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 #include <net/mptcp.h> 83 84 int sysctl_tcp_max_orphans __read_mostly = NR_FILE; 85 86 #define FLAG_DATA 0x01 /* Incoming frame contained data. */ 87 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */ 88 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */ 89 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */ 90 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */ 91 #define FLAG_DATA_SACKED 0x20 /* New SACK. */ 92 #define FLAG_ECE 0x40 /* ECE in this ACK */ 93 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */ 94 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/ 95 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */ 96 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */ 97 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */ 98 #define FLAG_SET_XMIT_TIMER 0x1000 /* Set TLP or RTO timer */ 99 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */ 100 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */ 101 #define FLAG_NO_CHALLENGE_ACK 0x8000 /* do not call tcp_send_challenge_ack() */ 102 #define FLAG_ACK_MAYBE_DELAYED 0x10000 /* Likely a delayed ACK */ 103 104 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) 105 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) 106 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK) 107 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) 108 109 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) 110 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH)) 111 112 #define REXMIT_NONE 0 /* no loss recovery to do */ 113 #define REXMIT_LOST 1 /* retransmit packets marked lost */ 114 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */ 115 116 #if IS_ENABLED(CONFIG_TLS_DEVICE) 117 static DEFINE_STATIC_KEY_DEFERRED_FALSE(clean_acked_data_enabled, HZ); 118 119 void clean_acked_data_enable(struct inet_connection_sock *icsk, 120 void (*cad)(struct sock *sk, u32 ack_seq)) 121 { 122 icsk->icsk_clean_acked = cad; 123 static_branch_deferred_inc(&clean_acked_data_enabled); 124 } 125 EXPORT_SYMBOL_GPL(clean_acked_data_enable); 126 127 void clean_acked_data_disable(struct inet_connection_sock *icsk) 128 { 129 static_branch_slow_dec_deferred(&clean_acked_data_enabled); 130 icsk->icsk_clean_acked = NULL; 131 } 132 EXPORT_SYMBOL_GPL(clean_acked_data_disable); 133 134 void clean_acked_data_flush(void) 135 { 136 static_key_deferred_flush(&clean_acked_data_enabled); 137 } 138 EXPORT_SYMBOL_GPL(clean_acked_data_flush); 139 #endif 140 141 static void tcp_gro_dev_warn(struct sock *sk, const struct sk_buff *skb, 142 unsigned int len) 143 { 144 static bool __once __read_mostly; 145 146 if (!__once) { 147 struct net_device *dev; 148 149 __once = true; 150 151 rcu_read_lock(); 152 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif); 153 if (!dev || len >= dev->mtu) 154 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n", 155 dev ? dev->name : "Unknown driver"); 156 rcu_read_unlock(); 157 } 158 } 159 160 /* Adapt the MSS value used to make delayed ack decision to the 161 * real world. 162 */ 163 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb) 164 { 165 struct inet_connection_sock *icsk = inet_csk(sk); 166 const unsigned int lss = icsk->icsk_ack.last_seg_size; 167 unsigned int len; 168 169 icsk->icsk_ack.last_seg_size = 0; 170 171 /* skb->len may jitter because of SACKs, even if peer 172 * sends good full-sized frames. 173 */ 174 len = skb_shinfo(skb)->gso_size ? : skb->len; 175 if (len >= icsk->icsk_ack.rcv_mss) { 176 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len, 177 tcp_sk(sk)->advmss); 178 /* Account for possibly-removed options */ 179 if (unlikely(len > icsk->icsk_ack.rcv_mss + 180 MAX_TCP_OPTION_SPACE)) 181 tcp_gro_dev_warn(sk, skb, len); 182 } else { 183 /* Otherwise, we make more careful check taking into account, 184 * that SACKs block is variable. 185 * 186 * "len" is invariant segment length, including TCP header. 187 */ 188 len += skb->data - skb_transport_header(skb); 189 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) || 190 /* If PSH is not set, packet should be 191 * full sized, provided peer TCP is not badly broken. 192 * This observation (if it is correct 8)) allows 193 * to handle super-low mtu links fairly. 194 */ 195 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && 196 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) { 197 /* Subtract also invariant (if peer is RFC compliant), 198 * tcp header plus fixed timestamp option length. 199 * Resulting "len" is MSS free of SACK jitter. 200 */ 201 len -= tcp_sk(sk)->tcp_header_len; 202 icsk->icsk_ack.last_seg_size = len; 203 if (len == lss) { 204 icsk->icsk_ack.rcv_mss = len; 205 return; 206 } 207 } 208 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) 209 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; 210 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 211 } 212 } 213 214 static void tcp_incr_quickack(struct sock *sk, unsigned int max_quickacks) 215 { 216 struct inet_connection_sock *icsk = inet_csk(sk); 217 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); 218 219 if (quickacks == 0) 220 quickacks = 2; 221 quickacks = min(quickacks, max_quickacks); 222 if (quickacks > icsk->icsk_ack.quick) 223 icsk->icsk_ack.quick = quickacks; 224 } 225 226 void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks) 227 { 228 struct inet_connection_sock *icsk = inet_csk(sk); 229 230 tcp_incr_quickack(sk, max_quickacks); 231 inet_csk_exit_pingpong_mode(sk); 232 icsk->icsk_ack.ato = TCP_ATO_MIN; 233 } 234 EXPORT_SYMBOL(tcp_enter_quickack_mode); 235 236 /* Send ACKs quickly, if "quick" count is not exhausted 237 * and the session is not interactive. 238 */ 239 240 static bool tcp_in_quickack_mode(struct sock *sk) 241 { 242 const struct inet_connection_sock *icsk = inet_csk(sk); 243 const struct dst_entry *dst = __sk_dst_get(sk); 244 245 return (dst && dst_metric(dst, RTAX_QUICKACK)) || 246 (icsk->icsk_ack.quick && !inet_csk_in_pingpong_mode(sk)); 247 } 248 249 static void tcp_ecn_queue_cwr(struct tcp_sock *tp) 250 { 251 if (tp->ecn_flags & TCP_ECN_OK) 252 tp->ecn_flags |= TCP_ECN_QUEUE_CWR; 253 } 254 255 static void tcp_ecn_accept_cwr(struct sock *sk, const struct sk_buff *skb) 256 { 257 if (tcp_hdr(skb)->cwr) { 258 tcp_sk(sk)->ecn_flags &= ~TCP_ECN_DEMAND_CWR; 259 260 /* If the sender is telling us it has entered CWR, then its 261 * cwnd may be very low (even just 1 packet), so we should ACK 262 * immediately. 263 */ 264 if (TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) 265 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW; 266 } 267 } 268 269 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp) 270 { 271 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 272 } 273 274 static void __tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb) 275 { 276 struct tcp_sock *tp = tcp_sk(sk); 277 278 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) { 279 case INET_ECN_NOT_ECT: 280 /* Funny extension: if ECT is not set on a segment, 281 * and we already seen ECT on a previous segment, 282 * it is probably a retransmit. 283 */ 284 if (tp->ecn_flags & TCP_ECN_SEEN) 285 tcp_enter_quickack_mode(sk, 2); 286 break; 287 case INET_ECN_CE: 288 if (tcp_ca_needs_ecn(sk)) 289 tcp_ca_event(sk, CA_EVENT_ECN_IS_CE); 290 291 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) { 292 /* Better not delay acks, sender can have a very low cwnd */ 293 tcp_enter_quickack_mode(sk, 2); 294 tp->ecn_flags |= TCP_ECN_DEMAND_CWR; 295 } 296 tp->ecn_flags |= TCP_ECN_SEEN; 297 break; 298 default: 299 if (tcp_ca_needs_ecn(sk)) 300 tcp_ca_event(sk, CA_EVENT_ECN_NO_CE); 301 tp->ecn_flags |= TCP_ECN_SEEN; 302 break; 303 } 304 } 305 306 static void tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb) 307 { 308 if (tcp_sk(sk)->ecn_flags & TCP_ECN_OK) 309 __tcp_ecn_check_ce(sk, skb); 310 } 311 312 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th) 313 { 314 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr)) 315 tp->ecn_flags &= ~TCP_ECN_OK; 316 } 317 318 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th) 319 { 320 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr)) 321 tp->ecn_flags &= ~TCP_ECN_OK; 322 } 323 324 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th) 325 { 326 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK)) 327 return true; 328 return false; 329 } 330 331 /* Buffer size and advertised window tuning. 332 * 333 * 1. Tuning sk->sk_sndbuf, when connection enters established state. 334 */ 335 336 static void tcp_sndbuf_expand(struct sock *sk) 337 { 338 const struct tcp_sock *tp = tcp_sk(sk); 339 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 340 int sndmem, per_mss; 341 u32 nr_segs; 342 343 /* Worst case is non GSO/TSO : each frame consumes one skb 344 * and skb->head is kmalloced using power of two area of memory 345 */ 346 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) + 347 MAX_TCP_HEADER + 348 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 349 350 per_mss = roundup_pow_of_two(per_mss) + 351 SKB_DATA_ALIGN(sizeof(struct sk_buff)); 352 353 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd); 354 nr_segs = max_t(u32, nr_segs, tp->reordering + 1); 355 356 /* Fast Recovery (RFC 5681 3.2) : 357 * Cubic needs 1.7 factor, rounded to 2 to include 358 * extra cushion (application might react slowly to EPOLLOUT) 359 */ 360 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2; 361 sndmem *= nr_segs * per_mss; 362 363 if (sk->sk_sndbuf < sndmem) 364 WRITE_ONCE(sk->sk_sndbuf, 365 min(sndmem, sock_net(sk)->ipv4.sysctl_tcp_wmem[2])); 366 } 367 368 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh) 369 * 370 * All tcp_full_space() is split to two parts: "network" buffer, allocated 371 * forward and advertised in receiver window (tp->rcv_wnd) and 372 * "application buffer", required to isolate scheduling/application 373 * latencies from network. 374 * window_clamp is maximal advertised window. It can be less than 375 * tcp_full_space(), in this case tcp_full_space() - window_clamp 376 * is reserved for "application" buffer. The less window_clamp is 377 * the smoother our behaviour from viewpoint of network, but the lower 378 * throughput and the higher sensitivity of the connection to losses. 8) 379 * 380 * rcv_ssthresh is more strict window_clamp used at "slow start" 381 * phase to predict further behaviour of this connection. 382 * It is used for two goals: 383 * - to enforce header prediction at sender, even when application 384 * requires some significant "application buffer". It is check #1. 385 * - to prevent pruning of receive queue because of misprediction 386 * of receiver window. Check #2. 387 * 388 * The scheme does not work when sender sends good segments opening 389 * window and then starts to feed us spaghetti. But it should work 390 * in common situations. Otherwise, we have to rely on queue collapsing. 391 */ 392 393 /* Slow part of check#2. */ 394 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb) 395 { 396 struct tcp_sock *tp = tcp_sk(sk); 397 /* Optimize this! */ 398 int truesize = tcp_win_from_space(sk, skb->truesize) >> 1; 399 int window = tcp_win_from_space(sk, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1; 400 401 while (tp->rcv_ssthresh <= window) { 402 if (truesize <= skb->len) 403 return 2 * inet_csk(sk)->icsk_ack.rcv_mss; 404 405 truesize >>= 1; 406 window >>= 1; 407 } 408 return 0; 409 } 410 411 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb) 412 { 413 struct tcp_sock *tp = tcp_sk(sk); 414 int room; 415 416 room = min_t(int, tp->window_clamp, tcp_space(sk)) - tp->rcv_ssthresh; 417 418 /* Check #1 */ 419 if (room > 0 && !tcp_under_memory_pressure(sk)) { 420 int incr; 421 422 /* Check #2. Increase window, if skb with such overhead 423 * will fit to rcvbuf in future. 424 */ 425 if (tcp_win_from_space(sk, skb->truesize) <= skb->len) 426 incr = 2 * tp->advmss; 427 else 428 incr = __tcp_grow_window(sk, skb); 429 430 if (incr) { 431 incr = max_t(int, incr, 2 * skb->len); 432 tp->rcv_ssthresh += min(room, incr); 433 inet_csk(sk)->icsk_ack.quick |= 1; 434 } 435 } 436 } 437 438 /* 3. Try to fixup all. It is made immediately after connection enters 439 * established state. 440 */ 441 static void tcp_init_buffer_space(struct sock *sk) 442 { 443 int tcp_app_win = sock_net(sk)->ipv4.sysctl_tcp_app_win; 444 struct tcp_sock *tp = tcp_sk(sk); 445 int maxwin; 446 447 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) 448 tcp_sndbuf_expand(sk); 449 450 tp->rcvq_space.space = min_t(u32, tp->rcv_wnd, TCP_INIT_CWND * tp->advmss); 451 tcp_mstamp_refresh(tp); 452 tp->rcvq_space.time = tp->tcp_mstamp; 453 tp->rcvq_space.seq = tp->copied_seq; 454 455 maxwin = tcp_full_space(sk); 456 457 if (tp->window_clamp >= maxwin) { 458 tp->window_clamp = maxwin; 459 460 if (tcp_app_win && maxwin > 4 * tp->advmss) 461 tp->window_clamp = max(maxwin - 462 (maxwin >> tcp_app_win), 463 4 * tp->advmss); 464 } 465 466 /* Force reservation of one segment. */ 467 if (tcp_app_win && 468 tp->window_clamp > 2 * tp->advmss && 469 tp->window_clamp + tp->advmss > maxwin) 470 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss); 471 472 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); 473 tp->snd_cwnd_stamp = tcp_jiffies32; 474 } 475 476 /* 4. Recalculate window clamp after socket hit its memory bounds. */ 477 static void tcp_clamp_window(struct sock *sk) 478 { 479 struct tcp_sock *tp = tcp_sk(sk); 480 struct inet_connection_sock *icsk = inet_csk(sk); 481 struct net *net = sock_net(sk); 482 483 icsk->icsk_ack.quick = 0; 484 485 if (sk->sk_rcvbuf < net->ipv4.sysctl_tcp_rmem[2] && 486 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && 487 !tcp_under_memory_pressure(sk) && 488 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) { 489 WRITE_ONCE(sk->sk_rcvbuf, 490 min(atomic_read(&sk->sk_rmem_alloc), 491 net->ipv4.sysctl_tcp_rmem[2])); 492 } 493 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) 494 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss); 495 } 496 497 /* Initialize RCV_MSS value. 498 * RCV_MSS is an our guess about MSS used by the peer. 499 * We haven't any direct information about the MSS. 500 * It's better to underestimate the RCV_MSS rather than overestimate. 501 * Overestimations make us ACKing less frequently than needed. 502 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). 503 */ 504 void tcp_initialize_rcv_mss(struct sock *sk) 505 { 506 const struct tcp_sock *tp = tcp_sk(sk); 507 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); 508 509 hint = min(hint, tp->rcv_wnd / 2); 510 hint = min(hint, TCP_MSS_DEFAULT); 511 hint = max(hint, TCP_MIN_MSS); 512 513 inet_csk(sk)->icsk_ack.rcv_mss = hint; 514 } 515 EXPORT_SYMBOL(tcp_initialize_rcv_mss); 516 517 /* Receiver "autotuning" code. 518 * 519 * The algorithm for RTT estimation w/o timestamps is based on 520 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. 521 * <https://public.lanl.gov/radiant/pubs.html#DRS> 522 * 523 * More detail on this code can be found at 524 * <http://staff.psc.edu/jheffner/>, 525 * though this reference is out of date. A new paper 526 * is pending. 527 */ 528 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) 529 { 530 u32 new_sample = tp->rcv_rtt_est.rtt_us; 531 long m = sample; 532 533 if (new_sample != 0) { 534 /* If we sample in larger samples in the non-timestamp 535 * case, we could grossly overestimate the RTT especially 536 * with chatty applications or bulk transfer apps which 537 * are stalled on filesystem I/O. 538 * 539 * Also, since we are only going for a minimum in the 540 * non-timestamp case, we do not smooth things out 541 * else with timestamps disabled convergence takes too 542 * long. 543 */ 544 if (!win_dep) { 545 m -= (new_sample >> 3); 546 new_sample += m; 547 } else { 548 m <<= 3; 549 if (m < new_sample) 550 new_sample = m; 551 } 552 } else { 553 /* No previous measure. */ 554 new_sample = m << 3; 555 } 556 557 tp->rcv_rtt_est.rtt_us = new_sample; 558 } 559 560 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) 561 { 562 u32 delta_us; 563 564 if (tp->rcv_rtt_est.time == 0) 565 goto new_measure; 566 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) 567 return; 568 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time); 569 if (!delta_us) 570 delta_us = 1; 571 tcp_rcv_rtt_update(tp, delta_us, 1); 572 573 new_measure: 574 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; 575 tp->rcv_rtt_est.time = tp->tcp_mstamp; 576 } 577 578 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, 579 const struct sk_buff *skb) 580 { 581 struct tcp_sock *tp = tcp_sk(sk); 582 583 if (tp->rx_opt.rcv_tsecr == tp->rcv_rtt_last_tsecr) 584 return; 585 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr; 586 587 if (TCP_SKB_CB(skb)->end_seq - 588 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) { 589 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr; 590 u32 delta_us; 591 592 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) { 593 if (!delta) 594 delta = 1; 595 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ); 596 tcp_rcv_rtt_update(tp, delta_us, 0); 597 } 598 } 599 } 600 601 /* 602 * This function should be called every time data is copied to user space. 603 * It calculates the appropriate TCP receive buffer space. 604 */ 605 void tcp_rcv_space_adjust(struct sock *sk) 606 { 607 struct tcp_sock *tp = tcp_sk(sk); 608 u32 copied; 609 int time; 610 611 trace_tcp_rcv_space_adjust(sk); 612 613 tcp_mstamp_refresh(tp); 614 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time); 615 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0) 616 return; 617 618 /* Number of bytes copied to user in last RTT */ 619 copied = tp->copied_seq - tp->rcvq_space.seq; 620 if (copied <= tp->rcvq_space.space) 621 goto new_measure; 622 623 /* A bit of theory : 624 * copied = bytes received in previous RTT, our base window 625 * To cope with packet losses, we need a 2x factor 626 * To cope with slow start, and sender growing its cwin by 100 % 627 * every RTT, we need a 4x factor, because the ACK we are sending 628 * now is for the next RTT, not the current one : 629 * <prev RTT . ><current RTT .. ><next RTT .... > 630 */ 631 632 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf && 633 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { 634 int rcvmem, rcvbuf; 635 u64 rcvwin, grow; 636 637 /* minimal window to cope with packet losses, assuming 638 * steady state. Add some cushion because of small variations. 639 */ 640 rcvwin = ((u64)copied << 1) + 16 * tp->advmss; 641 642 /* Accommodate for sender rate increase (eg. slow start) */ 643 grow = rcvwin * (copied - tp->rcvq_space.space); 644 do_div(grow, tp->rcvq_space.space); 645 rcvwin += (grow << 1); 646 647 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER); 648 while (tcp_win_from_space(sk, rcvmem) < tp->advmss) 649 rcvmem += 128; 650 651 do_div(rcvwin, tp->advmss); 652 rcvbuf = min_t(u64, rcvwin * rcvmem, 653 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]); 654 if (rcvbuf > sk->sk_rcvbuf) { 655 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf); 656 657 /* Make the window clamp follow along. */ 658 tp->window_clamp = tcp_win_from_space(sk, rcvbuf); 659 } 660 } 661 tp->rcvq_space.space = copied; 662 663 new_measure: 664 tp->rcvq_space.seq = tp->copied_seq; 665 tp->rcvq_space.time = tp->tcp_mstamp; 666 } 667 668 /* There is something which you must keep in mind when you analyze the 669 * behavior of the tp->ato delayed ack timeout interval. When a 670 * connection starts up, we want to ack as quickly as possible. The 671 * problem is that "good" TCP's do slow start at the beginning of data 672 * transmission. The means that until we send the first few ACK's the 673 * sender will sit on his end and only queue most of his data, because 674 * he can only send snd_cwnd unacked packets at any given time. For 675 * each ACK we send, he increments snd_cwnd and transmits more of his 676 * queue. -DaveM 677 */ 678 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb) 679 { 680 struct tcp_sock *tp = tcp_sk(sk); 681 struct inet_connection_sock *icsk = inet_csk(sk); 682 u32 now; 683 684 inet_csk_schedule_ack(sk); 685 686 tcp_measure_rcv_mss(sk, skb); 687 688 tcp_rcv_rtt_measure(tp); 689 690 now = tcp_jiffies32; 691 692 if (!icsk->icsk_ack.ato) { 693 /* The _first_ data packet received, initialize 694 * delayed ACK engine. 695 */ 696 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS); 697 icsk->icsk_ack.ato = TCP_ATO_MIN; 698 } else { 699 int m = now - icsk->icsk_ack.lrcvtime; 700 701 if (m <= TCP_ATO_MIN / 2) { 702 /* The fastest case is the first. */ 703 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; 704 } else if (m < icsk->icsk_ack.ato) { 705 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; 706 if (icsk->icsk_ack.ato > icsk->icsk_rto) 707 icsk->icsk_ack.ato = icsk->icsk_rto; 708 } else if (m > icsk->icsk_rto) { 709 /* Too long gap. Apparently sender failed to 710 * restart window, so that we send ACKs quickly. 711 */ 712 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS); 713 sk_mem_reclaim(sk); 714 } 715 } 716 icsk->icsk_ack.lrcvtime = now; 717 718 tcp_ecn_check_ce(sk, skb); 719 720 if (skb->len >= 128) 721 tcp_grow_window(sk, skb); 722 } 723 724 /* Called to compute a smoothed rtt estimate. The data fed to this 725 * routine either comes from timestamps, or from segments that were 726 * known _not_ to have been retransmitted [see Karn/Partridge 727 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 728 * piece by Van Jacobson. 729 * NOTE: the next three routines used to be one big routine. 730 * To save cycles in the RFC 1323 implementation it was better to break 731 * it up into three procedures. -- erics 732 */ 733 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us) 734 { 735 struct tcp_sock *tp = tcp_sk(sk); 736 long m = mrtt_us; /* RTT */ 737 u32 srtt = tp->srtt_us; 738 739 /* The following amusing code comes from Jacobson's 740 * article in SIGCOMM '88. Note that rtt and mdev 741 * are scaled versions of rtt and mean deviation. 742 * This is designed to be as fast as possible 743 * m stands for "measurement". 744 * 745 * On a 1990 paper the rto value is changed to: 746 * RTO = rtt + 4 * mdev 747 * 748 * Funny. This algorithm seems to be very broken. 749 * These formulae increase RTO, when it should be decreased, increase 750 * too slowly, when it should be increased quickly, decrease too quickly 751 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely 752 * does not matter how to _calculate_ it. Seems, it was trap 753 * that VJ failed to avoid. 8) 754 */ 755 if (srtt != 0) { 756 m -= (srtt >> 3); /* m is now error in rtt est */ 757 srtt += m; /* rtt = 7/8 rtt + 1/8 new */ 758 if (m < 0) { 759 m = -m; /* m is now abs(error) */ 760 m -= (tp->mdev_us >> 2); /* similar update on mdev */ 761 /* This is similar to one of Eifel findings. 762 * Eifel blocks mdev updates when rtt decreases. 763 * This solution is a bit different: we use finer gain 764 * for mdev in this case (alpha*beta). 765 * Like Eifel it also prevents growth of rto, 766 * but also it limits too fast rto decreases, 767 * happening in pure Eifel. 768 */ 769 if (m > 0) 770 m >>= 3; 771 } else { 772 m -= (tp->mdev_us >> 2); /* similar update on mdev */ 773 } 774 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */ 775 if (tp->mdev_us > tp->mdev_max_us) { 776 tp->mdev_max_us = tp->mdev_us; 777 if (tp->mdev_max_us > tp->rttvar_us) 778 tp->rttvar_us = tp->mdev_max_us; 779 } 780 if (after(tp->snd_una, tp->rtt_seq)) { 781 if (tp->mdev_max_us < tp->rttvar_us) 782 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2; 783 tp->rtt_seq = tp->snd_nxt; 784 tp->mdev_max_us = tcp_rto_min_us(sk); 785 786 tcp_bpf_rtt(sk); 787 } 788 } else { 789 /* no previous measure. */ 790 srtt = m << 3; /* take the measured time to be rtt */ 791 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */ 792 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk)); 793 tp->mdev_max_us = tp->rttvar_us; 794 tp->rtt_seq = tp->snd_nxt; 795 796 tcp_bpf_rtt(sk); 797 } 798 tp->srtt_us = max(1U, srtt); 799 } 800 801 static void tcp_update_pacing_rate(struct sock *sk) 802 { 803 const struct tcp_sock *tp = tcp_sk(sk); 804 u64 rate; 805 806 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */ 807 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3); 808 809 /* current rate is (cwnd * mss) / srtt 810 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate. 811 * In Congestion Avoidance phase, set it to 120 % the current rate. 812 * 813 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh) 814 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching 815 * end of slow start and should slow down. 816 */ 817 if (tp->snd_cwnd < tp->snd_ssthresh / 2) 818 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio; 819 else 820 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio; 821 822 rate *= max(tp->snd_cwnd, tp->packets_out); 823 824 if (likely(tp->srtt_us)) 825 do_div(rate, tp->srtt_us); 826 827 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate 828 * without any lock. We want to make sure compiler wont store 829 * intermediate values in this location. 830 */ 831 WRITE_ONCE(sk->sk_pacing_rate, min_t(u64, rate, 832 sk->sk_max_pacing_rate)); 833 } 834 835 /* Calculate rto without backoff. This is the second half of Van Jacobson's 836 * routine referred to above. 837 */ 838 static void tcp_set_rto(struct sock *sk) 839 { 840 const struct tcp_sock *tp = tcp_sk(sk); 841 /* Old crap is replaced with new one. 8) 842 * 843 * More seriously: 844 * 1. If rtt variance happened to be less 50msec, it is hallucination. 845 * It cannot be less due to utterly erratic ACK generation made 846 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ 847 * to do with delayed acks, because at cwnd>2 true delack timeout 848 * is invisible. Actually, Linux-2.4 also generates erratic 849 * ACKs in some circumstances. 850 */ 851 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp); 852 853 /* 2. Fixups made earlier cannot be right. 854 * If we do not estimate RTO correctly without them, 855 * all the algo is pure shit and should be replaced 856 * with correct one. It is exactly, which we pretend to do. 857 */ 858 859 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo 860 * guarantees that rto is higher. 861 */ 862 tcp_bound_rto(sk); 863 } 864 865 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst) 866 { 867 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); 868 869 if (!cwnd) 870 cwnd = TCP_INIT_CWND; 871 return min_t(__u32, cwnd, tp->snd_cwnd_clamp); 872 } 873 874 struct tcp_sacktag_state { 875 /* Timestamps for earliest and latest never-retransmitted segment 876 * that was SACKed. RTO needs the earliest RTT to stay conservative, 877 * but congestion control should still get an accurate delay signal. 