > sk->sk_forward_alloc) 5595 goto step5; 5596 5597 /* Predicted packet is in window by definition. 5598 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 5599 * Hence, check seq<=rcv_wup reduces to: 5600 */ 5601 if (tcp_header_len == 5602 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 5603 tp->rcv_nxt == tp->rcv_wup) 5604 tcp_store_ts_recent(tp); 5605 5606 tcp_rcv_rtt_measure_ts(sk, skb); 5607 5608 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS); 5609 5610 /* Bulk data transfer: receiver */ 5611 __skb_pull(skb, tcp_header_len); 5612 eaten = tcp_queue_rcv(sk, skb, &fragstolen); 5613 5614 tcp_event_data_recv(sk, skb); 5615 5616 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) { 5617 /* Well, only one small jumplet in fast path... */ 5618 tcp_ack(sk, skb, FLAG_DATA); 5619 tcp_data_snd_check(sk); 5620 if (!inet_csk_ack_scheduled(sk)) 5621 goto no_ack; 5622 } 5623 5624 __tcp_ack_snd_check(sk, 0); 5625 no_ack: 5626 if (eaten) 5627 kfree_skb_partial(skb, fragstolen); 5628 tcp_data_ready(sk); 5629 return; 5630 } 5631 } 5632 5633 slow_path: 5634 if (len < (th->doff << 2) || tcp_checksum_complete(skb)) 5635 goto csum_error; 5636 5637 if (!th->ack && !th->rst && !th->syn) 5638 goto discard; 5639 5640 /* 5641 * Standard slow path. 5642 */ 5643 5644 if (!tcp_validate_incoming(sk, skb, th, 1)) 5645 return; 5646 5647 step5: 5648 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0) 5649 goto discard; 5650 5651 tcp_rcv_rtt_measure_ts(sk, skb); 5652 5653 /* Process urgent data. */ 5654 tcp_urg(sk, skb, th); 5655 5656 /* step 7: process the segment text */ 5657 tcp_data_queue(sk, skb); 5658 5659 tcp_data_snd_check(sk); 5660 tcp_ack_snd_check(sk); 5661 return; 5662 5663 csum_error: 5664 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS); 5665 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS); 5666 5667 discard: 5668 tcp_drop(sk, skb); 5669 } 5670 EXPORT_SYMBOL(tcp_rcv_established); 5671 5672 void tcp_init_transfer(struct sock *sk, int bpf_op) 5673 { 5674 struct inet_connection_sock *icsk = inet_csk(sk); 5675 struct tcp_sock *tp = tcp_sk(sk); 5676 5677 tcp_mtup_init(sk); 5678 icsk->icsk_af_ops->rebuild_header(sk); 5679 tcp_init_metrics(sk); 5680 5681 /* Initialize the congestion window to start the transfer. 5682 * Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been 5683 * retransmitted. In light of RFC6298 more aggressive 1sec 5684 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK 5685 * retransmission has occurred. 5686 */ 5687 if (tp->total_retrans > 1 && tp->undo_marker) 5688 tp->snd_cwnd = 1; 5689 else 5690 tp->snd_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk)); 5691 tp->snd_cwnd_stamp = tcp_jiffies32; 5692 5693 tcp_call_bpf(sk, bpf_op, 0, NULL); 5694 tcp_init_congestion_control(sk); 5695 tcp_init_buffer_space(sk); 5696 } 5697 5698 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb) 5699 { 5700 struct tcp_sock *tp = tcp_sk(sk); 5701 struct inet_connection_sock *icsk = inet_csk(sk); 5702 5703 tcp_set_state(sk, TCP_ESTABLISHED); 5704 icsk->icsk_ack.lrcvtime = tcp_jiffies32; 5705 5706 if (skb) { 5707 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb); 5708 security_inet_conn_established(sk, skb); 5709 sk_mark_napi_id(sk, skb); 5710 } 5711 5712 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB); 5713 5714 /* Prevent spurious tcp_cwnd_restart() on first data 5715 * packet. 