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 #ifdef CONFIG_CGROUP_BPF 142 static void bpf_skops_parse_hdr(struct sock *sk, struct sk_buff *skb) 143 { 144 bool unknown_opt = tcp_sk(sk)->rx_opt.saw_unknown && 145 BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), 146 BPF_SOCK_OPS_PARSE_UNKNOWN_HDR_OPT_CB_FLAG); 147 bool parse_all_opt = BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), 148 BPF_SOCK_OPS_PARSE_ALL_HDR_OPT_CB_FLAG); 149 struct bpf_sock_ops_kern sock_ops; 150 151 if (likely(!unknown_opt && !parse_all_opt)) 152 return; 153 154 /* The skb will be handled in the 155 * bpf_skops_established() or 156 * bpf_skops_write_hdr_opt(). 157 */ 158 switch (sk->sk_state) { 159 case TCP_SYN_RECV: 160 case TCP_SYN_SENT: 161 case TCP_LISTEN: 162 return; 163 } 164 165 sock_owned_by_me(sk); 166 167 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp)); 168 sock_ops.op = BPF_SOCK_OPS_PARSE_HDR_OPT_CB; 169 sock_ops.is_fullsock = 1; 170 sock_ops.sk = sk; 171 bpf_skops_init_skb(&sock_ops, skb, tcp_hdrlen(skb)); 172 173 BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops); 174 } 175 176 static void bpf_skops_established(struct sock *sk, int bpf_op, 177 struct sk_buff *skb) 178 { 179 struct bpf_sock_ops_kern sock_ops; 180 181 sock_owned_by_me(sk); 182 183 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp)); 184 sock_ops.op = bpf_op; 185 sock_ops.is_fullsock = 1; 186 sock_ops.sk = sk; 187 /* sk with TCP_REPAIR_ON does not have skb in tcp_finish_connect */ 188 if (skb) 189 bpf_skops_init_skb(&sock_ops, skb, tcp_hdrlen(skb)); 190 191 BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops); 192 } 193 #else 194 static void bpf_skops_parse_hdr(struct sock *sk, struct sk_buff *skb) 195 { 196 } 197 198 static void bpf_skops_established(struct sock *sk, int bpf_op, 199 struct sk_buff *skb) 200 { 201 } 202 #endif 203 204 static void tcp_gro_dev_warn(struct sock *sk, const struct sk_buff *skb, 205 unsigned int len) 206 { 207 static bool __once __read_mostly; 208 209 if (!__once) { 210 struct net_device *dev; 211 212 __once = true; 213 214 rcu_read_lock(); 215 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif); 216 if (!dev || len >= dev->mtu) 217 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n", 218 dev ? dev->name : "Unknown driver"); 219 rcu_read_unlock(); 220 } 221 } 222 223 /* Adapt the MSS value used to make delayed ack decision to the 224 * real world. 225 */ 226 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb) 227 { 228 struct inet_connection_sock *icsk = inet_csk(sk); 229 const unsigned int lss = icsk->icsk_ack.last_seg_size; 230 unsigned int len; 231 232 icsk->icsk_ack.last_seg_size = 0; 233 234 /* skb->len may jitter because of SACKs, even if peer 235 * sends good full-sized frames. 236 */ 237 len = skb_shinfo(skb)->gso_size ? : skb->len; 238 if (len >= icsk->icsk_ack.rcv_mss) { 239 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len, 240 tcp_sk(sk)->advmss); 241 /* Account for possibly-removed options */ 242 if (unlikely(len > icsk->icsk_ack.rcv_mss + 243 MAX_TCP_OPTION_SPACE)) 244 tcp_gro_dev_warn(sk, skb, len); 245 } else { 246 /* Otherwise, we make more careful check taking into account, 247 * that SACKs block is variable. 248 * 249 * "len" is invariant segment length, including TCP header. 250 */ 251 len += skb->data - skb_transport_header(skb); 252 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) || 253 /* If PSH is not set, packet should be 254 * full sized, provided peer TCP is not badly broken. 255 * This observation (if it is correct 8)) allows 256 * to handle super-low mtu links fairly. 257 */ 258 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && 259 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) { 260 /* Subtract also invariant (if peer is RFC compliant), 261 * tcp header plus fixed timestamp option length. 262 * Resulting "len" is MSS free of SACK jitter. 263 */ 264 len -= tcp_sk(sk)->tcp_header_len; 265 icsk->icsk_ack.last_seg_size = len; 266 if (len == lss) { 267 icsk->icsk_ack.rcv_mss = len; 268 return; 269 } 270 } 271 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) 272 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; 273 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 274 } 275 } 276 277 static void tcp_incr_quickack(struct sock *sk, unsigned int max_quickacks) 278 { 279 struct inet_connection_sock *icsk = inet_csk(sk); 280 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); 281 282 if (quickacks == 0) 283 quickacks = 2; 284 quickacks = min(quickacks, max_quickacks); 285 if (quickacks > icsk->icsk_ack.quick) 286 icsk->icsk_ack.quick = quickacks; 287 } 288 289 void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks) 290 { 291 struct inet_connection_sock *icsk = inet_csk(sk); 292 293 tcp_incr_quickack(sk, max_quickacks); 294 inet_csk_exit_pingpong_mode(sk); 295 icsk->icsk_ack.ato = TCP_ATO_MIN; 296 } 297 EXPORT_SYMBOL(tcp_enter_quickack_mode); 298 299 /* Send ACKs quickly, if "quick" count is not exhausted 300 * and the session is not interactive. 301 */ 302 303 static bool tcp_in_quickack_mode(struct sock *sk) 304 { 305 const struct inet_connection_sock *icsk = inet_csk(sk); 306 const struct dst_entry *dst = __sk_dst_get(sk); 307 308 return (dst && dst_metric(dst, RTAX_QUICKACK)) || 309 (icsk->icsk_ack.quick && !inet_csk_in_pingpong_mode(sk)); 310 } 311 312 static void tcp_ecn_queue_cwr(struct tcp_sock *tp) 313 { 314 if (tp->ecn_flags & TCP_ECN_OK) 315 tp->ecn_flags |= TCP_ECN_QUEUE_CWR; 316 } 317 318 static void tcp_ecn_accept_cwr(struct sock *sk, const struct sk_buff *skb) 319 { 320 if (tcp_hdr(skb)->cwr) { 321 tcp_sk(sk)->ecn_flags &= ~TCP_ECN_DEMAND_CWR; 322 323 /* If the sender is telling us it has entered CWR, then its 324 * cwnd may be very low (even just 1 packet), so we should ACK 325 * immediately. 326 */ 327 if (TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) 328 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW; 329 } 330 } 331 332 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp) 333 { 334 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 335 } 336 337 static void __tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb) 338 { 339 struct tcp_sock *tp = tcp_sk(sk); 340 341 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) { 342 case INET_ECN_NOT_ECT: 343 /* Funny extension: if ECT is not set on a segment, 344 * and we already seen ECT on a previous segment, 345 * it is probably a retransmit. 346 */ 347 if (tp->ecn_flags & TCP_ECN_SEEN) 348 tcp_enter_quickack_mode(sk, 2); 349 break; 350 case INET_ECN_CE: 351 if (tcp_ca_needs_ecn(sk)) 352 tcp_ca_event(sk, CA_EVENT_ECN_IS_CE); 353 354 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) { 355 /* Better not delay acks, sender can have a very low cwnd */ 356 tcp_enter_quickack_mode(sk, 2); 357 tp->ecn_flags |= TCP_ECN_DEMAND_CWR; 358 } 359 tp->ecn_flags |= TCP_ECN_SEEN; 360 break; 361 default: 362 if (tcp_ca_needs_ecn(sk)) 363 tcp_ca_event(sk, CA_EVENT_ECN_NO_CE); 364 tp->ecn_flags |= TCP_ECN_SEEN; 365 break; 366 } 367 } 368 369 static void tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb) 370 { 371 if (tcp_sk(sk)->ecn_flags & TCP_ECN_OK) 372 __tcp_ecn_check_ce(sk, skb); 373 } 374 375 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th) 376 { 377 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr)) 378 tp->ecn_flags &= ~TCP_ECN_OK; 379 } 380 381 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th) 382 { 383 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr)) 384 tp->ecn_flags &= ~TCP_ECN_OK; 385 } 386 387 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th) 388 { 389 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK)) 390 return true; 391 return false; 392 } 393 394 /* Buffer size and advertised window tuning. 395 * 396 * 1. Tuning sk->sk_sndbuf, when connection enters established state. 397 */ 398 399 static void tcp_sndbuf_expand(struct sock *sk) 400 { 401 const struct tcp_sock *tp = tcp_sk(sk); 402 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 403 int sndmem, per_mss; 404 u32 nr_segs; 405 406 /* Worst case is non GSO/TSO : each frame consumes one skb 407 * and skb->head is kmalloced using power of two area of memory 408 */ 409 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) + 410 MAX_TCP_HEADER + 411 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 412 413 per_mss = roundup_pow_of_two(per_mss) + 414 SKB_DATA_ALIGN(sizeof(struct sk_buff)); 415 416 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd); 417 nr_segs = max_t(u32, nr_segs, tp->reordering + 1); 418 419 /* Fast Recovery (RFC 5681 3.2) : 420 * Cubic needs 1.7 factor, rounded to 2 to include 421 * extra cushion (application might react slowly to EPOLLOUT) 422 */ 423 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2; 424 sndmem *= nr_segs * per_mss; 425 426 if (sk->sk_sndbuf < sndmem) 427 WRITE_ONCE(sk->sk_sndbuf, 428 min(sndmem, sock_net(sk)->ipv4.sysctl_tcp_wmem[2])); 429 } 430 431 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh) 432 * 433 * All tcp_full_space() is split to two parts: "network" buffer, allocated 434 * forward and advertised in receiver window (tp->rcv_wnd) and 435 * "application buffer", required to isolate scheduling/application 436 * latencies from network. 437 * window_clamp is maximal advertised window. It can be less than 438 * tcp_full_space(), in this case tcp_full_space() - window_clamp 439 * is reserved for "application" buffer. The less window_clamp is 440 * the smoother our behaviour from viewpoint of network, but the lower 441 * throughput and the higher sensitivity of the connection to losses. 8) 442 * 443 * rcv_ssthresh is more strict window_clamp used at "slow start" 444 * phase to predict further behaviour of this connection. 445 * It is used for two goals: 446 * - to enforce header prediction at sender, even when application 447 * requires some significant "application buffer". It is check #1. 448 * - to prevent pruning of receive queue because of misprediction 449 * of receiver window. Check #2. 450 * 451 * The scheme does not work when sender sends good segments opening 452 * window and then starts to feed us spaghetti. But it should work 453 * in common situations. Otherwise, we have to rely on queue collapsing. 454 */ 455 456 /* Slow part of check#2. */ 457 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb) 458 { 459 struct tcp_sock *tp = tcp_sk(sk); 460 /* Optimize this! */ 461 int truesize = tcp_win_from_space(sk, skb->truesize) >> 1; 462 int window = tcp_win_from_space(sk, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1; 463 464 while (tp->rcv_ssthresh <= window) { 465 if (truesize <= skb->len) 466 return 2 * inet_csk(sk)->icsk_ack.rcv_mss; 467 468 truesize >>= 1; 469 window >>= 1; 470 } 471 return 0; 472 } 473 474 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb) 475 { 476 struct tcp_sock *tp = tcp_sk(sk); 477 int room; 478 479 room = min_t(int, tp->window_clamp, tcp_space(sk)) - tp->rcv_ssthresh; 480 481 /* Check #1 */ 482 if (room > 0 && !tcp_under_memory_pressure(sk)) { 483 int incr; 484 485 /* Check #2. Increase window, if skb with such overhead 486 * will fit to rcvbuf in future. 