878 */ 879 u64 first_sackt; 880 u64 last_sackt; 881 u32 reord; 882 u32 sack_delivered; 883 int flag; 884 unsigned int mss_now; 885 struct rate_sample *rate; 886 }; 887 888 /* Take a notice that peer is sending D-SACKs */ 889 static u32 tcp_dsack_seen(struct tcp_sock *tp, u32 start_seq, 890 u32 end_seq, struct tcp_sacktag_state *state) 891 { 892 u32 seq_len, dup_segs = 1; 893 894 if (before(start_seq, end_seq)) { 895 seq_len = end_seq - start_seq; 896 if (seq_len > tp->mss_cache) 897 dup_segs = DIV_ROUND_UP(seq_len, tp->mss_cache); 898 } 899 900 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN; 901 tp->rack.dsack_seen = 1; 902 tp->dsack_dups += dup_segs; 903 904 state->flag |= FLAG_DSACKING_ACK; 905 /* A spurious retransmission is delivered */ 906 state->sack_delivered += dup_segs; 907 908 return dup_segs; 909 } 910 911 /* It's reordering when higher sequence was delivered (i.e. sacked) before 912 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering 913 * distance is approximated in full-mss packet distance ("reordering"). 914 */ 915 static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq, 916 const int ts) 917 { 918 struct tcp_sock *tp = tcp_sk(sk); 919 const u32 mss = tp->mss_cache; 920 u32 fack, metric; 921 922 fack = tcp_highest_sack_seq(tp); 923 if (!before(low_seq, fack)) 924 return; 925 926 metric = fack - low_seq; 927 if ((metric > tp->reordering * mss) && mss) { 928 #if FASTRETRANS_DEBUG > 1 929 pr_debug("Disorder%d %d %u f%u s%u rr%d\n", 930 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, 931 tp->reordering, 932 0, 933 tp->sacked_out, 934 tp->undo_marker ? tp->undo_retrans : 0); 935 #endif 936 tp->reordering = min_t(u32, (metric + mss - 1) / mss, 937 sock_net(sk)->ipv4.sysctl_tcp_max_reordering); 938 } 939 940 /* This exciting event is worth to be remembered. 8) */ 941 tp->reord_seen++; 942 NET_INC_STATS(sock_net(sk), 943 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER); 944 } 945 946 /* This must be called before lost_out is incremented */ 947 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb) 948 { 949 if ((!tp->retransmit_skb_hint && tp->retrans_out >= tp->lost_out) || 950 (tp->retransmit_skb_hint && 951 before(TCP_SKB_CB(skb)->seq, 952 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))) 953 tp->retransmit_skb_hint = skb; 954 } 955 956 /* Sum the number of packets on the wire we have marked as lost. 957 * There are two cases we care about here: 958 * a) Packet hasn't been marked lost (nor retransmitted), 959 * and this is the first loss. 960 * b) Packet has been marked both lost and retransmitted, 961 * and this means we think it was lost again. 962 */ 963 static void tcp_sum_lost(struct tcp_sock *tp, struct sk_buff *skb) 964 { 965 __u8 sacked = TCP_SKB_CB(skb)->sacked; 966 967 if (!(sacked & TCPCB_LOST) || 968 ((sacked & TCPCB_LOST) && (sacked & TCPCB_SACKED_RETRANS))) 969 tp->lost += tcp_skb_pcount(skb); 970 } 971 972 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb) 973 { 974 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { 975 tcp_verify_retransmit_hint(tp, skb); 976 977 tp->lost_out += tcp_skb_pcount(skb); 978 tcp_sum_lost(tp, skb); 979 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 980 } 981 } 982 983 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb) 984 { 985 tcp_verify_retransmit_hint(tp, skb); 986 987 tcp_sum_lost(tp, skb); 988 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { 989 tp->lost_out += tcp_skb_pcount(skb); 990 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 991 } 992 } 993 994 /* Updates the delivered and delivered_ce counts */ 995 static void tcp_count_delivered(struct tcp_sock *tp, u32 delivered, 996 bool ece_ack) 997 { 998 tp->delivered += delivered; 999 if (ece_ack) 1000 tp->delivered_ce += delivered; 1001 } 1002 1003 /* This procedure tags the retransmission queue when SACKs arrive. 1004 * 1005 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). 1006 * Packets in queue with these bits set are counted in variables 1007 * sacked_out, retrans_out and lost_out, correspondingly. 1008 * 1009 * Valid combinations are: 1010 * Tag InFlight Description 1011 * 0 1 - orig segment is in flight. 1012 * S 0 - nothing flies, orig reached receiver. 1013 * L 0 - nothing flies, orig lost by net. 1014 * R 2 - both orig and retransmit are in flight. 1015 * L|R 1 - orig is lost, retransmit is in flight. 1016 * S|R 1 - orig reached receiver, retrans is still in flight. 1017 * (L|S|R is logically valid, it could occur when L|R is sacked, 1018 * but it is equivalent to plain S and code short-curcuits it to S. 1019 * L|S is logically invalid, it would mean -1 packet in flight 8)) 1020 * 1021 * These 6 states form finite state machine, controlled by the following events: 1022 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) 1023 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) 1024 * 3. Loss detection event of two flavors: 1025 * A. Scoreboard estimator decided the packet is lost. 1026 * A'. Reno "three dupacks" marks head of queue lost. 1027 * B. SACK arrives sacking SND.NXT at the moment, when the 1028 * segment was retransmitted. 1029 * 4. D-SACK added new rule: D-SACK changes any tag to S. 1030 * 1031 * It is pleasant to note, that state diagram turns out to be commutative, 1032 * so that we are allowed not to be bothered by order of our actions, 1033 * when multiple events arrive simultaneously. (see the function below). 1034 * 1035 * Reordering detection. 1036 * -------------------- 1037 * Reordering metric is maximal distance, which a packet can be displaced 1038 * in packet stream. With SACKs we can estimate it: 1039 * 1040 * 1. SACK fills old hole and the corresponding segment was not 1041 * ever retransmitted -> reordering. Alas, we cannot use it 1042 * when segment was retransmitted. 1043 * 2. The last flaw is solved with D-SACK. D-SACK arrives 1044 * for retransmitted and already SACKed segment -> reordering.. 1045 * Both of these heuristics are not used in Loss state, when we cannot 1046 * account for retransmits accurately. 1047 * 1048 * SACK block validation. 1049 * ---------------------- 1050 * 1051 * SACK block range validation checks that the received SACK block fits to 1052 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT. 1053 * Note that SND.UNA is not included to the range though being valid because 1054 * it means that the receiver is rather inconsistent with itself reporting 1055 * SACK reneging when it should advance SND.UNA. Such SACK block this is 1056 * perfectly valid, however, in light of RFC2018 which explicitly states 1057 * that "SACK block MUST reflect the newest segment. Even if the newest 1058 * segment is going to be discarded ...", not that it looks very clever 1059 * in case of head skb. Due to potentional receiver driven attacks, we 1060 * choose to avoid immediate execution of a walk in write queue due to 1061 * reneging and defer head skb's loss recovery to standard loss recovery 1062 * procedure that will eventually trigger (nothing forbids us doing this). 1063 * 1064 * Implements also blockage to start_seq wrap-around. Problem lies in the 1065 * fact that though start_seq (s) is before end_seq (i.e., not reversed), 1066 * there's no guarantee that it will be before snd_nxt (n). The problem 1067 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt 1068 * wrap (s_w): 1069 * 1070 * <- outs wnd -> <- wrapzone -> 1071 * u e n u_w e_w s n_w 1072 * | | | | | | | 1073 * |<------------+------+----- TCP seqno space --------------+---------->| 1074 * ...-- <2^31 ->| |<--------... 1075 * ...---- >2^31 ------>| |<--------... 1076 * 1077 * Current code wouldn't be vulnerable but it's better still to discard such 1078 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat 1079 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in 1080 * snd_nxt wrap -> snd_una region will then become "well defined", i.e., 1081 * equal to the ideal case (infinite seqno space without wrap caused issues). 1082 * 1083 * With D-SACK the lower bound is extended to cover sequence space below 1084 * SND.UNA down to undo_marker, which is the last point of interest. Yet 1085 * again, D-SACK block must not to go across snd_una (for the same reason as 1086 * for the normal SACK blocks, explained above). But there all simplicity 1087 * ends, TCP might receive valid D-SACKs below that. As long as they reside 1088 * fully below undo_marker they do not affect behavior in anyway and can 1089 * therefore be safely ignored. In rare cases (which are more or less 1090 * theoretical ones), the D-SACK will nicely cross that boundary due to skb 1091 * fragmentation and packet reordering past skb's retransmission. To consider 1092 * them correctly, the acceptable range must be extended even more though 1093 * the exact amount is rather hard to quantify. However, tp->max_window can 1094 * be used as an exaggerated estimate. 1095 */ 1096 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack, 1097 u32 start_seq, u32 end_seq) 1098 { 1099 /* Too far in future, or reversed (interpretation is ambiguous) */ 1100 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq)) 1101 return false; 1102 1103 /* Nasty start_seq wrap-around check (see comments above) */ 1104 if (!before(start_seq, tp->snd_nxt)) 1105 return false; 1106 1107 /* In outstanding window? ...This is valid exit for D-SACKs too. 1108 * start_seq == snd_una is non-sensical (see comments above) 1109 */ 1110 if (after(start_seq, tp->snd_una)) 1111 return true; 1112 1113 if (!is_dsack || !tp->undo_marker) 1114 return false; 1115 1116 /* ...Then it's D-SACK, and must reside below snd_una completely */ 1117 if (after(end_seq, tp->snd_una)) 1118 return false; 1119 1120 if (!before(start_seq, tp->undo_marker)) 1121 return true; 1122 1123 /* Too old */ 1124 if (!after(end_seq, tp->undo_marker)) 1125 return false; 1126 1127 /* Undo_marker boundary crossing (overestimates a lot). Known already: 1128 * start_seq < undo_marker and end_seq >= undo_marker. 1129 */ 1130 return !before(start_seq, end_seq - tp->max_window); 1131 } 1132 1133 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb, 1134 struct tcp_sack_block_wire *sp, int num_sacks, 1135 u32 prior_snd_una, struct tcp_sacktag_state *state) 1136 { 1137 struct tcp_sock *tp = tcp_sk(sk); 1138 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq); 1139 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq); 1140 u32 dup_segs; 1141 1142 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) { 1143 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV); 1144 } else if (num_sacks > 1) { 1145 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq); 1146 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq); 1147 1148 if (after(end_seq_0, end_seq_1) || before(start_seq_0, start_seq_1)) 1149 return false; 1150 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKOFORECV); 1151 } else { 1152 return false; 1153 } 1154 1155 dup_segs = tcp_dsack_seen(tp, start_seq_0, end_seq_0, state); 1156 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECVSEGS, dup_segs); 1157 1158 /* D-SACK for already forgotten data... Do dumb counting. */ 1159 if (tp->undo_marker && tp->undo_retrans > 0 && 1160 !after(end_seq_0, prior_snd_una) && 1161 after(end_seq_0, tp->undo_marker)) 1162 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - dup_segs); 1163 1164 return true; 1165 } 1166 1167 /* Check if skb is fully within the SACK block. In presence of GSO skbs, 1168 * the incoming SACK may not exactly match but we can find smaller MSS 1169 * aligned portion of it that matches. Therefore we might need to fragment 1170 * which may fail and creates some hassle (caller must handle error case 1171 * returns). 1172 * 1173 * FIXME: this could be merged to shift decision code 1174 */ 1175 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb, 1176 u32 start_seq, u32 end_seq) 1177 { 1178 int err; 1179 bool in_sack; 1180 unsigned int pkt_len; 1181 unsigned int mss; 1182 1183 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1184 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1185 1186 if (tcp_skb_pcount(skb) > 1 && !in_sack && 1187 after(TCP_SKB_CB(skb)->end_seq, start_seq)) { 1188 mss = tcp_skb_mss(skb); 1189 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); 1190 1191 if (!in_sack) { 1192 pkt_len = start_seq - TCP_SKB_CB(skb)->seq; 1193 if (pkt_len < mss) 1194 pkt_len = mss; 1195 } else { 1196 pkt_len = end_seq - TCP_SKB_CB(skb)->seq; 1197 if (pkt_len < mss) 1198 return -EINVAL; 1199 } 1200 1201 /* Round if necessary so that SACKs cover only full MSSes 1202 * and/or the remaining small portion (if present) 1203 */ 1204 if (pkt_len > mss) { 1205 unsigned int new_len = (pkt_len / mss) * mss; 1206 if (!in_sack && new_len < pkt_len) 1207 new_len += mss; 1208 pkt_len = new_len; 1209 } 1210 1211 if (pkt_len >= skb->len && !in_sack) 1212 return 0; 1213 1214 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, 1215 pkt_len, mss, GFP_ATOMIC); 1216 if (err < 0) 1217 return err; 1218 } 1219 1220 return in_sack; 1221 } 1222 1223 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */ 1224 static u8 tcp_sacktag_one(struct sock *sk, 1225 struct tcp_sacktag_state *state, u8 sacked, 1226 u32 start_seq, u32 end_seq, 1227 int dup_sack, int pcount, 1228 u64 xmit_time) 1229 { 1230 struct tcp_sock *tp = tcp_sk(sk); 1231 1232 /* Account D-SACK for retransmitted packet. */ 1233 if (dup_sack && (sacked & TCPCB_RETRANS)) { 1234 if (tp->undo_marker && tp->undo_retrans > 0 && 1235 after(end_seq, tp->undo_marker)) 1236 tp->undo_retrans--; 1237 if ((sacked & TCPCB_SACKED_ACKED) && 1238 before(start_seq, state->reord)) 1239 state->reord = start_seq; 1240 } 1241 1242 /* Nothing to do; acked frame is about to be dropped (was ACKed). */ 1243 if (!after(end_seq, tp->snd_una)) 1244 return sacked; 1245 1246 if (!(sacked & TCPCB_SACKED_ACKED)) { 1247 tcp_rack_advance(tp, sacked, end_seq, xmit_time); 1248 1249 if (sacked & TCPCB_SACKED_RETRANS) { 1250 /* If the segment is not tagged as lost, 1251 * we do not clear RETRANS, believing 1252 * that retransmission is still in flight. 1253 */ 1254 if (sacked & TCPCB_LOST) { 1255 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); 1256 tp->lost_out -= pcount; 1257 tp->retrans_out -= pcount; 1258 } 1259 } else { 1260 if (!(sacked & TCPCB_RETRANS)) { 1261 /* New sack for not retransmitted frame, 1262 * which was in hole. It is reordering. 1263 */ 1264 if (before(start_seq, 1265 tcp_highest_sack_seq(tp)) && 1266 before(start_seq, state->reord)) 1267 state->reord = start_seq; 1268 1269 if (!after(end_seq, tp->high_seq)) 1270 state->flag |= FLAG_ORIG_SACK_ACKED; 1271 if (state->first_sackt == 0) 1272 state->first_sackt = xmit_time; 1273 state->last_sackt = xmit_time; 1274 } 1275 1276 if (sacked & TCPCB_LOST) { 1277 sacked &= ~TCPCB_LOST; 1278 tp->lost_out -= pcount; 1279 } 1280 } 1281 1282 sacked |= TCPCB_SACKED_ACKED; 1283 state->flag |= FLAG_DATA_SACKED; 1284 tp->sacked_out += pcount; 1285 /* Out-of-order packets delivered */ 1286 state->sack_delivered += pcount; 1287 1288 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */ 1289 if (tp->lost_skb_hint && 1290 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq)) 1291 tp->lost_cnt_hint += pcount; 1292 } 1293 1294 /* D-SACK. We can detect redundant retransmission in S|R and plain R 1295 * frames and clear it. undo_retrans is decreased above, L|R frames 1296 * are accounted above as well. 1297 */ 1298 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) { 1299 sacked &= ~TCPCB_SACKED_RETRANS; 1300 tp->retrans_out -= pcount; 1301 } 1302 1303 return sacked; 1304 } 1305 1306 /* Shift newly-SACKed bytes from this skb to the immediately previous 1307 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such. 1308 */ 1309 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev, 1310 struct sk_buff *skb, 1311 struct tcp_sacktag_state *state, 1312 unsigned int pcount, int shifted, int mss, 1313 bool dup_sack) 1314 { 1315 struct tcp_sock *tp = tcp_sk(sk); 1316 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */ 1317 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */ 1318 1319 BUG_ON(!pcount); 1320 1321 /* Adjust counters and hints for the newly sacked sequence 1322 * range but discard the return value since prev is already 1323 * marked. We must tag the range first because the seq 1324 * advancement below implicitly advances 1325 * tcp_highest_sack_seq() when skb is highest_sack. 1326 */ 1327 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked, 1328 start_seq, end_seq, dup_sack, pcount, 1329 tcp_skb_timestamp_us(skb)); 1330 tcp_rate_skb_delivered(sk, skb, state->rate); 1331 1332 if (skb == tp->lost_skb_hint) 1333 tp->lost_cnt_hint += pcount; 1334 1335 TCP_SKB_CB(prev)->end_seq += shifted; 1336 TCP_SKB_CB(skb)->seq += shifted; 1337 1338 tcp_skb_pcount_add(prev, pcount); 1339 WARN_ON_ONCE(tcp_skb_pcount(skb) < pcount); 1340 tcp_skb_pcount_add(skb, -pcount); 1341 1342 /* When we're adding to gso_segs == 1, gso_size will be zero, 1343 * in theory this shouldn't be necessary but as long as DSACK 1344 * code can come after this skb later on it's better to keep 1345 * setting gso_size to something. 1346 */ 1347 if (!TCP_SKB_CB(prev)->tcp_gso_size) 1348 TCP_SKB_CB(prev)->tcp_gso_size = mss; 1349 1350 /* CHECKME: To clear or not to clear? Mimics normal skb currently */ 1351 if (tcp_skb_pcount(skb) <= 1) 1352 TCP_SKB_CB(skb)->tcp_gso_size = 0; 1353 1354 /* Difference in this won't matter, both ACKed by the same cumul. ACK */ 1355 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS); 1356 1357 if (skb->len > 0) { 1358 BUG_ON(!tcp_skb_pcount(skb)); 1359 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED); 1360 return false; 1361 } 1362 1363 /* Whole SKB was eaten :-) */ 1364 1365 if (skb == tp->retransmit_skb_hint) 1366 tp->retransmit_skb_hint = prev; 1367 if (skb == tp->lost_skb_hint) { 1368 tp->lost_skb_hint = prev; 1369 tp->lost_cnt_hint -= tcp_skb_pcount(prev); 1370 } 1371 1372 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 1373 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor; 1374 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1375 TCP_SKB_CB(prev)->end_seq++; 1376 1377 if (skb == tcp_highest_sack(sk)) 1378 tcp_advance_highest_sack(sk, skb); 1379 1380 tcp_skb_collapse_tstamp(prev, skb); 1381 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp)) 1382 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0; 1383 1384 tcp_rtx_queue_unlink_and_free(skb, sk); 1385 1386 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED); 1387 1388 return true; 1389 } 1390 1391 /* I wish gso_size would have a bit more sane initialization than 1392 * something-or-zero which complicates things 1393 */ 1394 static int tcp_skb_seglen(const struct sk_buff *skb) 1395 { 1396 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb); 1397 } 1398 1399 /* Shifting pages past head area doesn't work */ 1400 static int skb_can_shift(const struct sk_buff *skb) 1401 { 1402 return !skb_headlen(skb) && skb_is_nonlinear(skb); 1403 } 1404 1405 int tcp_skb_shift(struct sk_buff *to, struct sk_buff *from, 1406 int pcount, int shiftlen) 1407 { 1408 /* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE) 1409 * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need 1410 * to make sure not storing more than 65535 * 8 bytes per skb, 1411 * even if current MSS is bigger. 1412 */ 1413 if (unlikely(to->len + shiftlen >= 65535 * TCP_MIN_GSO_SIZE)) 1414 return 0; 1415 if (unlikely(tcp_skb_pcount(to) + pcount > 65535)) 1416 return 0; 1417 return skb_shift(to, from, shiftlen); 1418 } 1419 1420 /* Try collapsing SACK blocks spanning across multiple skbs to a single 1421 * skb. 1422 */ 1423 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb, 1424 struct tcp_sacktag_state *state, 1425 u32 start_seq, u32 end_seq, 1426 bool dup_sack) 1427 { 1428 struct tcp_sock *tp = tcp_sk(sk); 1429 struct sk_buff *prev; 1430 int mss; 1431 int pcount = 0; 1432 int len; 1433 int in_sack; 1434 1435 /* Normally R but no L won't result in plain S */ 1436 if (!dup_sack && 1437 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS) 1438 goto fallback; 1439 if (!skb_can_shift(skb)) 1440 goto fallback; 1441 /* This frame is about to be dropped (was ACKed). */ 1442 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1443 goto fallback; 1444 1445 /* Can only happen with delayed DSACK + discard craziness */ 1446 prev = skb_rb_prev(skb); 1447 if (!prev) 1448 goto fallback; 1449 1450 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) 1451 goto fallback; 1452 1453 if (!tcp_skb_can_collapse(prev, skb)) 1454 goto fallback; 1455 1456 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1457 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1458 1459 if (in_sack) { 1460 len = skb->len; 1461 pcount = tcp_skb_pcount(skb); 1462 mss = tcp_skb_seglen(skb); 1463 1464 /* TODO: Fix DSACKs to not fragment already SACKed and we can 1465 * drop this restriction as unnecessary 1466 */ 1467 if (mss != tcp_skb_seglen(prev)) 1468 goto fallback; 1469 } else { 1470 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq)) 1471 goto noop; 1472 /* CHECKME: This is non-MSS split case only?, this will 1473 * cause skipped skbs due to advancing loop btw, original 1474 * has that feature too 1475 */ 1476 if (tcp_skb_pcount(skb) <= 1) 1477 goto noop; 1478 1479 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); 1480 if (!in_sack) { 1481 /* TODO: head merge to next could be attempted here 1482 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)), 1483 * though it might not be worth of the additional hassle 1484 * 1485 * ...we can probably just fallback to what was done 1486 * previously. We could try merging non-SACKed ones 1487 * as well but it probably isn't going to buy off 1488 * because later SACKs might again split them, and 1489 * it would make skb timestamp tracking considerably 1490 * harder problem. 1491 */ 1492 goto fallback; 1493 } 1494 1495 len = end_seq - TCP_SKB_CB(skb)->seq; 1496 BUG_ON(len < 0); 1497 BUG_ON(len > skb->len); 1498 1499 /* MSS boundaries should be honoured or else pcount will 1500 * severely break even though it makes things bit trickier. 1501 * Optimize common case to avoid most of the divides 1502 */ 1503 mss = tcp_skb_mss(skb); 1504 1505 /* TODO: Fix DSACKs to not fragment already SACKed and we can 1506 * drop this restriction as unnecessary 1507 */ 1508 if (mss != tcp_skb_seglen(prev)) 1509 goto fallback; 1510 1511 if (len == mss) { 1512 pcount = 1; 1513 } else if (len < mss) { 1514 goto noop; 1515 } else { 1516 pcount = len / mss; 1517 len = pcount * mss; 1518 } 1519 } 1520 1521 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */ 1522 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una)) 1523 goto fallback; 1524 1525 if (!tcp_skb_shift(prev, skb, pcount, len)) 1526 goto fallback; 1527 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack)) 1528 goto out; 1529 1530 /* Hole filled allows collapsing with the next as well, this is very 1531 * useful when hole on every nth skb pattern happens 1532 */ 1533 skb = skb_rb_next(prev); 1534 if (!skb) 1535 goto out; 1536 1537 if (!skb_can_shift(skb) || 1538 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) || 1539 (mss != tcp_skb_seglen(skb))) 1540 goto out; 1541 1542 len = skb->len; 1543 pcount = tcp_skb_pcount(skb); 1544 if (tcp_skb_shift(prev, skb, pcount, len)) 1545 tcp_shifted_skb(sk, prev, skb, state, pcount, 1546 len, mss, 0); 1547 1548 out: 1549 return prev; 1550 1551 noop: 1552 return skb; 1553 1554 fallback: 1555 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK); 1556 return NULL; 1557 } 1558 1559 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk, 1560 struct tcp_sack_block *next_dup, 1561 struct tcp_sacktag_state *state, 1562 u32 start_seq, u32 end_seq, 1563 bool dup_sack_in) 1564 { 1565 struct tcp_sock *tp = tcp_sk(sk); 1566 struct sk_buff *tmp; 1567 1568 skb_rbtree_walk_from(skb) { 1569 int in_sack = 0; 1570 bool dup_sack = dup_sack_in; 1571 1572 /* queue is in-order => we can short-circuit the walk early */ 1573 if (!before(TCP_SKB_CB(skb)->seq, end_seq)) 1574 break; 1575 1576 if (next_dup && 1577 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) { 1578 in_sack = tcp_match_skb_to_sack(sk, skb, 1579 next_dup->start_seq, 1580 next_dup->end_seq); 1581 if (in_sack > 0) 1582 dup_sack = true; 1583 } 1584 1585 /* skb reference here is a bit tricky to get right, since 1586 * shifting can eat and free both this skb and the next, 1587 * so not even _safe variant of the loop is enough. 