5716 */ 5717 tp->lsndtime = tcp_jiffies32; 5718 5719 if (sock_flag(sk, SOCK_KEEPOPEN)) 5720 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp)); 5721 5722 if (!tp->rx_opt.snd_wscale) 5723 __tcp_fast_path_on(tp, tp->snd_wnd); 5724 else 5725 tp->pred_flags = 0; 5726 } 5727 5728 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack, 5729 struct tcp_fastopen_cookie *cookie) 5730 { 5731 struct tcp_sock *tp = tcp_sk(sk); 5732 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL; 5733 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0; 5734 bool syn_drop = false; 5735 5736 if (mss == tp->rx_opt.user_mss) { 5737 struct tcp_options_received opt; 5738 5739 /* Get original SYNACK MSS value if user MSS sets mss_clamp */ 5740 tcp_clear_options(&opt); 5741 opt.user_mss = opt.mss_clamp = 0; 5742 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL); 5743 mss = opt.mss_clamp; 5744 } 5745 5746 if (!tp->syn_fastopen) { 5747 /* Ignore an unsolicited cookie */ 5748 cookie->len = -1; 5749 } else if (tp->total_retrans) { 5750 /* SYN timed out and the SYN-ACK neither has a cookie nor 5751 * acknowledges data. Presumably the remote received only 5752 * the retransmitted (regular) SYNs: either the original 5753 * SYN-data or the corresponding SYN-ACK was dropped. 5754 */ 5755 syn_drop = (cookie->len < 0 && data); 5756 } else if (cookie->len < 0 && !tp->syn_data) { 5757 /* We requested a cookie but didn't get it. If we did not use 5758 * the (old) exp opt format then try so next time (try_exp=1). 5759 * Otherwise we go back to use the RFC7413 opt (try_exp=2). 5760 */ 5761 try_exp = tp->syn_fastopen_exp ? 2 : 1; 5762 } 5763 5764 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp); 5765 5766 if (data) { /* Retransmit unacked data in SYN */ 5767 skb_rbtree_walk_from(data) { 5768 if (__tcp_retransmit_skb(sk, data, 1)) 5769 break; 5770 } 5771 tcp_rearm_rto(sk); 5772 NET_INC_STATS(sock_net(sk), 5773 LINUX_MIB_TCPFASTOPENACTIVEFAIL); 5774 return true; 5775 } 5776 tp->syn_data_acked = tp->syn_data; 5777 if (tp->syn_data_acked) { 5778 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE); 5779 /* SYN-data is counted as two separate packets in tcp_ack() */ 5780 if (tp->delivered > 1) 5781 --tp->delivered; 5782 } 5783 5784 tcp_fastopen_add_skb(sk, synack); 5785 5786 return false; 5787 } 5788 5789 static void smc_check_reset_syn(struct tcp_sock *tp) 5790 { 5791 #if IS_ENABLED(CONFIG_SMC) 5792 if (static_branch_unlikely(&tcp_have_smc)) { 5793 if (tp->syn_smc && !tp->rx_opt.smc_ok) 5794 tp->syn_smc = 0; 5795 } 5796 #endif 5797 } 5798 5799 static void tcp_try_undo_spurious_syn(struct sock *sk) 5800 { 5801 struct tcp_sock *tp = tcp_sk(sk); 5802 u32 syn_stamp; 5803 5804 /* undo_marker is set when SYN or SYNACK times out. The timeout is 5805 * spurious if the ACK's timestamp option echo value matches the 5806 * original SYN timestamp. 5807 */ 5808 syn_stamp = tp->retrans_stamp; 5809 if (tp->undo_marker && syn_stamp && tp->rx_opt.saw_tstamp && 5810 syn_stamp == tp->rx_opt.rcv_tsecr) 5811 tp->undo_marker = 0; 5812 } 5813 5814 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb, 5815 const struct tcphdr *th) 5816 { 5817 struct inet_connection_sock *icsk = inet_csk(sk); 5818 struct tcp_sock *tp = tcp_sk(sk); 5819 struct tcp_fastopen_cookie foc = { .len = -1 }; 5820 int saved_clamp = tp->rx_opt.mss_clamp; 5821 bool fastopen_fail; 5822 5823 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc); 5824 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) 5825 tp->rx_opt.rcv_tsecr -= tp->tsoffset; 5826 5827 if (th->ack) { 5828 /* rfc793: 5829 * "If the state is SYN-SENT then 5830 * first check the ACK bit 5831 * If the ACK bit is set 5832 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send 5833 * a reset (unless the RST bit is set, if so drop 5834 * the segment and return)" 5835 */ 5836 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) || 5837 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) 5838 goto reset_and_undo; 5839 5840 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 5841 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp, 5842 tcp_time_stamp(tp))) { 5843 NET_INC_STATS(sock_net(sk), 5844 LINUX_MIB_PAWSACTIVEREJECTED); 5845 goto reset_and_undo; 5846 } 5847 5848 /* Now ACK is acceptable. 