487 */ 488 if (tcp_win_from_space(sk, skb->truesize) <= skb->len) 489 incr = 2 * tp->advmss; 490 else 491 incr = __tcp_grow_window(sk, skb); 492 493 if (incr) { 494 incr = max_t(int, incr, 2 * skb->len); 495 tp->rcv_ssthresh += min(room, incr); 496 inet_csk(sk)->icsk_ack.quick |= 1; 497 } 498 } 499 } 500 501 /* 3. Try to fixup all. It is made immediately after connection enters 502 * established state. 503 */ 504 static void tcp_init_buffer_space(struct sock *sk) 505 { 506 int tcp_app_win = sock_net(sk)->ipv4.sysctl_tcp_app_win; 507 struct tcp_sock *tp = tcp_sk(sk); 508 int maxwin; 509 510 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) 511 tcp_sndbuf_expand(sk); 512 513 tcp_mstamp_refresh(tp); 514 tp->rcvq_space.time = tp->tcp_mstamp; 515 tp->rcvq_space.seq = tp->copied_seq; 516 517 maxwin = tcp_full_space(sk); 518 519 if (tp->window_clamp >= maxwin) { 520 tp->window_clamp = maxwin; 521 522 if (tcp_app_win && maxwin > 4 * tp->advmss) 523 tp->window_clamp = max(maxwin - 524 (maxwin >> tcp_app_win), 525 4 * tp->advmss); 526 } 527 528 /* Force reservation of one segment. */ 529 if (tcp_app_win && 530 tp->window_clamp > 2 * tp->advmss && 531 tp->window_clamp + tp->advmss > maxwin) 532 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss); 533 534 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); 535 tp->snd_cwnd_stamp = tcp_jiffies32; 536 tp->rcvq_space.space = min3(tp->rcv_ssthresh, tp->rcv_wnd, 537 (u32)TCP_INIT_CWND * tp->advmss); 538 } 539 540 /* 4. Recalculate window clamp after socket hit its memory bounds. */ 541 static void tcp_clamp_window(struct sock *sk) 542 { 543 struct tcp_sock *tp = tcp_sk(sk); 544 struct inet_connection_sock *icsk = inet_csk(sk); 545 struct net *net = sock_net(sk); 546 547 icsk->icsk_ack.quick = 0; 548 549 if (sk->sk_rcvbuf < net->ipv4.sysctl_tcp_rmem[2] && 550 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && 551 !tcp_under_memory_pressure(sk) && 552 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) { 553 WRITE_ONCE(sk->sk_rcvbuf, 554 min(atomic_read(&sk->sk_rmem_alloc), 555 net->ipv4.sysctl_tcp_rmem[2])); 556 } 557 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) 558 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss); 559 } 560 561 /* Initialize RCV_MSS value. 562 * RCV_MSS is an our guess about MSS used by the peer. 563 * We haven't any direct information about the MSS. 564 * It's better to underestimate the RCV_MSS rather than overestimate. 565 * Overestimations make us ACKing less frequently than needed. 566 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). 567 */ 568 void tcp_initialize_rcv_mss(struct sock *sk) 569 { 570 const struct tcp_sock *tp = tcp_sk(sk); 571 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); 572 573 hint = min(hint, tp->rcv_wnd / 2); 574 hint = min(hint, TCP_MSS_DEFAULT); 575 hint = max(hint, TCP_MIN_MSS); 576 577 inet_csk(sk)->icsk_ack.rcv_mss = hint; 578 } 579 EXPORT_SYMBOL(tcp_initialize_rcv_mss); 580 581 /* Receiver "autotuning" code. 582 * 583 * The algorithm for RTT estimation w/o timestamps is based on 584 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. 585 * <https://public.lanl.gov/radiant/pubs.html#DRS> 586 * 587 * More detail on this code can be found at 588 * <http://staff.psc.edu/jheffner/>, 589 * though this reference is out of date. A new paper 590 * is pending. 591 */ 592 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) 593 { 594 u32 new_sample = tp->rcv_rtt_est.rtt_us; 595 long m = sample; 596 597 if (new_sample != 0) { 598 /* If we sample in larger samples in the non-timestamp 599 * case, we could grossly overestimate the RTT especially 600 * with chatty applications or bulk transfer apps which 601 * are stalled on filesystem I/O. 602 * 603 * Also, since we are only going for a minimum in the 604 * non-timestamp case, we do not smooth things out 605 * else with timestamps disabled convergence takes too 606 * long. 607 */ 608 if (!win_dep) { 609 m -= (new_sample >> 3); 610 new_sample += m; 611 } else { 612 m <<= 3; 613 if (m < new_sample) 614 new_sample = m; 615 } 616 } else { 617 /* No previous measure. */ 618 new_sample = m << 3; 619 } 620 621 tp->rcv_rtt_est.rtt_us = new_sample; 622 } 623 624 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) 625 { 626 u32 delta_us; 627 628 if (tp->rcv_rtt_est.time == 0) 629 goto new_measure; 630 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) 631 return; 632 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time); 633 if (!delta_us) 634 delta_us = 1; 635 tcp_rcv_rtt_update(tp, delta_us, 1); 636 637 new_measure: 638 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; 639 tp->rcv_rtt_est.time = tp->tcp_mstamp; 640 } 641 642 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, 643 const struct sk_buff *skb) 644 { 645 struct tcp_sock *tp = tcp_sk(sk); 646 647 if (tp->rx_opt.rcv_tsecr == tp->rcv_rtt_last_tsecr) 648 return; 649 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr; 650 651 if (TCP_SKB_CB(skb)->end_seq - 652 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) { 653 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr; 654 u32 delta_us; 655 656 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) { 657 if (!delta) 658 delta = 1; 659 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ); 660 tcp_rcv_rtt_update(tp, delta_us, 0); 661 } 662 } 663 } 664 665 /* 666 * This function should be called every time data is copied to user space. 667 * It calculates the appropriate TCP receive buffer space. 668 */ 669 void tcp_rcv_space_adjust(struct sock *sk) 670 { 671 struct tcp_sock *tp = tcp_sk(sk); 672 u32 copied; 673 int time; 674 675 trace_tcp_rcv_space_adjust(sk); 676 677 tcp_mstamp_refresh(tp); 678 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time); 679 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0) 680 return; 681 682 /* Number of bytes copied to user in last RTT */ 683 copied = tp->copied_seq - tp->rcvq_space.seq; 684 if (copied <= tp->rcvq_space.space) 685 goto new_measure; 686 687 /* A bit of theory : 688 * copied = bytes received in previous RTT, our base window 689 * To cope with packet losses, we need a 2x factor 690 * To cope with slow start, and sender growing its cwin by 100 % 691 * every RTT, we need a 4x factor, because the ACK we are sending 692 * now is for the next RTT, not the current one : 693 * <prev RTT . ><current RTT .. ><next RTT .... > 694 */ 695 696 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf && 697 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { 698 int rcvmem, rcvbuf; 699 u64 rcvwin, grow; 700 701 /* minimal window to cope with packet losses, assuming 702 * steady state. Add some cushion because of small variations. 703 */ 704 rcvwin = ((u64)copied << 1) + 16 * tp->advmss; 705 706 /* Accommodate for sender rate increase (eg. slow start) */ 707 grow = rcvwin * (copied - tp->rcvq_space.space); 708 do_div(grow, tp->rcvq_space.space); 709 rcvwin += (grow << 1); 710 711 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER); 712 while (tcp_win_from_space(sk, rcvmem) < tp->advmss) 713 rcvmem += 128; 714 715 do_div(rcvwin, tp->advmss); 716 rcvbuf = min_t(u64, rcvwin * rcvmem, 717 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]); 718 if (rcvbuf > sk->sk_rcvbuf) { 719 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf); 720 721 /* Make the window clamp follow along. */ 722 tp->window_clamp = tcp_win_from_space(sk, rcvbuf); 723 } 724 } 725 tp->rcvq_space.space = copied; 726 727 new_measure: 728 tp->rcvq_space.seq = tp->copied_seq; 729 tp->rcvq_space.time = tp->tcp_mstamp; 730 } 731 732 /* There is something which you must keep in mind when you analyze the 733 * behavior of the tp->ato delayed ack timeout interval. When a 734 * connection starts up, we want to ack as quickly as possible. The 735 * problem is that "good" TCP's do slow start at the beginning of data 736 * transmission. The means that until we send the first few ACK's the 737 * sender will sit on his end and only queue most of his data, because 738 * he can only send snd_cwnd unacked packets at any given time. For 739 * each ACK we send, he increments snd_cwnd and transmits more of his 740 * queue. -DaveM 741 */ 742 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb) 743 { 744 struct tcp_sock *tp = tcp_sk(sk); 745 struct inet_connection_sock *icsk = inet_csk(sk); 746 u32 now; 747 748 inet_csk_schedule_ack(sk); 749 750 tcp_measure_rcv_mss(sk, skb); 751 752 tcp_rcv_rtt_measure(tp); 753 754 now = tcp_jiffies32; 755 756 if (!icsk->icsk_ack.ato) { 757 /* The _first_ data packet received, initialize 758 * delayed ACK engine. 759 */ 760 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS); 761 icsk->icsk_ack.ato = TCP_ATO_MIN; 762 } else { 763 int m = now - icsk->icsk_ack.lrcvtime; 764 765 if (m <= TCP_ATO_MIN / 2) { 766 /* The fastest case is the first. */ 767 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; 768 } else if (m < icsk->icsk_ack.ato) { 769 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; 770 if (icsk->icsk_ack.ato > icsk->icsk_rto) 771 icsk->icsk_ack.ato = icsk->icsk_rto; 772 } else if (m > icsk->icsk_rto) { 773 /* Too long gap. Apparently sender failed to 774 * restart window, so that we send ACKs quickly. 775 */ 776 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS); 777 sk_mem_reclaim(sk); 778 } 779 } 780 icsk->icsk_ack.lrcvtime = now; 781 782 tcp_ecn_check_ce(sk, skb); 783 784 if (skb->len >= 128) 785 tcp_grow_window(sk, skb); 786 } 787 788 /* Called to compute a smoothed rtt estimate. The data fed to this 789 * routine either comes from timestamps, or from segments that were 790 * known _not_ to have been retransmitted [see Karn/Partridge 791 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 792 * piece by Van Jacobson. 793 * NOTE: the next three routines used to be one big routine. 794 * To save cycles in the RFC 1323 implementation it was better to break 795 * it up into three procedures. -- erics 796 */ 797 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us) 798 { 799 struct tcp_sock *tp = tcp_sk(sk); 800 long m = mrtt_us; /* RTT */ 801 u32 srtt = tp->srtt_us; 802 803 /* The following amusing code comes from Jacobson's 804 * article in SIGCOMM '88. Note that rtt and mdev 805 * are scaled versions of rtt and mean deviation. 806 * This is designed to be as fast as possible 807 * m stands for "measurement". 808 * 809 * On a 1990 paper the rto value is changed to: 810 * RTO = rtt + 4 * mdev 811 * 812 * Funny. This algorithm seems to be very broken. 813 * These formulae increase RTO, when it should be decreased, increase 814 * too slowly, when it should be increased quickly, decrease too quickly 815 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely 816 * does not matter how to _calculate_ it. Seems, it was trap 817 * that VJ failed to avoid. 8) 818 */ 819 if (srtt != 0) { 820 m -= (srtt >> 3); /* m is now error in rtt est */ 821 srtt += m; /* rtt = 7/8 rtt + 1/8 new */ 822 if (m < 0) { 823 m = -m; /* m is now abs(error) */ 824 m -= (tp->mdev_us >> 2); /* similar update on mdev */ 825 /* This is similar to one of Eifel findings. 