1588 */ 1589 if (in_sack <= 0) { 1590 tmp = tcp_shift_skb_data(sk, skb, state, 1591 start_seq, end_seq, dup_sack); 1592 if (tmp) { 1593 if (tmp != skb) { 1594 skb = tmp; 1595 continue; 1596 } 1597 1598 in_sack = 0; 1599 } else { 1600 in_sack = tcp_match_skb_to_sack(sk, skb, 1601 start_seq, 1602 end_seq); 1603 } 1604 } 1605 1606 if (unlikely(in_sack < 0)) 1607 break; 1608 1609 if (in_sack) { 1610 TCP_SKB_CB(skb)->sacked = 1611 tcp_sacktag_one(sk, 1612 state, 1613 TCP_SKB_CB(skb)->sacked, 1614 TCP_SKB_CB(skb)->seq, 1615 TCP_SKB_CB(skb)->end_seq, 1616 dup_sack, 1617 tcp_skb_pcount(skb), 1618 tcp_skb_timestamp_us(skb)); 1619 tcp_rate_skb_delivered(sk, skb, state->rate); 1620 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1621 list_del_init(&skb->tcp_tsorted_anchor); 1622 1623 if (!before(TCP_SKB_CB(skb)->seq, 1624 tcp_highest_sack_seq(tp))) 1625 tcp_advance_highest_sack(sk, skb); 1626 } 1627 } 1628 return skb; 1629 } 1630 1631 static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk, u32 seq) 1632 { 1633 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node; 1634 struct sk_buff *skb; 1635 1636 while (*p) { 1637 parent = *p; 1638 skb = rb_to_skb(parent); 1639 if (before(seq, TCP_SKB_CB(skb)->seq)) { 1640 p = &parent->rb_left; 1641 continue; 1642 } 1643 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) { 1644 p = &parent->rb_right; 1645 continue; 1646 } 1647 return skb; 1648 } 1649 return NULL; 1650 } 1651 1652 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk, 1653 u32 skip_to_seq) 1654 { 1655 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq)) 1656 return skb; 1657 1658 return tcp_sacktag_bsearch(sk, skip_to_seq); 1659 } 1660 1661 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb, 1662 struct sock *sk, 1663 struct tcp_sack_block *next_dup, 1664 struct tcp_sacktag_state *state, 1665 u32 skip_to_seq) 1666 { 1667 if (!next_dup) 1668 return skb; 1669 1670 if (before(next_dup->start_seq, skip_to_seq)) { 1671 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq); 1672 skb = tcp_sacktag_walk(skb, sk, NULL, state, 1673 next_dup->start_seq, next_dup->end_seq, 1674 1); 1675 } 1676 1677 return skb; 1678 } 1679 1680 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache) 1681 { 1682 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); 1683 } 1684 1685 static int 1686 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb, 1687 u32 prior_snd_una, struct tcp_sacktag_state *state) 1688 { 1689 struct tcp_sock *tp = tcp_sk(sk); 1690 const unsigned char *ptr = (skb_transport_header(ack_skb) + 1691 TCP_SKB_CB(ack_skb)->sacked); 1692 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2); 1693 struct tcp_sack_block sp[TCP_NUM_SACKS]; 1694 struct tcp_sack_block *cache; 1695 struct sk_buff *skb; 1696 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3); 1697 int used_sacks; 1698 bool found_dup_sack = false; 1699 int i, j; 1700 int first_sack_index; 1701 1702 state->flag = 0; 1703 state->reord = tp->snd_nxt; 1704 1705 if (!tp->sacked_out) 1706 tcp_highest_sack_reset(sk); 1707 1708 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire, 1709 num_sacks, prior_snd_una, state); 1710 1711 /* Eliminate too old ACKs, but take into 1712 * account more or less fresh ones, they can 1713 * contain valid SACK info. 1714 */ 1715 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) 1716 return 0; 1717 1718 if (!tp->packets_out) 1719 goto out; 1720 1721 used_sacks = 0; 1722 first_sack_index = 0; 1723 for (i = 0; i < num_sacks; i++) { 1724 bool dup_sack = !i && found_dup_sack; 1725 1726 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq); 1727 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq); 1728 1729 if (!tcp_is_sackblock_valid(tp, dup_sack, 1730 sp[used_sacks].start_seq, 1731 sp[used_sacks].end_seq)) { 1732 int mib_idx; 1733 1734 if (dup_sack) { 1735 if (!tp->undo_marker) 1736 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO; 1737 else 1738 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD; 1739 } else { 1740 /* Don't count olds caused by ACK reordering */ 1741 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) && 1742 !after(sp[used_sacks].end_seq, tp->snd_una)) 1743 continue; 1744 mib_idx = LINUX_MIB_TCPSACKDISCARD; 1745 } 1746 1747 NET_INC_STATS(sock_net(sk), mib_idx); 1748 if (i == 0) 1749 first_sack_index = -1; 1750 continue; 1751 } 1752 1753 /* Ignore very old stuff early */ 1754 if (!after(sp[used_sacks].end_seq, prior_snd_una)) { 1755 if (i == 0) 1756 first_sack_index = -1; 1757 continue; 1758 } 1759 1760 used_sacks++; 1761 } 1762 1763 /* order SACK blocks to allow in order walk of the retrans queue */ 1764 for (i = used_sacks - 1; i > 0; i--) { 1765 for (j = 0; j < i; j++) { 1766 if (after(sp[j].start_seq, sp[j + 1].start_seq)) { 1767 swap(sp[j], sp[j + 1]); 1768 1769 /* Track where the first SACK block goes to */ 1770 if (j == first_sack_index) 1771 first_sack_index = j + 1; 1772 } 1773 } 1774 } 1775 1776 state->mss_now = tcp_current_mss(sk); 1777 skb = NULL; 1778 i = 0; 1779 1780 if (!tp->sacked_out) { 1781 /* It's already past, so skip checking against it */ 1782 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); 1783 } else { 1784 cache = tp->recv_sack_cache; 1785 /* Skip empty blocks in at head of the cache */ 1786 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq && 1787 !cache->end_seq) 1788 cache++; 1789 } 1790 1791 while (i < used_sacks) { 1792 u32 start_seq = sp[i].start_seq; 1793 u32 end_seq = sp[i].end_seq; 1794 bool dup_sack = (found_dup_sack && (i == first_sack_index)); 1795 struct tcp_sack_block *next_dup = NULL; 1796 1797 if (found_dup_sack && ((i + 1) == first_sack_index)) 1798 next_dup = &sp[i + 1]; 1799 1800 /* Skip too early cached blocks */ 1801 while (tcp_sack_cache_ok(tp, cache) && 1802 !before(start_seq, cache->end_seq)) 1803 cache++; 1804 1805 /* Can skip some work by looking recv_sack_cache? */ 1806 if (tcp_sack_cache_ok(tp, cache) && !dup_sack && 1807 after(end_seq, cache->start_seq)) { 1808 1809 /* Head todo? */ 1810 if (before(start_seq, cache->start_seq)) { 1811 skb = tcp_sacktag_skip(skb, sk, start_seq); 1812 skb = tcp_sacktag_walk(skb, sk, next_dup, 1813 state, 1814 start_seq, 1815 cache->start_seq, 1816 dup_sack); 1817 } 1818 1819 /* Rest of the block already fully processed? */ 1820 if (!after(end_seq, cache->end_seq)) 1821 goto advance_sp; 1822 1823 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup, 1824 state, 1825 cache->end_seq); 1826 1827 /* ...tail remains todo... */ 1828 if (tcp_highest_sack_seq(tp) == cache->end_seq) { 1829 /* ...but better entrypoint exists! */ 1830 skb = tcp_highest_sack(sk); 1831 if (!skb) 1832 break; 1833 cache++; 1834 goto walk; 1835 } 1836 1837 skb = tcp_sacktag_skip(skb, sk, cache->end_seq); 1838 /* Check overlap against next cached too (past this one already) */ 1839 cache++; 1840 continue; 1841 } 1842 1843 if (!before(start_seq, tcp_highest_sack_seq(tp))) { 1844 skb = tcp_highest_sack(sk); 1845 if (!skb) 1846 break; 1847 } 1848 skb = tcp_sacktag_skip(skb, sk, start_seq); 1849 1850 walk: 1851 skb = tcp_sacktag_walk(skb, sk, next_dup, state, 1852 start_seq, end_seq, dup_sack); 1853 1854 advance_sp: 1855 i++; 1856 } 1857 1858 /* Clear the head of the cache sack blocks so we can skip it next time */ 1859 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) { 1860 tp->recv_sack_cache[i].start_seq = 0; 1861 tp->recv_sack_cache[i].end_seq = 0; 1862 } 1863 for (j = 0; j < used_sacks; j++) 1864 tp->recv_sack_cache[i++] = sp[j]; 1865 1866 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker) 1867 tcp_check_sack_reordering(sk, state->reord, 0); 1868 1869 tcp_verify_left_out(tp); 1870 out: 1871 1872 #if FASTRETRANS_DEBUG > 0 1873 WARN_ON((int)tp->sacked_out < 0); 1874 WARN_ON((int)tp->lost_out < 0); 1875 WARN_ON((int)tp->retrans_out < 0); 1876 WARN_ON((int)tcp_packets_in_flight(tp) < 0); 1877 #endif 1878 return state->flag; 1879 } 1880 1881 /* Limits sacked_out so that sum with lost_out isn't ever larger than 1882 * packets_out. Returns false if sacked_out adjustement wasn't necessary. 1883 */ 1884 static bool tcp_limit_reno_sacked(struct tcp_sock *tp) 1885 { 1886 u32 holes; 1887 1888 holes = max(tp->lost_out, 1U); 1889 holes = min(holes, tp->packets_out); 1890 1891 if ((tp->sacked_out + holes) > tp->packets_out) { 1892 tp->sacked_out = tp->packets_out - holes; 1893 return true; 1894 } 1895 return false; 1896 } 1897 1898 /* If we receive more dupacks than we expected counting segments 1899 * in assumption of absent reordering, interpret this as reordering. 1900 * The only another reason could be bug in receiver TCP. 1901 */ 1902 static void tcp_check_reno_reordering(struct sock *sk, const int addend) 1903 { 1904 struct tcp_sock *tp = tcp_sk(sk); 1905 1906 if (!tcp_limit_reno_sacked(tp)) 1907 return; 1908 1909 tp->reordering = min_t(u32, tp->packets_out + addend, 1910 sock_net(sk)->ipv4.sysctl_tcp_max_reordering); 1911 tp->reord_seen++; 1912 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER); 1913 } 1914 1915 /* Emulate SACKs for SACKless connection: account for a new dupack. */ 1916 1917 static void tcp_add_reno_sack(struct sock *sk, int num_dupack, bool ece_ack) 1918 { 1919 if (num_dupack) { 1920 struct tcp_sock *tp = tcp_sk(sk); 1921 u32 prior_sacked = tp->sacked_out; 1922 s32 delivered; 1923 1924 tp->sacked_out += num_dupack; 1925 tcp_check_reno_reordering(sk, 0); 1926 delivered = tp->sacked_out - prior_sacked; 1927 if (delivered > 0) 1928 tcp_count_delivered(tp, delivered, ece_ack); 1929 tcp_verify_left_out(tp); 1930 } 1931 } 1932 1933 /* Account for ACK, ACKing some data in Reno Recovery phase. */ 1934 1935 static void tcp_remove_reno_sacks(struct sock *sk, int acked, bool ece_ack) 1936 { 1937 struct tcp_sock *tp = tcp_sk(sk); 1938 1939 if (acked > 0) { 1940 /* One ACK acked hole. The rest eat duplicate ACKs. */ 1941 tcp_count_delivered(tp, max_t(int, acked - tp->sacked_out, 1), 1942 ece_ack); 1943 if (acked - 1 >= tp->sacked_out) 1944 tp->sacked_out = 0; 1945 else 1946 tp->sacked_out -= acked - 1; 1947 } 1948 tcp_check_reno_reordering(sk, acked); 1949 tcp_verify_left_out(tp); 1950 } 1951 1952 static inline void tcp_reset_reno_sack(struct tcp_sock *tp) 1953 { 1954 tp->sacked_out = 0; 1955 } 1956 1957 void tcp_clear_retrans(struct tcp_sock *tp) 1958 { 1959 tp->retrans_out = 0; 1960 tp->lost_out = 0; 1961 tp->undo_marker = 0; 1962 tp->undo_retrans = -1; 1963 tp->sacked_out = 0; 1964 } 1965 1966 static inline void tcp_init_undo(struct tcp_sock *tp) 1967 { 1968 tp->undo_marker = tp->snd_una; 1969 /* Retransmission still in flight may cause DSACKs later. */ 1970 tp->undo_retrans = tp->retrans_out ? : -1; 1971 } 1972 1973 static bool tcp_is_rack(const struct sock *sk) 1974 { 1975 return sock_net(sk)->ipv4.sysctl_tcp_recovery & TCP_RACK_LOSS_DETECTION; 1976 } 1977 1978 /* If we detect SACK reneging, forget all SACK information 1979 * and reset tags completely, otherwise preserve SACKs. If receiver 1980 * dropped its ofo queue, we will know this due to reneging detection. 1981 */ 1982 static void tcp_timeout_mark_lost(struct sock *sk) 1983 { 1984 struct tcp_sock *tp = tcp_sk(sk); 1985 struct sk_buff *skb, *head; 1986 bool is_reneg; /* is receiver reneging on SACKs? */ 1987 1988 head = tcp_rtx_queue_head(sk); 1989 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED); 1990 if (is_reneg) { 1991 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING); 1992 tp->sacked_out = 0; 1993 /* Mark SACK reneging until we recover from this loss event. */ 1994 tp->is_sack_reneg = 1; 1995 } else if (tcp_is_reno(tp)) { 1996 tcp_reset_reno_sack(tp); 1997 } 1998 1999 skb = head; 2000 skb_rbtree_walk_from(skb) { 2001 if (is_reneg) 2002 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED; 2003 else if (tcp_is_rack(sk) && skb != head && 2004 tcp_rack_skb_timeout(tp, skb, 0) > 0) 2005 continue; /* Don't mark recently sent ones lost yet */ 2006 tcp_mark_skb_lost(sk, skb); 2007 } 2008 tcp_verify_left_out(tp); 2009 tcp_clear_all_retrans_hints(tp); 2010 } 2011 2012 /* Enter Loss state. */ 2013 void tcp_enter_loss(struct sock *sk) 2014 { 2015 const struct inet_connection_sock *icsk = inet_csk(sk); 2016 struct tcp_sock *tp = tcp_sk(sk); 2017 struct net *net = sock_net(sk); 2018 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery; 2019 2020 tcp_timeout_mark_lost(sk); 2021 2022 /* Reduce ssthresh if it has not yet been made inside this window. */ 2023 if (icsk->icsk_ca_state <= TCP_CA_Disorder || 2024 !after(tp->high_seq, tp->snd_una) || 2025 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) { 2026 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2027 tp->prior_cwnd = tp->snd_cwnd; 2028 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 2029 tcp_ca_event(sk, CA_EVENT_LOSS); 2030 tcp_init_undo(tp); 2031 } 2032 tp->snd_cwnd = tcp_packets_in_flight(tp) + 1; 2033 tp->snd_cwnd_cnt = 0; 2034 tp->snd_cwnd_stamp = tcp_jiffies32; 2035 2036 /* Timeout in disordered state after receiving substantial DUPACKs 2037 * suggests that the degree of reordering is over-estimated. 2038 */ 2039 if (icsk->icsk_ca_state <= TCP_CA_Disorder && 2040 tp->sacked_out >= net->ipv4.sysctl_tcp_reordering) 2041 tp->reordering = min_t(unsigned int, tp->reordering, 2042 net->ipv4.sysctl_tcp_reordering); 2043 tcp_set_ca_state(sk, TCP_CA_Loss); 2044 tp->high_seq = tp->snd_nxt; 2045 tcp_ecn_queue_cwr(tp); 2046 2047 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous 2048 * loss recovery is underway except recurring timeout(s) on 2049 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing 2050 */ 2051 tp->frto = net->ipv4.sysctl_tcp_frto && 2052 (new_recovery || icsk->icsk_retransmits) && 2053 !inet_csk(sk)->icsk_mtup.probe_size; 2054 } 2055 2056 /* If ACK arrived pointing to a remembered SACK, it means that our 2057 * remembered SACKs do not reflect real state of receiver i.e. 2058 * receiver _host_ is heavily congested (or buggy). 2059 * 2060 * To avoid big spurious retransmission bursts due to transient SACK 2061 * scoreboard oddities that look like reneging, we give the receiver a 2062 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will 2063 * restore sanity to the SACK scoreboard. If the apparent reneging 2064 * persists until this RTO then we'll clear the SACK scoreboard. 2065 */ 2066 static bool tcp_check_sack_reneging(struct sock *sk, int flag) 2067 { 2068 if (flag & FLAG_SACK_RENEGING) { 2069 struct tcp_sock *tp = tcp_sk(sk); 2070 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4), 2071 msecs_to_jiffies(10)); 2072 2073 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 2074 delay, TCP_RTO_MAX); 2075 return true; 2076 } 2077 return false; 2078 } 2079 2080 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs 2081 * counter when SACK is enabled (without SACK, sacked_out is used for 2082 * that purpose). 2083 * 2084 * With reordering, holes may still be in flight, so RFC3517 recovery 2085 * uses pure sacked_out (total number of SACKed segments) even though 2086 * it violates the RFC that uses duplicate ACKs, often these are equal 2087 * but when e.g. out-of-window ACKs or packet duplication occurs, 2088 * they differ. Since neither occurs due to loss, TCP should really 2089 * ignore them. 2090 */ 2091 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp) 2092 { 2093 return tp->sacked_out + 1; 2094 } 2095 2096 /* Linux NewReno/SACK/ECN state machine. 2097 * -------------------------------------- 2098 * 2099 * "Open" Normal state, no dubious events, fast path. 2100 * "Disorder" In all the respects it is "Open", 2101 * but requires a bit more attention. It is entered when 2102 * we see some SACKs or dupacks. It is split of "Open" 2103 * mainly to move some processing from fast path to slow one. 2104 * "CWR" CWND was reduced due to some Congestion Notification event. 2105 * It can be ECN, ICMP source quench, local device congestion. 2106 * "Recovery" CWND was reduced, we are fast-retransmitting. 2107 * "Loss" CWND was reduced due to RTO timeout or SACK reneging. 2108 * 2109 * tcp_fastretrans_alert() is entered: 2110 * - each incoming ACK, if state is not "Open" 2111 * - when arrived ACK is unusual, namely: 2112 * * SACK 2113 * * Duplicate ACK. 2114 * * ECN ECE. 2115 * 2116 * Counting packets in flight is pretty simple. 2117 * 2118 * in_flight = packets_out - left_out + retrans_out 2119 * 2120 * packets_out is SND.NXT-SND.UNA counted in packets. 2121 * 2122 * retrans_out is number of retransmitted segments. 2123 * 2124 * left_out is number of segments left network, but not ACKed yet. 2125 * 2126 * left_out = sacked_out + lost_out 2127 * 2128 * sacked_out: Packets, which arrived to receiver out of order 2129 * and hence not ACKed. With SACKs this number is simply 2130 * amount of SACKed data. Even without SACKs 2131 * it is easy to give pretty reliable estimate of this number, 2132 * counting duplicate ACKs. 2133 * 2134 * lost_out: Packets lost by network. TCP has no explicit 2135 * "loss notification" feedback from network (for now). 2136 * It means that this number can be only _guessed_. 2137 * Actually, it is the heuristics to predict lossage that 2138 * distinguishes different algorithms. 2139 * 2140 * F.e. after RTO, when all the queue is considered as lost, 2141 * lost_out = packets_out and in_flight = retrans_out. 2142 * 2143 * Essentially, we have now a few algorithms detecting 2144 * lost packets. 2145 * 2146 * If the receiver supports SACK: 2147 * 2148 * RFC6675/3517: It is the conventional algorithm. A packet is 2149 * considered lost if the number of higher sequence packets 2150 * SACKed is greater than or equal the DUPACK thoreshold 2151 * (reordering). This is implemented in tcp_mark_head_lost and 2152 * tcp_update_scoreboard. 2153 * 2154 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm 2155 * (2017-) that checks timing instead of counting DUPACKs. 2156 * Essentially a packet is considered lost if it's not S/ACKed 2157 * after RTT + reordering_window, where both metrics are 2158 * dynamically measured and adjusted. This is implemented in 2159 * tcp_rack_mark_lost. 2160 * 2161 * If the receiver does not support SACK: 2162 * 2163 * NewReno (RFC6582): in Recovery we assume that one segment 2164 * is lost (classic Reno). While we are in Recovery and 2165 * a partial ACK arrives, we assume that one more packet 2166 * is lost (NewReno). This heuristics are the same in NewReno 2167 * and SACK. 2168 * 2169 * Really tricky (and requiring careful tuning) part of algorithm 2170 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue(). 2171 * The first determines the moment _when_ we should reduce CWND and, 2172 * hence, slow down forward transmission. In fact, it determines the moment 2173 * when we decide that hole is caused by loss, rather than by a reorder. 2174 * 2175 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill 2176 * holes, caused by lost packets. 2177 * 2178 * And the most logically complicated part of algorithm is undo 2179 * heuristics. We detect false retransmits due to both too early 2180 * fast retransmit (reordering) and underestimated RTO, analyzing 2181 * timestamps and D-SACKs. When we detect that some segments were 2182 * retransmitted by mistake and CWND reduction was wrong, we undo 2183 * window reduction and abort recovery phase. This logic is hidden 2184 * inside several functions named tcp_try_undo_<something>. 2185 */ 2186 2187 /* This function decides, when we should leave Disordered state 2188 * and enter Recovery phase, reducing congestion window. 2189 * 2190 * Main question: may we further continue forward transmission 2191 * with the same cwnd? 2192 */ 2193 static bool tcp_time_to_recover(struct sock *sk, int flag) 2194 { 2195 struct tcp_sock *tp = tcp_sk(sk); 2196 2197 /* Trick#1: The loss is proven. */ 2198 if (tp->lost_out) 2199 return true; 2200 2201 /* Not-A-Trick#2 : Classic rule... */ 2202 if (!tcp_is_rack(sk) && tcp_dupack_heuristics(tp) > tp->reordering) 2203 return true; 2204 2205 return false; 2206 } 2207 2208 /* Detect loss in event "A" above by marking head of queue up as lost. 2209 * For RFC3517 SACK, a segment is considered lost if it 2210 * has at least tp->reordering SACKed seqments above it; "packets" refers to 2211 * the maximum SACKed segments to pass before reaching this limit. 2212 */ 2213 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head) 2214 { 2215 struct tcp_sock *tp = tcp_sk(sk); 2216 struct sk_buff *skb; 2217 int cnt; 2218 /* Use SACK to deduce losses of new sequences sent during recovery */ 2219 const u32 loss_high = tp->snd_nxt; 2220 2221 WARN_ON(packets > tp->packets_out); 2222 skb = tp->lost_skb_hint; 2223 if (skb) { 2224 /* Head already handled? */ 2225 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una)) 2226 return; 2227 cnt = tp->lost_cnt_hint; 2228 } else { 2229 skb = tcp_rtx_queue_head(sk); 2230 cnt = 0; 2231 } 2232 2233 skb_rbtree_walk_from(skb) { 2234 /* TODO: do this better */ 2235 /* this is not the most efficient way to do this... */ 2236 tp->lost_skb_hint = skb; 2237 tp->lost_cnt_hint = cnt; 2238 2239 if (after(TCP_SKB_CB(skb)->end_seq, loss_high)) 2240 break; 2241 2242 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2243 cnt += tcp_skb_pcount(skb); 2244 2245 if (cnt > packets) 2246 break; 2247 2248 tcp_skb_mark_lost(tp, skb); 2249 2250 if (mark_head) 2251 break; 2252 } 2253 tcp_verify_left_out(tp); 2254 } 2255 2256 /* Account newly detected lost packet(s) */ 2257 2258 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit) 2259 { 2260 struct tcp_sock *tp = tcp_sk(sk); 2261 2262 if (tcp_is_sack(tp)) { 2263 int sacked_upto = tp->sacked_out - tp->reordering; 2264 if (sacked_upto >= 0) 2265 tcp_mark_head_lost(sk, sacked_upto, 0); 2266 else if (fast_rexmit) 2267 tcp_mark_head_lost(sk, 1, 1); 2268 } 2269 } 2270 2271 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when) 2272 { 2273 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 2274 before(tp->rx_opt.rcv_tsecr, when); 2275 } 2276 2277 /* skb is spurious retransmitted if the returned timestamp echo 2278 * reply is prior to the skb transmission time 2279 */ 2280 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp, 2281 const struct sk_buff *skb) 2282 { 2283 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) && 2284 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb)); 2285 } 2286 2287 /* Nothing was retransmitted or returned timestamp is less 2288 * than timestamp of the first retransmission. 2289 */ 2290 static inline bool tcp_packet_delayed(const struct tcp_sock *tp) 2291 { 2292 return tp->retrans_stamp && 2293 tcp_tsopt_ecr_before(tp, tp->retrans_stamp); 2294 } 2295 2296 /* Undo procedures. */ 2297 2298 /* We can clear retrans_stamp when there are no retransmissions in the 2299 * window. It would seem that it is trivially available for us in 2300 * tp->retrans_out, however, that kind of assumptions doesn't consider 2301 * what will happen if errors occur when sending retransmission for the 2302 * second time. ...It could the that such segment has only 2303 * TCPCB_EVER_RETRANS set at the present time. It seems that checking 2304 * the head skb is enough except for some reneging corner cases that 2305 * are not worth the effort. 2306 * 2307 * Main reason for all this complexity is the fact that connection dying 2308 * time now depends on the validity of the retrans_stamp, in particular, 2309 * that successive retransmissions of a segment must not advance 2310 * retrans_stamp under any conditions. 2311 */ 2312 static bool tcp_any_retrans_done(const struct sock *sk) 2313 { 2314 const struct tcp_sock *tp = tcp_sk(sk); 2315 struct sk_buff *skb; 2316 2317 if (tp->retrans_out) 2318 return true; 2319 2320 skb = tcp_rtx_queue_head(sk); 2321 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS)) 2322 return true; 2323 2324 return false; 2325 } 2326 2327 static void DBGUNDO(struct sock *sk, const char *msg) 2328 { 2329 #if FASTRETRANS_DEBUG > 1 2330 struct tcp_sock *tp = tcp_sk(sk); 2331 struct inet_sock *inet = inet_sk(sk); 2332 2333 if (sk->sk_family == AF_INET) { 2334 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n", 2335 msg, 2336 &inet->inet_daddr, ntohs(inet->inet_dport), 2337 tp->snd_cwnd, tcp_left_out(tp), 2338 tp->snd_ssthresh, tp->prior_ssthresh, 2339 tp->packets_out); 2340 } 2341 #if IS_ENABLED(CONFIG_IPV6) 2342 else if (sk->sk_family == AF_INET6) { 2343 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n", 2344 msg, 2345 &sk->sk_v6_daddr, ntohs(inet->inet_dport), 2346 tp->snd_cwnd, tcp_left_out(tp), 2347 tp->snd_ssthresh, tp->prior_ssthresh, 2348 tp->packets_out); 2349 } 2350 #endif 2351 #endif 2352 } 2353 2354 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss) 2355 { 2356 struct tcp_sock *tp = tcp_sk(sk); 2357 2358 if (unmark_loss) { 2359 struct sk_buff *skb; 2360 2361 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) { 2362 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 2363 } 2364 tp->lost_out = 0; 2365 tcp_clear_all_retrans_hints(tp); 2366 } 2367 2368 if (tp->prior_ssthresh) { 2369 const struct inet_connection_sock *icsk = inet_csk(sk); 2370 2371 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk); 2372 2373 if (tp->prior_ssthresh > tp->snd_ssthresh) { 2374 tp->snd_ssthresh = tp->prior_ssthresh; 2375 tcp_ecn_withdraw_cwr(tp); 2376 } 2377 } 2378 tp->snd_cwnd_stamp = tcp_jiffies32; 2379 tp->undo_marker = 0; 2380 tp->rack.advanced = 1; /* Force RACK to re-exam losses */ 2381 } 2382 2383 static inline bool tcp_may_undo(const struct tcp_sock *tp) 2384 { 2385 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp)); 2386 } 2387 2388 /* People celebrate: "We love our President!" */ 2389 static bool tcp_try_undo_recovery(struct sock *sk) 2390 { 2391 struct tcp_sock *tp = tcp_sk(sk); 2392 2393 if (tcp_may_undo(tp)) { 2394 int mib_idx; 2395 2396 /* Happy end! We did not retransmit anything 2397 * or our original transmission succeeded. 2398 */ 2399 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans"); 2400 tcp_undo_cwnd_reduction(sk, false); 2401 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss) 2402 mib_idx = LINUX_MIB_TCPLOSSUNDO; 2403 else 2404 mib_idx = LINUX_MIB_TCPFULLUNDO; 2405 2406 NET_INC_STATS(sock_net(sk), mib_idx); 2407 } else if (tp->rack.reo_wnd_persist) { 2408 tp->rack.reo_wnd_persist--; 2409 } 2410 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) { 2411 /* Hold old state until something *above* high_seq 2412 * is ACKed. For Reno it is MUST to prevent false 2413 * fast retransmits (RFC2582). SACK TCP is safe. */ 2414 if (!tcp_any_retrans_done(sk)) 2415 tp->retrans_stamp = 0; 2416 return true; 2417 } 2418 tcp_set_ca_state(sk, TCP_CA_Open); 2419 tp->is_sack_reneg = 0; 2420 return false; 2421 } 2422 2423 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */ 2424 static bool tcp_try_undo_dsack(struct sock *sk) 2425 { 2426 struct tcp_sock *tp = tcp_sk(sk); 2427 2428 if (tp->undo_marker && !tp->undo_retrans) { 2429 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH, 2430 tp->rack.reo_wnd_persist + 1); 2431 DBGUNDO(sk, "D-SACK"); 2432 tcp_undo_cwnd_reduction(sk, false); 2433 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO); 2434 return true; 2435 } 2436 return false; 2437 } 2438 2439 /* Undo during loss recovery after partial ACK or using F-RTO. */ 2440 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo) 2441 { 2442 struct tcp_sock *tp = tcp_sk(sk); 2443 2444 if (frto_undo || tcp_may_undo(tp)) { 2445 tcp_undo_cwnd_reduction(sk, true); 2446 2447 DBGUNDO(sk, "partial loss"); 2448 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO); 2449 if (frto_undo) 2450 NET_INC_STATS(sock_net(sk), 2451 LINUX_MIB_TCPSPURIOUSRTOS); 2452 inet_csk(sk)->icsk_retransmits = 0; 2453 if (frto_undo || tcp_is_sack(tp)) { 2454 tcp_set_ca_state(sk, TCP_CA_Open); 2455 tp->is_sack_reneg = 0; 2456 } 2457 return true; 2458 } 2459 return false; 2460 } 2461 2462 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937. 2463 * It computes the number of packets to send (sndcnt) based on packets newly 2464 * delivered: 2465 * 1) If the packets in flight is larger than ssthresh, PRR spreads the 2466 * cwnd reductions across a full RTT. 2467 * 2) Otherwise PRR uses packet conservation to send as much as delivered. 2468 * But when the retransmits are acked without further losses, PRR 2469 * slow starts cwnd up to ssthresh to speed up the recovery. 