5849 * 5850 * "If the RST bit is set 5851 * If the ACK was acceptable then signal the user "error: 5852 * connection reset", drop the segment, enter CLOSED state, 5853 * delete TCB, and return." 5854 */ 5855 5856 if (th->rst) { 5857 tcp_reset(sk); 5858 goto discard; 5859 } 5860 5861 /* rfc793: 5862 * "fifth, if neither of the SYN or RST bits is set then 5863 * drop the segment and return." 5864 * 5865 * See note below! 5866 * --ANK(990513) 5867 */ 5868 if (!th->syn) 5869 goto discard_and_undo; 5870 5871 /* rfc793: 5872 * "If the SYN bit is on ... 5873 * are acceptable then ... 5874 * (our SYN has been ACKed), change the connection 5875 * state to ESTABLISHED..." 5876 */ 5877 5878 tcp_ecn_rcv_synack(tp, th); 5879 5880 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); 5881 tcp_try_undo_spurious_syn(sk); 5882 tcp_ack(sk, skb, FLAG_SLOWPATH); 5883 5884 /* Ok.. it's good. Set up sequence numbers and 5885 * move to established. 5886 */ 5887 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 5888 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 5889 5890 /* RFC1323: The window in SYN & SYN/ACK segments is 5891 * never scaled. 5892 */ 5893 tp->snd_wnd = ntohs(th->window); 5894 5895 if (!tp->rx_opt.wscale_ok) { 5896 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0; 5897 tp->window_clamp = min(tp->window_clamp, 65535U); 5898 } 5899 5900 if (tp->rx_opt.saw_tstamp) { 5901 tp->rx_opt.tstamp_ok = 1; 5902 tp->tcp_header_len = 5903 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 5904 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 5905 tcp_store_ts_recent(tp); 5906 } else { 5907 tp->tcp_header_len = sizeof(struct tcphdr); 5908 } 5909 5910 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 5911 tcp_initialize_rcv_mss(sk); 5912 5913 /* Remember, tcp_poll() does not lock socket! 5914 * Change state from SYN-SENT only after copied_seq 5915 * is initialized. */ 5916 tp->copied_seq = tp->rcv_nxt; 5917 5918 smc_check_reset_syn(tp); 5919 5920 smp_mb(); 5921 5922 tcp_finish_connect(sk, skb); 5923 5924 fastopen_fail = (tp->syn_fastopen || tp->syn_data) && 5925 tcp_rcv_fastopen_synack(sk, skb, &foc); 5926 5927 if (!sock_flag(sk, SOCK_DEAD)) { 5928 sk->sk_state_change(sk); 5929 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT); 5930 } 5931 if (fastopen_fail) 5932 return -1; 5933 if (sk->sk_write_pending || 5934 icsk->icsk_accept_queue.rskq_defer_accept || 5935 inet_csk_in_pingpong_mode(sk)) { 5936 /* Save one ACK. Data will be ready after 5937 * several ticks, if write_pending is set. 5938 * 5939 * It may be deleted, but with this feature tcpdumps 5940 * look so _wonderfully_ clever, that I was not able 5941 * to stand against the temptation 8) --ANK 5942 */ 5943 inet_csk_schedule_ack(sk); 5944 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS); 5945 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 5946 TCP_DELACK_MAX, TCP_RTO_MAX); 5947 5948 discard: 5949 tcp_drop(sk, skb); 5950 return 0; 5951 } else { 5952 tcp_send_ack(sk); 5953 } 5954 return -1; 5955 } 5956 5957 /* No ACK in the segment */ 5958 5959 if (th->rst) { 5960 /* rfc793: 5961 * "If the RST bit is set 5962 * 5963 * Otherwise (no ACK) drop the segment and return." 