826 * Eifel blocks mdev updates when rtt decreases. 827 * This solution is a bit different: we use finer gain 828 * for mdev in this case (alpha*beta). 829 * Like Eifel it also prevents growth of rto, 830 * but also it limits too fast rto decreases, 831 * happening in pure Eifel. 832 */ 833 if (m > 0) 834 m >>= 3; 835 } else { 836 m -= (tp->mdev_us >> 2); /* similar update on mdev */ 837 } 838 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */ 839 if (tp->mdev_us > tp->mdev_max_us) { 840 tp->mdev_max_us = tp->mdev_us; 841 if (tp->mdev_max_us > tp->rttvar_us) 842 tp->rttvar_us = tp->mdev_max_us; 843 } 844 if (after(tp->snd_una, tp->rtt_seq)) { 845 if (tp->mdev_max_us < tp->rttvar_us) 846 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2; 847 tp->rtt_seq = tp->snd_nxt; 848 tp->mdev_max_us = tcp_rto_min_us(sk); 849 850 tcp_bpf_rtt(sk); 851 } 852 } else { 853 /* no previous measure. */ 854 srtt = m << 3; /* take the measured time to be rtt */ 855 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */ 856 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk)); 857 tp->mdev_max_us = tp->rttvar_us; 858 tp->rtt_seq = tp->snd_nxt; 859 860 tcp_bpf_rtt(sk); 861 } 862 tp->srtt_us = max(1U, srtt); 863 } 864 865 static void tcp_update_pacing_rate(struct sock *sk) 866 { 867 const struct tcp_sock *tp = tcp_sk(sk); 868 u64 rate; 869 870 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */ 871 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3); 872 873 /* current rate is (cwnd * mss) / srtt 874 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate. 875 * In Congestion Avoidance phase, set it to 120 % the current rate. 876 * 877 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh) 878 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching 879 * end of slow start and should slow down. 880 */ 881 if (tp->snd_cwnd < tp->snd_ssthresh / 2) 882 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio; 883 else 884 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio; 885 886 rate *= max(tp->snd_cwnd, tp->packets_out); 887 888 if (likely(tp->srtt_us)) 889 do_div(rate, tp->srtt_us); 890 891 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate 892 * without any lock. We want to make sure compiler wont store 893 * intermediate values in this location. 894 */ 895 WRITE_ONCE(sk->sk_pacing_rate, min_t(u64, rate, 896 sk->sk_max_pacing_rate)); 897 } 898 899 /* Calculate rto without backoff. This is the second half of Van Jacobson's 900 * routine referred to above. 901 */ 902 static void tcp_set_rto(struct sock *sk) 903 { 904 const struct tcp_sock *tp = tcp_sk(sk); 905 /* Old crap is replaced with new one. 8) 906 * 907 * More seriously: 908 * 1. If rtt variance happened to be less 50msec, it is hallucination. 909 * It cannot be less due to utterly erratic ACK generation made 910 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ 911 * to do with delayed acks, because at cwnd>2 true delack timeout 912 * is invisible. Actually, Linux-2.4 also generates erratic 913 * ACKs in some circumstances. 914 */ 915 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp); 916 917 /* 2. Fixups made earlier cannot be right. 918 * If we do not estimate RTO correctly without them, 919 * all the algo is pure shit and should be replaced 920 * with correct one. It is exactly, which we pretend to do. 921 */ 922 923 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo 924 * guarantees that rto is higher. 925 */ 926 tcp_bound_rto(sk); 927 } 928 929 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst) 930 { 931 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); 932 933 if (!cwnd) 934 cwnd = TCP_INIT_CWND; 935 return min_t(__u32, cwnd, tp->snd_cwnd_clamp); 936 } 937 938 struct tcp_sacktag_state { 939 /* Timestamps for earliest and latest never-retransmitted segment 940 * that was SACKed. RTO needs the earliest RTT to stay conservative, 941 * but congestion control should still get an accurate delay signal. 942 */ 943 u64 first_sackt; 944 u64 last_sackt; 945 u32 reord; 946 u32 sack_delivered; 947 int flag; 948 unsigned int mss_now; 949 struct rate_sample *rate; 950 }; 951 952 /* Take a notice that peer is sending D-SACKs. Skip update of data delivery 953 * and spurious retransmission information if this DSACK is unlikely caused by 954 * sender's action: 955 * - DSACKed sequence range is larger than maximum receiver's window. 956 * - Total no. of DSACKed segments exceed the total no. of retransmitted segs. 957 */ 958 static u32 tcp_dsack_seen(struct tcp_sock *tp, u32 start_seq, 959 u32 end_seq, struct tcp_sacktag_state *state) 960 { 961 u32 seq_len, dup_segs = 1; 962 963 if (!before(start_seq, end_seq)) 964 return 0; 965 966 seq_len = end_seq - start_seq; 967 /* Dubious DSACK: DSACKed range greater than maximum advertised rwnd */ 968 if (seq_len > tp->max_window) 969 return 0; 970 if (seq_len > tp->mss_cache) 971 dup_segs = DIV_ROUND_UP(seq_len, tp->mss_cache); 972 973 tp->dsack_dups += dup_segs; 974 /* Skip the DSACK if dup segs weren't retransmitted by sender */ 975 if (tp->dsack_dups > tp->total_retrans) 976 return 0; 977 978 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN; 979 tp->rack.dsack_seen = 1; 980 981 state->flag |= FLAG_DSACKING_ACK; 982 /* A spurious retransmission is delivered */ 983 state->sack_delivered += dup_segs; 984 985 return dup_segs; 986 } 987 988 /* It's reordering when higher sequence was delivered (i.e. sacked) before 989 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering 990 * distance is approximated in full-mss packet distance ("reordering"). 991 */ 992 static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq, 993 const int ts) 994 { 995 struct tcp_sock *tp = tcp_sk(sk); 996 const u32 mss = tp->mss_cache; 997 u32 fack, metric; 998 999 fack = tcp_highest_sack_seq(tp); 1000 if (!before(low_seq, fack)) 1001 return; 1002 1003 metric = fack - low_seq; 1004 if ((metric > tp->reordering * mss) && mss) { 1005 #if FASTRETRANS_DEBUG > 1 1006 pr_debug("Disorder%d %d %u f%u s%u rr%d\n", 1007 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, 1008 tp->reordering, 1009 0, 1010 tp->sacked_out, 1011 tp->undo_marker ? tp->undo_retrans : 0); 1012 #endif 1013 tp->reordering = min_t(u32, (metric + mss - 1) / mss, 1014 sock_net(sk)->ipv4.sysctl_tcp_max_reordering); 1015 } 1016 1017 /* This exciting event is worth to be remembered. 8) */ 1018 tp->reord_seen++; 1019 NET_INC_STATS(sock_net(sk), 1020 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER); 1021 } 1022 1023 /* This must be called before lost_out or retrans_out are updated 1024 * on a new loss, because we want to know if all skbs previously 1025 * known to be lost have already been retransmitted, indicating 1026 * that this newly lost skb is our next skb to retransmit. 1027 */ 1028 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb) 1029 { 1030 if ((!tp->retransmit_skb_hint && tp->retrans_out >= tp->lost_out) || 1031 (tp->retransmit_skb_hint && 1032 before(TCP_SKB_CB(skb)->seq, 1033 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))) 1034 tp->retransmit_skb_hint = skb; 1035 } 1036 1037 /* Sum the number of packets on the wire we have marked as lost, and 1038 * notify the congestion control module that the given skb was marked lost. 1039 */ 1040 static void tcp_notify_skb_loss_event(struct tcp_sock *tp, const struct sk_buff *skb) 1041 { 1042 tp->lost += tcp_skb_pcount(skb); 1043 } 1044 1045 void tcp_mark_skb_lost(struct sock *sk, struct sk_buff *skb) 1046 { 1047 __u8 sacked = TCP_SKB_CB(skb)->sacked; 1048 struct tcp_sock *tp = tcp_sk(sk); 1049 1050 if (sacked & TCPCB_SACKED_ACKED) 1051 return; 1052 1053 tcp_verify_retransmit_hint(tp, skb); 1054 if (sacked & TCPCB_LOST) { 1055 if (sacked & TCPCB_SACKED_RETRANS) { 1056 /* Account for retransmits that are lost again */ 1057 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1058 tp->retrans_out -= tcp_skb_pcount(skb); 1059 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT, 1060 tcp_skb_pcount(skb)); 1061 tcp_notify_skb_loss_event(tp, skb); 1062 } 1063 } else { 1064 tp->lost_out += tcp_skb_pcount(skb); 1065 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1066 tcp_notify_skb_loss_event(tp, skb); 1067 } 1068 } 1069 1070 /* Updates the delivered and delivered_ce counts */ 1071 static void tcp_count_delivered(struct tcp_sock *tp, u32 delivered, 1072 bool ece_ack) 1073 { 1074 tp->delivered += delivered; 1075 if (ece_ack) 1076 tp->delivered_ce += delivered; 1077 } 1078 1079 /* This procedure tags the retransmission queue when SACKs arrive. 1080 * 1081 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). 1082 * Packets in queue with these bits set are counted in variables 1083 * sacked_out, retrans_out and lost_out, correspondingly. 1084 * 1085 * Valid combinations are: 1086 * Tag InFlight Description 1087 * 0 1 - orig segment is in flight. 1088 * S 0 - nothing flies, orig reached receiver. 1089 * L 0 - nothing flies, orig lost by net. 1090 * R 2 - both orig and retransmit are in flight. 1091 * L|R 1 - orig is lost, retransmit is in flight. 1092 * S|R 1 - orig reached receiver, retrans is still in flight. 1093 * (L|S|R is logically valid, it could occur when L|R is sacked, 1094 * but it is equivalent to plain S and code short-curcuits it to S. 1095 * L|S is logically invalid, it would mean -1 packet in flight 8)) 1096 * 1097 * These 6 states form finite state machine, controlled by the following events: 1098 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) 1099 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) 1100 * 3. Loss detection event of two flavors: 1101 * A. Scoreboard estimator decided the packet is lost. 1102 * A'. Reno "three dupacks" marks head of queue lost. 1103 * B. SACK arrives sacking SND.NXT at the moment, when the 1104 * segment was retransmitted. 1105 * 4. D-SACK added new rule: D-SACK changes any tag to S. 1106 * 1107 * It is pleasant to note, that state diagram turns out to be commutative, 1108 * so that we are allowed not to be bothered by order of our actions, 1109 * when multiple events arrive simultaneously. (see the function below). 1110 * 1111 * Reordering detection. 1112 * -------------------- 1113 * Reordering metric is maximal distance, which a packet can be displaced 1114 * in packet stream. With SACKs we can estimate it: 1115 * 1116 * 1. SACK fills old hole and the corresponding segment was not 1117 * ever retransmitted -> reordering. Alas, we cannot use it 1118 * when segment was retransmitted. 1119 * 2. The last flaw is solved with D-SACK. D-SACK arrives 1120 * for retransmitted and already SACKed segment -> reordering.. 1121 * Both of these heuristics are not used in Loss state, when we cannot 1122 * account for retransmits accurately. 1123 * 1124 * SACK block validation. 