2470 */ 2471 static void tcp_init_cwnd_reduction(struct sock *sk) 2472 { 2473 struct tcp_sock *tp = tcp_sk(sk); 2474 2475 tp->high_seq = tp->snd_nxt; 2476 tp->tlp_high_seq = 0; 2477 tp->snd_cwnd_cnt = 0; 2478 tp->prior_cwnd = tp->snd_cwnd; 2479 tp->prr_delivered = 0; 2480 tp->prr_out = 0; 2481 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk); 2482 tcp_ecn_queue_cwr(tp); 2483 } 2484 2485 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int flag) 2486 { 2487 struct tcp_sock *tp = tcp_sk(sk); 2488 int sndcnt = 0; 2489 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp); 2490 2491 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd)) 2492 return; 2493 2494 tp->prr_delivered += newly_acked_sacked; 2495 if (delta < 0) { 2496 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered + 2497 tp->prior_cwnd - 1; 2498 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out; 2499 } else if ((flag & (FLAG_RETRANS_DATA_ACKED | FLAG_LOST_RETRANS)) == 2500 FLAG_RETRANS_DATA_ACKED) { 2501 sndcnt = min_t(int, delta, 2502 max_t(int, tp->prr_delivered - tp->prr_out, 2503 newly_acked_sacked) + 1); 2504 } else { 2505 sndcnt = min(delta, newly_acked_sacked); 2506 } 2507 /* Force a fast retransmit upon entering fast recovery */ 2508 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1)); 2509 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt; 2510 } 2511 2512 static inline void tcp_end_cwnd_reduction(struct sock *sk) 2513 { 2514 struct tcp_sock *tp = tcp_sk(sk); 2515 2516 if (inet_csk(sk)->icsk_ca_ops->cong_control) 2517 return; 2518 2519 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */ 2520 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH && 2521 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) { 2522 tp->snd_cwnd = tp->snd_ssthresh; 2523 tp->snd_cwnd_stamp = tcp_jiffies32; 2524 } 2525 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR); 2526 } 2527 2528 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */ 2529 void tcp_enter_cwr(struct sock *sk) 2530 { 2531 struct tcp_sock *tp = tcp_sk(sk); 2532 2533 tp->prior_ssthresh = 0; 2534 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) { 2535 tp->undo_marker = 0; 2536 tcp_init_cwnd_reduction(sk); 2537 tcp_set_ca_state(sk, TCP_CA_CWR); 2538 } 2539 } 2540 EXPORT_SYMBOL(tcp_enter_cwr); 2541 2542 static void tcp_try_keep_open(struct sock *sk) 2543 { 2544 struct tcp_sock *tp = tcp_sk(sk); 2545 int state = TCP_CA_Open; 2546 2547 if (tcp_left_out(tp) || tcp_any_retrans_done(sk)) 2548 state = TCP_CA_Disorder; 2549 2550 if (inet_csk(sk)->icsk_ca_state != state) { 2551 tcp_set_ca_state(sk, state); 2552 tp->high_seq = tp->snd_nxt; 2553 } 2554 } 2555 2556 static void tcp_try_to_open(struct sock *sk, int flag) 2557 { 2558 struct tcp_sock *tp = tcp_sk(sk); 2559 2560 tcp_verify_left_out(tp); 2561 2562 if (!tcp_any_retrans_done(sk)) 2563 tp->retrans_stamp = 0; 2564 2565 if (flag & FLAG_ECE) 2566 tcp_enter_cwr(sk); 2567 2568 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) { 2569 tcp_try_keep_open(sk); 2570 } 2571 } 2572 2573 static void tcp_mtup_probe_failed(struct sock *sk) 2574 { 2575 struct inet_connection_sock *icsk = inet_csk(sk); 2576 2577 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1; 2578 icsk->icsk_mtup.probe_size = 0; 2579 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL); 2580 } 2581 2582 static void tcp_mtup_probe_success(struct sock *sk) 2583 { 2584 struct tcp_sock *tp = tcp_sk(sk); 2585 struct inet_connection_sock *icsk = inet_csk(sk); 2586 2587 /* FIXME: breaks with very large cwnd */ 2588 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2589 tp->snd_cwnd = tp->snd_cwnd * 2590 tcp_mss_to_mtu(sk, tp->mss_cache) / 2591 icsk->icsk_mtup.probe_size; 2592 tp->snd_cwnd_cnt = 0; 2593 tp->snd_cwnd_stamp = tcp_jiffies32; 2594 tp->snd_ssthresh = tcp_current_ssthresh(sk); 2595 2596 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size; 2597 icsk->icsk_mtup.probe_size = 0; 2598 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 2599 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS); 2600 } 2601 2602 /* Do a simple retransmit without using the backoff mechanisms in 2603 * tcp_timer. This is used for path mtu discovery. 2604 * The socket is already locked here. 2605 */ 2606 void tcp_simple_retransmit(struct sock *sk) 2607 { 2608 const struct inet_connection_sock *icsk = inet_csk(sk); 2609 struct tcp_sock *tp = tcp_sk(sk); 2610 struct sk_buff *skb; 2611 unsigned int mss = tcp_current_mss(sk); 2612 2613 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) { 2614 if (tcp_skb_seglen(skb) > mss && 2615 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { 2616 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2617 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 2618 tp->retrans_out -= tcp_skb_pcount(skb); 2619 } 2620 tcp_skb_mark_lost_uncond_verify(tp, skb); 2621 } 2622 } 2623 2624 tcp_clear_retrans_hints_partial(tp); 2625 2626 if (!tp->lost_out) 2627 return; 2628 2629 if (tcp_is_reno(tp)) 2630 tcp_limit_reno_sacked(tp); 2631 2632 tcp_verify_left_out(tp); 2633 2634 /* Don't muck with the congestion window here. 2635 * Reason is that we do not increase amount of _data_ 2636 * in network, but units changed and effective 2637 * cwnd/ssthresh really reduced now. 2638 */ 2639 if (icsk->icsk_ca_state != TCP_CA_Loss) { 2640 tp->high_seq = tp->snd_nxt; 2641 tp->snd_ssthresh = tcp_current_ssthresh(sk); 2642 tp->prior_ssthresh = 0; 2643 tp->undo_marker = 0; 2644 tcp_set_ca_state(sk, TCP_CA_Loss); 2645 } 2646 tcp_xmit_retransmit_queue(sk); 2647 } 2648 EXPORT_SYMBOL(tcp_simple_retransmit); 2649 2650 void tcp_enter_recovery(struct sock *sk, bool ece_ack) 2651 { 2652 struct tcp_sock *tp = tcp_sk(sk); 2653 int mib_idx; 2654 2655 if (tcp_is_reno(tp)) 2656 mib_idx = LINUX_MIB_TCPRENORECOVERY; 2657 else 2658 mib_idx = LINUX_MIB_TCPSACKRECOVERY; 2659 2660 NET_INC_STATS(sock_net(sk), mib_idx); 2661 2662 tp->prior_ssthresh = 0; 2663 tcp_init_undo(tp); 2664 2665 if (!tcp_in_cwnd_reduction(sk)) { 2666 if (!ece_ack) 2667 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2668 tcp_init_cwnd_reduction(sk); 2669 } 2670 tcp_set_ca_state(sk, TCP_CA_Recovery); 2671 } 2672 2673 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are 2674 * recovered or spurious. Otherwise retransmits more on partial ACKs. 2675 */ 2676 static void tcp_process_loss(struct sock *sk, int flag, int num_dupack, 2677 int *rexmit) 2678 { 2679 struct tcp_sock *tp = tcp_sk(sk); 2680 bool recovered = !before(tp->snd_una, tp->high_seq); 2681 2682 if ((flag & FLAG_SND_UNA_ADVANCED || rcu_access_pointer(tp->fastopen_rsk)) && 2683 tcp_try_undo_loss(sk, false)) 2684 return; 2685 2686 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */ 2687 /* Step 3.b. A timeout is spurious if not all data are 2688 * lost, i.e., never-retransmitted data are (s)acked. 2689 */ 2690 if ((flag & FLAG_ORIG_SACK_ACKED) && 2691 tcp_try_undo_loss(sk, true)) 2692 return; 2693 2694 if (after(tp->snd_nxt, tp->high_seq)) { 2695 if (flag & FLAG_DATA_SACKED || num_dupack) 2696 tp->frto = 0; /* Step 3.a. loss was real */ 2697 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) { 2698 tp->high_seq = tp->snd_nxt; 2699 /* Step 2.b. Try send new data (but deferred until cwnd 2700 * is updated in tcp_ack()). Otherwise fall back to 2701 * the conventional recovery. 2702 */ 2703 if (!tcp_write_queue_empty(sk) && 2704 after(tcp_wnd_end(tp), tp->snd_nxt)) { 2705 *rexmit = REXMIT_NEW; 2706 return; 2707 } 2708 tp->frto = 0; 2709 } 2710 } 2711 2712 if (recovered) { 2713 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */ 2714 tcp_try_undo_recovery(sk); 2715 return; 2716 } 2717 if (tcp_is_reno(tp)) { 2718 /* A Reno DUPACK means new data in F-RTO step 2.b above are 2719 * delivered. Lower inflight to clock out (re)tranmissions. 2720 */ 2721 if (after(tp->snd_nxt, tp->high_seq) && num_dupack) 2722 tcp_add_reno_sack(sk, num_dupack, flag & FLAG_ECE); 2723 else if (flag & FLAG_SND_UNA_ADVANCED) 2724 tcp_reset_reno_sack(tp); 2725 } 2726 *rexmit = REXMIT_LOST; 2727 } 2728 2729 /* Undo during fast recovery after partial ACK. */ 2730 static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una) 2731 { 2732 struct tcp_sock *tp = tcp_sk(sk); 2733 2734 if (tp->undo_marker && tcp_packet_delayed(tp)) { 2735 /* Plain luck! Hole if filled with delayed 2736 * packet, rather than with a retransmit. Check reordering. 2737 */ 2738 tcp_check_sack_reordering(sk, prior_snd_una, 1); 2739 2740 /* We are getting evidence that the reordering degree is higher 2741 * than we realized. If there are no retransmits out then we 2742 * can undo. Otherwise we clock out new packets but do not 2743 * mark more packets lost or retransmit more. 2744 */ 2745 if (tp->retrans_out) 2746 return true; 2747 2748 if (!tcp_any_retrans_done(sk)) 2749 tp->retrans_stamp = 0; 2750 2751 DBGUNDO(sk, "partial recovery"); 2752 tcp_undo_cwnd_reduction(sk, true); 2753 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO); 2754 tcp_try_keep_open(sk); 2755 return true; 2756 } 2757 return false; 2758 } 2759 2760 static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag) 2761 { 2762 struct tcp_sock *tp = tcp_sk(sk); 2763 2764 if (tcp_rtx_queue_empty(sk)) 2765 return; 2766 2767 if (unlikely(tcp_is_reno(tp))) { 2768 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED); 2769 } else if (tcp_is_rack(sk)) { 2770 u32 prior_retrans = tp->retrans_out; 2771 2772 tcp_rack_mark_lost(sk); 2773 if (prior_retrans > tp->retrans_out) 2774 *ack_flag |= FLAG_LOST_RETRANS; 2775 } 2776 } 2777 2778 static bool tcp_force_fast_retransmit(struct sock *sk) 2779 { 2780 struct tcp_sock *tp = tcp_sk(sk); 2781 2782 return after(tcp_highest_sack_seq(tp), 2783 tp->snd_una + tp->reordering * tp->mss_cache); 2784 } 2785 2786 /* Process an event, which can update packets-in-flight not trivially. 2787 * Main goal of this function is to calculate new estimate for left_out, 2788 * taking into account both packets sitting in receiver's buffer and 2789 * packets lost by network. 2790 * 2791 * Besides that it updates the congestion state when packet loss or ECN 2792 * is detected. But it does not reduce the cwnd, it is done by the 2793 * congestion control later. 2794 * 2795 * It does _not_ decide what to send, it is made in function 2796 * tcp_xmit_retransmit_queue(). 2797 */ 2798 static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una, 2799 int num_dupack, int *ack_flag, int *rexmit) 2800 { 2801 struct inet_connection_sock *icsk = inet_csk(sk); 2802 struct tcp_sock *tp = tcp_sk(sk); 2803 int fast_rexmit = 0, flag = *ack_flag; 2804 bool ece_ack = flag & FLAG_ECE; 2805 bool do_lost = num_dupack || ((flag & FLAG_DATA_SACKED) && 2806 tcp_force_fast_retransmit(sk)); 2807 2808 if (!tp->packets_out && tp->sacked_out) 2809 tp->sacked_out = 0; 2810 2811 /* Now state machine starts. 2812 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */ 2813 if (ece_ack) 2814 tp->prior_ssthresh = 0; 2815 2816 /* B. In all the states check for reneging SACKs. */ 2817 if (tcp_check_sack_reneging(sk, flag)) 2818 return; 2819 2820 /* C. Check consistency of the current state. */ 2821 tcp_verify_left_out(tp); 2822 2823 /* D. Check state exit conditions. State can be terminated 2824 * when high_seq is ACKed. */ 2825 if (icsk->icsk_ca_state == TCP_CA_Open) { 2826 WARN_ON(tp->retrans_out != 0); 2827 tp->retrans_stamp = 0; 2828 } else if (!before(tp->snd_una, tp->high_seq)) { 2829 switch (icsk->icsk_ca_state) { 2830 case TCP_CA_CWR: 2831 /* CWR is to be held something *above* high_seq 2832 * is ACKed for CWR bit to reach receiver. */ 2833 if (tp->snd_una != tp->high_seq) { 2834 tcp_end_cwnd_reduction(sk); 2835 tcp_set_ca_state(sk, TCP_CA_Open); 2836 } 2837 break; 2838 2839 case TCP_CA_Recovery: 2840 if (tcp_is_reno(tp)) 2841 tcp_reset_reno_sack(tp); 2842 if (tcp_try_undo_recovery(sk)) 2843 return; 2844 tcp_end_cwnd_reduction(sk); 2845 break; 2846 } 2847 } 2848 2849 /* E. Process state. */ 2850 switch (icsk->icsk_ca_state) { 2851 case TCP_CA_Recovery: 2852 if (!(flag & FLAG_SND_UNA_ADVANCED)) { 2853 if (tcp_is_reno(tp)) 2854 tcp_add_reno_sack(sk, num_dupack, ece_ack); 2855 } else { 2856 if (tcp_try_undo_partial(sk, prior_snd_una)) 2857 return; 2858 /* Partial ACK arrived. Force fast retransmit. */ 2859 do_lost = tcp_force_fast_retransmit(sk); 2860 } 2861 if (tcp_try_undo_dsack(sk)) { 2862 tcp_try_keep_open(sk); 2863 return; 2864 } 2865 tcp_identify_packet_loss(sk, ack_flag); 2866 break; 2867 case TCP_CA_Loss: 2868 tcp_process_loss(sk, flag, num_dupack, rexmit); 2869 tcp_identify_packet_loss(sk, ack_flag); 2870 if (!(icsk->icsk_ca_state == TCP_CA_Open || 2871 (*ack_flag & FLAG_LOST_RETRANS))) 2872 return; 2873 /* Change state if cwnd is undone or retransmits are lost */ 2874 fallthrough; 2875 default: 2876 if (tcp_is_reno(tp)) { 2877 if (flag & FLAG_SND_UNA_ADVANCED) 2878 tcp_reset_reno_sack(tp); 2879 tcp_add_reno_sack(sk, num_dupack, ece_ack); 2880 } 2881 2882 if (icsk->icsk_ca_state <= TCP_CA_Disorder) 2883 tcp_try_undo_dsack(sk); 2884 2885 tcp_identify_packet_loss(sk, ack_flag); 2886 if (!tcp_time_to_recover(sk, flag)) { 2887 tcp_try_to_open(sk, flag); 2888 return; 2889 } 2890 2891 /* MTU probe failure: don't reduce cwnd */ 2892 if (icsk->icsk_ca_state < TCP_CA_CWR && 2893 icsk->icsk_mtup.probe_size && 2894 tp->snd_una == tp->mtu_probe.probe_seq_start) { 2895 tcp_mtup_probe_failed(sk); 2896 /* Restores the reduction we did in tcp_mtup_probe() */ 2897 tp->snd_cwnd++; 2898 tcp_simple_retransmit(sk); 2899 return; 2900 } 2901 2902 /* Otherwise enter Recovery state */ 2903 tcp_enter_recovery(sk, ece_ack); 2904 fast_rexmit = 1; 2905 } 2906 2907 if (!tcp_is_rack(sk) && do_lost) 2908 tcp_update_scoreboard(sk, fast_rexmit); 2909 *rexmit = REXMIT_LOST; 2910 } 2911 2912 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag) 2913 { 2914 u32 wlen = sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen * HZ; 2915 struct tcp_sock *tp = tcp_sk(sk); 2916 2917 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) { 2918 /* If the remote keeps returning delayed ACKs, eventually 2919 * the min filter would pick it up and overestimate the 2920 * prop. delay when it expires. Skip suspected delayed ACKs. 2921 */ 2922 return; 2923 } 2924 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32, 2925 rtt_us ? : jiffies_to_usecs(1)); 2926 } 2927 2928 static bool tcp_ack_update_rtt(struct sock *sk, const int flag, 2929 long seq_rtt_us, long sack_rtt_us, 2930 long ca_rtt_us, struct rate_sample *rs) 2931 { 2932 const struct tcp_sock *tp = tcp_sk(sk); 2933 2934 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because 2935 * broken middle-boxes or peers may corrupt TS-ECR fields. But 2936 * Karn's algorithm forbids taking RTT if some retransmitted data 2937 * is acked (RFC6298). 2938 */ 2939 if (seq_rtt_us < 0) 2940 seq_rtt_us = sack_rtt_us; 2941 2942 /* RTTM Rule: A TSecr value received in a segment is used to 2943 * update the averaged RTT measurement only if the segment 2944 * acknowledges some new data, i.e., only if it advances the 2945 * left edge of the send window. 2946 * See draft-ietf-tcplw-high-performance-00, section 3.3. 2947 */ 2948 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 2949 flag & FLAG_ACKED) { 2950 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr; 2951 2952 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) { 2953 if (!delta) 2954 delta = 1; 2955 seq_rtt_us = delta * (USEC_PER_SEC / TCP_TS_HZ); 2956 ca_rtt_us = seq_rtt_us; 2957 } 2958 } 2959 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */ 2960 if (seq_rtt_us < 0) 2961 return false; 2962 2963 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is 2964 * always taken together with ACK, SACK, or TS-opts. Any negative 2965 * values will be skipped with the seq_rtt_us < 0 check above. 2966 */ 2967 tcp_update_rtt_min(sk, ca_rtt_us, flag); 2968 tcp_rtt_estimator(sk, seq_rtt_us); 2969 tcp_set_rto(sk); 2970 2971 /* RFC6298: only reset backoff on valid RTT measurement. */ 2972 inet_csk(sk)->icsk_backoff = 0; 2973 return true; 2974 } 2975 2976 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */ 2977 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req) 2978 { 2979 struct rate_sample rs; 2980 long rtt_us = -1L; 2981 2982 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack) 2983 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack); 2984 2985 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs); 2986 } 2987 2988 2989 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked) 2990 { 2991 const struct inet_connection_sock *icsk = inet_csk(sk); 2992 2993 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked); 2994 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32; 2995 } 2996 2997 /* Restart timer after forward progress on connection. 2998 * RFC2988 recommends to restart timer to now+rto. 2999 */ 3000 void tcp_rearm_rto(struct sock *sk) 3001 { 3002 const struct inet_connection_sock *icsk = inet_csk(sk); 3003 struct tcp_sock *tp = tcp_sk(sk); 3004 3005 /* If the retrans timer is currently being used by Fast Open 3006 * for SYN-ACK retrans purpose, stay put. 3007 */ 3008 if (rcu_access_pointer(tp->fastopen_rsk)) 3009 return; 3010 3011 if (!tp->packets_out) { 3012 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS); 3013 } else { 3014 u32 rto = inet_csk(sk)->icsk_rto; 3015 /* Offset the time elapsed after installing regular RTO */ 3016 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT || 3017 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) { 3018 s64 delta_us = tcp_rto_delta_us(sk); 3019 /* delta_us may not be positive if the socket is locked 3020 * when the retrans timer fires and is rescheduled. 3021 */ 3022 rto = usecs_to_jiffies(max_t(int, delta_us, 1)); 3023 } 3024 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto, 3025 TCP_RTO_MAX); 3026 } 3027 } 3028 3029 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */ 3030 static void tcp_set_xmit_timer(struct sock *sk) 3031 { 3032 if (!tcp_schedule_loss_probe(sk, true)) 3033 tcp_rearm_rto(sk); 3034 } 3035 3036 /* If we get here, the whole TSO packet has not been acked. */ 3037 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb) 3038 { 3039 struct tcp_sock *tp = tcp_sk(sk); 3040 u32 packets_acked; 3041 3042 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)); 3043 3044 packets_acked = tcp_skb_pcount(skb); 3045 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 3046 return 0; 3047 packets_acked -= tcp_skb_pcount(skb); 3048 3049 if (packets_acked) { 3050 BUG_ON(tcp_skb_pcount(skb) == 0); 3051 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)); 3052 } 3053 3054 return packets_acked; 3055 } 3056 3057 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb, 3058 u32 prior_snd_una) 3059 { 3060 const struct skb_shared_info *shinfo; 3061 3062 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */ 3063 if (likely(!TCP_SKB_CB(skb)->txstamp_ack)) 3064 return; 3065 3066 shinfo = skb_shinfo(skb); 3067 if (!before(shinfo->tskey, prior_snd_una) && 3068 before(shinfo->tskey, tcp_sk(sk)->snd_una)) { 3069 tcp_skb_tsorted_save(skb) { 3070 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK); 3071 } tcp_skb_tsorted_restore(skb); 3072 } 3073 } 3074 3075 /* Remove acknowledged frames from the retransmission queue. If our packet 3076 * is before the ack sequence we can discard it as it's confirmed to have 3077 * arrived at the other end. 3078 */ 3079 static int tcp_clean_rtx_queue(struct sock *sk, u32 prior_fack, 3080 u32 prior_snd_una, 3081 struct tcp_sacktag_state *sack, bool ece_ack) 3082 { 3083 const struct inet_connection_sock *icsk = inet_csk(sk); 3084 u64 first_ackt, last_ackt; 3085 struct tcp_sock *tp = tcp_sk(sk); 3086 u32 prior_sacked = tp->sacked_out; 3087 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */ 3088 struct sk_buff *skb, *next; 3089 bool fully_acked = true; 3090 long sack_rtt_us = -1L; 3091 long seq_rtt_us = -1L; 3092 long ca_rtt_us = -1L; 3093 u32 pkts_acked = 0; 3094 u32 last_in_flight = 0; 3095 bool rtt_update; 3096 int flag = 0; 3097 3098 first_ackt = 0; 3099 3100 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) { 3101 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 3102 const u32 start_seq = scb->seq; 3103 u8 sacked = scb->sacked; 3104 u32 acked_pcount; 3105 3106 /* Determine how many packets and what bytes were acked, tso and else */ 3107 if (after(scb->end_seq, tp->snd_una)) { 3108 if (tcp_skb_pcount(skb) == 1 || 3109 !after(tp->snd_una, scb->seq)) 3110 break; 3111 3112 acked_pcount = tcp_tso_acked(sk, skb); 3113 if (!acked_pcount) 3114 break; 3115 fully_acked = false; 3116 } else { 3117 acked_pcount = tcp_skb_pcount(skb); 3118 } 3119 3120 if (unlikely(sacked & TCPCB_RETRANS)) { 3121 if (sacked & TCPCB_SACKED_RETRANS) 3122 tp->retrans_out -= acked_pcount; 3123 flag |= FLAG_RETRANS_DATA_ACKED; 3124 } else if (!(sacked & TCPCB_SACKED_ACKED)) { 3125 last_ackt = tcp_skb_timestamp_us(skb); 3126 WARN_ON_ONCE(last_ackt == 0); 3127 if (!first_ackt) 3128 first_ackt = last_ackt; 3129 3130 last_in_flight = TCP_SKB_CB(skb)->tx.in_flight; 3131 if (before(start_seq, reord)) 3132 reord = start_seq; 3133 if (!after(scb->end_seq, tp->high_seq)) 3134 flag |= FLAG_ORIG_SACK_ACKED; 3135 } 3136 3137 if (sacked & TCPCB_SACKED_ACKED) { 3138 tp->sacked_out -= acked_pcount; 3139 } else if (tcp_is_sack(tp)) { 3140 tcp_count_delivered(tp, acked_pcount, ece_ack); 3141 if (!tcp_skb_spurious_retrans(tp, skb)) 3142 tcp_rack_advance(tp, sacked, scb->end_seq, 3143 tcp_skb_timestamp_us(skb)); 3144 } 3145 if (sacked & TCPCB_LOST) 3146 tp->lost_out -= acked_pcount; 3147 3148 tp->packets_out -= acked_pcount; 3149 pkts_acked += acked_pcount; 3150 tcp_rate_skb_delivered(sk, skb, sack->rate); 3151 3152 /* Initial outgoing SYN's get put onto the write_queue 3153 * just like anything else we transmit. It is not 3154 * true data, and if we misinform our callers that 3155 * this ACK acks real data, we will erroneously exit 3156 * connection startup slow start one packet too 3157 * quickly. This is severely frowned upon behavior. 3158 */ 3159 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) { 3160 flag |= FLAG_DATA_ACKED; 3161 } else { 3162 flag |= FLAG_SYN_ACKED; 3163 tp->retrans_stamp = 0; 3164 } 3165 3166 if (!fully_acked) 3167 break; 3168 3169 tcp_ack_tstamp(sk, skb, prior_snd_una); 3170 3171 next = skb_rb_next(skb); 3172 if (unlikely(skb == tp->retransmit_skb_hint)) 3173 tp->retransmit_skb_hint = NULL; 3174 if (unlikely(skb == tp->lost_skb_hint)) 3175 tp->lost_skb_hint = NULL; 3176 tcp_highest_sack_replace(sk, skb, next); 3177 tcp_rtx_queue_unlink_and_free(skb, sk); 3178 } 3179 3180 if (!skb) 3181 tcp_chrono_stop(sk, TCP_CHRONO_BUSY); 3182 3183 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una))) 3184 tp->snd_up = tp->snd_una; 3185 3186 if (skb) { 3187 tcp_ack_tstamp(sk, skb, prior_snd_una); 3188 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 3189 flag |= FLAG_SACK_RENEGING; 3190 } 3191 3192 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) { 3193 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt); 3194 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt); 3195 3196 if (pkts_acked == 1 && last_in_flight < tp->mss_cache && 3197 last_in_flight && !prior_sacked && fully_acked && 3198 sack->rate->prior_delivered + 1 == tp->delivered && 3199 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) { 3200 /* Conservatively mark a delayed ACK. It's typically 3201 * from a lone runt packet over the round trip to 3202 * a receiver w/o out-of-order or CE events. 3203 */ 3204 flag |= FLAG_ACK_MAYBE_DELAYED; 3205 } 3206 } 3207 if (sack->first_sackt) { 3208 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt); 3209 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt); 3210 } 3211 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us, 3212 ca_rtt_us, sack->rate); 3213 3214 if (flag & FLAG_ACKED) { 3215 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */ 3216 if (unlikely(icsk->icsk_mtup.probe_size && 3217 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) { 3218 tcp_mtup_probe_success(sk); 3219 } 3220 3221 if (tcp_is_reno(tp)) { 3222 tcp_remove_reno_sacks(sk, pkts_acked, ece_ack); 3223 3224 /* If any of the cumulatively ACKed segments was 3225 * retransmitted, non-SACK case cannot confirm that 3226 * progress was due to original transmission due to 3227 * lack of TCPCB_SACKED_ACKED bits even if some of 3228 * the packets may have been never retransmitted. 3229 */ 3230 if (flag & FLAG_RETRANS_DATA_ACKED) 3231 flag &= ~FLAG_ORIG_SACK_ACKED; 3232 } else { 3233 int delta; 3234 3235 /* Non-retransmitted hole got filled? That's reordering */ 3236 if (before(reord, prior_fack)) 3237 tcp_check_sack_reordering(sk, reord, 0); 3238 3239 delta = prior_sacked - tp->sacked_out; 3240 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta); 3241 } 3242 } else if (skb && rtt_update && sack_rtt_us >= 0 && 3243 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp, 3244 tcp_skb_timestamp_us(skb))) { 3245 /* Do not re-arm RTO if the sack RTT is measured from data sent 3246 * after when the head was last (re)transmitted. Otherwise the 3247 * timeout may continue to extend in loss recovery. 3248 */ 3249 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */ 3250 } 3251 3252 if (icsk->icsk_ca_ops->pkts_acked) { 3253 struct ack_sample sample = { .pkts_acked = pkts_acked, 3254 .rtt_us = sack->rate->rtt_us, 3255 .in_flight = last_in_flight }; 3256 3257 icsk->icsk_ca_ops->pkts_acked(sk, &sample); 3258 } 3259 3260 #if FASTRETRANS_DEBUG > 0 3261 WARN_ON((int)tp->sacked_out < 0); 3262 WARN_ON((int)tp->lost_out < 0); 3263 WARN_ON((int)tp->retrans_out < 0); 3264 if (!tp->packets_out && tcp_is_sack(tp)) { 3265 icsk = inet_csk(sk); 3266 if (tp->lost_out) { 3267 pr_debug("Leak l=%u %d\n", 3268 tp->lost_out, icsk->icsk_ca_state); 3269 tp->lost_out = 0; 3270 } 3271 if (tp->sacked_out) { 3272 pr_debug("Leak s=%u %d\n", 3273 tp->sacked_out, icsk->icsk_ca_state); 3274 tp->sacked_out = 0; 3275 } 3276 if (tp->retrans_out) { 3277 pr_debug("Leak r=%u %d\n", 3278 tp->retrans_out, icsk->icsk_ca_state); 3279 tp->retrans_out = 0; 3280 } 3281 } 3282 #endif 3283 return flag; 3284 } 3285 3286 static void tcp_ack_probe(struct sock *sk) 3287 { 3288 struct inet_connection_sock *icsk = inet_csk(sk); 3289 struct sk_buff *head = tcp_send_head(sk); 3290 const struct tcp_sock *tp = tcp_sk(sk); 3291 3292 /* Was it a usable window open? */ 3293 if (!head) 3294 return; 3295 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) { 3296 icsk->icsk_backoff = 0; 3297 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0); 3298 /* Socket must be waked up by subsequent tcp_data_snd_check(). 3299 * This function is not for random using! 3300 */ 3301 } else { 3302 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX); 3303 3304 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 3305 when, TCP_RTO_MAX); 3306 } 3307 } 3308 3309 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag) 3310 { 3311 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) || 3312 inet_csk(sk)->icsk_ca_state != TCP_CA_Open; 3313 } 3314 3315 /* Decide wheather to run the increase function of congestion control. */ 3316 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag) 3317 { 3318 /* If reordering is high then always grow cwnd whenever data is 3319 * delivered regardless of its ordering. Otherwise stay conservative 3320 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/ 3321 * new SACK or ECE mark may first advance cwnd here and later reduce 3322 * cwnd in tcp_fastretrans_alert() based on more states. 