5964 */ 5965 5966 goto discard_and_undo; 5967 } 5968 5969 /* PAWS check. */ 5970 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && 5971 tcp_paws_reject(&tp->rx_opt, 0)) 5972 goto discard_and_undo; 5973 5974 if (th->syn) { 5975 /* We see SYN without ACK. It is attempt of 5976 * simultaneous connect with crossed SYNs. 5977 * Particularly, it can be connect to self. 5978 */ 5979 tcp_set_state(sk, TCP_SYN_RECV); 5980 5981 if (tp->rx_opt.saw_tstamp) { 5982 tp->rx_opt.tstamp_ok = 1; 5983 tcp_store_ts_recent(tp); 5984 tp->tcp_header_len = 5985 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 5986 } else { 5987 tp->tcp_header_len = sizeof(struct tcphdr); 5988 } 5989 5990 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 5991 tp->copied_seq = tp->rcv_nxt; 5992 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 5993 5994 /* RFC1323: The window in SYN & SYN/ACK segments is 5995 * never scaled. 5996 */ 5997 tp->snd_wnd = ntohs(th->window); 5998 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 5999 tp->max_window = tp->snd_wnd; 6000 6001 tcp_ecn_rcv_syn(tp, th); 6002 6003 tcp_mtup_init(sk); 6004 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 6005 tcp_initialize_rcv_mss(sk); 6006 6007 tcp_send_synack(sk); 6008 #if 0 6009 /* Note, we could accept data and URG from this segment. 6010 * There are no obstacles to make this (except that we must 6011 * either change tcp_recvmsg() to prevent it from returning data 6012 * before 3WHS completes per RFC793, or employ TCP Fast Open). 6013 * 6014 * However, if we ignore data in ACKless segments sometimes, 6015 * we have no reasons to accept it sometimes. 6016 * Also, seems the code doing it in step6 of tcp_rcv_state_process 6017 * is not flawless. So, discard packet for sanity. 6018 * Uncomment this return to process the data. 6019 */ 6020 return -1; 6021 #else 6022 goto discard; 6023 #endif 6024 } 6025 /* "fifth, if neither of the SYN or RST bits is set then 6026 * drop the segment and return." 6027 */ 6028 6029 discard_and_undo: 6030 tcp_clear_options(&tp->rx_opt); 6031 tp->rx_opt.mss_clamp = saved_clamp; 6032 goto discard; 6033 6034 reset_and_undo: 6035 tcp_clear_options(&tp->rx_opt); 6036 tp->rx_opt.mss_clamp = saved_clamp; 6037 return 1; 6038 } 6039 6040 static void tcp_rcv_synrecv_state_fastopen(struct sock *sk) 6041 { 6042 tcp_try_undo_loss(sk, false); 6043 6044 /* Reset rtx states to prevent spurious retransmits_timed_out() */ 6045 tcp_sk(sk)->retrans_stamp = 0; 6046 inet_csk(sk)->icsk_retransmits = 0; 6047 6048 /* Once we leave TCP_SYN_RECV or TCP_FIN_WAIT_1, 6049 * we no longer need req so release it. 6050 */ 6051 reqsk_fastopen_remove(sk, tcp_sk(sk)->fastopen_rsk, false); 6052 6053 /* Re-arm the timer because data may have been sent out. 6054 * This is similar to the regular data transmission case 6055 * when new data has just been ack'ed. 6056 * 6057 * (TFO) - we could try to be more aggressive and 6058 * retransmitting any data sooner based on when they 6059 * are sent out. 6060 */ 6061 tcp_rearm_rto(sk); 6062 } 6063 6064 /* 6065 * This function implements the receiving procedure of RFC 793 for 6066 * all states except ESTABLISHED and TIME_WAIT. 6067 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be 6068 * address independent. 6069 */ 6070 6071 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb) 6072 { 6073 struct tcp_sock *tp = tcp_sk(sk); 6074 struct inet_connection_sock *icsk = inet_csk(sk); 6075 const struct tcphdr *th = tcp_hdr(skb); 6076 struct request_sock *req; 6077 int queued = 0; 6078 bool acceptable; 6079 6080 switch (sk->sk_state) { 6081 case TCP_CLOSE: 6082 goto discard; 6083 6084 case TCP_LISTEN: 6085 if (th->ack) 6086 return 1; 6087 6088 if (th->rst) 6089 goto discard; 6090 6091 if (th->syn) { 6092 if (th->fin) 6093 goto discard; 6094 /* It is possible that we process SYN packets from backlog, 6095 * so we need to make sure to disable BH and RCU right there. 