1125 * ---------------------- 1126 * 1127 * SACK block range validation checks that the received SACK block fits to 1128 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT. 1129 * Note that SND.UNA is not included to the range though being valid because 1130 * it means that the receiver is rather inconsistent with itself reporting 1131 * SACK reneging when it should advance SND.UNA. Such SACK block this is 1132 * perfectly valid, however, in light of RFC2018 which explicitly states 1133 * that "SACK block MUST reflect the newest segment. Even if the newest 1134 * segment is going to be discarded ...", not that it looks very clever 1135 * in case of head skb. Due to potentional receiver driven attacks, we 1136 * choose to avoid immediate execution of a walk in write queue due to 1137 * reneging and defer head skb's loss recovery to standard loss recovery 1138 * procedure that will eventually trigger (nothing forbids us doing this). 1139 * 1140 * Implements also blockage to start_seq wrap-around. Problem lies in the 1141 * fact that though start_seq (s) is before end_seq (i.e., not reversed), 1142 * there's no guarantee that it will be before snd_nxt (n). The problem 1143 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt 1144 * wrap (s_w): 1145 * 1146 * <- outs wnd -> <- wrapzone -> 1147 * u e n u_w e_w s n_w 1148 * | | | | | | | 1149 * |<------------+------+----- TCP seqno space --------------+---------->| 1150 * ...-- <2^31 ->| |<--------... 1151 * ...---- >2^31 ------>| |<--------... 1152 * 1153 * Current code wouldn't be vulnerable but it's better still to discard such 1154 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat 1155 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in 1156 * snd_nxt wrap -> snd_una region will then become "well defined", i.e., 1157 * equal to the ideal case (infinite seqno space without wrap caused issues). 1158 * 1159 * With D-SACK the lower bound is extended to cover sequence space below 1160 * SND.UNA down to undo_marker, which is the last point of interest. Yet 1161 * again, D-SACK block must not to go across snd_una (for the same reason as 1162 * for the normal SACK blocks, explained above). But there all simplicity 1163 * ends, TCP might receive valid D-SACKs below that. As long as they reside 1164 * fully below undo_marker they do not affect behavior in anyway and can 1165 * therefore be safely ignored. In rare cases (which are more or less 1166 * theoretical ones), the D-SACK will nicely cross that boundary due to skb 1167 * fragmentation and packet reordering past skb's retransmission. To consider 1168 * them correctly, the acceptable range must be extended even more though 1169 * the exact amount is rather hard to quantify. However, tp->max_window can 1170 * be used as an exaggerated estimate. 1171 */ 1172 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack, 1173 u32 start_seq, u32 end_seq) 1174 { 1175 /* Too far in future, or reversed (interpretation is ambiguous) */ 1176 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq)) 1177 return false; 1178 1179 /* Nasty start_seq wrap-around check (see comments above) */ 1180 if (!before(start_seq, tp->snd_nxt)) 1181 return false; 1182 1183 /* In outstanding window? ...This is valid exit for D-SACKs too. 1184 * start_seq == snd_una is non-sensical (see comments above) 1185 */ 1186 if (after(start_seq, tp->snd_una)) 1187 return true; 1188 1189 if (!is_dsack || !tp->undo_marker) 1190 return false; 1191 1192 /* ...Then it's D-SACK, and must reside below snd_una completely */ 1193 if (after(end_seq, tp->snd_una)) 1194 return false; 1195 1196 if (!before(start_seq, tp->undo_marker)) 1197 return true; 1198 1199 /* Too old */ 1200 if (!after(end_seq, tp->undo_marker)) 1201 return false; 1202 1203 /* Undo_marker boundary crossing (overestimates a lot). Known already: 1204 * start_seq < undo_marker and end_seq >= undo_marker. 1205 */ 1206 return !before(start_seq, end_seq - tp->max_window); 1207 } 1208 1209 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb, 1210 struct tcp_sack_block_wire *sp, int num_sacks, 1211 u32 prior_snd_una, struct tcp_sacktag_state *state) 1212 { 1213 struct tcp_sock *tp = tcp_sk(sk); 1214 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq); 1215 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq); 1216 u32 dup_segs; 1217 1218 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) { 1219 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV); 1220 } else if (num_sacks > 1) { 1221 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq); 1222 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq); 1223 1224 if (after(end_seq_0, end_seq_1) || before(start_seq_0, start_seq_1)) 1225 return false; 1226 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKOFORECV); 1227 } else { 1228 return false; 1229 } 1230 1231 dup_segs = tcp_dsack_seen(tp, start_seq_0, end_seq_0, state); 1232 if (!dup_segs) { /* Skip dubious DSACK */ 1233 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKIGNOREDDUBIOUS); 1234 return false; 1235 } 1236 1237 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECVSEGS, dup_segs); 1238 1239 /* D-SACK for already forgotten data... Do dumb counting. */ 1240 if (tp->undo_marker && tp->undo_retrans > 0 && 1241 !after(end_seq_0, prior_snd_una) && 1242 after(end_seq_0, tp->undo_marker)) 1243 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - dup_segs); 1244 1245 return true; 1246 } 1247 1248 /* Check if skb is fully within the SACK block. In presence of GSO skbs, 1249 * the incoming SACK may not exactly match but we can find smaller MSS 1250 * aligned portion of it that matches. Therefore we might need to fragment 1251 * which may fail and creates some hassle (caller must handle error case 1252 * returns). 1253 * 1254 * FIXME: this could be merged to shift decision code 1255 */ 1256 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb, 1257 u32 start_seq, u32 end_seq) 1258 { 1259 int err; 1260 bool in_sack; 1261 unsigned int pkt_len; 1262 unsigned int mss; 1263 1264 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1265 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1266 1267 if (tcp_skb_pcount(skb) > 1 && !in_sack && 1268 after(TCP_SKB_CB(skb)->end_seq, start_seq)) { 1269 mss = tcp_skb_mss(skb); 1270 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); 1271 1272 if (!in_sack) { 1273 pkt_len = start_seq - TCP_SKB_CB(skb)->seq; 1274 if (pkt_len < mss) 1275 pkt_len = mss; 1276 } else { 1277 pkt_len = end_seq - TCP_SKB_CB(skb)->seq; 1278 if (pkt_len < mss) 1279 return -EINVAL; 1280 } 1281 1282 /* Round if necessary so that SACKs cover only full MSSes 1283 * and/or the remaining small portion (if present) 1284 */ 1285 if (pkt_len > mss) { 1286 unsigned int new_len = (pkt_len / mss) * mss; 1287 if (!in_sack && new_len < pkt_len) 1288 new_len += mss; 1289 pkt_len = new_len; 1290 } 1291 1292 if (pkt_len >= skb->len && !in_sack) 1293 return 0; 1294 1295 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, 1296 pkt_len, mss, GFP_ATOMIC); 1297 if (err < 0) 1298 return err; 1299 } 1300 1301 return in_sack; 1302 } 1303 1304 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */ 1305 static u8 tcp_sacktag_one(struct sock *sk, 1306 struct tcp_sacktag_state *state, u8 sacked, 1307 u32 start_seq, u32 end_seq, 1308 int dup_sack, int pcount, 1309 u64 xmit_time) 1310 { 1311 struct tcp_sock *tp = tcp_sk(sk); 1312 1313 /* Account D-SACK for retransmitted packet. */ 1314 if (dup_sack && (sacked & TCPCB_RETRANS)) { 1315 if (tp->undo_marker && tp->undo_retrans > 0 && 1316 after(end_seq, tp->undo_marker)) 1317 tp->undo_retrans--; 1318 if ((sacked & TCPCB_SACKED_ACKED) && 1319 before(start_seq, state->reord)) 1320 state->reord = start_seq; 1321 } 1322 1323 /* Nothing to do; acked frame is about to be dropped (was ACKed). */ 1324 if (!after(end_seq, tp->snd_una)) 1325 return sacked; 1326 1327 if (!(sacked & TCPCB_SACKED_ACKED)) { 1328 tcp_rack_advance(tp, sacked, end_seq, xmit_time); 1329 1330 if (sacked & TCPCB_SACKED_RETRANS) { 1331 /* If the segment is not tagged as lost, 1332 * we do not clear RETRANS, believing 1333 * that retransmission is still in flight. 1334 */ 1335 if (sacked & TCPCB_LOST) { 1336 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); 1337 tp->lost_out -= pcount; 1338 tp->retrans_out -= pcount; 1339 } 1340 } else { 1341 if (!(sacked & TCPCB_RETRANS)) { 1342 /* New sack for not retransmitted frame, 1343 * which was in hole. It is reordering. 1344 */ 1345 if (before(start_seq, 1346 tcp_highest_sack_seq(tp)) && 1347 before(start_seq, state->reord)) 1348 state->reord = start_seq; 1349 1350 if (!after(end_seq, tp->high_seq)) 1351 state->flag |= FLAG_ORIG_SACK_ACKED; 1352 if (state->first_sackt == 0) 1353 state->first_sackt = xmit_time; 1354 state->last_sackt = xmit_time; 1355 } 1356 1357 if (sacked & TCPCB_LOST) { 1358 sacked &= ~TCPCB_LOST; 1359 tp->lost_out -= pcount; 1360 } 1361 } 1362 1363 sacked |= TCPCB_SACKED_ACKED; 1364 state->flag |= FLAG_DATA_SACKED; 1365 tp->sacked_out += pcount; 1366 /* Out-of-order packets delivered */ 1367 state->sack_delivered += pcount; 1368 1369 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */ 1370 if (tp->lost_skb_hint && 1371 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq)) 1372 tp->lost_cnt_hint += pcount; 1373 } 1374 1375 /* D-SACK. We can detect redundant retransmission in S|R and plain R 1376 * frames and clear it. undo_retrans is decreased above, L|R frames 1377 * are accounted above as well. 1378 */ 1379 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) { 1380 sacked &= ~TCPCB_SACKED_RETRANS; 1381 tp->retrans_out -= pcount; 1382 } 1383 1384 return sacked; 1385 } 1386 1387 /* Shift newly-SACKed bytes from this skb to the immediately previous 1388 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such. 1389 */ 1390 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev, 1391 struct sk_buff *skb, 1392 struct tcp_sacktag_state *state, 1393 unsigned int pcount, int shifted, int mss, 1394 bool dup_sack) 1395 { 1396 struct tcp_sock *tp = tcp_sk(sk); 1397 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */ 1398 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */ 1399 1400 BUG_ON(!pcount); 1401 1402 /* Adjust counters and hints for the newly sacked sequence 1403 * range but discard the return value since prev is already 1404 * marked. We must tag the range first because the seq 1405 * advancement below implicitly advances 1406 * tcp_highest_sack_seq() when skb is highest_sack. 1407 */ 1408 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked, 1409 start_seq, end_seq, dup_sack, pcount, 1410 tcp_skb_timestamp_us(skb)); 1411 tcp_rate_skb_delivered(sk, skb, state->rate); 1412 1413 if (skb == tp->lost_skb_hint) 1414 tp->lost_cnt_hint += pcount; 1415 1416 TCP_SKB_CB(prev)->end_seq += shifted; 1417 TCP_SKB_CB(skb)->seq += shifted; 1418 1419 tcp_skb_pcount_add(prev, pcount); 1420 WARN_ON_ONCE(tcp_skb_pcount(skb) < pcount); 1421 tcp_skb_pcount_add(skb, -pcount); 1422 1423 /* When we're adding to gso_segs == 1, gso_size will be zero, 1424 * in theory this shouldn't be necessary but as long as DSACK 1425 * code can come after this skb later on it's better to keep 1426 * setting gso_size to something. 