3323 */ 3324 if (tcp_sk(sk)->reordering > sock_net(sk)->ipv4.sysctl_tcp_reordering) 3325 return flag & FLAG_FORWARD_PROGRESS; 3326 3327 return flag & FLAG_DATA_ACKED; 3328 } 3329 3330 /* The "ultimate" congestion control function that aims to replace the rigid 3331 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction). 3332 * It's called toward the end of processing an ACK with precise rate 3333 * information. All transmission or retransmission are delayed afterwards. 3334 */ 3335 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked, 3336 int flag, const struct rate_sample *rs) 3337 { 3338 const struct inet_connection_sock *icsk = inet_csk(sk); 3339 3340 if (icsk->icsk_ca_ops->cong_control) { 3341 icsk->icsk_ca_ops->cong_control(sk, rs); 3342 return; 3343 } 3344 3345 if (tcp_in_cwnd_reduction(sk)) { 3346 /* Reduce cwnd if state mandates */ 3347 tcp_cwnd_reduction(sk, acked_sacked, flag); 3348 } else if (tcp_may_raise_cwnd(sk, flag)) { 3349 /* Advance cwnd if state allows */ 3350 tcp_cong_avoid(sk, ack, acked_sacked); 3351 } 3352 tcp_update_pacing_rate(sk); 3353 } 3354 3355 /* Check that window update is acceptable. 3356 * The function assumes that snd_una<=ack<=snd_next. 3357 */ 3358 static inline bool tcp_may_update_window(const struct tcp_sock *tp, 3359 const u32 ack, const u32 ack_seq, 3360 const u32 nwin) 3361 { 3362 return after(ack, tp->snd_una) || 3363 after(ack_seq, tp->snd_wl1) || 3364 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd); 3365 } 3366 3367 /* If we update tp->snd_una, also update tp->bytes_acked */ 3368 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack) 3369 { 3370 u32 delta = ack - tp->snd_una; 3371 3372 sock_owned_by_me((struct sock *)tp); 3373 tp->bytes_acked += delta; 3374 tp->snd_una = ack; 3375 } 3376 3377 /* If we update tp->rcv_nxt, also update tp->bytes_received */ 3378 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq) 3379 { 3380 u32 delta = seq - tp->rcv_nxt; 3381 3382 sock_owned_by_me((struct sock *)tp); 3383 tp->bytes_received += delta; 3384 WRITE_ONCE(tp->rcv_nxt, seq); 3385 } 3386 3387 /* Update our send window. 3388 * 3389 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2 3390 * and in FreeBSD. NetBSD's one is even worse.) is wrong. 3391 */ 3392 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack, 3393 u32 ack_seq) 3394 { 3395 struct tcp_sock *tp = tcp_sk(sk); 3396 int flag = 0; 3397 u32 nwin = ntohs(tcp_hdr(skb)->window); 3398 3399 if (likely(!tcp_hdr(skb)->syn)) 3400 nwin <<= tp->rx_opt.snd_wscale; 3401 3402 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) { 3403 flag |= FLAG_WIN_UPDATE; 3404 tcp_update_wl(tp, ack_seq); 3405 3406 if (tp->snd_wnd != nwin) { 3407 tp->snd_wnd = nwin; 3408 3409 /* Note, it is the only place, where 3410 * fast path is recovered for sending TCP. 3411 */ 3412 tp->pred_flags = 0; 3413 tcp_fast_path_check(sk); 3414 3415 if (!tcp_write_queue_empty(sk)) 3416 tcp_slow_start_after_idle_check(sk); 3417 3418 if (nwin > tp->max_window) { 3419 tp->max_window = nwin; 3420 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie); 3421 } 3422 } 3423 } 3424 3425 tcp_snd_una_update(tp, ack); 3426 3427 return flag; 3428 } 3429 3430 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx, 3431 u32 *last_oow_ack_time) 3432 { 3433 if (*last_oow_ack_time) { 3434 s32 elapsed = (s32)(tcp_jiffies32 - *last_oow_ack_time); 3435 3436 if (0 <= elapsed && elapsed < net->ipv4.sysctl_tcp_invalid_ratelimit) { 3437 NET_INC_STATS(net, mib_idx); 3438 return true; /* rate-limited: don't send yet! */ 3439 } 3440 } 3441 3442 *last_oow_ack_time = tcp_jiffies32; 3443 3444 return false; /* not rate-limited: go ahead, send dupack now! */ 3445 } 3446 3447 /* Return true if we're currently rate-limiting out-of-window ACKs and 3448 * thus shouldn't send a dupack right now. We rate-limit dupacks in 3449 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS 3450 * attacks that send repeated SYNs or ACKs for the same connection. To 3451 * do this, we do not send a duplicate SYNACK or ACK if the remote 3452 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate. 3453 */ 3454 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb, 3455 int mib_idx, u32 *last_oow_ack_time) 3456 { 3457 /* Data packets without SYNs are not likely part of an ACK loop. */ 3458 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) && 3459 !tcp_hdr(skb)->syn) 3460 return false; 3461 3462 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time); 3463 } 3464 3465 /* RFC 5961 7 [ACK Throttling] */ 3466 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb) 3467 { 3468 /* unprotected vars, we dont care of overwrites */ 3469 static u32 challenge_timestamp; 3470 static unsigned int challenge_count; 3471 struct tcp_sock *tp = tcp_sk(sk); 3472 struct net *net = sock_net(sk); 3473 u32 count, now; 3474 3475 /* First check our per-socket dupack rate limit. */ 3476 if (__tcp_oow_rate_limited(net, 3477 LINUX_MIB_TCPACKSKIPPEDCHALLENGE, 3478 &tp->last_oow_ack_time)) 3479 return; 3480 3481 /* Then check host-wide RFC 5961 rate limit. */ 3482 now = jiffies / HZ; 3483 if (now != challenge_timestamp) { 3484 u32 ack_limit = net->ipv4.sysctl_tcp_challenge_ack_limit; 3485 u32 half = (ack_limit + 1) >> 1; 3486 3487 challenge_timestamp = now; 3488 WRITE_ONCE(challenge_count, half + prandom_u32_max(ack_limit)); 3489 } 3490 count = READ_ONCE(challenge_count); 3491 if (count > 0) { 3492 WRITE_ONCE(challenge_count, count - 1); 3493 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK); 3494 tcp_send_ack(sk); 3495 } 3496 } 3497 3498 static void tcp_store_ts_recent(struct tcp_sock *tp) 3499 { 3500 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval; 3501 tp->rx_opt.ts_recent_stamp = ktime_get_seconds(); 3502 } 3503 3504 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq) 3505 { 3506 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) { 3507 /* PAWS bug workaround wrt. ACK frames, the PAWS discard 3508 * extra check below makes sure this can only happen 3509 * for pure ACK frames. -DaveM 3510 * 3511 * Not only, also it occurs for expired timestamps. 3512 */ 3513 3514 if (tcp_paws_check(&tp->rx_opt, 0)) 3515 tcp_store_ts_recent(tp); 3516 } 3517 } 3518 3519 /* This routine deals with acks during a TLP episode and ends an episode by 3520 * resetting tlp_high_seq. Ref: TLP algorithm in draft-ietf-tcpm-rack 3521 */ 3522 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag) 3523 { 3524 struct tcp_sock *tp = tcp_sk(sk); 3525 3526 if (before(ack, tp->tlp_high_seq)) 3527 return; 3528 3529 if (!tp->tlp_retrans) { 3530 /* TLP of new data has been acknowledged */ 3531 tp->tlp_high_seq = 0; 3532 } else if (flag & FLAG_DSACKING_ACK) { 3533 /* This DSACK means original and TLP probe arrived; no loss */ 3534 tp->tlp_high_seq = 0; 3535 } else if (after(ack, tp->tlp_high_seq)) { 3536 /* ACK advances: there was a loss, so reduce cwnd. Reset 3537 * tlp_high_seq in tcp_init_cwnd_reduction() 3538 */ 3539 tcp_init_cwnd_reduction(sk); 3540 tcp_set_ca_state(sk, TCP_CA_CWR); 3541 tcp_end_cwnd_reduction(sk); 3542 tcp_try_keep_open(sk); 3543 NET_INC_STATS(sock_net(sk), 3544 LINUX_MIB_TCPLOSSPROBERECOVERY); 3545 } else if (!(flag & (FLAG_SND_UNA_ADVANCED | 3546 FLAG_NOT_DUP | FLAG_DATA_SACKED))) { 3547 /* Pure dupack: original and TLP probe arrived; no loss */ 3548 tp->tlp_high_seq = 0; 3549 } 3550 } 3551 3552 static inline void tcp_in_ack_event(struct sock *sk, u32 flags) 3553 { 3554 const struct inet_connection_sock *icsk = inet_csk(sk); 3555 3556 if (icsk->icsk_ca_ops->in_ack_event) 3557 icsk->icsk_ca_ops->in_ack_event(sk, flags); 3558 } 3559 3560 /* Congestion control has updated the cwnd already. So if we're in 3561 * loss recovery then now we do any new sends (for FRTO) or 3562 * retransmits (for CA_Loss or CA_recovery) that make sense. 3563 */ 3564 static void tcp_xmit_recovery(struct sock *sk, int rexmit) 3565 { 3566 struct tcp_sock *tp = tcp_sk(sk); 3567 3568 if (rexmit == REXMIT_NONE || sk->sk_state == TCP_SYN_SENT) 3569 return; 3570 3571 if (unlikely(rexmit == REXMIT_NEW)) { 3572 __tcp_push_pending_frames(sk, tcp_current_mss(sk), 3573 TCP_NAGLE_OFF); 3574 if (after(tp->snd_nxt, tp->high_seq)) 3575 return; 3576 tp->frto = 0; 3577 } 3578 tcp_xmit_retransmit_queue(sk); 3579 } 3580 3581 /* Returns the number of packets newly acked or sacked by the current ACK */ 3582 static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, int flag) 3583 { 3584 const struct net *net = sock_net(sk); 3585 struct tcp_sock *tp = tcp_sk(sk); 3586 u32 delivered; 3587 3588 delivered = tp->delivered - prior_delivered; 3589 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered); 3590 if (flag & FLAG_ECE) 3591 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, delivered); 3592 3593 return delivered; 3594 } 3595 3596 /* This routine deals with incoming acks, but not outgoing ones. */ 3597 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag) 3598 { 3599 struct inet_connection_sock *icsk = inet_csk(sk); 3600 struct tcp_sock *tp = tcp_sk(sk); 3601 struct tcp_sacktag_state sack_state; 3602 struct rate_sample rs = { .prior_delivered = 0 }; 3603 u32 prior_snd_una = tp->snd_una; 3604 bool is_sack_reneg = tp->is_sack_reneg; 3605 u32 ack_seq = TCP_SKB_CB(skb)->seq; 3606 u32 ack = TCP_SKB_CB(skb)->ack_seq; 3607 int num_dupack = 0; 3608 int prior_packets = tp->packets_out; 3609 u32 delivered = tp->delivered; 3610 u32 lost = tp->lost; 3611 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */ 3612 u32 prior_fack; 3613 3614 sack_state.first_sackt = 0; 3615 sack_state.rate = &rs; 3616 sack_state.sack_delivered = 0; 3617 3618 /* We very likely will need to access rtx queue. */ 3619 prefetch(sk->tcp_rtx_queue.rb_node); 3620 3621 /* If the ack is older than previous acks 3622 * then we can probably ignore it. 3623 */ 3624 if (before(ack, prior_snd_una)) { 3625 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */ 3626 if (before(ack, prior_snd_una - tp->max_window)) { 3627 if (!(flag & FLAG_NO_CHALLENGE_ACK)) 3628 tcp_send_challenge_ack(sk, skb); 3629 return -1; 3630 } 3631 goto old_ack; 3632 } 3633 3634 /* If the ack includes data we haven't sent yet, discard 3635 * this segment (RFC793 Section 3.9). 3636 */ 3637 if (after(ack, tp->snd_nxt)) 3638 return -1; 3639 3640 if (after(ack, prior_snd_una)) { 3641 flag |= FLAG_SND_UNA_ADVANCED; 3642 icsk->icsk_retransmits = 0; 3643 3644 #if IS_ENABLED(CONFIG_TLS_DEVICE) 3645 if (static_branch_unlikely(&clean_acked_data_enabled.key)) 3646 if (icsk->icsk_clean_acked) 3647 icsk->icsk_clean_acked(sk, ack); 3648 #endif 3649 } 3650 3651 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una; 3652 rs.prior_in_flight = tcp_packets_in_flight(tp); 3653 3654 /* ts_recent update must be made after we are sure that the packet 3655 * is in window. 3656 */ 3657 if (flag & FLAG_UPDATE_TS_RECENT) 3658 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); 3659 3660 if ((flag & (FLAG_SLOWPATH | FLAG_SND_UNA_ADVANCED)) == 3661 FLAG_SND_UNA_ADVANCED) { 3662 /* Window is constant, pure forward advance. 3663 * No more checks are required. 3664 * Note, we use the fact that SND.UNA>=SND.WL2. 3665 */ 3666 tcp_update_wl(tp, ack_seq); 3667 tcp_snd_una_update(tp, ack); 3668 flag |= FLAG_WIN_UPDATE; 3669 3670 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE); 3671 3672 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS); 3673 } else { 3674 u32 ack_ev_flags = CA_ACK_SLOWPATH; 3675 3676 if (ack_seq != TCP_SKB_CB(skb)->end_seq) 3677 flag |= FLAG_DATA; 3678 else 3679 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS); 3680 3681 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq); 3682 3683 if (TCP_SKB_CB(skb)->sacked) 3684 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una, 3685 &sack_state); 3686 3687 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) { 3688 flag |= FLAG_ECE; 3689 ack_ev_flags |= CA_ACK_ECE; 3690 } 3691 3692 if (sack_state.sack_delivered) 3693 tcp_count_delivered(tp, sack_state.sack_delivered, 3694 flag & FLAG_ECE); 3695 3696 if (flag & FLAG_WIN_UPDATE) 3697 ack_ev_flags |= CA_ACK_WIN_UPDATE; 3698 3699 tcp_in_ack_event(sk, ack_ev_flags); 3700 } 3701 3702 /* This is a deviation from RFC3168 since it states that: 3703 * "When the TCP data sender is ready to set the CWR bit after reducing 3704 * the congestion window, it SHOULD set the CWR bit only on the first 3705 * new data packet that it transmits." 3706 * We accept CWR on pure ACKs to be more robust 3707 * with widely-deployed TCP implementations that do this. 3708 */ 3709 tcp_ecn_accept_cwr(sk, skb); 3710 3711 /* We passed data and got it acked, remove any soft error 3712 * log. Something worked... 3713 */ 3714 sk->sk_err_soft = 0; 3715 icsk->icsk_probes_out = 0; 3716 tp->rcv_tstamp = tcp_jiffies32; 3717 if (!prior_packets) 3718 goto no_queue; 3719 3720 /* See if we can take anything off of the retransmit queue. */ 3721 flag |= tcp_clean_rtx_queue(sk, prior_fack, prior_snd_una, &sack_state, 3722 flag & FLAG_ECE); 3723 3724 tcp_rack_update_reo_wnd(sk, &rs); 3725 3726 if (tp->tlp_high_seq) 3727 tcp_process_tlp_ack(sk, ack, flag); 3728 /* If needed, reset TLP/RTO timer; RACK may later override this. */ 3729 if (flag & FLAG_SET_XMIT_TIMER) 3730 tcp_set_xmit_timer(sk); 3731 3732 if (tcp_ack_is_dubious(sk, flag)) { 3733 if (!(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP))) { 3734 num_dupack = 1; 3735 /* Consider if pure acks were aggregated in tcp_add_backlog() */ 3736 if (!(flag & FLAG_DATA)) 3737 num_dupack = max_t(u16, 1, skb_shinfo(skb)->gso_segs); 3738 } 3739 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag, 3740 &rexmit); 3741 } 3742 3743 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) 3744 sk_dst_confirm(sk); 3745 3746 delivered = tcp_newly_delivered(sk, delivered, flag); 3747 lost = tp->lost - lost; /* freshly marked lost */ 3748 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED); 3749 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate); 3750 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate); 3751 tcp_xmit_recovery(sk, rexmit); 3752 return 1; 3753 3754 no_queue: 3755 /* If data was DSACKed, see if we can undo a cwnd reduction. */ 3756 if (flag & FLAG_DSACKING_ACK) { 3757 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag, 3758 &rexmit); 3759 tcp_newly_delivered(sk, delivered, flag); 3760 } 3761 /* If this ack opens up a zero window, clear backoff. It was 3762 * being used to time the probes, and is probably far higher than 3763 * it needs to be for normal retransmission. 3764 */ 3765 tcp_ack_probe(sk); 3766 3767 if (tp->tlp_high_seq) 3768 tcp_process_tlp_ack(sk, ack, flag); 3769 return 1; 3770 3771 old_ack: 3772 /* If data was SACKed, tag it and see if we should send more data. 3773 * If data was DSACKed, see if we can undo a cwnd reduction. 3774 */ 3775 if (TCP_SKB_CB(skb)->sacked) { 3776 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una, 3777 &sack_state); 3778 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag, 3779 &rexmit); 3780 tcp_newly_delivered(sk, delivered, flag); 3781 tcp_xmit_recovery(sk, rexmit); 3782 } 3783 3784 return 0; 3785 } 3786 3787 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie, 3788 bool syn, struct tcp_fastopen_cookie *foc, 3789 bool exp_opt) 3790 { 3791 /* Valid only in SYN or SYN-ACK with an even length. */ 3792 if (!foc || !syn || len < 0 || (len & 1)) 3793 return; 3794 3795 if (len >= TCP_FASTOPEN_COOKIE_MIN && 3796 len <= TCP_FASTOPEN_COOKIE_MAX) 3797 memcpy(foc->val, cookie, len); 3798 else if (len != 0) 3799 len = -1; 3800 foc->len = len; 3801 foc->exp = exp_opt; 3802 } 3803 3804 static void smc_parse_options(const struct tcphdr *th, 3805 struct tcp_options_received *opt_rx, 3806 const unsigned char *ptr, 3807 int opsize) 3808 { 3809 #if IS_ENABLED(CONFIG_SMC) 3810 if (static_branch_unlikely(&tcp_have_smc)) { 3811 if (th->syn && !(opsize & 1) && 3812 opsize >= TCPOLEN_EXP_SMC_BASE && 3813 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC) 3814 opt_rx->smc_ok = 1; 3815 } 3816 #endif 3817 } 3818 3819 /* Try to parse the MSS option from the TCP header. Return 0 on failure, clamped 3820 * value on success. 3821 */ 3822 static u16 tcp_parse_mss_option(const struct tcphdr *th, u16 user_mss) 3823 { 3824 const unsigned char *ptr = (const unsigned char *)(th + 1); 3825 int length = (th->doff * 4) - sizeof(struct tcphdr); 3826 u16 mss = 0; 3827 3828 while (length > 0) { 3829 int opcode = *ptr++; 3830 int opsize; 3831 3832 switch (opcode) { 3833 case TCPOPT_EOL: 3834 return mss; 3835 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ 3836 length--; 3837 continue; 3838 default: 3839 if (length < 2) 3840 return mss; 3841 opsize = *ptr++; 3842 if (opsize < 2) /* "silly options" */ 3843 return mss; 3844 if (opsize > length) 3845 return mss; /* fail on partial options */ 3846 if (opcode == TCPOPT_MSS && opsize == TCPOLEN_MSS) { 3847 u16 in_mss = get_unaligned_be16(ptr); 3848 3849 if (in_mss) { 3850 if (user_mss && user_mss < in_mss) 3851 in_mss = user_mss; 3852 mss = in_mss; 3853 } 3854 } 3855 ptr += opsize - 2; 3856 length -= opsize; 3857 } 3858 } 3859 return mss; 3860 } 3861 3862 /* Look for tcp options. Normally only called on SYN and SYNACK packets. 3863 * But, this can also be called on packets in the established flow when 3864 * the fast version below fails. 3865 */ 3866 void tcp_parse_options(const struct net *net, 3867 const struct sk_buff *skb, 3868 struct tcp_options_received *opt_rx, int estab, 3869 struct tcp_fastopen_cookie *foc) 3870 { 3871 const unsigned char *ptr; 3872 const struct tcphdr *th = tcp_hdr(skb); 3873 int length = (th->doff * 4) - sizeof(struct tcphdr); 3874 3875 ptr = (const unsigned char *)(th + 1); 3876 opt_rx->saw_tstamp = 0; 3877 3878 while (length > 0) { 3879 int opcode = *ptr++; 3880 int opsize; 3881 3882 switch (opcode) { 3883 case TCPOPT_EOL: 3884 return; 3885 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ 3886 length--; 3887 continue; 3888 default: 3889 if (length < 2) 3890 return; 3891 opsize = *ptr++; 3892 if (opsize < 2) /* "silly options" */ 3893 return; 3894 if (opsize > length) 3895 return; /* don't parse partial options */ 3896 switch (opcode) { 3897 case TCPOPT_MSS: 3898 if (opsize == TCPOLEN_MSS && th->syn && !estab) { 3899 u16 in_mss = get_unaligned_be16(ptr); 3900 if (in_mss) { 3901 if (opt_rx->user_mss && 3902 opt_rx->user_mss < in_mss) 3903 in_mss = opt_rx->user_mss; 3904 opt_rx->mss_clamp = in_mss; 3905 } 3906 } 3907 break; 3908 case TCPOPT_WINDOW: 3909 if (opsize == TCPOLEN_WINDOW && th->syn && 3910 !estab && net->ipv4.sysctl_tcp_window_scaling) { 3911 __u8 snd_wscale = *(__u8 *)ptr; 3912 opt_rx->wscale_ok = 1; 3913 if (snd_wscale > TCP_MAX_WSCALE) { 3914 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n", 3915 __func__, 3916 snd_wscale, 3917 TCP_MAX_WSCALE); 3918 snd_wscale = TCP_MAX_WSCALE; 3919 } 3920 opt_rx->snd_wscale = snd_wscale; 3921 } 3922 break; 3923 case TCPOPT_TIMESTAMP: 3924 if ((opsize == TCPOLEN_TIMESTAMP) && 3925 ((estab && opt_rx->tstamp_ok) || 3926 (!estab && net->ipv4.sysctl_tcp_timestamps))) { 3927 opt_rx->saw_tstamp = 1; 3928 opt_rx->rcv_tsval = get_unaligned_be32(ptr); 3929 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4); 3930 } 3931 break; 3932 case TCPOPT_SACK_PERM: 3933 if (opsize == TCPOLEN_SACK_PERM && th->syn && 3934 !estab && net->ipv4.sysctl_tcp_sack) { 3935 opt_rx->sack_ok = TCP_SACK_SEEN; 3936 tcp_sack_reset(opt_rx); 3937 } 3938 break; 3939 3940 case TCPOPT_SACK: 3941 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) && 3942 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) && 3943 opt_rx->sack_ok) { 3944 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th; 3945 } 3946 break; 3947 #ifdef CONFIG_TCP_MD5SIG 3948 case TCPOPT_MD5SIG: 3949 /* 3950 * The MD5 Hash has already been 3951 * checked (see tcp_v{4,6}_do_rcv()). 3952 */ 3953 break; 3954 #endif 3955 case TCPOPT_FASTOPEN: 3956 tcp_parse_fastopen_option( 3957 opsize - TCPOLEN_FASTOPEN_BASE, 3958 ptr, th->syn, foc, false); 3959 break; 3960 3961 case TCPOPT_EXP: 3962 /* Fast Open option shares code 254 using a 3963 * 16 bits magic number. 3964 */ 3965 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE && 3966 get_unaligned_be16(ptr) == 3967 TCPOPT_FASTOPEN_MAGIC) 3968 tcp_parse_fastopen_option(opsize - 3969 TCPOLEN_EXP_FASTOPEN_BASE, 3970 ptr + 2, th->syn, foc, true); 3971 else 3972 smc_parse_options(th, opt_rx, ptr, 3973 opsize); 3974 break; 3975 3976 } 3977 ptr += opsize-2; 3978 length -= opsize; 3979 } 3980 } 3981 } 3982 EXPORT_SYMBOL(tcp_parse_options); 3983 3984 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th) 3985 { 3986 const __be32 *ptr = (const __be32 *)(th + 1); 3987 3988 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) 3989 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) { 3990 tp->rx_opt.saw_tstamp = 1; 3991 ++ptr; 3992 tp->rx_opt.rcv_tsval = ntohl(*ptr); 3993 ++ptr; 3994 if (*ptr) 3995 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset; 3996 else 3997 tp->rx_opt.rcv_tsecr = 0; 3998 return true; 3999 } 4000 return false; 4001 } 4002 4003 /* Fast parse options. This hopes to only see timestamps. 4004 * If it is wrong it falls back on tcp_parse_options(). 4005 */ 4006 static bool tcp_fast_parse_options(const struct net *net, 4007 const struct sk_buff *skb, 4008 const struct tcphdr *th, struct tcp_sock *tp) 4009 { 4010 /* In the spirit of fast parsing, compare doff directly to constant 4011 * values. Because equality is used, short doff can be ignored here. 4012 */ 4013 if (th->doff == (sizeof(*th) / 4)) { 4014 tp->rx_opt.saw_tstamp = 0; 4015 return false; 4016 } else if (tp->rx_opt.tstamp_ok && 4017 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) { 4018 if (tcp_parse_aligned_timestamp(tp, th)) 4019 return true; 4020 } 4021 4022 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL); 4023 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) 4024 tp->rx_opt.rcv_tsecr -= tp->tsoffset; 4025 4026 return true; 4027 } 4028 4029 #ifdef CONFIG_TCP_MD5SIG 4030 /* 4031 * Parse MD5 Signature option 4032 */ 4033 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th) 4034 { 4035 int length = (th->doff << 2) - sizeof(*th); 4036 const u8 *ptr = (const u8 *)(th + 1); 4037 4038 /* If not enough data remaining, we can short cut */ 4039 while (length >= TCPOLEN_MD5SIG) { 4040 int opcode = *ptr++; 4041 int opsize; 4042 4043 switch (opcode) { 4044 case TCPOPT_EOL: 4045 return NULL; 4046 case TCPOPT_NOP: 4047 length--; 4048 continue; 4049 default: 4050 opsize = *ptr++; 4051 if (opsize < 2 || opsize > length) 4052 return NULL; 4053 if (opcode == TCPOPT_MD5SIG) 4054 return opsize == TCPOLEN_MD5SIG ? ptr : NULL; 4055 } 4056 ptr += opsize - 2; 4057 length -= opsize; 4058 } 4059 return NULL; 4060 } 4061 EXPORT_SYMBOL(tcp_parse_md5sig_option); 4062 #endif 4063 4064 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM 4065 * 4066 * It is not fatal. If this ACK does _not_ change critical state (seqs, window) 4067 * it can pass through stack. So, the following predicate verifies that 4068 * this segment is not used for anything but congestion avoidance or 4069 * fast retransmit. Moreover, we even are able to eliminate most of such 4070 * second order effects, if we apply some small "replay" window (~RTO) 4071 * to timestamp space. 4072 * 4073 * All these measures still do not guarantee that we reject wrapped ACKs 4074 * on networks with high bandwidth, when sequence space is recycled fastly, 4075 * but it guarantees that such events will be very rare and do not affect 4076 * connection seriously. This doesn't look nice, but alas, PAWS is really 4077 * buggy extension. 4078 * 4079 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC 4080 * states that events when retransmit arrives after original data are rare. 4081 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is 4082 * the biggest problem on large power networks even with minor reordering. 4083 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe 4084 * up to bandwidth of 18Gigabit/sec. 8) ] 4085 */ 4086 4087 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb) 4088 { 4089 const struct tcp_sock *tp = tcp_sk(sk); 4090 const struct tcphdr *th = tcp_hdr(skb); 4091 u32 seq = TCP_SKB_CB(skb)->seq; 4092 u32 ack = TCP_SKB_CB(skb)->ack_seq; 4093 4094 return (/* 1. Pure ACK with correct sequence number. */ 4095 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) && 4096 4097 /* 2. ... and duplicate ACK. */ 4098 ack == tp->snd_una && 4099 4100 /* 3. ... and does not update window. */ 4101 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) && 4102 4103 /* 4. ... and sits in replay window. */ 4104 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ); 4105 } 4106 4107 static inline bool tcp_paws_discard(const struct sock *sk, 4108 const struct sk_buff *skb) 4109 { 4110 const struct tcp_sock *tp = tcp_sk(sk); 4111 4112 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) && 4113 !tcp_disordered_ack(sk, skb); 4114 } 4115 4116 /* Check segment sequence number for validity. 4117 * 4118 * Segment controls are considered valid, if the segment 4119 * fits to the window after truncation to the window. Acceptability 4120 * of data (and SYN, FIN, of course) is checked separately. 4121 * See tcp_data_queue(), for example. 4122 * 4123 * Also, controls (RST is main one) are accepted using RCV.WUP instead 4124 * of RCV.NXT. Peer still did not advance his SND.UNA when we 4125 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP. 4126 * (borrowed from freebsd) 4127 */ 4128 4129 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq) 4130 { 4131 return !before(end_seq, tp->rcv_wup) && 4132 !after(seq, tp->rcv_nxt + tcp_receive_window(tp)); 4133 } 4134 4135 /* When we get a reset we do this. */ 4136 void tcp_reset(struct sock *sk) 4137 { 4138 trace_tcp_receive_reset(sk); 4139 4140 /* We want the right error as BSD sees it (and indeed as we do). */ 4141 switch (sk->sk_state) { 4142 case TCP_SYN_SENT: 4143 sk->sk_err = ECONNREFUSED; 4144 break; 4145 case TCP_CLOSE_WAIT: 4146 sk->sk_err = EPIPE; 4147 break; 4148 case TCP_CLOSE: 4149 return; 4150 default: 4151 sk->sk_err = ECONNRESET; 4152 } 4153 /* This barrier is coupled with smp_rmb() in tcp_poll() */ 4154 smp_wmb(); 4155 4156 tcp_write_queue_purge(sk); 4157 tcp_done(sk); 4158 4159 if (!sock_flag(sk, SOCK_DEAD)) 4160 sk->sk_error_report(sk); 4161 } 4162 4163 /* 4164 * Process the FIN bit. This now behaves as it is supposed to work 4165 * and the FIN takes effect when it is validly part of sequence 4166 * space. Not before when we get holes. 