6096 */ 6097 rcu_read_lock(); 6098 local_bh_disable(); 6099 acceptable = icsk->icsk_af_ops->conn_request(sk, skb) >= 0; 6100 local_bh_enable(); 6101 rcu_read_unlock(); 6102 6103 if (!acceptable) 6104 return 1; 6105 consume_skb(skb); 6106 return 0; 6107 } 6108 goto discard; 6109 6110 case TCP_SYN_SENT: 6111 tp->rx_opt.saw_tstamp = 0; 6112 tcp_mstamp_refresh(tp); 6113 queued = tcp_rcv_synsent_state_process(sk, skb, th); 6114 if (queued >= 0) 6115 return queued; 6116 6117 /* Do step6 onward by hand. */ 6118 tcp_urg(sk, skb, th); 6119 __kfree_skb(skb); 6120 tcp_data_snd_check(sk); 6121 return 0; 6122 } 6123 6124 tcp_mstamp_refresh(tp); 6125 tp->rx_opt.saw_tstamp = 0; 6126 req = tp->fastopen_rsk; 6127 if (req) { 6128 bool req_stolen; 6129 6130 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV && 6131 sk->sk_state != TCP_FIN_WAIT1); 6132 6133 if (!tcp_check_req(sk, skb, req, true, &req_stolen)) 6134 goto discard; 6135 } 6136 6137 if (!th->ack && !th->rst && !th->syn) 6138 goto discard; 6139 6140 if (!tcp_validate_incoming(sk, skb, th, 0)) 6141 return 0; 6142 6143 /* step 5: check the ACK field */ 6144 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH | 6145 FLAG_UPDATE_TS_RECENT | 6146 FLAG_NO_CHALLENGE_ACK) > 0; 6147 6148 if (!acceptable) { 6149 if (sk->sk_state == TCP_SYN_RECV) 6150 return 1; /* send one RST */ 6151 tcp_send_challenge_ack(sk, skb); 6152 goto discard; 6153 } 6154 switch (sk->sk_state) { 6155 case TCP_SYN_RECV: 6156 tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */ 6157 if (!tp->srtt_us) 6158 tcp_synack_rtt_meas(sk, req); 6159 6160 if (req) { 6161 tcp_rcv_synrecv_state_fastopen(sk); 6162 } else { 6163 tcp_try_undo_spurious_syn(sk); 6164 tp->retrans_stamp = 0; 6165 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB); 6166 tp->copied_seq = tp->rcv_nxt; 6167 } 6168 smp_mb(); 6169 tcp_set_state(sk, TCP_ESTABLISHED); 6170 sk->sk_state_change(sk); 6171 6172 /* Note, that this wakeup is only for marginal crossed SYN case. 6173 * Passively open sockets are not waked up, because 6174 * sk->sk_sleep == NULL and sk->sk_socket == NULL. 6175 */ 6176 if (sk->sk_socket) 6177 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT); 6178 6179 tp->snd_una = TCP_SKB_CB(skb)->ack_seq; 6180 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale; 6181 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); 6182 6183 if (tp->rx_opt.tstamp_ok) 6184 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 6185 6186 if (!inet_csk(sk)->icsk_ca_ops->cong_control) 6187 tcp_update_pacing_rate(sk); 6188 6189 /* Prevent spurious tcp_cwnd_restart() on first data packet */ 6190 tp->lsndtime = tcp_jiffies32; 6191 6192 tcp_initialize_rcv_mss(sk); 6193 tcp_fast_path_on(tp); 6194 break; 6195 6196 case TCP_FIN_WAIT1: { 6197 int tmo; 6198 6199 if (req) 6200 tcp_rcv_synrecv_state_fastopen(sk); 6201 6202 if (tp->snd_una != tp->write_seq) 6203 break; 6204 6205 tcp_set_state(sk, TCP_FIN_WAIT2); 6206 sk->sk_shutdown |= SEND_SHUTDOWN; 6207 6208 sk_dst_confirm(sk); 6209 6210 if (!sock_flag(sk, SOCK_DEAD)) { 6211 /* Wake up lingering close() */ 6212 sk->sk_state_change(sk); 6213 break; 6214 } 6215 6216 if (tp->linger2 < 0) { 6217 tcp_done(sk); 6218 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 6219 return 1; 6220 } 6221 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 6222 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { 6223 /* Receive out of order FIN after close() */ 6224 if (tp->syn_fastopen && th->fin) 6225 tcp_fastopen_active_disable(sk); 6226 tcp_done(sk); 6227 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 6228 return 1; 6229 } 6230 6231 tmo = tcp_fin_time(sk); 6232 if (tmo > TCP_TIMEWAIT_LEN) { 6233 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN); 6234 } else if (th->fin || sock_owned_by_user(sk)) { 6235 /* Bad case. We could lose such FIN otherwise. 