1427 */ 1428 if (!TCP_SKB_CB(prev)->tcp_gso_size) 1429 TCP_SKB_CB(prev)->tcp_gso_size = mss; 1430 1431 /* CHECKME: To clear or not to clear? Mimics normal skb currently */ 1432 if (tcp_skb_pcount(skb) <= 1) 1433 TCP_SKB_CB(skb)->tcp_gso_size = 0; 1434 1435 /* Difference in this won't matter, both ACKed by the same cumul. ACK */ 1436 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS); 1437 1438 if (skb->len > 0) { 1439 BUG_ON(!tcp_skb_pcount(skb)); 1440 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED); 1441 return false; 1442 } 1443 1444 /* Whole SKB was eaten :-) */ 1445 1446 if (skb == tp->retransmit_skb_hint) 1447 tp->retransmit_skb_hint = prev; 1448 if (skb == tp->lost_skb_hint) { 1449 tp->lost_skb_hint = prev; 1450 tp->lost_cnt_hint -= tcp_skb_pcount(prev); 1451 } 1452 1453 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 1454 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor; 1455 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1456 TCP_SKB_CB(prev)->end_seq++; 1457 1458 if (skb == tcp_highest_sack(sk)) 1459 tcp_advance_highest_sack(sk, skb); 1460 1461 tcp_skb_collapse_tstamp(prev, skb); 1462 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp)) 1463 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0; 1464 1465 tcp_rtx_queue_unlink_and_free(skb, sk); 1466 1467 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED); 1468 1469 return true; 1470 } 1471 1472 /* I wish gso_size would have a bit more sane initialization than 1473 * something-or-zero which complicates things 1474 */ 1475 static int tcp_skb_seglen(const struct sk_buff *skb) 1476 { 1477 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb); 1478 } 1479 1480 /* Shifting pages past head area doesn't work */ 1481 static int skb_can_shift(const struct sk_buff *skb) 1482 { 1483 return !skb_headlen(skb) && skb_is_nonlinear(skb); 1484 } 1485 1486 int tcp_skb_shift(struct sk_buff *to, struct sk_buff *from, 1487 int pcount, int shiftlen) 1488 { 1489 /* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE) 1490 * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need 1491 * to make sure not storing more than 65535 * 8 bytes per skb, 1492 * even if current MSS is bigger. 1493 */ 1494 if (unlikely(to->len + shiftlen >= 65535 * TCP_MIN_GSO_SIZE)) 1495 return 0; 1496 if (unlikely(tcp_skb_pcount(to) + pcount > 65535)) 1497 return 0; 1498 return skb_shift(to, from, shiftlen); 1499 } 1500 1501 /* Try collapsing SACK blocks spanning across multiple skbs to a single 1502 * skb. 1503 */ 1504 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb, 1505 struct tcp_sacktag_state *state, 1506 u32 start_seq, u32 end_seq, 1507 bool dup_sack) 1508 { 1509 struct tcp_sock *tp = tcp_sk(sk); 1510 struct sk_buff *prev; 1511 int mss; 1512 int pcount = 0; 1513 int len; 1514 int in_sack; 1515 1516 /* Normally R but no L won't result in plain S */ 1517 if (!dup_sack && 1518 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS) 1519 goto fallback; 1520 if (!skb_can_shift(skb)) 1521 goto fallback; 1522 /* This frame is about to be dropped (was ACKed). */ 1523 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1524 goto fallback; 1525 1526 /* Can only happen with delayed DSACK + discard craziness */ 1527 prev = skb_rb_prev(skb); 1528 if (!prev) 1529 goto fallback; 1530 1531 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) 1532 goto fallback; 1533 1534 if (!tcp_skb_can_collapse(prev, skb)) 1535 goto fallback; 1536 1537 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1538 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1539 1540 if (in_sack) { 1541 len = skb->len; 1542 pcount = tcp_skb_pcount(skb); 1543 mss = tcp_skb_seglen(skb); 1544 1545 /* TODO: Fix DSACKs to not fragment already SACKed and we can 1546 * drop this restriction as unnecessary 1547 */ 1548 if (mss != tcp_skb_seglen(prev)) 1549 goto fallback; 1550 } else { 1551 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq)) 1552 goto noop; 1553 /* CHECKME: This is non-MSS split case only?, this will 1554 * cause skipped skbs due to advancing loop btw, original 1555 * has that feature too 1556 */ 1557 if (tcp_skb_pcount(skb) <= 1) 1558 goto noop; 1559 1560 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); 1561 if (!in_sack) { 1562 /* TODO: head merge to next could be attempted here 1563 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)), 1564 * though it might not be worth of the additional hassle 1565 * 1566 * ...we can probably just fallback to what was done 1567 * previously. We could try merging non-SACKed ones 1568 * as well but it probably isn't going to buy off 1569 * because later SACKs might again split them, and 1570 * it would make skb timestamp tracking considerably 1571 * harder problem. 1572 */ 1573 goto fallback; 1574 } 1575 1576 len = end_seq - TCP_SKB_CB(skb)->seq; 1577 BUG_ON(len < 0); 1578 BUG_ON(len > skb->len); 1579 1580 /* MSS boundaries should be honoured or else pcount will 1581 * severely break even though it makes things bit trickier. 1582 * Optimize common case to avoid most of the divides 1583 */ 1584 mss = tcp_skb_mss(skb); 1585 1586 /* TODO: Fix DSACKs to not fragment already SACKed and we can 1587 * drop this restriction as unnecessary 1588 */ 1589 if (mss != tcp_skb_seglen(prev)) 1590 goto fallback; 1591 1592 if (len == mss) { 1593 pcount = 1; 1594 } else if (len < mss) { 1595 goto noop; 1596 } else { 1597 pcount = len / mss; 1598 len = pcount * mss; 1599 } 1600 } 1601 1602 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */ 1603 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una)) 1604 goto fallback; 1605 1606 if (!tcp_skb_shift(prev, skb, pcount, len)) 1607 goto fallback; 1608 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack)) 1609 goto out; 1610 1611 /* Hole filled allows collapsing with the next as well, this is very 1612 * useful when hole on every nth skb pattern happens 1613 */ 1614 skb = skb_rb_next(prev); 1615 if (!skb) 1616 goto out; 1617 1618 if (!skb_can_shift(skb) || 1619 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) || 1620 (mss != tcp_skb_seglen(skb))) 1621 goto out; 1622 1623 len = skb->len; 1624 pcount = tcp_skb_pcount(skb); 1625 if (tcp_skb_shift(prev, skb, pcount, len)) 1626 tcp_shifted_skb(sk, prev, skb, state, pcount, 1627 len, mss, 0); 1628 1629 out: 1630 return prev; 1631 1632 noop: 1633 return skb; 1634 1635 fallback: 1636 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK); 1637 return NULL; 1638 } 1639 1640 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk, 1641 struct tcp_sack_block *next_dup, 1642 struct tcp_sacktag_state *state, 1643 u32 start_seq, u32 end_seq, 1644 bool dup_sack_in) 1645 { 1646 struct tcp_sock *tp = tcp_sk(sk); 1647 struct sk_buff *tmp; 1648 1649 skb_rbtree_walk_from(skb) { 1650 int in_sack = 0; 1651 bool dup_sack = dup_sack_in; 1652 1653 /* queue is in-order => we can short-circuit the walk early */ 1654 if (!before(TCP_SKB_CB(skb)->seq, end_seq)) 1655 break; 1656 1657 if (next_dup && 1658 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) { 1659 in_sack = tcp_match_skb_to_sack(sk, skb, 1660 next_dup->start_seq, 1661 next_dup->end_seq); 1662 if (in_sack > 0) 1663 dup_sack = true; 1664 } 1665 1666 /* skb reference here is a bit tricky to get right, since 1667 * shifting can eat and free both this skb and the next, 1668 * so not even _safe variant of the loop is enough. 1669 */ 1670 if (in_sack <= 0) { 1671 tmp = tcp_shift_skb_data(sk, skb, state, 1672 start_seq, end_seq, dup_sack); 1673 if (tmp) { 1674 if (tmp != skb) { 1675 skb = tmp; 1676 continue; 1677 } 1678 1679 in_sack = 0; 1680 } else { 1681 in_sack = tcp_match_skb_to_sack(sk, skb, 1682 start_seq, 1683 end_seq); 1684 } 1685 } 1686 1687 if (unlikely(in_sack < 0)) 1688 break; 1689 1690 if (in_sack) { 1691 TCP_SKB_CB(skb)->sacked = 1692 tcp_sacktag_one(sk, 1693 state, 1694 TCP_SKB_CB(skb)->sacked, 1695 TCP_SKB_CB(skb)->seq, 1696 TCP_SKB_CB(skb)->end_seq, 1697 dup_sack, 1698 tcp_skb_pcount(skb), 1699 tcp_skb_timestamp_us(skb)); 1700 tcp_rate_skb_delivered(sk, skb, state->rate); 1701 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1702 list_del_init(&skb->tcp_tsorted_anchor); 1703 1704 if (!before(TCP_SKB_CB(skb)->seq, 1705 tcp_highest_sack_seq(tp))) 1706 tcp_advance_highest_sack(sk, skb); 1707 } 1708 } 1709 return skb; 1710 } 1711 1712 static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk, u32 seq) 1713 { 1714 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node; 1715 struct sk_buff *skb; 1716 1717 while (*p) { 1718 parent = *p; 1719 skb = rb_to_skb(parent); 1720 if (before(seq, TCP_SKB_CB(skb)->seq)) { 1721 p = &parent->rb_left; 1722 continue; 1723 } 1724 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) { 1725 p = &parent->rb_right; 1726 continue; 1727 } 1728 return skb; 1729 } 1730 return NULL; 1731 } 1732 1733 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk, 1734 u32 skip_to_seq) 1735 { 1736 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq)) 1737 return skb; 1738 1739 return tcp_sacktag_bsearch(sk, skip_to_seq); 1740 } 1741 1742 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb, 1743 struct sock *sk, 1744 struct tcp_sack_block *next_dup, 1745 struct tcp_sacktag_state *state, 1746 u32 skip_to_seq) 1747 { 1748 if (!next_dup) 1749 return skb; 1750 1751 if (before(next_dup->start_seq, skip_to_seq)) { 1752 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq); 1753 skb = tcp_sacktag_walk(skb, sk, NULL, state, 1754 next_dup->start_seq, next_dup->end_seq, 1755 1); 1756 } 1757 1758 return skb; 1759 } 1760 1761 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache) 1762 { 1763 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); 1764 } 1765 1766 static int 1767 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb, 1768 u32 prior_snd_una, struct tcp_sacktag_state *state) 1769 { 1770 struct tcp_sock *tp = tcp_sk(sk); 1771 const unsigned char *ptr = (skb_transport_header(ack_skb) + 1772 TCP_SKB_CB(ack_skb)->sacked); 1773 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2); 1774 struct tcp_sack_block sp[TCP_NUM_SACKS]; 1775 struct tcp_sack_block *cache; 1776 struct sk_buff *skb; 1777 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3); 1778 int used_sacks; 1779 bool found_dup_sack = false; 1780 int i, j; 1781 int first_sack_index; 1782 1783 state->flag = 0; 1784 state->reord = tp->snd_nxt; 1785 1786 if (!tp->sacked_out) 1787 tcp_highest_sack_reset(sk); 1788 1789 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire, 1790 num_sacks, prior_snd_una, state); 1791 1792 /* Eliminate too old ACKs, but take into 1793 * account more or less fresh ones, they can 1794 * contain valid SACK info. 1795 */ 1796 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) 1797 return 0; 1798 1799 if (!tp->packets_out) 1800 goto out; 1801 1802 used_sacks = 0; 1803 first_sack_index = 0; 1804 for (i = 0; i < num_sacks; i++) { 1805 bool dup_sack = !i && found_dup_sack; 1806 1807 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq); 1808 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq); 1809 1810 if (!tcp_is_sackblock_valid(tp, dup_sack, 1811 sp[used_sacks].start_seq, 1812 sp[used_sacks].end_seq)) { 1813 int mib_idx; 1814 1815 if (dup_sack) { 1816 if (!tp->undo_marker) 1817 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO; 1818 else 1819 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD; 1820 } else { 1821 /* Don't count olds caused by ACK reordering */ 1822 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) && 