4167 * 4168 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT 4169 * (and thence onto LAST-ACK and finally, CLOSE, we never enter 4170 * TIME-WAIT) 4171 * 4172 * If we are in FINWAIT-1, a received FIN indicates simultaneous 4173 * close and we go into CLOSING (and later onto TIME-WAIT) 4174 * 4175 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT. 4176 */ 4177 void tcp_fin(struct sock *sk) 4178 { 4179 struct tcp_sock *tp = tcp_sk(sk); 4180 4181 inet_csk_schedule_ack(sk); 4182 4183 sk->sk_shutdown |= RCV_SHUTDOWN; 4184 sock_set_flag(sk, SOCK_DONE); 4185 4186 switch (sk->sk_state) { 4187 case TCP_SYN_RECV: 4188 case TCP_ESTABLISHED: 4189 /* Move to CLOSE_WAIT */ 4190 tcp_set_state(sk, TCP_CLOSE_WAIT); 4191 inet_csk_enter_pingpong_mode(sk); 4192 break; 4193 4194 case TCP_CLOSE_WAIT: 4195 case TCP_CLOSING: 4196 /* Received a retransmission of the FIN, do 4197 * nothing. 4198 */ 4199 break; 4200 case TCP_LAST_ACK: 4201 /* RFC793: Remain in the LAST-ACK state. */ 4202 break; 4203 4204 case TCP_FIN_WAIT1: 4205 /* This case occurs when a simultaneous close 4206 * happens, we must ack the received FIN and 4207 * enter the CLOSING state. 4208 */ 4209 tcp_send_ack(sk); 4210 tcp_set_state(sk, TCP_CLOSING); 4211 break; 4212 case TCP_FIN_WAIT2: 4213 /* Received a FIN -- send ACK and enter TIME_WAIT. */ 4214 tcp_send_ack(sk); 4215 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 4216 break; 4217 default: 4218 /* Only TCP_LISTEN and TCP_CLOSE are left, in these 4219 * cases we should never reach this piece of code. 4220 */ 4221 pr_err("%s: Impossible, sk->sk_state=%d\n", 4222 __func__, sk->sk_state); 4223 break; 4224 } 4225 4226 /* It _is_ possible, that we have something out-of-order _after_ FIN. 4227 * Probably, we should reset in this case. For now drop them. 4228 */ 4229 skb_rbtree_purge(&tp->out_of_order_queue); 4230 if (tcp_is_sack(tp)) 4231 tcp_sack_reset(&tp->rx_opt); 4232 sk_mem_reclaim(sk); 4233 4234 if (!sock_flag(sk, SOCK_DEAD)) { 4235 sk->sk_state_change(sk); 4236 4237 /* Do not send POLL_HUP for half duplex close. */ 4238 if (sk->sk_shutdown == SHUTDOWN_MASK || 4239 sk->sk_state == TCP_CLOSE) 4240 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP); 4241 else 4242 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN); 4243 } 4244 } 4245 4246 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, 4247 u32 end_seq) 4248 { 4249 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) { 4250 if (before(seq, sp->start_seq)) 4251 sp->start_seq = seq; 4252 if (after(end_seq, sp->end_seq)) 4253 sp->end_seq = end_seq; 4254 return true; 4255 } 4256 return false; 4257 } 4258 4259 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq) 4260 { 4261 struct tcp_sock *tp = tcp_sk(sk); 4262 4263 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) { 4264 int mib_idx; 4265 4266 if (before(seq, tp->rcv_nxt)) 4267 mib_idx = LINUX_MIB_TCPDSACKOLDSENT; 4268 else 4269 mib_idx = LINUX_MIB_TCPDSACKOFOSENT; 4270 4271 NET_INC_STATS(sock_net(sk), mib_idx); 4272 4273 tp->rx_opt.dsack = 1; 4274 tp->duplicate_sack[0].start_seq = seq; 4275 tp->duplicate_sack[0].end_seq = end_seq; 4276 } 4277 } 4278 4279 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq) 4280 { 4281 struct tcp_sock *tp = tcp_sk(sk); 4282 4283 if (!tp->rx_opt.dsack) 4284 tcp_dsack_set(sk, seq, end_seq); 4285 else 4286 tcp_sack_extend(tp->duplicate_sack, seq, end_seq); 4287 } 4288 4289 static void tcp_rcv_spurious_retrans(struct sock *sk, const struct sk_buff *skb) 4290 { 4291 /* When the ACK path fails or drops most ACKs, the sender would 4292 * timeout and spuriously retransmit the same segment repeatedly. 4293 * The receiver remembers and reflects via DSACKs. Leverage the 4294 * DSACK state and change the txhash to re-route speculatively. 4295 */ 4296 if (TCP_SKB_CB(skb)->seq == tcp_sk(sk)->duplicate_sack[0].start_seq) { 4297 sk_rethink_txhash(sk); 4298 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDUPLICATEDATAREHASH); 4299 } 4300 } 4301 4302 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb) 4303 { 4304 struct tcp_sock *tp = tcp_sk(sk); 4305 4306 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 4307 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4308 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); 4309 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS); 4310 4311 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) { 4312 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 4313 4314 tcp_rcv_spurious_retrans(sk, skb); 4315 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) 4316 end_seq = tp->rcv_nxt; 4317 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq); 4318 } 4319 } 4320 4321 tcp_send_ack(sk); 4322 } 4323 4324 /* These routines update the SACK block as out-of-order packets arrive or 4325 * in-order packets close up the sequence space. 4326 */ 4327 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp) 4328 { 4329 int this_sack; 4330 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4331 struct tcp_sack_block *swalk = sp + 1; 4332 4333 /* See if the recent change to the first SACK eats into 4334 * or hits the sequence space of other SACK blocks, if so coalesce. 4335 */ 4336 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) { 4337 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) { 4338 int i; 4339 4340 /* Zap SWALK, by moving every further SACK up by one slot. 4341 * Decrease num_sacks. 4342 */ 4343 tp->rx_opt.num_sacks--; 4344 for (i = this_sack; i < tp->rx_opt.num_sacks; i++) 4345 sp[i] = sp[i + 1]; 4346 continue; 4347 } 4348 this_sack++, swalk++; 4349 } 4350 } 4351 4352 static void tcp_sack_compress_send_ack(struct sock *sk) 4353 { 4354 struct tcp_sock *tp = tcp_sk(sk); 4355 4356 if (!tp->compressed_ack) 4357 return; 4358 4359 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1) 4360 __sock_put(sk); 4361 4362 /* Since we have to send one ack finally, 4363 * substract one from tp->compressed_ack to keep 4364 * LINUX_MIB_TCPACKCOMPRESSED accurate. 4365 */ 4366 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED, 4367 tp->compressed_ack - 1); 4368 4369 tp->compressed_ack = 0; 4370 tcp_send_ack(sk); 4371 } 4372 4373 /* Reasonable amount of sack blocks included in TCP SACK option 4374 * The max is 4, but this becomes 3 if TCP timestamps are there. 4375 * Given that SACK packets might be lost, be conservative and use 2. 4376 */ 4377 #define TCP_SACK_BLOCKS_EXPECTED 2 4378 4379 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq) 4380 { 4381 struct tcp_sock *tp = tcp_sk(sk); 4382 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4383 int cur_sacks = tp->rx_opt.num_sacks; 4384 int this_sack; 4385 4386 if (!cur_sacks) 4387 goto new_sack; 4388 4389 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) { 4390 if (tcp_sack_extend(sp, seq, end_seq)) { 4391 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED) 4392 tcp_sack_compress_send_ack(sk); 4393 /* Rotate this_sack to the first one. */ 4394 for (; this_sack > 0; this_sack--, sp--) 4395 swap(*sp, *(sp - 1)); 4396 if (cur_sacks > 1) 4397 tcp_sack_maybe_coalesce(tp); 4398 return; 4399 } 4400 } 4401 4402 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED) 4403 tcp_sack_compress_send_ack(sk); 4404 4405 /* Could not find an adjacent existing SACK, build a new one, 4406 * put it at the front, and shift everyone else down. We 4407 * always know there is at least one SACK present already here. 4408 * 4409 * If the sack array is full, forget about the last one. 4410 */ 4411 if (this_sack >= TCP_NUM_SACKS) { 4412 this_sack--; 4413 tp->rx_opt.num_sacks--; 4414 sp--; 4415 } 4416 for (; this_sack > 0; this_sack--, sp--) 4417 *sp = *(sp - 1); 4418 4419 new_sack: 4420 /* Build the new head SACK, and we're done. */ 4421 sp->start_seq = seq; 4422 sp->end_seq = end_seq; 4423 tp->rx_opt.num_sacks++; 4424 } 4425 4426 /* RCV.NXT advances, some SACKs should be eaten. */ 4427 4428 static void tcp_sack_remove(struct tcp_sock *tp) 4429 { 4430 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4431 int num_sacks = tp->rx_opt.num_sacks; 4432 int this_sack; 4433 4434 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */ 4435 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) { 4436 tp->rx_opt.num_sacks = 0; 4437 return; 4438 } 4439 4440 for (this_sack = 0; this_sack < num_sacks;) { 4441 /* Check if the start of the sack is covered by RCV.NXT. */ 4442 if (!before(tp->rcv_nxt, sp->start_seq)) { 4443 int i; 4444 4445 /* RCV.NXT must cover all the block! */ 4446 WARN_ON(before(tp->rcv_nxt, sp->end_seq)); 4447 4448 /* Zap this SACK, by moving forward any other SACKS. */ 4449 for (i = this_sack+1; i < num_sacks; i++) 4450 tp->selective_acks[i-1] = tp->selective_acks[i]; 4451 num_sacks--; 4452 continue; 4453 } 4454 this_sack++; 4455 sp++; 4456 } 4457 tp->rx_opt.num_sacks = num_sacks; 4458 } 4459 4460 /** 4461 * tcp_try_coalesce - try to merge skb to prior one 4462 * @sk: socket 4463 * @to: prior buffer 4464 * @from: buffer to add in queue 4465 * @fragstolen: pointer to boolean 4466 * 4467 * Before queueing skb @from after @to, try to merge them 4468 * to reduce overall memory use and queue lengths, if cost is small. 4469 * Packets in ofo or receive queues can stay a long time. 4470 * Better try to coalesce them right now to avoid future collapses. 4471 * Returns true if caller should free @from instead of queueing it 4472 */ 4473 static bool tcp_try_coalesce(struct sock *sk, 4474 struct sk_buff *to, 4475 struct sk_buff *from, 4476 bool *fragstolen) 4477 { 4478 int delta; 4479 4480 *fragstolen = false; 4481 4482 /* Its possible this segment overlaps with prior segment in queue */ 4483 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq) 4484 return false; 4485 4486 if (!mptcp_skb_can_collapse(to, from)) 4487 return false; 4488 4489 #ifdef CONFIG_TLS_DEVICE 4490 if (from->decrypted != to->decrypted) 4491 return false; 4492 #endif 4493 4494 if (!skb_try_coalesce(to, from, fragstolen, &delta)) 4495 return false; 4496 4497 atomic_add(delta, &sk->sk_rmem_alloc); 4498 sk_mem_charge(sk, delta); 4499 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE); 4500 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq; 4501 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq; 4502 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags; 4503 4504 if (TCP_SKB_CB(from)->has_rxtstamp) { 4505 TCP_SKB_CB(to)->has_rxtstamp = true; 4506 to->tstamp = from->tstamp; 4507 skb_hwtstamps(to)->hwtstamp = skb_hwtstamps(from)->hwtstamp; 4508 } 4509 4510 return true; 4511 } 4512 4513 static bool tcp_ooo_try_coalesce(struct sock *sk, 4514 struct sk_buff *to, 4515 struct sk_buff *from, 4516 bool *fragstolen) 4517 { 4518 bool res = tcp_try_coalesce(sk, to, from, fragstolen); 4519 4520 /* In case tcp_drop() is called later, update to->gso_segs */ 4521 if (res) { 4522 u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) + 4523 max_t(u16, 1, skb_shinfo(from)->gso_segs); 4524 4525 skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF); 4526 } 4527 return res; 4528 } 4529 4530 static void tcp_drop(struct sock *sk, struct sk_buff *skb) 4531 { 4532 sk_drops_add(sk, skb); 4533 __kfree_skb(skb); 4534 } 4535 4536 /* This one checks to see if we can put data from the 4537 * out_of_order queue into the receive_queue. 4538 */ 4539 static void tcp_ofo_queue(struct sock *sk) 4540 { 4541 struct tcp_sock *tp = tcp_sk(sk); 4542 __u32 dsack_high = tp->rcv_nxt; 4543 bool fin, fragstolen, eaten; 4544 struct sk_buff *skb, *tail; 4545 struct rb_node *p; 4546 4547 p = rb_first(&tp->out_of_order_queue); 4548 while (p) { 4549 skb = rb_to_skb(p); 4550 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 4551 break; 4552 4553 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) { 4554 __u32 dsack = dsack_high; 4555 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high)) 4556 dsack_high = TCP_SKB_CB(skb)->end_seq; 4557 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack); 4558 } 4559 p = rb_next(p); 4560 rb_erase(&skb->rbnode, &tp->out_of_order_queue); 4561 4562 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) { 4563 tcp_drop(sk, skb); 4564 continue; 4565 } 4566 4567 tail = skb_peek_tail(&sk->sk_receive_queue); 4568 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen); 4569 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq); 4570 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN; 4571 if (!eaten) 4572 __skb_queue_tail(&sk->sk_receive_queue, skb); 4573 else 4574 kfree_skb_partial(skb, fragstolen); 4575 4576 if (unlikely(fin)) { 4577 tcp_fin(sk); 4578 /* tcp_fin() purges tp->out_of_order_queue, 4579 * so we must end this loop right now. 4580 */ 4581 break; 4582 } 4583 } 4584 } 4585 4586 static bool tcp_prune_ofo_queue(struct sock *sk); 4587 static int tcp_prune_queue(struct sock *sk); 4588 4589 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb, 4590 unsigned int size) 4591 { 4592 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || 4593 !sk_rmem_schedule(sk, skb, size)) { 4594 4595 if (tcp_prune_queue(sk) < 0) 4596 return -1; 4597 4598 while (!sk_rmem_schedule(sk, skb, size)) { 4599 if (!tcp_prune_ofo_queue(sk)) 4600 return -1; 4601 } 4602 } 4603 return 0; 4604 } 4605 4606 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb) 4607 { 4608 struct tcp_sock *tp = tcp_sk(sk); 4609 struct rb_node **p, *parent; 4610 struct sk_buff *skb1; 4611 u32 seq, end_seq; 4612 bool fragstolen; 4613 4614 tcp_ecn_check_ce(sk, skb); 4615 4616 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) { 4617 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP); 4618 sk->sk_data_ready(sk); 4619 tcp_drop(sk, skb); 4620 return; 4621 } 4622 4623 /* Disable header prediction. */ 4624 tp->pred_flags = 0; 4625 inet_csk_schedule_ack(sk); 4626 4627 tp->rcv_ooopack += max_t(u16, 1, skb_shinfo(skb)->gso_segs); 4628 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE); 4629 seq = TCP_SKB_CB(skb)->seq; 4630 end_seq = TCP_SKB_CB(skb)->end_seq; 4631 4632 p = &tp->out_of_order_queue.rb_node; 4633 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) { 4634 /* Initial out of order segment, build 1 SACK. */ 4635 if (tcp_is_sack(tp)) { 4636 tp->rx_opt.num_sacks = 1; 4637 tp->selective_acks[0].start_seq = seq; 4638 tp->selective_acks[0].end_seq = end_seq; 4639 } 4640 rb_link_node(&skb->rbnode, NULL, p); 4641 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue); 4642 tp->ooo_last_skb = skb; 4643 goto end; 4644 } 4645 4646 /* In the typical case, we are adding an skb to the end of the list. 4647 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup. 4648 */ 4649 if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb, 4650 skb, &fragstolen)) { 4651 coalesce_done: 4652 /* For non sack flows, do not grow window to force DUPACK 4653 * and trigger fast retransmit. 4654 */ 4655 if (tcp_is_sack(tp)) 4656 tcp_grow_window(sk, skb); 4657 kfree_skb_partial(skb, fragstolen); 4658 skb = NULL; 4659 goto add_sack; 4660 } 4661 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */ 4662 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) { 4663 parent = &tp->ooo_last_skb->rbnode; 4664 p = &parent->rb_right; 4665 goto insert; 4666 } 4667 4668 /* Find place to insert this segment. Handle overlaps on the way. */ 4669 parent = NULL; 4670 while (*p) { 4671 parent = *p; 4672 skb1 = rb_to_skb(parent); 4673 if (before(seq, TCP_SKB_CB(skb1)->seq)) { 4674 p = &parent->rb_left; 4675 continue; 4676 } 4677 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) { 4678 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 4679 /* All the bits are present. Drop. */ 4680 NET_INC_STATS(sock_net(sk), 4681 LINUX_MIB_TCPOFOMERGE); 4682 tcp_drop(sk, skb); 4683 skb = NULL; 4684 tcp_dsack_set(sk, seq, end_seq); 4685 goto add_sack; 4686 } 4687 if (after(seq, TCP_SKB_CB(skb1)->seq)) { 4688 /* Partial overlap. */ 4689 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq); 4690 } else { 4691 /* skb's seq == skb1's seq and skb covers skb1. 4692 * Replace skb1 with skb. 4693 */ 4694 rb_replace_node(&skb1->rbnode, &skb->rbnode, 4695 &tp->out_of_order_queue); 4696 tcp_dsack_extend(sk, 4697 TCP_SKB_CB(skb1)->seq, 4698 TCP_SKB_CB(skb1)->end_seq); 4699 NET_INC_STATS(sock_net(sk), 4700 LINUX_MIB_TCPOFOMERGE); 4701 tcp_drop(sk, skb1); 4702 goto merge_right; 4703 } 4704 } else if (tcp_ooo_try_coalesce(sk, skb1, 4705 skb, &fragstolen)) { 4706 goto coalesce_done; 4707 } 4708 p = &parent->rb_right; 4709 } 4710 insert: 4711 /* Insert segment into RB tree. */ 4712 rb_link_node(&skb->rbnode, parent, p); 4713 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue); 4714 4715 merge_right: 4716 /* Remove other segments covered by skb. */ 4717 while ((skb1 = skb_rb_next(skb)) != NULL) { 4718 if (!after(end_seq, TCP_SKB_CB(skb1)->seq)) 4719 break; 4720 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 4721 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, 4722 end_seq); 4723 break; 4724 } 4725 rb_erase(&skb1->rbnode, &tp->out_of_order_queue); 4726 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, 4727 TCP_SKB_CB(skb1)->end_seq); 4728 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE); 4729 tcp_drop(sk, skb1); 4730 } 4731 /* If there is no skb after us, we are the last_skb ! */ 4732 if (!skb1) 4733 tp->ooo_last_skb = skb; 4734 4735 add_sack: 4736 if (tcp_is_sack(tp)) 4737 tcp_sack_new_ofo_skb(sk, seq, end_seq); 4738 end: 4739 if (skb) { 4740 /* For non sack flows, do not grow window to force DUPACK 4741 * and trigger fast retransmit. 4742 */ 4743 if (tcp_is_sack(tp)) 4744 tcp_grow_window(sk, skb); 4745 skb_condense(skb); 4746 skb_set_owner_r(skb, sk); 4747 } 4748 } 4749 4750 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, 4751 bool *fragstolen) 4752 { 4753 int eaten; 4754 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue); 4755 4756 eaten = (tail && 4757 tcp_try_coalesce(sk, tail, 4758 skb, fragstolen)) ? 1 : 0; 4759 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq); 4760 if (!eaten) { 4761 __skb_queue_tail(&sk->sk_receive_queue, skb); 4762 skb_set_owner_r(skb, sk); 4763 } 4764 return eaten; 4765 } 4766 4767 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size) 4768 { 4769 struct sk_buff *skb; 4770 int err = -ENOMEM; 4771 int data_len = 0; 4772 bool fragstolen; 4773 4774 if (size == 0) 4775 return 0; 4776 4777 if (size > PAGE_SIZE) { 4778 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS); 4779 4780 data_len = npages << PAGE_SHIFT; 4781 size = data_len + (size & ~PAGE_MASK); 4782 } 4783 skb = alloc_skb_with_frags(size - data_len, data_len, 4784 PAGE_ALLOC_COSTLY_ORDER, 4785 &err, sk->sk_allocation); 4786 if (!skb) 4787 goto err; 4788 4789 skb_put(skb, size - data_len); 4790 skb->data_len = data_len; 4791 skb->len = size; 4792 4793 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) { 4794 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP); 4795 goto err_free; 4796 } 4797 4798 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size); 4799 if (err) 4800 goto err_free; 4801 4802 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt; 4803 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size; 4804 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1; 4805 4806 if (tcp_queue_rcv(sk, skb, &fragstolen)) { 4807 WARN_ON_ONCE(fragstolen); /* should not happen */ 4808 __kfree_skb(skb); 4809 } 4810 return size; 4811 4812 err_free: 4813 kfree_skb(skb); 4814 err: 4815 return err; 4816 4817 } 4818 4819 void tcp_data_ready(struct sock *sk) 4820 { 4821 const struct tcp_sock *tp = tcp_sk(sk); 4822 int avail = tp->rcv_nxt - tp->copied_seq; 4823 4824 if (avail < sk->sk_rcvlowat && !tcp_rmem_pressure(sk) && 4825 !sock_flag(sk, SOCK_DONE)) 4826 return; 4827 4828 sk->sk_data_ready(sk); 4829 } 4830 4831 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb) 4832 { 4833 struct tcp_sock *tp = tcp_sk(sk); 4834 bool fragstolen; 4835 int eaten; 4836 4837 if (sk_is_mptcp(sk)) 4838 mptcp_incoming_options(sk, skb, &tp->rx_opt); 4839 4840 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) { 4841 __kfree_skb(skb); 4842 return; 4843 } 4844 skb_dst_drop(skb); 4845 __skb_pull(skb, tcp_hdr(skb)->doff * 4); 4846 4847 tp->rx_opt.dsack = 0; 4848 4849 /* Queue data for delivery to the user. 4850 * Packets in sequence go to the receive queue. 4851 * Out of sequence packets to the out_of_order_queue. 4852 */ 4853 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { 4854 if (tcp_receive_window(tp) == 0) { 4855 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP); 4856 goto out_of_window; 4857 } 4858 4859 /* Ok. In sequence. In window. */ 4860 queue_and_out: 4861 if (skb_queue_len(&sk->sk_receive_queue) == 0) 4862 sk_forced_mem_schedule(sk, skb->truesize); 4863 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) { 4864 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP); 4865 sk->sk_data_ready(sk); 4866 goto drop; 4867 } 4868 4869 eaten = tcp_queue_rcv(sk, skb, &fragstolen); 4870 if (skb->len) 4871 tcp_event_data_recv(sk, skb); 4872 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 4873 tcp_fin(sk); 4874 4875 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) { 4876 tcp_ofo_queue(sk); 4877 4878 /* RFC5681. 4.2. SHOULD send immediate ACK, when 4879 * gap in queue is filled. 4880 */ 4881 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) 4882 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW; 4883 } 4884 4885 if (tp->rx_opt.num_sacks) 4886 tcp_sack_remove(tp); 4887 4888 tcp_fast_path_check(sk); 4889 4890 if (eaten > 0) 4891 kfree_skb_partial(skb, fragstolen); 4892 if (!sock_flag(sk, SOCK_DEAD)) 4893 tcp_data_ready(sk); 4894 return; 4895 } 4896 4897 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 4898 tcp_rcv_spurious_retrans(sk, skb); 4899 /* A retransmit, 2nd most common case. Force an immediate ack. */ 4900 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); 4901 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 4902 4903 out_of_window: 4904 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS); 4905 inet_csk_schedule_ack(sk); 4906 drop: 4907 tcp_drop(sk, skb); 4908 return; 4909 } 4910 4911 /* Out of window. F.e. zero window probe. */ 4912 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp))) 4913 goto out_of_window; 4914 4915 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4916 /* Partial packet, seq < rcv_next < end_seq */ 4917 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt); 4918 4919 /* If window is closed, drop tail of packet. But after 4920 * remembering D-SACK for its head made in previous line. 4921 */ 4922 if (!tcp_receive_window(tp)) { 4923 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP); 4924 goto out_of_window; 4925 } 4926 goto queue_and_out; 4927 } 4928 4929 tcp_data_queue_ofo(sk, skb); 4930 } 4931 4932 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list) 4933 { 4934 if (list) 4935 return !skb_queue_is_last(list, skb) ? skb->next : NULL; 4936 4937 return skb_rb_next(skb); 4938 } 4939 4940 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb, 4941 struct sk_buff_head *list, 4942 struct rb_root *root) 4943 { 4944 struct sk_buff *next = tcp_skb_next(skb, list); 4945 4946 if (list) 4947 __skb_unlink(skb, list); 4948 else 4949 rb_erase(&skb->rbnode, root); 4950 4951 __kfree_skb(skb); 4952 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED); 4953 4954 return next; 4955 } 4956 4957 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */ 4958 void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb) 4959 { 4960 struct rb_node **p = &root->rb_node; 4961 struct rb_node *parent = NULL; 4962 struct sk_buff *skb1; 4963 4964 while (*p) { 4965 parent = *p; 4966 skb1 = rb_to_skb(parent); 4967 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq)) 4968 p = &parent->rb_left; 4969 else 4970 p = &parent->rb_right; 4971 } 4972 rb_link_node(&skb->rbnode, parent, p); 4973 rb_insert_color(&skb->rbnode, root); 4974 } 4975 4976 /* Collapse contiguous sequence of skbs head..tail with 4977 * sequence numbers start..end. 4978 * 4979 * If tail is NULL, this means until the end of the queue. 4980 * 4981 * Segments with FIN/SYN are not collapsed (only because this 4982 * simplifies code) 4983 */ 4984 static void 4985 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root, 4986 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end) 4987 { 4988 struct sk_buff *skb = head, *n; 4989 struct sk_buff_head tmp; 4990 bool end_of_skbs; 4991 4992 /* First, check that queue is collapsible and find 4993 * the point where collapsing can be useful. 4994 */ 4995 restart: 4996 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) { 4997 n = tcp_skb_next(skb, list); 4998 4999 /* No new bits? It is possible on ofo queue. */ 5000 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 5001 skb = tcp_collapse_one(sk, skb, list, root); 5002 if (!skb) 5003 break; 5004 goto restart; 5005 } 5006 5007 /* The first skb to collapse is: 5008 * - not SYN/FIN and 5009 * - bloated or contains data before "start" or 5010 * overlaps to the next one and mptcp allow collapsing. 5011 */ 5012 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) && 5013 (tcp_win_from_space(sk, skb->truesize) > skb->len || 5014 before(TCP_SKB_CB(skb)->seq, start))) { 5015 end_of_skbs = false; 5016 break; 5017 } 5018 5019 if (n && n != tail && mptcp_skb_can_collapse(skb, n) && 5020 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) { 5021 end_of_skbs = false; 5022 break; 5023 } 5024 5025 /* Decided to skip this, advance start seq. */ 5026 start = TCP_SKB_CB(skb)->end_seq; 5027 } 5028 if (end_of_skbs || 5029 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN))) 5030 return; 5031 5032 __skb_queue_head_init(&tmp); 5033 5034 while (before(start, end)) { 5035 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start); 5036 struct sk_buff *nskb; 5037 5038 nskb = alloc_skb(copy, GFP_ATOMIC); 5039 if (!nskb) 5040 break; 5041 5042 memcpy(nskb->cb, skb->cb, sizeof(skb->cb)); 5043 #ifdef CONFIG_TLS_DEVICE 5044 nskb->decrypted = skb->decrypted; 5045 #endif 5046 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start; 5047 if (list) 5048 __skb_queue_before(list, skb, nskb); 5049 else 5050 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */ 5051 skb_set_owner_r(nskb, sk); 5052 mptcp_skb_ext_move(nskb, skb); 5053 5054 /* Copy data, releasing collapsed skbs. */ 5055 while (copy > 0) { 5056 int offset = start - TCP_SKB_CB(skb)->seq; 5057 int size = TCP_SKB_CB(skb)->end_seq - start; 5058 5059 BUG_ON(offset < 0); 5060 if (size > 0) { 5061 size = min(copy, size); 5062 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size)) 5063 BUG(); 5064 TCP_SKB_CB(nskb)->end_seq += size; 5065 copy -= size; 5066 start += size; 5067 } 5068 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 5069 skb = tcp_collapse_one(sk, skb, list, root); 5070 if (!skb || 5071 skb == tail || 5072 !mptcp_skb_can_collapse(nskb, skb) || 5073 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN))) 5074 goto end; 5075 #ifdef CONFIG_TLS_DEVICE 5076 if (skb->decrypted != nskb->decrypted) 5077 goto end; 5078 #endif 5079 } 5080 } 5081 } 5082 end: 5083 skb_queue_walk_safe(&tmp, skb, n) 5084 tcp_rbtree_insert(root, skb); 5085 } 5086 5087 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs 5088 * and tcp_collapse() them until all the queue is collapsed. 