6236 * It is not a big problem, but it looks confusing 6237 * and not so rare event. We still can lose it now, 6238 * if it spins in bh_lock_sock(), but it is really 6239 * marginal case. 6240 */ 6241 inet_csk_reset_keepalive_timer(sk, tmo); 6242 } else { 6243 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 6244 goto discard; 6245 } 6246 break; 6247 } 6248 6249 case TCP_CLOSING: 6250 if (tp->snd_una == tp->write_seq) { 6251 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 6252 goto discard; 6253 } 6254 break; 6255 6256 case TCP_LAST_ACK: 6257 if (tp->snd_una == tp->write_seq) { 6258 tcp_update_metrics(sk); 6259 tcp_done(sk); 6260 goto discard; 6261 } 6262 break; 6263 } 6264 6265 /* step 6: check the URG bit */ 6266 tcp_urg(sk, skb, th); 6267 6268 /* step 7: process the segment text */ 6269 switch (sk->sk_state) { 6270 case TCP_CLOSE_WAIT: 6271 case TCP_CLOSING: 6272 case TCP_LAST_ACK: 6273 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 6274 break; 6275 /* fall through */ 6276 case TCP_FIN_WAIT1: 6277 case TCP_FIN_WAIT2: 6278 /* RFC 793 says to queue data in these states, 6279 * RFC 1122 says we MUST send a reset. 6280 * BSD 4.4 also does reset. 6281 */ 6282 if (sk->sk_shutdown & RCV_SHUTDOWN) { 6283 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 6284 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { 6285 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 6286 tcp_reset(sk); 6287 return 1; 6288 } 6289 } 6290 /* Fall through */ 6291 case TCP_ESTABLISHED: 6292 tcp_data_queue(sk, skb); 6293 queued = 1; 6294 break; 6295 } 6296 6297 /* tcp_data could move socket to TIME-WAIT */ 6298 if (sk->sk_state != TCP_CLOSE) { 6299 tcp_data_snd_check(sk); 6300 tcp_ack_snd_check(sk); 6301 } 6302 6303 if (!queued) { 6304 discard: 6305 tcp_drop(sk, skb); 6306 } 6307 return 0; 6308 } 6309 EXPORT_SYMBOL(tcp_rcv_state_process); 6310 6311 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family) 6312 { 6313 struct inet_request_sock *ireq = inet_rsk(req); 6314 6315 if (family == AF_INET) 6316 net_dbg_ratelimited("drop open request from %pI4/%u\n", 6317 &ireq->ir_rmt_addr, port); 6318 #if IS_ENABLED(CONFIG_IPV6) 6319 else if (family == AF_INET6) 6320 net_dbg_ratelimited("drop open request from %pI6/%u\n", 6321 &ireq->ir_v6_rmt_addr, port); 6322 #endif 6323 } 6324 6325 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set 6326 * 6327 * If we receive a SYN packet with these bits set, it means a 6328 * network is playing bad games with TOS bits. In order to 6329 * avoid possible false congestion notifications, we disable 6330 * TCP ECN negotiation. 6331 * 6332 * Exception: tcp_ca wants ECN. This is required for DCTCP 6333 * congestion control: Linux DCTCP asserts ECT on all packets, 6334 * including SYN, which is most optimal solution; however, 6335 * others, such as FreeBSD do not. 6336 * 6337 * Exception: At least one of the reserved bits of the TCP header (th->res1) is 6338 * set, indicating the use of a future TCP extension (such as AccECN). See 6339 * RFC8311 §4.3 which updates RFC3168 to allow the development of such 6340 * extensions. 6341 */ 6342 static void tcp_ecn_create_request(struct request_sock *req, 6343 const struct sk_buff *skb, 6344 const struct sock *listen_sk, 6345 const struct dst_entry *dst) 6346 { 6347 const struct tcphdr *th = tcp_hdr(skb); 6348 const struct net *net = sock_net(listen_sk); 6349 bool th_ecn = th->ece && th->cwr; 6350 bool ect, ecn_ok; 6351 u32 ecn_ok_dst; 6352 6353 if (!th_ecn) 6354 return; 6355 6356 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield); 6357 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK); 6358 ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst; 6359 6360 if (((!ect || th->res1) && ecn_ok) || tcp_ca_needs_ecn(listen_sk) || 6361 (ecn_ok_dst & DST_FEATURE_ECN_CA) || 6362 tcp_bpf_ca_needs_ecn((struct sock *)req)) 6363 inet_rsk(req)->ecn_ok = 1; 6364 } 6365 6366 static void tcp_openreq_init(struct request_sock *req, 6367 const struct tcp_options_received *rx_opt, 6368 struct sk_buff *skb, const struct sock *sk) 6369 { 6370 struct inet_request_sock *ireq = inet_rsk(req); 6371 6372 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */ 6373 req->cookie_ts = 0; 6374 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq; 6375 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 6376 tcp_rsk(req)->snt_synack = 0; 6377 tcp_rsk(req)->last_oow_ack_time = 0; 6378 req->mss = rx_opt->mss_clamp; 6379 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0; 6380 ireq->tstamp_ok = rx_opt->tstamp_ok; 6381 ireq->sack_ok = rx_opt->sack_ok; 6382 ireq->snd_wscale = rx_opt->snd_wscale; 6383 ireq->wscale_ok = rx_opt->wscale_ok; 6384 ireq->acked = 0; 6385 ireq->ecn_ok = 0; 6386 ireq->ir_rmt_port = tcp_hdr(skb)->source; 6387 ireq->ir_num = ntohs(tcp_hdr(skb)->dest); 6388 ireq->ir_mark = inet_request_mark(sk, skb); 6389 #if IS_ENABLED(CONFIG_SMC) 6390 ireq->smc_ok = rx_opt->smc_ok; 6391 #endif 6392 } 6393 6394 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops, 6395 struct sock *sk_listener, 6396 bool attach_listener) 6397 { 6398 struct request_sock *req = reqsk_alloc(ops, sk_listener, 6399 attach_listener); 6400 6401 if (req) { 6402 struct inet_request_sock *ireq = inet_rsk(req); 6403 6404 ireq->ireq_opt = NULL; 6405 #if IS_ENABLED(CONFIG_IPV6) 6406 ireq->pktopts = NULL; 6407 #endif 6408 atomic64_set(&ireq->ir_cookie, 0); 6409 ireq->ireq_state = TCP_NEW_SYN_RECV; 6410 write_pnet(&ireq->ireq_net, sock_net(sk_listener)); 6411 ireq->ireq_family = sk_listener->sk_family; 6412 } 6413 6414 return req; 6415 } 6416 EXPORT_SYMBOL(inet_reqsk_alloc); 6417 6418 /* 6419 * Return true if a syncookie should be sent 6420 */ 6421 static bool tcp_syn_flood_action(const struct sock *sk, 6422 const struct sk_buff *skb, 6423 const char *proto) 6424 { 6425 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue; 6426 const char *msg = "Dropping request"; 6427 bool want_cookie = false; 6428 struct net *net = sock_net(sk); 6429 6430 #ifdef CONFIG_SYN_COOKIES 6431 if (net->ipv4.sysctl_tcp_syncookies) { 6432 msg = "Sending cookies"; 6433 want_cookie = true; 6434 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES); 6435 } else 6436 #endif 6437 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP); 6438 6439 if (!queue->synflood_warned && 6440 net->ipv4.sysctl_tcp_syncookies != 2 && 6441 xchg(&queue->synflood_warned, 1) == 0) 6442 net_info_ratelimited("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n", 6443 proto, ntohs(tcp_hdr(skb)->dest), msg); 6444 6445 return want_cookie; 6446 } 6447 6448 static void tcp_reqsk_record_syn(const struct sock *sk, 6449 struct request_sock *req, 6450 const struct sk_buff *skb) 6451 { 6452 if (tcp_sk(sk)->save_syn) { 6453 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb); 6454 u32 *copy; 6455 6456 copy = kmalloc(len + sizeof(u32), GFP_ATOMIC); 6457 if (copy) { 6458 copy[0] = len; 6459 memcpy(©[1], skb_network_header(skb), len); 6460 req->saved_syn = copy; 6461 } 6462 } 6463 } 6464 6465 int tcp_conn_request(struct request_sock_ops *rsk_ops, 6466 const struct tcp_request_sock_ops *af_ops, 6467 struct sock *sk, struct sk_buff *skb) 6468 { 6469 struct tcp_fastopen_cookie foc = { .len = -1 }; 6470 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn; 6471 struct tcp_options_received tmp_opt; 6472 struct tcp_sock *tp = tcp_sk(sk); 6473 struct net *net = sock_net(sk); 6474 struct sock *fastopen_sk = NULL; 6475 struct request_sock *req; 6476 bool want_cookie = false; 6477 struct dst_entry *dst; 6478 struct flowi fl; 6479 6480 /* TW buckets are converted to open requests without 6481 * limitations, they conserve resources and peer is 6482 * evidently real one. 6483 */ 6484 if ((net->ipv4.sysctl_tcp_syncookies == 2 || 6485 