1823 !after(sp[used_sacks].end_seq, tp->snd_una)) 1824 continue; 1825 mib_idx = LINUX_MIB_TCPSACKDISCARD; 1826 } 1827 1828 NET_INC_STATS(sock_net(sk), mib_idx); 1829 if (i == 0) 1830 first_sack_index = -1; 1831 continue; 1832 } 1833 1834 /* Ignore very old stuff early */ 1835 if (!after(sp[used_sacks].end_seq, prior_snd_una)) { 1836 if (i == 0) 1837 first_sack_index = -1; 1838 continue; 1839 } 1840 1841 used_sacks++; 1842 } 1843 1844 /* order SACK blocks to allow in order walk of the retrans queue */ 1845 for (i = used_sacks - 1; i > 0; i--) { 1846 for (j = 0; j < i; j++) { 1847 if (after(sp[j].start_seq, sp[j + 1].start_seq)) { 1848 swap(sp[j], sp[j + 1]); 1849 1850 /* Track where the first SACK block goes to */ 1851 if (j == first_sack_index) 1852 first_sack_index = j + 1; 1853 } 1854 } 1855 } 1856 1857 state->mss_now = tcp_current_mss(sk); 1858 skb = NULL; 1859 i = 0; 1860 1861 if (!tp->sacked_out) { 1862 /* It's already past, so skip checking against it */ 1863 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); 1864 } else { 1865 cache = tp->recv_sack_cache; 1866 /* Skip empty blocks in at head of the cache */ 1867 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq && 1868 !cache->end_seq) 1869 cache++; 1870 } 1871 1872 while (i < used_sacks) { 1873 u32 start_seq = sp[i].start_seq; 1874 u32 end_seq = sp[i].end_seq; 1875 bool dup_sack = (found_dup_sack && (i == first_sack_index)); 1876 struct tcp_sack_block *next_dup = NULL; 1877 1878 if (found_dup_sack && ((i + 1) == first_sack_index)) 1879 next_dup = &sp[i + 1]; 1880 1881 /* Skip too early cached blocks */ 1882 while (tcp_sack_cache_ok(tp, cache) && 1883 !before(start_seq, cache->end_seq)) 1884 cache++; 1885 1886 /* Can skip some work by looking recv_sack_cache? */ 1887 if (tcp_sack_cache_ok(tp, cache) && !dup_sack && 1888 after(end_seq, cache->start_seq)) { 1889 1890 /* Head todo? */ 1891 if (before(start_seq, cache->start_seq)) { 1892 skb = tcp_sacktag_skip(skb, sk, start_seq); 1893 skb = tcp_sacktag_walk(skb, sk, next_dup, 1894 state, 1895 start_seq, 1896 cache->start_seq, 1897 dup_sack); 1898 } 1899 1900 /* Rest of the block already fully processed? */ 1901 if (!after(end_seq, cache->end_seq)) 1902 goto advance_sp; 1903 1904 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup, 1905 state, 1906 cache->end_seq); 1907 1908 /* ...tail remains todo... */ 1909 if (tcp_highest_sack_seq(tp) == cache->end_seq) { 1910 /* ...but better entrypoint exists! */ 1911 skb = tcp_highest_sack(sk); 1912 if (!skb) 1913 break; 1914 cache++; 1915 goto walk; 1916 } 1917 1918 skb = tcp_sacktag_skip(skb, sk, cache->end_seq); 1919 /* Check overlap against next cached too (past this one already) */ 1920 cache++; 1921 continue; 1922 } 1923 1924 if (!before(start_seq, tcp_highest_sack_seq(tp))) { 1925 skb = tcp_highest_sack(sk); 1926 if (!skb) 1927 break; 1928 } 1929 skb = tcp_sacktag_skip(skb, sk, start_seq); 1930 1931 walk: 1932 skb = tcp_sacktag_walk(skb, sk, next_dup, state, 1933 start_seq, end_seq, dup_sack); 1934 1935 advance_sp: 1936 i++; 1937 } 1938 1939 /* Clear the head of the cache sack blocks so we can skip it next time */ 1940 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) { 1941 tp->recv_sack_cache[i].start_seq = 0; 1942 tp->recv_sack_cache[i].end_seq = 0; 1943 } 1944 for (j = 0; j < used_sacks; j++) 1945 tp->recv_sack_cache[i++] = sp[j]; 1946 1947 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker) 1948 tcp_check_sack_reordering(sk, state->reord, 0); 1949 1950 tcp_verify_left_out(tp); 1951 out: 1952 1953 #if FASTRETRANS_DEBUG > 0 1954 WARN_ON((int)tp->sacked_out < 0); 1955 WARN_ON((int)tp->lost_out < 0); 1956 WARN_ON((int)tp->retrans_out < 0); 1957 WARN_ON((int)tcp_packets_in_flight(tp) < 0); 1958 #endif 1959 return state->flag; 1960 } 1961 1962 /* Limits sacked_out so that sum with lost_out isn't ever larger than 1963 * packets_out. Returns false if sacked_out adjustement wasn't necessary. 1964 */ 1965 static bool tcp_limit_reno_sacked(struct tcp_sock *tp) 1966 { 1967 u32 holes; 1968 1969 holes = max(tp->lost_out, 1U); 1970 holes = min(holes, tp->packets_out); 1971 1972 if ((tp->sacked_out + holes) > tp->packets_out) { 1973 tp->sacked_out = tp->packets_out - holes; 1974 return true; 1975 } 1976 return false; 1977 } 1978 1979 /* If we receive more dupacks than we expected counting segments 1980 * in assumption of absent reordering, interpret this as reordering. 1981 * The only another reason could be bug in receiver TCP. 1982 */ 1983 static void tcp_check_reno_reordering(struct sock *sk, const int addend) 1984 { 1985 struct tcp_sock *tp = tcp_sk(sk); 1986 1987 if (!tcp_limit_reno_sacked(tp)) 1988 return; 1989 1990 tp->reordering = min_t(u32, tp->packets_out + addend, 1991 sock_net(sk)->ipv4.sysctl_tcp_max_reordering); 1992 tp->reord_seen++; 1993 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER); 1994 } 1995 1996 /* Emulate SACKs for SACKless connection: account for a new dupack. */ 1997 1998 static void tcp_add_reno_sack(struct sock *sk, int num_dupack, bool ece_ack) 1999 { 2000 if (num_dupack) { 2001 struct tcp_sock *tp = tcp_sk(sk); 2002 u32 prior_sacked = tp->sacked_out; 2003 s32 delivered; 2004 2005 tp->sacked_out += num_dupack; 2006 tcp_check_reno_reordering(sk, 0); 2007 delivered = tp->sacked_out - prior_sacked; 2008 if (delivered > 0) 2009 tcp_count_delivered(tp, delivered, ece_ack); 2010 tcp_verify_left_out(tp); 2011 } 2012 } 2013 2014 /* Account for ACK, ACKing some data in Reno Recovery phase. */ 2015 2016 static void tcp_remove_reno_sacks(struct sock *sk, int acked, bool ece_ack) 2017 { 2018 struct tcp_sock *tp = tcp_sk(sk); 2019 2020 if (acked > 0) { 2021 /* One ACK acked hole. The rest eat duplicate ACKs. */ 2022 tcp_count_delivered(tp, max_t(int, acked - tp->sacked_out, 1), 2023 ece_ack); 2024 if (acked - 1 >= tp->sacked_out) 2025 tp->sacked_out = 0; 2026 else 2027 tp->sacked_out -= acked - 1; 2028 } 2029 tcp_check_reno_reordering(sk, acked); 2030 tcp_verify_left_out(tp); 2031 } 2032 2033 static inline void tcp_reset_reno_sack(struct tcp_sock *tp) 2034 { 2035 tp->sacked_out = 0; 2036 } 2037 2038 void tcp_clear_retrans(struct tcp_sock *tp) 2039 { 2040 tp->retrans_out = 0; 2041 tp->lost_out = 0; 2042 tp->undo_marker = 0; 2043 tp->undo_retrans = -1; 2044 tp->sacked_out = 0; 2045 } 2046 2047 static inline void tcp_init_undo(struct tcp_sock *tp) 2048 { 2049 tp->undo_marker = tp->snd_una; 2050 /* Retransmission still in flight may cause DSACKs later. */ 2051 tp->undo_retrans = tp->retrans_out ? : -1; 2052 } 2053 2054 static bool tcp_is_rack(const struct sock *sk) 2055 { 2056 return sock_net(sk)->ipv4.sysctl_tcp_recovery & TCP_RACK_LOSS_DETECTION; 2057 } 2058 2059 /* If we detect SACK reneging, forget all SACK information 2060 * and reset tags completely, otherwise preserve SACKs. If receiver 2061 * dropped its ofo queue, we will know this due to reneging detection. 2062 */ 2063 static void tcp_timeout_mark_lost(struct sock *sk) 2064 { 2065 struct tcp_sock *tp = tcp_sk(sk); 2066 struct sk_buff *skb, *head; 2067 bool is_reneg; /* is receiver reneging on SACKs? */ 2068 2069 head = tcp_rtx_queue_head(sk); 2070 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED); 2071 if (is_reneg) { 2072 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING); 2073 tp->sacked_out = 0; 2074 /* Mark SACK reneging until we recover from this loss event. */ 2075 tp->is_sack_reneg = 1; 2076 } else if (tcp_is_reno(tp)) { 2077 tcp_reset_reno_sack(tp); 2078 } 2079 2080 skb = head; 2081 skb_rbtree_walk_from(skb) { 2082 if (is_reneg) 2083 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED; 2084 else if (tcp_is_rack(sk) && skb != head && 2085 tcp_rack_skb_timeout(tp, skb, 0) > 0) 2086 continue; /* Don't mark recently sent ones lost yet */ 2087 tcp_mark_skb_lost(sk, skb); 2088 } 2089 tcp_verify_left_out(tp); 2090 tcp_clear_all_retrans_hints(tp); 2091 } 2092 2093 /* Enter Loss state. */ 2094 void tcp_enter_loss(struct sock *sk) 2095 { 2096 const struct inet_connection_sock *icsk = inet_csk(sk); 2097 struct tcp_sock *tp = tcp_sk(sk); 2098 struct net *net = sock_net(sk); 2099 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery; 2100 2101 tcp_timeout_mark_lost(sk); 2102 2103 /* Reduce ssthresh if it has not yet been made inside this window. */ 2104 if (icsk->icsk_ca_state <= TCP_CA_Disorder || 2105 !after(tp->high_seq, tp->snd_una) || 2106 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) { 2107 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2108 tp->prior_cwnd = tp->snd_cwnd; 2109 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 2110 tcp_ca_event(sk, CA_EVENT_LOSS); 2111 tcp_init_undo(tp); 2112 } 2113 tp->snd_cwnd = tcp_packets_in_flight(tp) + 1; 2114 tp->snd_cwnd_cnt = 0; 2115 tp->snd_cwnd_stamp = tcp_jiffies32; 2116 2117 /* Timeout in disordered state after receiving substantial DUPACKs 2118 * suggests that the degree of reordering is over-estimated. 2119 */ 2120 if (icsk->icsk_ca_state <= TCP_CA_Disorder && 2121 tp->sacked_out >= net->ipv4.sysctl_tcp_reordering) 2122 tp->reordering = min_t(unsigned int, tp->reordering, 2123 net->ipv4.sysctl_tcp_reordering); 2124 tcp_set_ca_state(sk, TCP_CA_Loss); 2125 tp->high_seq = tp->snd_nxt; 2126 tcp_ecn_queue_cwr(tp); 2127 2128 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous 2129 * loss recovery is underway except recurring timeout(s) on 2130 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing 2131 */ 2132 tp->frto = net->ipv4.sysctl_tcp_frto && 2133 (new_recovery || icsk->icsk_retransmits) && 2134 !inet_csk(sk)->icsk_mtup.probe_size; 2135 } 2136 2137 /* If ACK arrived pointing to a remembered SACK, it means that our 2138 * remembered SACKs do not reflect real state of receiver i.e. 2139 * receiver _host_ is heavily congested (or buggy). 2140 * 2141 * To avoid big spurious retransmission bursts due to transient SACK 2142 * scoreboard oddities that look like reneging, we give the receiver a 2143 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will 2144 * restore sanity to the SACK scoreboard. If the apparent reneging 2145 * persists until this RTO then we'll clear the SACK scoreboard. 2146 */ 2147 static bool tcp_check_sack_reneging(struct sock *sk, int flag) 2148 { 2149 if (flag & FLAG_SACK_RENEGING) { 2150 struct tcp_sock *tp = tcp_sk(sk); 2151 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4), 2152 msecs_to_jiffies(10)); 2153 2154 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 2155 delay, TCP_RTO_MAX); 2156 return true; 2157 } 2158 return false; 2159 } 2160 2161 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs 2162 * counter when SACK is enabled (without SACK, sacked_out is used for 2163 * that purpose). 2164 * 2165 * With reordering, holes may still be in flight, so RFC3517 recovery 2166 * uses pure sacked_out (total number of SACKed segments) even though 2167 * it violates the RFC that uses duplicate ACKs, often these are equal 2168 * but when e.g. out-of-window ACKs or packet duplication occurs, 2169 * they differ. Since neither occurs due to loss, TCP should really 2170 * ignore them. 2171 */ 2172 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp) 2173 { 2174 return tp->sacked_out + 1; 2175 } 2176 2177 /* Linux NewReno/SACK/ECN state machine. 