5089 */ 5090 static void tcp_collapse_ofo_queue(struct sock *sk) 5091 { 5092 struct tcp_sock *tp = tcp_sk(sk); 5093 u32 range_truesize, sum_tiny = 0; 5094 struct sk_buff *skb, *head; 5095 u32 start, end; 5096 5097 skb = skb_rb_first(&tp->out_of_order_queue); 5098 new_range: 5099 if (!skb) { 5100 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue); 5101 return; 5102 } 5103 start = TCP_SKB_CB(skb)->seq; 5104 end = TCP_SKB_CB(skb)->end_seq; 5105 range_truesize = skb->truesize; 5106 5107 for (head = skb;;) { 5108 skb = skb_rb_next(skb); 5109 5110 /* Range is terminated when we see a gap or when 5111 * we are at the queue end. 5112 */ 5113 if (!skb || 5114 after(TCP_SKB_CB(skb)->seq, end) || 5115 before(TCP_SKB_CB(skb)->end_seq, start)) { 5116 /* Do not attempt collapsing tiny skbs */ 5117 if (range_truesize != head->truesize || 5118 end - start >= SKB_WITH_OVERHEAD(SK_MEM_QUANTUM)) { 5119 tcp_collapse(sk, NULL, &tp->out_of_order_queue, 5120 head, skb, start, end); 5121 } else { 5122 sum_tiny += range_truesize; 5123 if (sum_tiny > sk->sk_rcvbuf >> 3) 5124 return; 5125 } 5126 goto new_range; 5127 } 5128 5129 range_truesize += skb->truesize; 5130 if (unlikely(before(TCP_SKB_CB(skb)->seq, start))) 5131 start = TCP_SKB_CB(skb)->seq; 5132 if (after(TCP_SKB_CB(skb)->end_seq, end)) 5133 end = TCP_SKB_CB(skb)->end_seq; 5134 } 5135 } 5136 5137 /* 5138 * Clean the out-of-order queue to make room. 5139 * We drop high sequences packets to : 5140 * 1) Let a chance for holes to be filled. 5141 * 2) not add too big latencies if thousands of packets sit there. 5142 * (But if application shrinks SO_RCVBUF, we could still end up 5143 * freeing whole queue here) 5144 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks. 5145 * 5146 * Return true if queue has shrunk. 5147 */ 5148 static bool tcp_prune_ofo_queue(struct sock *sk) 5149 { 5150 struct tcp_sock *tp = tcp_sk(sk); 5151 struct rb_node *node, *prev; 5152 int goal; 5153 5154 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) 5155 return false; 5156 5157 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED); 5158 goal = sk->sk_rcvbuf >> 3; 5159 node = &tp->ooo_last_skb->rbnode; 5160 do { 5161 prev = rb_prev(node); 5162 rb_erase(node, &tp->out_of_order_queue); 5163 goal -= rb_to_skb(node)->truesize; 5164 tcp_drop(sk, rb_to_skb(node)); 5165 if (!prev || goal <= 0) { 5166 sk_mem_reclaim(sk); 5167 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf && 5168 !tcp_under_memory_pressure(sk)) 5169 break; 5170 goal = sk->sk_rcvbuf >> 3; 5171 } 5172 node = prev; 5173 } while (node); 5174 tp->ooo_last_skb = rb_to_skb(prev); 5175 5176 /* Reset SACK state. A conforming SACK implementation will 5177 * do the same at a timeout based retransmit. When a connection 5178 * is in a sad state like this, we care only about integrity 5179 * of the connection not performance. 5180 */ 5181 if (tp->rx_opt.sack_ok) 5182 tcp_sack_reset(&tp->rx_opt); 5183 return true; 5184 } 5185 5186 /* Reduce allocated memory if we can, trying to get 5187 * the socket within its memory limits again. 5188 * 5189 * Return less than zero if we should start dropping frames 5190 * until the socket owning process reads some of the data 5191 * to stabilize the situation. 5192 */ 5193 static int tcp_prune_queue(struct sock *sk) 5194 { 5195 struct tcp_sock *tp = tcp_sk(sk); 5196 5197 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED); 5198 5199 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf) 5200 tcp_clamp_window(sk); 5201 else if (tcp_under_memory_pressure(sk)) 5202 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss); 5203 5204 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 5205 return 0; 5206 5207 tcp_collapse_ofo_queue(sk); 5208 if (!skb_queue_empty(&sk->sk_receive_queue)) 5209 tcp_collapse(sk, &sk->sk_receive_queue, NULL, 5210 skb_peek(&sk->sk_receive_queue), 5211 NULL, 5212 tp->copied_seq, tp->rcv_nxt); 5213 sk_mem_reclaim(sk); 5214 5215 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 5216 return 0; 5217 5218 /* Collapsing did not help, destructive actions follow. 5219 * This must not ever occur. */ 5220 5221 tcp_prune_ofo_queue(sk); 5222 5223 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 5224 return 0; 5225 5226 /* If we are really being abused, tell the caller to silently 5227 * drop receive data on the floor. It will get retransmitted 5228 * and hopefully then we'll have sufficient space. 5229 */ 5230 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED); 5231 5232 /* Massive buffer overcommit. */ 5233 tp->pred_flags = 0; 5234 return -1; 5235 } 5236 5237 static bool tcp_should_expand_sndbuf(const struct sock *sk) 5238 { 5239 const struct tcp_sock *tp = tcp_sk(sk); 5240 5241 /* If the user specified a specific send buffer setting, do 5242 * not modify it. 5243 */ 5244 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK) 5245 return false; 5246 5247 /* If we are under global TCP memory pressure, do not expand. */ 5248 if (tcp_under_memory_pressure(sk)) 5249 return false; 5250 5251 /* If we are under soft global TCP memory pressure, do not expand. */ 5252 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0)) 5253 return false; 5254 5255 /* If we filled the congestion window, do not expand. */ 5256 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 5257 return false; 5258 5259 return true; 5260 } 5261 5262 /* When incoming ACK allowed to free some skb from write_queue, 5263 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket 5264 * on the exit from tcp input handler. 5265 * 5266 * PROBLEM: sndbuf expansion does not work well with largesend. 5267 */ 5268 static void tcp_new_space(struct sock *sk) 5269 { 5270 struct tcp_sock *tp = tcp_sk(sk); 5271 5272 if (tcp_should_expand_sndbuf(sk)) { 5273 tcp_sndbuf_expand(sk); 5274 tp->snd_cwnd_stamp = tcp_jiffies32; 5275 } 5276 5277 sk->sk_write_space(sk); 5278 } 5279 5280 static void tcp_check_space(struct sock *sk) 5281 { 5282 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) { 5283 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK); 5284 /* pairs with tcp_poll() */ 5285 smp_mb(); 5286 if (sk->sk_socket && 5287 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 5288 tcp_new_space(sk); 5289 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) 5290 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED); 5291 } 5292 } 5293 } 5294 5295 static inline void tcp_data_snd_check(struct sock *sk) 5296 { 5297 tcp_push_pending_frames(sk); 5298 tcp_check_space(sk); 5299 } 5300 5301 /* 5302 * Check if sending an ack is needed. 5303 */ 5304 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible) 5305 { 5306 struct tcp_sock *tp = tcp_sk(sk); 5307 unsigned long rtt, delay; 5308 5309 /* More than one full frame received... */ 5310 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss && 5311 /* ... and right edge of window advances far enough. 5312 * (tcp_recvmsg() will send ACK otherwise). 5313 * If application uses SO_RCVLOWAT, we want send ack now if 5314 * we have not received enough bytes to satisfy the condition. 5315 */ 5316 (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat || 5317 __tcp_select_window(sk) >= tp->rcv_wnd)) || 5318 /* We ACK each frame or... */ 5319 tcp_in_quickack_mode(sk) || 5320 /* Protocol state mandates a one-time immediate ACK */ 5321 inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) { 5322 send_now: 5323 tcp_send_ack(sk); 5324 return; 5325 } 5326 5327 if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) { 5328 tcp_send_delayed_ack(sk); 5329 return; 5330 } 5331 5332 if (!tcp_is_sack(tp) || 5333 tp->compressed_ack >= sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr) 5334 goto send_now; 5335 5336 if (tp->compressed_ack_rcv_nxt != tp->rcv_nxt) { 5337 tp->compressed_ack_rcv_nxt = tp->rcv_nxt; 5338 tp->dup_ack_counter = 0; 5339 } 5340 if (tp->dup_ack_counter < TCP_FASTRETRANS_THRESH) { 5341 tp->dup_ack_counter++; 5342 goto send_now; 5343 } 5344 tp->compressed_ack++; 5345 if (hrtimer_is_queued(&tp->compressed_ack_timer)) 5346 return; 5347 5348 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */ 5349 5350 rtt = tp->rcv_rtt_est.rtt_us; 5351 if (tp->srtt_us && tp->srtt_us < rtt) 5352 rtt = tp->srtt_us; 5353 5354 delay = min_t(unsigned long, sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns, 5355 rtt * (NSEC_PER_USEC >> 3)/20); 5356 sock_hold(sk); 5357 hrtimer_start_range_ns(&tp->compressed_ack_timer, ns_to_ktime(delay), 5358 sock_net(sk)->ipv4.sysctl_tcp_comp_sack_slack_ns, 5359 HRTIMER_MODE_REL_PINNED_SOFT); 5360 } 5361 5362 static inline void tcp_ack_snd_check(struct sock *sk) 5363 { 5364 if (!inet_csk_ack_scheduled(sk)) { 5365 /* We sent a data segment already. */ 5366 return; 5367 } 5368 __tcp_ack_snd_check(sk, 1); 5369 } 5370 5371 /* 5372 * This routine is only called when we have urgent data 5373 * signaled. Its the 'slow' part of tcp_urg. It could be 5374 * moved inline now as tcp_urg is only called from one 5375 * place. We handle URGent data wrong. We have to - as 5376 * BSD still doesn't use the correction from RFC961. 5377 * For 1003.1g we should support a new option TCP_STDURG to permit 5378 * either form (or just set the sysctl tcp_stdurg). 5379 */ 5380 5381 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th) 5382 { 5383 struct tcp_sock *tp = tcp_sk(sk); 5384 u32 ptr = ntohs(th->urg_ptr); 5385 5386 if (ptr && !sock_net(sk)->ipv4.sysctl_tcp_stdurg) 5387 ptr--; 5388 ptr += ntohl(th->seq); 5389 5390 /* Ignore urgent data that we've already seen and read. */ 5391 if (after(tp->copied_seq, ptr)) 5392 return; 5393 5394 /* Do not replay urg ptr. 5395 * 5396 * NOTE: interesting situation not covered by specs. 5397 * Misbehaving sender may send urg ptr, pointing to segment, 5398 * which we already have in ofo queue. We are not able to fetch 5399 * such data and will stay in TCP_URG_NOTYET until will be eaten 5400 * by recvmsg(). Seems, we are not obliged to handle such wicked 5401 * situations. But it is worth to think about possibility of some 5402 * DoSes using some hypothetical application level deadlock. 5403 */ 5404 if (before(ptr, tp->rcv_nxt)) 5405 return; 5406 5407 /* Do we already have a newer (or duplicate) urgent pointer? */ 5408 if (tp->urg_data && !after(ptr, tp->urg_seq)) 5409 return; 5410 5411 /* Tell the world about our new urgent pointer. */ 5412 sk_send_sigurg(sk); 5413 5414 /* We may be adding urgent data when the last byte read was 5415 * urgent. To do this requires some care. We cannot just ignore 5416 * tp->copied_seq since we would read the last urgent byte again 5417 * as data, nor can we alter copied_seq until this data arrives 5418 * or we break the semantics of SIOCATMARK (and thus sockatmark()) 5419 * 5420 * NOTE. Double Dutch. Rendering to plain English: author of comment 5421 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB); 5422 * and expect that both A and B disappear from stream. This is _wrong_. 5423 * Though this happens in BSD with high probability, this is occasional. 5424 * Any application relying on this is buggy. Note also, that fix "works" 5425 * only in this artificial test. Insert some normal data between A and B and we will 5426 * decline of BSD again. Verdict: it is better to remove to trap 5427 * buggy users. 5428 */ 5429 if (tp->urg_seq == tp->copied_seq && tp->urg_data && 5430 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) { 5431 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 5432 tp->copied_seq++; 5433 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) { 5434 __skb_unlink(skb, &sk->sk_receive_queue); 5435 __kfree_skb(skb); 5436 } 5437 } 5438 5439 tp->urg_data = TCP_URG_NOTYET; 5440 WRITE_ONCE(tp->urg_seq, ptr); 5441 5442 /* Disable header prediction. */ 5443 tp->pred_flags = 0; 5444 } 5445 5446 /* This is the 'fast' part of urgent handling. */ 5447 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th) 5448 { 5449 struct tcp_sock *tp = tcp_sk(sk); 5450 5451 /* Check if we get a new urgent pointer - normally not. */ 5452 if (th->urg) 5453 tcp_check_urg(sk, th); 5454 5455 /* Do we wait for any urgent data? - normally not... */ 5456 if (tp->urg_data == TCP_URG_NOTYET) { 5457 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) - 5458 th->syn; 5459 5460 /* Is the urgent pointer pointing into this packet? */ 5461 if (ptr < skb->len) { 5462 u8 tmp; 5463 if (skb_copy_bits(skb, ptr, &tmp, 1)) 5464 BUG(); 5465 tp->urg_data = TCP_URG_VALID | tmp; 5466 if (!sock_flag(sk, SOCK_DEAD)) 5467 sk->sk_data_ready(sk); 5468 } 5469 } 5470 } 5471 5472 /* Accept RST for rcv_nxt - 1 after a FIN. 5473 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a 5474 * FIN is sent followed by a RST packet. The RST is sent with the same 5475 * sequence number as the FIN, and thus according to RFC 5961 a challenge 5476 * ACK should be sent. However, Mac OSX rate limits replies to challenge 5477 * ACKs on the closed socket. In addition middleboxes can drop either the 5478 * challenge ACK or a subsequent RST. 5479 */ 5480 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb) 5481 { 5482 struct tcp_sock *tp = tcp_sk(sk); 5483 5484 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) && 5485 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK | 5486 TCPF_CLOSING)); 5487 } 5488 5489 /* Does PAWS and seqno based validation of an incoming segment, flags will 5490 * play significant role here. 5491 */ 5492 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb, 5493 const struct tcphdr *th, int syn_inerr) 5494 { 5495 struct tcp_sock *tp = tcp_sk(sk); 5496 bool rst_seq_match = false; 5497 5498 /* RFC1323: H1. Apply PAWS check first. */ 5499 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) && 5500 tp->rx_opt.saw_tstamp && 5501 tcp_paws_discard(sk, skb)) { 5502 if (!th->rst) { 5503 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED); 5504 if (!tcp_oow_rate_limited(sock_net(sk), skb, 5505 LINUX_MIB_TCPACKSKIPPEDPAWS, 5506 &tp->last_oow_ack_time)) 5507 tcp_send_dupack(sk, skb); 5508 goto discard; 5509 } 5510 /* Reset is accepted even if it did not pass PAWS. */ 5511 } 5512 5513 /* Step 1: check sequence number */ 5514 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { 5515 /* RFC793, page 37: "In all states except SYN-SENT, all reset 5516 * (RST) segments are validated by checking their SEQ-fields." 5517 * And page 69: "If an incoming segment is not acceptable, 5518 * an acknowledgment should be sent in reply (unless the RST 5519 * bit is set, if so drop the segment and return)". 5520 */ 5521 if (!th->rst) { 5522 if (th->syn) 5523 goto syn_challenge; 5524 if (!tcp_oow_rate_limited(sock_net(sk), skb, 5525 LINUX_MIB_TCPACKSKIPPEDSEQ, 5526 &tp->last_oow_ack_time)) 5527 tcp_send_dupack(sk, skb); 5528 } else if (tcp_reset_check(sk, skb)) { 5529 tcp_reset(sk); 5530 } 5531 goto discard; 5532 } 5533 5534 /* Step 2: check RST bit */ 5535 if (th->rst) { 5536 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a 5537 * FIN and SACK too if available): 5538 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or 5539 * the right-most SACK block, 5540 * then 5541 * RESET the connection 5542 * else 5543 * Send a challenge ACK 5544 */ 5545 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt || 5546 tcp_reset_check(sk, skb)) { 5547 rst_seq_match = true; 5548 } else if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) { 5549 struct tcp_sack_block *sp = &tp->selective_acks[0]; 5550 int max_sack = sp[0].end_seq; 5551 int this_sack; 5552 5553 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; 5554 ++this_sack) { 5555 max_sack = after(sp[this_sack].end_seq, 5556 max_sack) ? 5557 sp[this_sack].end_seq : max_sack; 5558 } 5559 5560 if (TCP_SKB_CB(skb)->seq == max_sack) 5561 rst_seq_match = true; 5562 } 5563 5564 if (rst_seq_match) 5565 tcp_reset(sk); 5566 else { 5567 /* Disable TFO if RST is out-of-order 5568 * and no data has been received 5569 * for current active TFO socket 5570 */ 5571 if (tp->syn_fastopen && !tp->data_segs_in && 5572 sk->sk_state == TCP_ESTABLISHED) 5573 tcp_fastopen_active_disable(sk); 5574 tcp_send_challenge_ack(sk, skb); 5575 } 5576 goto discard; 5577 } 5578 5579 /* step 3: check security and precedence [ignored] */ 5580 5581 /* step 4: Check for a SYN 5582 * RFC 5961 4.2 : Send a challenge ack 5583 */ 5584 if (th->syn) { 5585 syn_challenge: 5586 if (syn_inerr) 5587 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS); 5588 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE); 5589 tcp_send_challenge_ack(sk, skb); 5590 goto discard; 5591 } 5592 5593 return true; 5594 5595 discard: 5596 tcp_drop(sk, skb); 5597 return false; 5598 } 5599 5600 /* 5601 * TCP receive function for the ESTABLISHED state. 5602 * 5603 * It is split into a fast path and a slow path. The fast path is 5604 * disabled when: 5605 * - A zero window was announced from us - zero window probing 5606 * is only handled properly in the slow path. 5607 * - Out of order segments arrived. 5608 * - Urgent data is expected. 5609 * - There is no buffer space left 5610 * - Unexpected TCP flags/window values/header lengths are received 5611 * (detected by checking the TCP header against pred_flags) 5612 * - Data is sent in both directions. Fast path only supports pure senders 5613 * or pure receivers (this means either the sequence number or the ack 5614 * value must stay constant) 5615 * - Unexpected TCP option. 5616 * 5617 * When these conditions are not satisfied it drops into a standard 5618 * receive procedure patterned after RFC793 to handle all cases. 5619 * The first three cases are guaranteed by proper pred_flags setting, 5620 * the rest is checked inline. Fast processing is turned on in 5621 * tcp_data_queue when everything is OK. 5622 */ 5623 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb) 5624 { 5625 const struct tcphdr *th = (const struct tcphdr *)skb->data; 5626 struct tcp_sock *tp = tcp_sk(sk); 5627 unsigned int len = skb->len; 5628 5629 /* TCP congestion window tracking */ 5630 trace_tcp_probe(sk, skb); 5631 5632 tcp_mstamp_refresh(tp); 5633 if (unlikely(!sk->sk_rx_dst)) 5634 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb); 5635 /* 5636 * Header prediction. 5637 * The code loosely follows the one in the famous 5638 * "30 instruction TCP receive" Van Jacobson mail. 5639 * 5640 * Van's trick is to deposit buffers into socket queue 5641 * on a device interrupt, to call tcp_recv function 5642 * on the receive process context and checksum and copy 5643 * the buffer to user space. smart... 5644 * 5645 * Our current scheme is not silly either but we take the 5646 * extra cost of the net_bh soft interrupt processing... 5647 * We do checksum and copy also but from device to kernel. 5648 */ 5649 5650 tp->rx_opt.saw_tstamp = 0; 5651 5652 /* pred_flags is 0xS?10 << 16 + snd_wnd 5653 * if header_prediction is to be made 5654 * 'S' will always be tp->tcp_header_len >> 2 5655 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to 5656 * turn it off (when there are holes in the receive 5657 * space for instance) 5658 * PSH flag is ignored. 5659 */ 5660 5661 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags && 5662 TCP_SKB_CB(skb)->seq == tp->rcv_nxt && 5663 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) { 5664 int tcp_header_len = tp->tcp_header_len; 5665 5666 /* Timestamp header prediction: tcp_header_len 5667 * is automatically equal to th->doff*4 due to pred_flags 5668 * match. 5669 */ 5670 5671 /* Check timestamp */ 5672 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) { 5673 /* No? Slow path! */ 5674 if (!tcp_parse_aligned_timestamp(tp, th)) 5675 goto slow_path; 5676 5677 /* If PAWS failed, check it more carefully in slow path */ 5678 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0) 5679 goto slow_path; 5680 5681 /* DO NOT update ts_recent here, if checksum fails 5682 * and timestamp was corrupted part, it will result 5683 * in a hung connection since we will drop all 5684 * future packets due to the PAWS test. 5685 */ 5686 } 5687 5688 if (len <= tcp_header_len) { 5689 /* Bulk data transfer: sender */ 5690 if (len == tcp_header_len) { 5691 /* Predicted packet is in window by definition. 5692 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 5693 * Hence, check seq<=rcv_wup reduces to: 5694 */ 5695 if (tcp_header_len == 5696 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 5697 tp->rcv_nxt == tp->rcv_wup) 5698 tcp_store_ts_recent(tp); 5699 5700 /* We know that such packets are checksummed 5701 * on entry. 5702 */ 5703 tcp_ack(sk, skb, 0); 5704 __kfree_skb(skb); 5705 tcp_data_snd_check(sk); 5706 /* When receiving pure ack in fast path, update 5707 * last ts ecr directly instead of calling 5708 * tcp_rcv_rtt_measure_ts() 5709 */ 5710 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr; 5711 return; 5712 } else { /* Header too small */ 5713 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS); 5714 goto discard; 5715 } 5716 } else { 5717 int eaten = 0; 5718 bool fragstolen = false; 5719 5720 if (tcp_checksum_complete(skb)) 5721 goto csum_error; 5722 5723 if ((int)skb->truesize > sk->sk_forward_alloc) 5724 goto step5; 5725 5726 /* Predicted packet is in window by definition. 5727 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 5728 * Hence, check seq<=rcv_wup reduces to: 5729 */ 5730 if (tcp_header_len == 5731 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 5732 tp->rcv_nxt == tp->rcv_wup) 5733 tcp_store_ts_recent(tp); 5734 5735 tcp_rcv_rtt_measure_ts(sk, skb); 5736 5737 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS); 5738 5739 /* Bulk data transfer: receiver */ 5740 __skb_pull(skb, tcp_header_len); 5741 eaten = tcp_queue_rcv(sk, skb, &fragstolen); 5742 5743 tcp_event_data_recv(sk, skb); 5744 5745 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) { 5746 /* Well, only one small jumplet in fast path... */ 5747 tcp_ack(sk, skb, FLAG_DATA); 5748 tcp_data_snd_check(sk); 5749 if (!inet_csk_ack_scheduled(sk)) 5750 goto no_ack; 5751 } 5752 5753 __tcp_ack_snd_check(sk, 0); 5754 no_ack: 5755 if (eaten) 5756 kfree_skb_partial(skb, fragstolen); 5757 tcp_data_ready(sk); 5758 return; 5759 } 5760 } 5761 5762 slow_path: 5763 if (len < (th->doff << 2) || tcp_checksum_complete(skb)) 5764 goto csum_error; 5765 5766 if (!th->ack && !th->rst && !th->syn) 5767 goto discard; 5768 5769 /* 5770 * Standard slow path. 5771 */ 5772 5773 if (!tcp_validate_incoming(sk, skb, th, 1)) 5774 return; 5775 5776 step5: 5777 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0) 5778 goto discard; 5779 5780 tcp_rcv_rtt_measure_ts(sk, skb); 5781 5782 /* Process urgent data. */ 5783 tcp_urg(sk, skb, th); 5784 5785 /* step 7: process the segment text */ 5786 tcp_data_queue(sk, skb); 5787 5788 tcp_data_snd_check(sk); 5789 tcp_ack_snd_check(sk); 5790 return; 5791 5792 csum_error: 5793 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS); 5794 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS); 5795 5796 discard: 5797 tcp_drop(sk, skb); 5798 } 5799 EXPORT_SYMBOL(tcp_rcv_established); 5800 5801 void tcp_init_transfer(struct sock *sk, int bpf_op) 5802 { 5803 struct inet_connection_sock *icsk = inet_csk(sk); 5804 struct tcp_sock *tp = tcp_sk(sk); 5805 5806 tcp_mtup_init(sk); 5807 icsk->icsk_af_ops->rebuild_header(sk); 5808 tcp_init_metrics(sk); 5809 5810 /* Initialize the congestion window to start the transfer. 5811 * Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been 5812 * retransmitted. In light of RFC6298 more aggressive 1sec 5813 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK 5814 * retransmission has occurred. 5815 */ 5816 if (tp->total_retrans > 1 && tp->undo_marker) 5817 tp->snd_cwnd = 1; 5818 else 5819 tp->snd_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk)); 5820 tp->snd_cwnd_stamp = tcp_jiffies32; 5821 5822 tcp_call_bpf(sk, bpf_op, 0, NULL); 5823 tcp_init_congestion_control(sk); 5824 tcp_init_buffer_space(sk); 5825 } 5826 5827 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb) 5828 { 5829 struct tcp_sock *tp = tcp_sk(sk); 5830 struct inet_connection_sock *icsk = inet_csk(sk); 5831 5832 tcp_set_state(sk, TCP_ESTABLISHED); 5833 icsk->icsk_ack.lrcvtime = tcp_jiffies32; 5834 5835 if (skb) { 5836 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb); 5837 security_inet_conn_established(sk, skb); 5838 sk_mark_napi_id(sk, skb); 5839 } 5840 5841 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB); 5842 5843 /* Prevent spurious tcp_cwnd_restart() on first data 5844 * packet. 5845 */ 5846 tp->lsndtime = tcp_jiffies32; 5847 5848 if (sock_flag(sk, SOCK_KEEPOPEN)) 5849 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp)); 5850 5851 if (!tp->rx_opt.snd_wscale) 5852 __tcp_fast_path_on(tp, tp->snd_wnd); 5853 else 5854 tp->pred_flags = 0; 5855 } 5856 5857 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack, 5858 struct tcp_fastopen_cookie *cookie) 5859 { 5860 struct tcp_sock *tp = tcp_sk(sk); 5861 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL; 5862 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0; 5863 bool syn_drop = false; 5864 5865 if (mss == tp->rx_opt.user_mss) { 5866 struct tcp_options_received opt; 5867 5868 /* Get original SYNACK MSS value if user MSS sets mss_clamp */ 5869 tcp_clear_options(&opt); 5870 opt.user_mss = opt.mss_clamp = 0; 5871 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL); 5872 mss = opt.mss_clamp; 5873 } 5874 5875 if (!tp->syn_fastopen) { 5876 /* Ignore an unsolicited cookie */ 5877 cookie->len = -1; 5878 } else if (tp->total_retrans) { 5879 /* SYN timed out and the SYN-ACK neither has a cookie nor 5880 * acknowledges data. Presumably the remote received only 5881 * the retransmitted (regular) SYNs: either the original 5882 * SYN-data or the corresponding SYN-ACK was dropped. 5883 */ 5884 syn_drop = (cookie->len < 0 && data); 5885 } else if (cookie->len < 0 && !tp->syn_data) { 5886 /* We requested a cookie but didn't get it. If we did not use 5887 * the (old) exp opt format then try so next time (try_exp=1). 5888 * Otherwise we go back to use the RFC7413 opt (try_exp=2). 5889 */ 5890 try_exp = tp->syn_fastopen_exp ? 