inet_csk_reqsk_queue_is_full(sk)) && !isn) { 6486 want_cookie = tcp_syn_flood_action(sk, skb, rsk_ops->slab_name); 6487 if (!want_cookie) 6488 goto drop; 6489 } 6490 6491 if (sk_acceptq_is_full(sk)) { 6492 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS); 6493 goto drop; 6494 } 6495 6496 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie); 6497 if (!req) 6498 goto drop; 6499 6500 tcp_rsk(req)->af_specific = af_ops; 6501 tcp_rsk(req)->ts_off = 0; 6502 6503 tcp_clear_options(&tmp_opt); 6504 tmp_opt.mss_clamp = af_ops->mss_clamp; 6505 tmp_opt.user_mss = tp->rx_opt.user_mss; 6506 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0, 6507 want_cookie ? NULL : &foc); 6508 6509 if (want_cookie && !tmp_opt.saw_tstamp) 6510 tcp_clear_options(&tmp_opt); 6511 6512 if (IS_ENABLED(CONFIG_SMC) && want_cookie) 6513 tmp_opt.smc_ok = 0; 6514 6515 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp; 6516 tcp_openreq_init(req, &tmp_opt, skb, sk); 6517 inet_rsk(req)->no_srccheck = inet_sk(sk)->transparent; 6518 6519 /* Note: tcp_v6_init_req() might override ir_iif for link locals */ 6520 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb); 6521 6522 af_ops->init_req(req, sk, skb); 6523 6524 if (security_inet_conn_request(sk, skb, req)) 6525 goto drop_and_free; 6526 6527 if (tmp_opt.tstamp_ok) 6528 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb); 6529 6530 dst = af_ops->route_req(sk, &fl, req); 6531 if (!dst) 6532 goto drop_and_free; 6533 6534 if (!want_cookie && !isn) { 6535 /* Kill the following clause, if you dislike this way. */ 6536 if (!net->ipv4.sysctl_tcp_syncookies && 6537 (net->ipv4.sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) < 6538 (net->ipv4.sysctl_max_syn_backlog >> 2)) && 6539 !tcp_peer_is_proven(req, dst)) { 6540 /* Without syncookies last quarter of 6541 * backlog is filled with destinations, 6542 * proven to be alive. 6543 * It means that we continue to communicate 6544 * to destinations, already remembered 6545 * to the moment of synflood. 6546 */ 6547 pr_drop_req(req, ntohs(tcp_hdr(skb)->source), 6548 rsk_ops->family); 6549 goto drop_and_release; 6550 } 6551 6552 isn = af_ops->init_seq(skb); 6553 } 6554 6555 tcp_ecn_create_request(req, skb, sk, dst); 6556 6557 if (want_cookie) { 6558 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss); 6559 req->cookie_ts = tmp_opt.tstamp_ok; 6560 if (!tmp_opt.tstamp_ok) 6561 inet_rsk(req)->ecn_ok = 0; 6562 } 6563 6564 tcp_rsk(req)->snt_isn = isn; 6565 tcp_rsk(req)->txhash = net_tx_rndhash(); 6566 tcp_openreq_init_rwin(req, sk, dst); 6567 sk_rx_queue_set(req_to_sk(req), skb); 6568 if (!want_cookie) { 6569 tcp_reqsk_record_syn(sk, req, skb); 6570 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst); 6571 } 6572 if (fastopen_sk) { 6573 af_ops->send_synack(fastopen_sk, dst, &fl, req, 6574 &foc, TCP_SYNACK_FASTOPEN); 6575 /* Add the child socket directly into the accept queue */ 6576 if (!inet_csk_reqsk_queue_add(sk, req, fastopen_sk)) { 6577 reqsk_fastopen_remove(fastopen_sk, req, false); 6578 bh_unlock_sock(fastopen_sk); 6579 sock_put(fastopen_sk); 6580 goto drop_and_free; 6581 } 6582 sk->sk_data_ready(sk); 6583 bh_unlock_sock(fastopen_sk); 6584 sock_put(fastopen_sk); 6585 } else { 6586 tcp_rsk(req)->tfo_listener = false; 6587 if (!want_cookie) 6588 inet_csk_reqsk_queue_hash_add(sk, req, 6589 tcp_timeout_init((struct sock *)req)); 6590 af_ops->send_synack(sk, dst, &fl, req, &foc, 6591 !want_cookie ? TCP_SYNACK_NORMAL : 6592 TCP_SYNACK_COOKIE); 6593 if (want_cookie) { 6594 reqsk_free(req); 6595 return 0; 6596 } 6597 } 6598 reqsk_put(req); 6599 return 0; 6600 6601 drop_and_release: 6602 dst_release(dst); 6603 drop_and_free: 6604 __reqsk_free(req); 6605 drop: 6606 tcp_listendrop(sk); 6607 return 0; 6608 } 6609 EXPORT_SYMBOL(tcp_conn_request); 6610