2178 * -------------------------------------- 2179 * 2180 * "Open" Normal state, no dubious events, fast path. 2181 * "Disorder" In all the respects it is "Open", 2182 * but requires a bit more attention. It is entered when 2183 * we see some SACKs or dupacks. It is split of "Open" 2184 * mainly to move some processing from fast path to slow one. 2185 * "CWR" CWND was reduced due to some Congestion Notification event. 2186 * It can be ECN, ICMP source quench, local device congestion. 2187 * "Recovery" CWND was reduced, we are fast-retransmitting. 2188 * "Loss" CWND was reduced due to RTO timeout or SACK reneging. 2189 * 2190 * tcp_fastretrans_alert() is entered: 2191 * - each incoming ACK, if state is not "Open" 2192 * - when arrived ACK is unusual, namely: 2193 * * SACK 2194 * * Duplicate ACK. 2195 * * ECN ECE. 2196 * 2197 * Counting packets in flight is pretty simple. 2198 * 2199 * in_flight = packets_out - left_out + retrans_out 2200 * 2201 * packets_out is SND.NXT-SND.UNA counted in packets. 2202 * 2203 * retrans_out is number of retransmitted segments. 2204 * 2205 * left_out is number of segments left network, but not ACKed yet. 2206 * 2207 * left_out = sacked_out + lost_out 2208 * 2209 * sacked_out: Packets, which arrived to receiver out of order 2210 * and hence not ACKed. With SACKs this number is simply 2211 * amount of SACKed data. Even without SACKs 2212 * it is easy to give pretty reliable estimate of this number, 2213 * counting duplicate ACKs. 2214 * 2215 * lost_out: Packets lost by network. TCP has no explicit 2216 * "loss notification" feedback from network (for now). 2217 * It means that this number can be only _guessed_. 2218 * Actually, it is the heuristics to predict lossage that 2219 * distinguishes different algorithms. 2220 * 2221 * F.e. after RTO, when all the queue is considered as lost, 2222 * lost_out = packets_out and in_flight = retrans_out. 2223 * 2224 * Essentially, we have now a few algorithms detecting 2225 * lost packets. 2226 * 2227 * If the receiver supports SACK: 2228 * 2229 * RFC6675/3517: It is the conventional algorithm. A packet is 2230 * considered lost if the number of higher sequence packets 2231 * SACKed is greater than or equal the DUPACK thoreshold 2232 * (reordering). This is implemented in tcp_mark_head_lost and 2233 * tcp_update_scoreboard. 2234 * 2235 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm 2236 * (2017-) that checks timing instead of counting DUPACKs. 2237 * Essentially a packet is considered lost if it's not S/ACKed 2238 * after RTT + reordering_window, where both metrics are 2239 * dynamically measured and adjusted. This is implemented in 2240 * tcp_rack_mark_lost. 2241 * 2242 * If the receiver does not support SACK: 2243 * 2244 * NewReno (RFC6582): in Recovery we assume that one segment 2245 * is lost (classic Reno). While we are in Recovery and 2246 * a partial ACK arrives, we assume that one more packet 2247 * is lost (NewReno). This heuristics are the same in NewReno 2248 * and SACK. 2249 * 2250 * Really tricky (and requiring careful tuning) part of algorithm 2251 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue(). 2252 * The first determines the moment _when_ we should reduce CWND and, 2253 * hence, slow down forward transmission. In fact, it determines the moment 2254 * when we decide that hole is caused by loss, rather than by a reorder. 2255 * 2256 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill 2257 * holes, caused by lost packets. 2258 * 2259 * And the most logically complicated part of algorithm is undo 2260 * heuristics. We detect false retransmits due to both too early 2261 * fast retransmit (reordering) and underestimated RTO, analyzing 2262 * timestamps and D-SACKs. When we detect that some segments were 2263 * retransmitted by mistake and CWND reduction was wrong, we undo 2264 * window reduction and abort recovery phase. This logic is hidden 2265 * inside several functions named tcp_try_undo_<something>. 2266 */ 2267 2268 /* This function decides, when we should leave Disordered state 2269 * and enter Recovery phase, reducing congestion window. 2270 * 2271 * Main question: may we further continue forward transmission 2272 * with the same cwnd? 2273 */ 2274 static bool tcp_time_to_recover(struct sock *sk, int flag) 2275 { 2276 struct tcp_sock *tp = tcp_sk(sk); 2277 2278 /* Trick#1: The loss is proven. */ 2279 if (tp->lost_out) 2280 return true; 2281 2282 /* Not-A-Trick#2 : Classic rule... */ 2283 if (!tcp_is_rack(sk) && tcp_dupack_heuristics(tp) > tp->reordering) 2284 return true; 2285 2286 return false; 2287 } 2288 2289 /* Detect loss in event "A" above by marking head of queue up as lost. 2290 * For RFC3517 SACK, a segment is considered lost if it 2291 * has at least tp->reordering SACKed seqments above it; "packets" refers to 2292 * the maximum SACKed segments to pass before reaching this limit. 2293 */ 2294 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head) 2295 { 2296 struct tcp_sock *tp = tcp_sk(sk); 2297 struct sk_buff *skb; 2298 int cnt; 2299 /* Use SACK to deduce losses of new sequences sent during recovery */ 2300 const u32 loss_high = tp->snd_nxt; 2301 2302 WARN_ON(packets > tp->packets_out); 2303 skb = tp->lost_skb_hint; 2304 if (skb) { 2305 /* Head already handled? */ 2306 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una)) 2307 return; 2308 cnt = tp->lost_cnt_hint; 2309 } else { 2310 skb = tcp_rtx_queue_head(sk); 2311 cnt = 0; 2312 } 2313 2314 skb_rbtree_walk_from(skb) { 2315 /* TODO: do this better */ 2316 /* this is not the most efficient way to do this... */ 2317 tp->lost_skb_hint = skb; 2318 tp->lost_cnt_hint = cnt; 2319 2320 if (after(TCP_SKB_CB(skb)->end_seq, loss_high)) 2321 break; 2322 2323 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2324 cnt += tcp_skb_pcount(skb); 2325 2326 if (cnt > packets) 2327 break; 2328 2329 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_LOST)) 2330 tcp_mark_skb_lost(sk, skb); 2331 2332 if (mark_head) 2333 break; 2334 } 2335 tcp_verify_left_out(tp); 2336 } 2337 2338 /* Account newly detected lost packet(s) */ 2339 2340 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit) 2341 { 2342 struct tcp_sock *tp = tcp_sk(sk); 2343 2344 if (tcp_is_sack(tp)) { 2345 int sacked_upto = tp->sacked_out - tp->reordering; 2346 if (sacked_upto >= 0) 2347 tcp_mark_head_lost(sk, sacked_upto, 0); 2348 else if (fast_rexmit) 2349 tcp_mark_head_lost(sk, 1, 1); 2350 } 2351 } 2352 2353 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when) 2354 { 2355 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 2356 before(tp->rx_opt.rcv_tsecr, when); 2357 } 2358 2359 /* skb is spurious retransmitted if the returned timestamp echo 2360 * reply is prior to the skb transmission time 2361 */ 2362 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp, 2363 const struct sk_buff *skb) 2364 { 2365 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) && 2366 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb)); 2367 } 2368 2369 /* Nothing was retransmitted or returned timestamp is less 2370 * than timestamp of the first retransmission. 2371 */ 2372 static inline bool tcp_packet_delayed(const struct tcp_sock *tp) 2373 { 2374 return tp->retrans_stamp && 2375 tcp_tsopt_ecr_before(tp, tp->retrans_stamp); 2376 } 2377 2378 /* Undo procedures. */ 2379 2380 /* We can clear retrans_stamp when there are no retransmissions in the 2381 * window. It would seem that it is trivially available for us in 2382 * tp->retrans_out, however, that kind of assumptions doesn't consider 2383 * what will happen if errors occur when sending retransmission for the 2384 * second time. ...It could the that such segment has only 2385 * TCPCB_EVER_RETRANS set at the present time. It seems that checking 2386 * the head skb is enough except for some reneging corner cases that 2387 * are not worth the effort. 2388 * 2389 * Main reason for all this complexity is the fact that connection dying 2390 * time now depends on the validity of the retrans_stamp, in particular, 2391 * that successive retransmissions of a segment must not advance 2392 * retrans_stamp under any conditions. 2393 */ 2394 static bool tcp_any_retrans_done(const struct sock *sk) 2395 { 2396 const struct tcp_sock *tp = tcp_sk(sk); 2397 struct sk_buff *skb; 2398 2399 if (tp->retrans_out) 2400 return true; 2401 2402 skb = tcp_rtx_queue_head(sk); 2403 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS)) 2404 return true; 2405 2406 return false; 2407 } 2408 2409 static void DBGUNDO(struct sock *sk, const char *msg) 2410 { 2411 #if FASTRETRANS_DEBUG > 1 2412 struct tcp_sock *tp = tcp_sk(sk); 2413 struct inet_sock *inet = inet_sk(sk); 2414 2415 if (sk->sk_family == AF_INET) { 2416 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n", 2417 msg, 2418 &inet->inet_daddr, ntohs(inet->inet_dport), 2419 tp->snd_cwnd, tcp_left_out(tp), 2420 tp->snd_ssthresh, tp->prior_ssthresh, 2421 tp->packets_out); 2422 } 2423 #if IS_ENABLED(CONFIG_IPV6) 2424 else if (sk->sk_family == AF_INET6) { 2425 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n", 2426 msg, 2427 &sk->sk_v6_daddr, ntohs(inet->inet_dport), 2428 tp->snd_cwnd, tcp_left_out(tp), 2429 tp->snd_ssthresh, tp->prior_ssthresh, 2430 tp->packets_out); 2431 } 2432 #endif 2433 #endif 2434 } 2435 2436 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss) 2437 { 2438 struct tcp_sock *tp = tcp_sk(sk); 2439 2440 if (unmark_loss) { 2441 struct sk_buff *skb; 2442 2443 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) { 2444 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 2445 } 2446 tp->lost_out = 0; 2447 tcp_clear_all_retrans_hints(tp); 2448 } 2449 2450 if (tp->prior_ssthresh) { 2451 const struct inet_connection_sock *icsk = inet_csk(sk); 2452 2453 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk); 2454 2455 if (tp->prior_ssthresh > tp->snd_ssthresh) { 2456 tp->snd_ssthresh = tp->prior_ssthresh; 2457 tcp_ecn_withdraw_cwr(tp); 2458 } 2459 } 2460 tp->snd_cwnd_stamp = tcp_jiffies32; 2461 tp->undo_marker = 0; 2462 tp->rack.advanced = 1; /* Force RACK to re-exam losses */ 2463 } 2464 2465 static inline bool tcp_may_undo(const struct tcp_sock *tp) 2466 { 2467 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp)); 2468 } 2469 2470 /* People celebrate: "We love our President!" */ 2471 static bool tcp_try_undo_recovery(struct sock *sk) 2472 { 2473 struct tcp_sock *tp = tcp_sk(sk); 2474 2475 if (tcp_may_undo(tp)) { 2476 int mib_idx; 2477 2478 /* Happy end! We did not retransmit anything 2479 * or our original transmission succeeded. 