2 : 1; 5891 } 5892 5893 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp); 5894 5895 if (data) { /* Retransmit unacked data in SYN */ 5896 if (tp->total_retrans) 5897 tp->fastopen_client_fail = TFO_SYN_RETRANSMITTED; 5898 else 5899 tp->fastopen_client_fail = TFO_DATA_NOT_ACKED; 5900 skb_rbtree_walk_from(data) { 5901 if (__tcp_retransmit_skb(sk, data, 1)) 5902 break; 5903 } 5904 tcp_rearm_rto(sk); 5905 NET_INC_STATS(sock_net(sk), 5906 LINUX_MIB_TCPFASTOPENACTIVEFAIL); 5907 return true; 5908 } 5909 tp->syn_data_acked = tp->syn_data; 5910 if (tp->syn_data_acked) { 5911 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE); 5912 /* SYN-data is counted as two separate packets in tcp_ack() */ 5913 if (tp->delivered > 1) 5914 --tp->delivered; 5915 } 5916 5917 tcp_fastopen_add_skb(sk, synack); 5918 5919 return false; 5920 } 5921 5922 static void smc_check_reset_syn(struct tcp_sock *tp) 5923 { 5924 #if IS_ENABLED(CONFIG_SMC) 5925 if (static_branch_unlikely(&tcp_have_smc)) { 5926 if (tp->syn_smc && !tp->rx_opt.smc_ok) 5927 tp->syn_smc = 0; 5928 } 5929 #endif 5930 } 5931 5932 static void tcp_try_undo_spurious_syn(struct sock *sk) 5933 { 5934 struct tcp_sock *tp = tcp_sk(sk); 5935 u32 syn_stamp; 5936 5937 /* undo_marker is set when SYN or SYNACK times out. The timeout is 5938 * spurious if the ACK's timestamp option echo value matches the 5939 * original SYN timestamp. 5940 */ 5941 syn_stamp = tp->retrans_stamp; 5942 if (tp->undo_marker && syn_stamp && tp->rx_opt.saw_tstamp && 5943 syn_stamp == tp->rx_opt.rcv_tsecr) 5944 tp->undo_marker = 0; 5945 } 5946 5947 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb, 5948 const struct tcphdr *th) 5949 { 5950 struct inet_connection_sock *icsk = inet_csk(sk); 5951 struct tcp_sock *tp = tcp_sk(sk); 5952 struct tcp_fastopen_cookie foc = { .len = -1 }; 5953 int saved_clamp = tp->rx_opt.mss_clamp; 5954 bool fastopen_fail; 5955 5956 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc); 5957 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) 5958 tp->rx_opt.rcv_tsecr -= tp->tsoffset; 5959 5960 if (th->ack) { 5961 /* rfc793: 5962 * "If the state is SYN-SENT then 5963 * first check the ACK bit 5964 * If the ACK bit is set 5965 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send 5966 * a reset (unless the RST bit is set, if so drop 5967 * the segment and return)" 5968 */ 5969 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) || 5970 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) { 5971 /* Previous FIN/ACK or RST/ACK might be ignored. */ 5972 if (icsk->icsk_retransmits == 0) 5973 inet_csk_reset_xmit_timer(sk, 5974 ICSK_TIME_RETRANS, 5975 TCP_TIMEOUT_MIN, TCP_RTO_MAX); 5976 goto reset_and_undo; 5977 } 5978 5979 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 5980 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp, 5981 tcp_time_stamp(tp))) { 5982 NET_INC_STATS(sock_net(sk), 5983 LINUX_MIB_PAWSACTIVEREJECTED); 5984 goto reset_and_undo; 5985 } 5986 5987 /* Now ACK is acceptable. 5988 * 5989 * "If the RST bit is set 5990 * If the ACK was acceptable then signal the user "error: 5991 * connection reset", drop the segment, enter CLOSED state, 5992 * delete TCB, and return." 5993 */ 5994 5995 if (th->rst) { 5996 tcp_reset(sk); 5997 goto discard; 5998 } 5999 6000 /* rfc793: 6001 * "fifth, if neither of the SYN or RST bits is set then 6002 * drop the segment and return." 6003 * 6004 * See note below! 6005 * --ANK(990513) 6006 */ 6007 if (!th->syn) 6008 goto discard_and_undo; 6009 6010 /* rfc793: 6011 * "If the SYN bit is on ... 6012 * are acceptable then ... 6013 * (our SYN has been ACKed), change the connection 6014 * state to ESTABLISHED..." 6015 */ 6016 6017 tcp_ecn_rcv_synack(tp, th); 6018 6019 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); 6020 tcp_try_undo_spurious_syn(sk); 6021 tcp_ack(sk, skb, FLAG_SLOWPATH); 6022 6023 /* Ok.. it's good. Set up sequence numbers and 6024 * move to established. 6025 */ 6026 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1); 6027 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 6028 6029 /* RFC1323: The window in SYN & SYN/ACK segments is 6030 * never scaled. 6031 */ 6032 tp->snd_wnd = ntohs(th->window); 6033 6034 if (!tp->rx_opt.wscale_ok) { 6035 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0; 6036 tp->window_clamp = min(tp->window_clamp, 65535U); 6037 } 6038 6039 if (tp->rx_opt.saw_tstamp) { 6040 tp->rx_opt.tstamp_ok = 1; 6041 tp->tcp_header_len = 6042 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 6043 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 6044 tcp_store_ts_recent(tp); 6045 } else { 6046 tp->tcp_header_len = sizeof(struct tcphdr); 6047 } 6048 6049 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 6050 tcp_initialize_rcv_mss(sk); 6051 6052 /* Remember, tcp_poll() does not lock socket! 6053 * Change state from SYN-SENT only after copied_seq 6054 * is initialized. */ 6055 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 6056 6057 smc_check_reset_syn(tp); 6058 6059 smp_mb(); 6060 6061 tcp_finish_connect(sk, skb); 6062 6063 fastopen_fail = (tp->syn_fastopen || tp->syn_data) && 6064 tcp_rcv_fastopen_synack(sk, skb, &foc); 6065 6066 if (!sock_flag(sk, SOCK_DEAD)) { 6067 sk->sk_state_change(sk); 6068 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT); 6069 } 6070 if (fastopen_fail) 6071 return -1; 6072 if (sk->sk_write_pending || 6073 icsk->icsk_accept_queue.rskq_defer_accept || 6074 inet_csk_in_pingpong_mode(sk)) { 6075 /* Save one ACK. Data will be ready after 6076 * several ticks, if write_pending is set. 6077 * 6078 * It may be deleted, but with this feature tcpdumps 6079 * look so _wonderfully_ clever, that I was not able 6080 * to stand against the temptation 8) --ANK 6081 */ 6082 inet_csk_schedule_ack(sk); 6083 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS); 6084 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 6085 TCP_DELACK_MAX, TCP_RTO_MAX); 6086 6087 discard: 6088 tcp_drop(sk, skb); 6089 return 0; 6090 } else { 6091 tcp_send_ack(sk); 6092 } 6093 return -1; 6094 } 6095 6096 /* No ACK in the segment */ 6097 6098 if (th->rst) { 6099 /* rfc793: 6100 * "If the RST bit is set 6101 * 6102 * Otherwise (no ACK) drop the segment and return." 6103 */ 6104 6105 goto discard_and_undo; 6106 } 6107 6108 /* PAWS check. */ 6109 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && 6110 tcp_paws_reject(&tp->rx_opt, 0)) 6111 goto discard_and_undo; 6112 6113 if (th->syn) { 6114 /* We see SYN without ACK. It is attempt of 6115 * simultaneous connect with crossed SYNs. 6116 * Particularly, it can be connect to self. 6117 */ 6118 tcp_set_state(sk, TCP_SYN_RECV); 6119 6120 if (tp->rx_opt.saw_tstamp) { 6121 tp->rx_opt.tstamp_ok = 1; 6122 tcp_store_ts_recent(tp); 6123 tp->tcp_header_len = 6124 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 6125 } else { 6126 tp->tcp_header_len = sizeof(struct tcphdr); 6127 } 6128 6129 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1); 6130 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 6131 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 6132 6133 /* RFC1323: The window in SYN & SYN/ACK segments is 6134 * never scaled. 6135 */ 6136 tp->snd_wnd = ntohs(th->window); 6137 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 6138 tp->max_window = tp->snd_wnd; 6139 6140 tcp_ecn_rcv_syn(tp, th); 6141 6142 tcp_mtup_init(sk); 6143 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 6144 tcp_initialize_rcv_mss(sk); 6145 6146 tcp_send_synack(sk); 6147 #if 0 6148 /* Note, we could accept data and URG from this segment. 6149 * There are no obstacles to make this (except that we must 6150 * either change tcp_recvmsg() to prevent it from returning data 6151 * before 3WHS completes per RFC793, or employ TCP Fast Open). 6152 * 6153 * However, if we ignore data in ACKless segments sometimes, 6154 * we have no reasons to accept it sometimes. 6155 * Also, seems the code doing it in step6 of tcp_rcv_state_process 6156 * is not flawless. So, discard packet for sanity. 6157 * Uncomment this return to process the data. 6158 */ 6159 return -1; 6160 #else 6161 goto discard; 6162 #endif 6163 } 6164 /* "fifth, if neither of the SYN or RST bits is set then 6165 * drop the segment and return." 6166 */ 6167 6168 discard_and_undo: 6169 tcp_clear_options(&tp->rx_opt); 6170 tp->rx_opt.mss_clamp = saved_clamp; 6171 goto discard; 6172 6173 reset_and_undo: 6174 tcp_clear_options(&tp->rx_opt); 6175 tp->rx_opt.mss_clamp = saved_clamp; 6176 return 1; 6177 } 6178 6179 static void tcp_rcv_synrecv_state_fastopen(struct sock *sk) 6180 { 6181 struct request_sock *req; 6182 6183 /* If we are still handling the SYNACK RTO, see if timestamp ECR allows 6184 * undo. If peer SACKs triggered fast recovery, we can't undo here. 6185 */ 6186 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss) 6187 tcp_try_undo_loss(sk, false); 6188 6189 /* Reset rtx states to prevent spurious retransmits_timed_out() */ 6190 tcp_sk(sk)->retrans_stamp = 0; 6191 inet_csk(sk)->icsk_retransmits = 0; 6192 6193 /* Once we leave TCP_SYN_RECV or TCP_FIN_WAIT_1, 6194 * we no longer need req so release it. 6195 */ 6196 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk, 6197 lockdep_sock_is_held(sk)); 6198 reqsk_fastopen_remove(sk, req, false); 6199 6200 /* Re-arm the timer because data may have been sent out. 6201 * This is similar to the regular data transmission case 6202 * when new data has just been ack'ed. 6203 * 6204 * (TFO) - we could try to be more aggressive and 6205 * retransmitting any data sooner based on when they 6206 * are sent out. 6207 */ 6208 tcp_rearm_rto(sk); 6209 } 6210 6211 /* 6212 * This function implements the receiving procedure of RFC 793 for 6213 * all states except ESTABLISHED and TIME_WAIT. 6214 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be 6215 * address independent. 6216 */ 6217 6218 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb) 6219 { 6220 struct tcp_sock *tp = tcp_sk(sk); 6221 struct inet_connection_sock *icsk = inet_csk(sk); 6222 const struct tcphdr *th = tcp_hdr(skb); 6223 struct request_sock *req; 6224 int queued = 0; 6225 bool acceptable; 6226 6227 switch (sk->sk_state) { 6228 case TCP_CLOSE: 6229 goto discard; 6230 6231 case TCP_LISTEN: 6232 if (th->ack) 6233 return 1; 6234 6235 if (th->rst) 6236 goto discard; 6237 6238 if (th->syn) { 6239 if (th->fin) 6240 goto discard; 6241 /* It is possible that we process SYN packets from backlog, 6242 * so we need to make sure to disable BH and RCU right there. 6243 */ 6244 rcu_read_lock(); 6245 local_bh_disable(); 6246 acceptable = icsk->icsk_af_ops->conn_request(sk, skb) >= 0; 6247 local_bh_enable(); 6248 rcu_read_unlock(); 6249 6250 if (!acceptable) 6251 return 1; 6252 consume_skb(skb); 6253 return 0; 6254 } 6255 goto discard; 6256 6257 case TCP_SYN_SENT: 6258 tp->rx_opt.saw_tstamp = 0; 6259 tcp_mstamp_refresh(tp); 6260 queued = tcp_rcv_synsent_state_process(sk, skb, th); 6261 if (queued >= 0) 6262 return queued; 6263 6264 /* Do step6 onward by hand. */ 6265 tcp_urg(sk, skb, th); 6266 __kfree_skb(skb); 6267 tcp_data_snd_check(sk); 6268 return 0; 6269 } 6270 6271 tcp_mstamp_refresh(tp); 6272 tp->rx_opt.saw_tstamp = 0; 6273 req = rcu_dereference_protected(tp->fastopen_rsk, 6274 lockdep_sock_is_held(sk)); 6275 if (req) { 6276 bool req_stolen; 6277 6278 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV && 6279 sk->sk_state != TCP_FIN_WAIT1); 6280 6281 if (!tcp_check_req(sk, skb, req, true, &req_stolen)) 6282 goto discard; 6283 } 6284 6285 if (!th->ack && !th->rst && !th->syn) 6286 goto discard; 6287 6288 if (!tcp_validate_incoming(sk, skb, th, 0)) 6289 return 0; 6290 6291 /* step 5: check the ACK field */ 6292 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH | 6293 FLAG_UPDATE_TS_RECENT | 6294 FLAG_NO_CHALLENGE_ACK) > 0; 6295 6296 if (!acceptable) { 6297 if (sk->sk_state == TCP_SYN_RECV) 6298 return 1; /* send one RST */ 6299 tcp_send_challenge_ack(sk, skb); 6300 goto discard; 6301 } 6302 switch (sk->sk_state) { 6303 case TCP_SYN_RECV: 6304 tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */ 6305 if (!tp->srtt_us) 6306 tcp_synack_rtt_meas(sk, req); 6307 6308 if (req) { 6309 tcp_rcv_synrecv_state_fastopen(sk); 6310 } else { 6311 tcp_try_undo_spurious_syn(sk); 6312 tp->retrans_stamp = 0; 6313 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB); 6314 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 6315 } 6316 smp_mb(); 6317 tcp_set_state(sk, TCP_ESTABLISHED); 6318 sk->sk_state_change(sk); 6319 6320 /* Note, that this wakeup is only for marginal crossed SYN case. 6321 * Passively open sockets are not waked up, because 6322 * sk->sk_sleep == NULL and sk->sk_socket == NULL. 6323 */ 6324 if (sk->sk_socket) 6325 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT); 6326 6327 tp->snd_una = TCP_SKB_CB(skb)->ack_seq; 6328 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale; 6329 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); 6330 6331 if (tp->rx_opt.tstamp_ok) 6332 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 6333 6334 if (!inet_csk(sk)->icsk_ca_ops->cong_control) 6335 tcp_update_pacing_rate(sk); 6336 6337 /* Prevent spurious tcp_cwnd_restart() on first data packet */ 6338 tp->lsndtime = tcp_jiffies32; 6339 6340 tcp_initialize_rcv_mss(sk); 6341 tcp_fast_path_on(tp); 6342 break; 6343 6344 case TCP_FIN_WAIT1: { 6345 int tmo; 6346 6347 if (req) 6348 tcp_rcv_synrecv_state_fastopen(sk); 6349 6350 if (tp->snd_una != tp->write_seq) 6351 break; 6352 6353 tcp_set_state(sk, TCP_FIN_WAIT2); 6354 sk->sk_shutdown |= SEND_SHUTDOWN; 6355 6356 sk_dst_confirm(sk); 6357 6358 if (!sock_flag(sk, SOCK_DEAD)) { 6359 /* Wake up lingering close() */ 6360 sk->sk_state_change(sk); 6361 break; 6362 } 6363 6364 if (tp->linger2 < 0) { 6365 tcp_done(sk); 6366 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 6367 return 1; 6368 } 6369 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 6370 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { 6371 /* Receive out of order FIN after close() */ 6372 if (tp->syn_fastopen && th->fin) 6373 tcp_fastopen_active_disable(sk); 6374 tcp_done(sk); 6375 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 6376 return 1; 6377 } 6378 6379 tmo = tcp_fin_time(sk); 6380 if (tmo > TCP_TIMEWAIT_LEN) { 6381 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN); 6382 } else if (th->fin || sock_owned_by_user(sk)) { 6383 /* Bad case. We could lose such FIN otherwise. 6384 * It is not a big problem, but it looks confusing 6385 * and not so rare event. We still can lose it now, 6386 * if it spins in bh_lock_sock(), but it is really 6387 * marginal case. 6388 */ 6389 inet_csk_reset_keepalive_timer(sk, tmo); 6390 } else { 6391 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 6392 goto discard; 6393 } 6394 break; 6395 } 6396 6397 case TCP_CLOSING: 6398 if (tp->snd_una == tp->write_seq) { 6399 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 6400 goto discard; 6401 } 6402 break; 6403 6404 case TCP_LAST_ACK: 6405 if (tp->snd_una == tp->write_seq) { 6406 tcp_update_metrics(sk); 6407 tcp_done(sk); 6408 goto discard; 6409 } 6410 break; 6411 } 6412 6413 /* step 6: check the URG bit */ 6414 tcp_urg(sk, skb, th); 6415 6416 /* step 7: process the segment text */ 6417 switch (sk->sk_state) { 6418 case TCP_CLOSE_WAIT: 6419 case TCP_CLOSING: 6420 case TCP_LAST_ACK: 6421 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 6422 if (sk_is_mptcp(sk)) 6423 mptcp_incoming_options(sk, skb, &tp->rx_opt); 6424 break; 6425 } 6426 fallthrough; 6427 case TCP_FIN_WAIT1: 6428 case TCP_FIN_WAIT2: 6429 /* RFC 793 says to queue data in these states, 6430 * RFC 1122 says we MUST send a reset. 6431 * BSD 4.4 also does reset. 6432 */ 6433 if (sk->sk_shutdown & RCV_SHUTDOWN) { 6434 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 6435 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { 6436 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 6437 tcp_reset(sk); 6438 return 1; 6439 } 6440 } 6441 fallthrough; 6442 case TCP_ESTABLISHED: 6443 tcp_data_queue(sk, skb); 6444 queued = 1; 6445 break; 6446 } 6447 6448 /* tcp_data could move socket to TIME-WAIT */ 6449 if (sk->sk_state != TCP_CLOSE) { 6450 tcp_data_snd_check(sk); 6451 tcp_ack_snd_check(sk); 6452 } 6453 6454 if (!queued) { 6455 discard: 6456 tcp_drop(sk, skb); 6457 } 6458 return 0; 6459 } 6460 EXPORT_SYMBOL(tcp_rcv_state_process); 6461 6462 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family) 6463 { 6464 struct inet_request_sock *ireq = inet_rsk(req); 6465 6466 if (family == AF_INET) 6467 net_dbg_ratelimited("drop open request from %pI4/%u\n", 6468 &ireq->ir_rmt_addr, port); 6469 #if IS_ENABLED(CONFIG_IPV6) 6470 else if (family == AF_INET6) 6471 net_dbg_ratelimited("drop open request from %pI6/%u\n", 6472 &ireq->ir_v6_rmt_addr, port); 6473 #endif 6474 } 6475 6476 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set 6477 * 6478 * If we receive a SYN packet with these bits set, it means a 6479 * network is playing bad games with TOS bits. In order to 6480 * avoid possible false congestion notifications, we disable 6481 * TCP ECN negotiation. 6482 * 6483 * Exception: tcp_ca wants ECN. This is required for DCTCP 6484 * congestion control: Linux DCTCP asserts ECT on all packets, 6485 * including SYN, which is most optimal solution; however, 6486 * others, such as FreeBSD do not. 6487 * 6488 * Exception: At least one of the reserved bits of the TCP header (th->res1) is 6489 * set, indicating the use of a future TCP extension (such as AccECN). See 6490 * RFC8311 §4.3 which updates RFC3168 to allow the development of such 6491 * extensions. 6492 */ 6493 static void tcp_ecn_create_request(struct request_sock *req, 6494 const struct sk_buff *skb, 6495 const struct sock *listen_sk, 6496 const struct dst_entry *dst) 6497 { 6498 const struct tcphdr *th = tcp_hdr(skb); 6499 const struct net *net = sock_net(listen_sk); 6500 bool th_ecn = th->ece && th->cwr; 6501 bool ect, ecn_ok; 6502 u32 ecn_ok_dst; 6503 6504 if (!th_ecn) 6505 return; 6506 6507 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield); 6508 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK); 6509 ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst; 6510 6511 if (((!ect || th->res1) && ecn_ok) || tcp_ca_needs_ecn(listen_sk) || 6512 (ecn_ok_dst & DST_FEATURE_ECN_CA) || 6513 tcp_bpf_ca_needs_ecn((struct sock *)req)) 6514 inet_rsk(req)->ecn_ok = 1; 6515 } 6516 6517 static void tcp_openreq_init(struct request_sock *req, 6518 const struct tcp_options_received *rx_opt, 6519 struct sk_buff *skb, const struct sock *sk) 6520 { 6521 struct inet_request_sock *ireq = inet_rsk(req); 6522 6523 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */ 6524 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq; 6525 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 6526 tcp_rsk(req)->snt_synack = 0; 6527 tcp_rsk(req)->last_oow_ack_time = 0; 6528 req->mss = rx_opt->mss_clamp; 6529 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0; 6530 ireq->tstamp_ok = rx_opt->tstamp_ok; 6531 ireq->sack_ok = rx_opt->sack_ok; 6532 ireq->snd_wscale = rx_opt->snd_wscale; 6533 ireq->wscale_ok = rx_opt->wscale_ok; 6534 ireq->acked = 0; 6535 ireq->ecn_ok = 0; 6536 ireq->ir_rmt_port = tcp_hdr(skb)->source; 6537 ireq->ir_num = ntohs(tcp_hdr(skb)->dest); 6538 ireq->ir_mark = inet_request_mark(sk, skb); 6539 #if IS_ENABLED(CONFIG_SMC) 6540 ireq->smc_ok = rx_opt->smc_ok; 6541 #endif 6542 } 6543 6544 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops, 6545 struct sock *sk_listener, 6546 bool attach_listener) 6547 { 6548 struct request_sock *req = reqsk_alloc(ops, sk_listener, 6549 attach_listener); 6550 6551 if (req) { 6552 struct inet_request_sock *ireq = inet_rsk(req); 6553 6554 ireq->ireq_opt = NULL; 6555 #if IS_ENABLED(CONFIG_IPV6) 6556 ireq->pktopts = NULL; 6557 #endif 6558 atomic64_set(&ireq->ir_cookie, 0); 6559 ireq->ireq_state = TCP_NEW_SYN_RECV; 6560 write_pnet(&ireq->ireq_net, sock_net(sk_listener)); 6561 ireq->ireq_family = sk_listener->sk_family; 6562 } 6563 6564 return req; 6565 } 6566 EXPORT_SYMBOL(inet_reqsk_alloc); 6567 6568 /* 6569 * Return true if a syncookie should be sent 6570 */ 6571 static bool tcp_syn_flood_action(const struct sock *sk, const char *proto) 6572 { 6573 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue; 6574 const char *msg = "Dropping request"; 6575 bool want_cookie = false; 6576 struct net *net = sock_net(sk); 6577 6578 #ifdef CONFIG_SYN_COOKIES 6579 if (net->ipv4.sysctl_tcp_syncookies) { 6580 msg = "Sending cookies"; 6581 want_cookie = true; 6582 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES); 6583 } else 6584 #endif 6585 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP); 6586 6587 if (!queue->synflood_warned && 6588 net->ipv4.sysctl_tcp_syncookies != 2 && 6589 xchg(&queue->synflood_warned, 1) == 0) 6590 net_info_ratelimited("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n", 6591 proto, sk->sk_num, msg); 6592 6593 return want_cookie; 6594 } 6595 6596 static void tcp_reqsk_record_syn(const struct sock *sk, 6597 struct request_sock *req, 6598 const struct sk_buff *skb) 6599 { 6600 if (tcp_sk(sk)->save_syn) { 6601 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb); 6602 u32 *copy; 6603 6604 copy = kmalloc(len + sizeof(u32), GFP_ATOMIC); 6605 if (copy) { 6606 copy[0] = len; 6607 memcpy(©[1], skb_network_header(skb), len); 6608 req->saved_syn = copy; 6609 } 6610 } 6611 } 6612 6613 /* If a SYN cookie is required and supported, returns a clamped MSS value to be 6614 * used for SYN cookie generation. 6615 */ 6616 u16 tcp_get_syncookie_mss(struct request_sock_ops *rsk_ops, 6617 const struct tcp_request_sock_ops *af_ops, 6618 struct sock *sk, struct tcphdr *th) 6619 { 6620 struct tcp_sock *tp = tcp_sk(sk); 6621 u16 mss; 6622 6623 if (sock_net(sk)->ipv4.sysctl_tcp_syncookies != 2 && 6624 !inet_csk_reqsk_queue_is_full(sk)) 6625 return 0; 6626 6627 if (!tcp_syn_flood_action(sk, rsk_ops->slab_name)) 6628 return 0; 6629 6630 if (sk_acceptq_is_full(sk)) { 6631 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS); 6632 return 0; 6633 } 6634 6635 mss = tcp_parse_mss_option(th, tp->rx_opt.user_mss); 6636 if (!mss) 6637 mss = af_ops->mss_clamp; 6638 6639 return mss; 6640 } 6641 EXPORT_SYMBOL_GPL(tcp_get_syncookie_mss); 6642 6643 int tcp_conn_request(struct request_sock_ops *rsk_ops, 6644 const struct tcp_request_sock_ops *af_ops, 6645 struct sock *sk, struct sk_buff *skb) 6646 { 6647 struct tcp_fastopen_cookie foc = { .len = -1 }; 6648 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn; 6649 struct tcp_options_received tmp_opt; 6650 struct tcp_sock *tp = tcp_sk(sk); 6651 struct net *net = sock_net(sk); 6652 struct sock *fastopen_sk = NULL; 6653 struct request_sock *req; 6654 bool want_cookie = false; 6655 struct dst_entry *dst; 6656 struct flowi fl; 6657 6658 /* TW buckets are converted to open requests without 6659 * limitations, they conserve resources and peer is 6660 * evidently real one. 6661 */ 6662 if ((net->ipv4.sysctl_tcp_syncookies == 2 || 6663 inet_csk_reqsk_queue_is_full(sk)) && !isn) { 6664 want_cookie = tcp_syn_flood_action(sk, rsk_ops->slab_name); 6665 if (!want_cookie) 6666 goto drop; 6667 } 6668 6669 if (sk_acceptq_is_full(sk)) { 6670 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS); 6671 goto drop; 6672 } 6673 6674 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie); 6675 if (!req) 6676 goto drop; 6677 6678 req->syncookie = want_cookie; 6679 tcp_rsk(req)->af_specific = af_ops; 6680 tcp_rsk(req)->ts_off = 0; 6681 #if IS_ENABLED(CONFIG_MPTCP) 6682 tcp_rsk(req)->is_mptcp = 0; 6683 #endif 6684 6685 tcp_clear_options(&tmp_opt); 6686 tmp_opt.mss_clamp = af_ops->mss_clamp; 6687 tmp_opt.user_mss = tp->rx_opt.user_mss; 6688 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0, 6689 want_cookie ? NULL : &foc); 6690 6691 if (want_cookie && !tmp_opt.saw_tstamp) 6692 tcp_clear_options(&tmp_opt); 6693 6694 if (IS_ENABLED(CONFIG_SMC) && want_cookie) 6695 tmp_opt.smc_ok = 0; 6696 6697 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp; 6698 tcp_openreq_init(req, &tmp_opt, skb, sk); 6699 inet_rsk(req)->no_srccheck = inet_sk(sk)->transparent; 6700 6701 /* Note: tcp_v6_init_req() might override ir_iif for link locals */ 6702 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb); 6703 6704 af_ops->init_req(req, sk, skb); 6705 6706 if (security_inet_conn_request(sk, skb, req)) 6707 goto drop_and_free; 6708 6709 if (tmp_opt.tstamp_ok) 6710 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb); 6711 6712 dst = af_ops->route_req(sk, &fl, req); 6713 if (!dst) 6714 goto drop_and_free; 6715 6716 if (!want_cookie && !isn) { 6717 /* Kill the following clause, if you dislike this way. */ 6718 if (!net->ipv4.sysctl_tcp_syncookies && 6719 (net->ipv4.sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) < 6720 (net->ipv4.sysctl_max_syn_backlog >> 2)) && 6721 !tcp_peer_is_proven(req, dst)) { 6722 /* Without syncookies last quarter of 6723 * backlog is filled with destinations, 6724 * proven to be alive. 6725 * It means that we continue to communicate 6726 * to destinations, already remembered 6727 * to the moment of synflood. 6728 */ 6729 pr_drop_req(req, ntohs(tcp_hdr(skb)->source), 6730 rsk_ops->family); 6731 goto drop_and_release; 6732 } 6733 6734 isn = af_ops->init_seq(skb); 6735 } 6736 6737 tcp_ecn_create_request(req, skb, sk, dst); 6738 6739 if (want_cookie) { 6740 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss); 6741 if (!tmp_opt.tstamp_ok) 6742 inet_rsk(req)->ecn_ok = 0; 6743 } 6744 6745 tcp_rsk(req)->snt_isn = isn; 6746 tcp_rsk(req)->txhash = net_tx_rndhash(); 6747 tcp_openreq_init_rwin(req, sk, dst); 6748 sk_rx_queue_set(req_to_sk(req), skb); 6749 if (!want_cookie) { 6750 tcp_reqsk_record_syn(sk, req, skb); 6751 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst); 6752 } 6753 if (fastopen_sk) { 6754 af_ops->send_synack(fastopen_sk, dst, &fl, req, 6755 &foc, TCP_SYNACK_FASTOPEN); 6756 /* Add the child socket directly into the accept queue */ 6757 if (!inet_csk_reqsk_queue_add(sk, req, fastopen_sk)) { 6758 reqsk_fastopen_remove(fastopen_sk, req, false); 6759 bh_unlock_sock(fastopen_sk); 6760 sock_put(fastopen_sk); 6761 goto drop_and_free; 6762 } 6763 sk->sk_data_ready(sk); 6764 bh_unlock_sock(fastopen_sk); 6765 sock_put(fastopen_sk); 6766 } else { 6767 tcp_rsk(req)->tfo_listener = false; 6768 if (!want_cookie) 6769 inet_csk_reqsk_queue_hash_add(sk, req, 6770 tcp_timeout_init((struct sock *)req)); 6771 af_ops->send_synack(sk, dst, &fl, req, &foc, 6772 !want_cookie ? TCP_SYNACK_NORMAL : 6773 TCP_SYNACK_COOKIE); 6774 if (want_cookie) { 6775 reqsk_free(req); 6776 return 0; 6777 } 6778 } 6779 reqsk_put(req); 6780 return 0; 6781 6782 drop_and_release: 6783 dst_release(dst); 6784 drop_and_free: 6785 __reqsk_free(req); 6786 drop: 6787 tcp_listendrop(sk); 6788 return 0; 6789 } 6790 EXPORT_SYMBOL(tcp_conn_request); 6791