2480 */ 2481 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans"); 2482 tcp_undo_cwnd_reduction(sk, false); 2483 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss) 2484 mib_idx = LINUX_MIB_TCPLOSSUNDO; 2485 else 2486 mib_idx = LINUX_MIB_TCPFULLUNDO; 2487 2488 NET_INC_STATS(sock_net(sk), mib_idx); 2489 } else if (tp->rack.reo_wnd_persist) { 2490 tp->rack.reo_wnd_persist--; 2491 } 2492 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) { 2493 /* Hold old state until something *above* high_seq 2494 * is ACKed. For Reno it is MUST to prevent false 2495 * fast retransmits (RFC2582). SACK TCP is safe. */ 2496 if (!tcp_any_retrans_done(sk)) 2497 tp->retrans_stamp = 0; 2498 return true; 2499 } 2500 tcp_set_ca_state(sk, TCP_CA_Open); 2501 tp->is_sack_reneg = 0; 2502 return false; 2503 } 2504 2505 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */ 2506 static bool tcp_try_undo_dsack(struct sock *sk) 2507 { 2508 struct tcp_sock *tp = tcp_sk(sk); 2509 2510 if (tp->undo_marker && !tp->undo_retrans) { 2511 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH, 2512 tp->rack.reo_wnd_persist + 1); 2513 DBGUNDO(sk, "D-SACK"); 2514 tcp_undo_cwnd_reduction(sk, false); 2515 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO); 2516 return true; 2517 } 2518 return false; 2519 } 2520 2521 /* Undo during loss recovery after partial ACK or using F-RTO. */ 2522 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo) 2523 { 2524 struct tcp_sock *tp = tcp_sk(sk); 2525 2526 if (frto_undo || tcp_may_undo(tp)) { 2527 tcp_undo_cwnd_reduction(sk, true); 2528 2529 DBGUNDO(sk, "partial loss"); 2530 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO); 2531 if (frto_undo) 2532 NET_INC_STATS(sock_net(sk), 2533 LINUX_MIB_TCPSPURIOUSRTOS); 2534 inet_csk(sk)->icsk_retransmits = 0; 2535 if (frto_undo || tcp_is_sack(tp)) { 2536 tcp_set_ca_state(sk, TCP_CA_Open); 2537 tp->is_sack_reneg = 0; 2538 } 2539 return true; 2540 } 2541 return false; 2542 } 2543 2544 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937. 2545 * It computes the number of packets to send (sndcnt) based on packets newly 2546 * delivered: 2547 * 1) If the packets in flight is larger than ssthresh, PRR spreads the 2548 * cwnd reductions across a full RTT. 2549 * 2) Otherwise PRR uses packet conservation to send as much as delivered. 2550 * But when SND_UNA is acked without further losses, 2551 * slow starts cwnd up to ssthresh to speed up the recovery. 2552 */ 2553 static void tcp_init_cwnd_reduction(struct sock *sk) 2554 { 2555 struct tcp_sock *tp = tcp_sk(sk); 2556 2557 tp->high_seq = tp->snd_nxt; 2558 tp->tlp_high_seq = 0; 2559 tp->snd_cwnd_cnt = 0; 2560 tp->prior_cwnd = tp->snd_cwnd; 2561 tp->prr_delivered = 0; 2562 tp->prr_out = 0; 2563 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk); 2564 tcp_ecn_queue_cwr(tp); 2565 } 2566 2567 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int newly_lost, int flag) 2568 { 2569 struct tcp_sock *tp = tcp_sk(sk); 2570 int sndcnt = 0; 2571 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp); 2572 2573 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd)) 2574 return; 2575 2576 tp->prr_delivered += newly_acked_sacked; 2577 if (delta < 0) { 2578 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered + 2579 tp->prior_cwnd - 1; 2580 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out; 2581 } else if (flag & FLAG_SND_UNA_ADVANCED && !newly_lost) { 2582 sndcnt = min_t(int, delta, 2583 max_t(int, tp->prr_delivered - tp->prr_out, 2584 newly_acked_sacked) + 1); 2585 } else { 2586 sndcnt = min(delta, newly_acked_sacked); 2587 } 2588 /* Force a fast retransmit upon entering fast recovery */ 2589 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1)); 2590 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt; 2591 } 2592 2593 static inline void tcp_end_cwnd_reduction(struct sock *sk) 2594 { 2595 struct tcp_sock *tp = tcp_sk(sk); 2596 2597 if (inet_csk(sk)->icsk_ca_ops->cong_control) 2598 return; 2599 2600 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */ 2601 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH && 2602 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) { 2603 tp->snd_cwnd = tp->snd_ssthresh; 2604 tp->snd_cwnd_stamp = tcp_jiffies32; 2605 } 2606 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR); 2607 } 2608 2609 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */ 2610 void tcp_enter_cwr(struct sock *sk) 2611 { 2612 struct tcp_sock *tp = tcp_sk(sk); 2613 2614 tp->prior_ssthresh = 0; 2615 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) { 2616 tp->undo_marker = 0; 2617 tcp_init_cwnd_reduction(sk); 2618 tcp_set_ca_state(sk, TCP_CA_CWR); 2619 } 2620 } 2621 EXPORT_SYMBOL(tcp_enter_cwr); 2622 2623 static void tcp_try_keep_open(struct sock *sk) 2624 { 2625 struct tcp_sock *tp = tcp_sk(sk); 2626 int state = TCP_CA_Open; 2627 2628 if (tcp_left_out(tp) || tcp_any_retrans_done(sk)) 2629 state = TCP_CA_Disorder; 2630 2631 if (inet_csk(sk)->icsk_ca_state != state) { 2632 tcp_set_ca_state(sk, state); 2633 tp->high_seq = tp->snd_nxt; 2634 } 2635 } 2636 2637 static void tcp_try_to_open(struct sock *sk, int flag) 2638 { 2639 struct tcp_sock *tp = tcp_sk(sk); 2640 2641 tcp_verify_left_out(tp); 2642 2643 if (!tcp_any_retrans_done(sk)) 2644 tp->retrans_stamp = 0; 2645 2646 if (flag & FLAG_ECE) 2647 tcp_enter_cwr(sk); 2648 2649 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) { 2650 tcp_try_keep_open(sk); 2651 } 2652 } 2653 2654 static void tcp_mtup_probe_failed(struct sock *sk) 2655 { 2656 struct inet_connection_sock *icsk = inet_csk(sk); 2657 2658 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1; 2659 icsk->icsk_mtup.probe_size = 0; 2660 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL); 2661 } 2662 2663 static void tcp_mtup_probe_success(struct sock *sk) 2664 { 2665 struct tcp_sock *tp = tcp_sk(sk); 2666 struct inet_connection_sock *icsk = inet_csk(sk); 2667 2668 /* FIXME: breaks with very large cwnd */ 2669 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2670 tp->snd_cwnd = tp->snd_cwnd * 2671 tcp_mss_to_mtu(sk, tp->mss_cache) / 2672 icsk->icsk_mtup.probe_size; 2673 tp->snd_cwnd_cnt = 0; 2674 tp->snd_cwnd_stamp = tcp_jiffies32; 2675 tp->snd_ssthresh = tcp_current_ssthresh(sk); 2676 2677 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size; 2678 icsk->icsk_mtup.probe_size = 0; 2679 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 2680 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS); 2681 } 2682 2683 /* Do a simple retransmit without using the backoff mechanisms in 2684 * tcp_timer. This is used for path mtu discovery. 2685 * The socket is already locked here. 2686 */ 2687 void tcp_simple_retransmit(struct sock *sk) 2688 { 2689 const struct inet_connection_sock *icsk = inet_csk(sk); 2690 struct tcp_sock *tp = tcp_sk(sk); 2691 struct sk_buff *skb; 2692 int mss; 2693 2694 /* A fastopen SYN request is stored as two separate packets within 2695 * the retransmit queue, this is done by tcp_send_syn_data(). 2696 * As a result simply checking the MSS of the frames in the queue 2697 * will not work for the SYN packet. 2698 * 2699 * Us being here is an indication of a path MTU issue so we can 2700 * assume that the fastopen SYN was lost and just mark all the 2701 * frames in the retransmit queue as lost. We will use an MSS of 2702 * -1 to mark all frames as lost, otherwise compute the current MSS. 2703 */ 2704 if (tp->syn_data && sk->sk_state == TCP_SYN_SENT) 2705 mss = -1; 2706 else 2707 mss = tcp_current_mss(sk); 2708 2709 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) { 2710 if (tcp_skb_seglen(skb) > mss) 2711 tcp_mark_skb_lost(sk, skb); 2712 } 2713 2714 tcp_clear_retrans_hints_partial(tp); 2715 2716 if (!tp->lost_out) 2717 return; 2718 2719 if (tcp_is_reno(tp)) 2720 tcp_limit_reno_sacked(tp); 2721 2722 tcp_verify_left_out(tp); 2723 2724 /* Don't muck with the congestion window here. 2725 * Reason is that we do not increase amount of _data_ 2726 * in network, but units changed and effective 2727 * cwnd/ssthresh really reduced now. 2728 */ 2729 if (icsk->icsk_ca_state != TCP_CA_Loss) { 2730 tp->high_seq = tp->snd_nxt; 2731 tp->snd_ssthresh = tcp_current_ssthresh(sk); 2732 tp->prior_ssthresh = 0; 2733 tp->undo_marker = 0; 2734 tcp_set_ca_state(sk, TCP_CA_Loss); 2735 } 2736 tcp_xmit_retransmit_queue(sk); 2737 } 2738 EXPORT_SYMBOL(tcp_simple_retransmit); 2739 2740 void tcp_enter_recovery(struct sock *sk, bool ece_ack) 2741 { 2742 struct tcp_sock *tp = tcp_sk(sk); 2743 int mib_idx; 2744 2745 if (tcp_is_reno(tp)) 2746 mib_idx = LINUX_MIB_TCPRENORECOVERY; 2747 else 2748 mib_idx = LINUX_MIB_TCPSACKRECOVERY; 2749 2750 NET_INC_STATS(sock_net(sk), mib_idx); 2751 2752 tp->prior_ssthresh = 0; 2753 tcp_init_undo(tp); 2754 2755 if (!tcp_in_cwnd_reduction(sk)) { 2756 if (!ece_ack) 2757 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2758 tcp_init_cwnd_reduction(sk); 2759 } 2760 tcp_set_ca_state(sk, TCP_CA_Recovery); 2761 } 2762 2763 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are 2764 * recovered or spurious. Otherwise retransmits more on partial ACKs. 2765 */ 2766 static void tcp_process_loss(struct sock *sk, int flag, int num_dupack, 2767 int *rexmit) 2768 { 2769 struct tcp_sock *tp = tcp_sk(sk); 2770 bool recovered = !before(tp->snd_una, tp->high_seq); 2771 2772 if ((flag & FLAG_SND_UNA_ADVANCED || rcu_access_pointer(tp->fastopen_rsk)) && 2773 tcp_try_undo_loss(sk, false)) 2774 return; 2775 2776 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */ 2777 /* Step 3.b. A timeout is spurious if not all data are 2778 * lost, i.e., never-retransmitted data are (s)acked. 2779 */ 2780 if ((flag & FLAG_ORIG_SACK_ACKED) && 2781 tcp_try_undo_loss(sk, true)) 2782 return; 2783 2784 if (after(tp->snd_nxt, tp->high_seq)) { 2785 if (flag & FLAG_DATA_SACKED || num_dupack) 2786 tp->frto = 0; /* Step 3.a. loss was real */ 2787 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) { 2788 tp->high_seq = tp->snd_nxt; 2789 /* Step 2.b. Try send new data (but deferred until cwnd 2790 * is updated in tcp_ack()). Otherwise fall back to 2791 * the conventional recovery. 2792 */ 2793 if (!tcp_write_queue_empty(sk) && 2794 after(tcp_wnd_end(tp), tp->snd_nxt)) { 2795 *rexmit = REXMIT_NEW; 2796 return; 2797 } 2798 tp->frto = 0; 2799 } 2800 } 2801 2802 if (recovered) { 2803 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */ 2804 tcp_try_undo_recovery(sk); 2805 return; 2806 } 2807 if (tcp_is_reno(tp)) { 2808 /* A Reno DUPACK means new data in F-RTO step 2.b above are 2809 * delivered. Lower inflight to clock out (re)tranmissions. 2810 */ 2811 if (after(tp->snd_nxt, tp->high_seq) && num_dupack) 2812 tcp_add_reno_sack(sk, num_dupack, flag & FLAG_ECE); 2813 else if (flag & FLAG_SND_UNA_ADVANCED) 2814 tcp_reset_reno_sack(tp); 2815 } 2816 *rexmit = REXMIT_LOST; 2817 } 2818 2819 /* Undo during fast recovery after partial ACK. */ 2820 static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una) 2821 { 2822 struct tcp_sock *tp = tcp_sk(sk); 2823 2824 if (tp->undo_marker && tcp_packet_delayed(tp)) { 2825 /* Plain luck! Hole if filled with delayed 2826 * packet, rather than with a retransmit. Check reordering. 2827 */ 2828 tcp_check_sack_reordering(sk, prior_snd_una, 1); 2829 2830 /* We are getting evidence that the reordering degree is higher 2831 * than we realized. If there are no retransmits out then we 2832 * can undo. Otherwise we clock out new packets but do not 2833 * mark more packets lost or retransmit more. 2834 */ 2835 if (tp->retrans_out) 2836 return true; 2837 2838 if (!tcp_any_retrans_done(sk)) 2839 tp->retrans_stamp = 0; 2840 2841 DBGUNDO(sk, "partial recovery"); 2842 tcp_undo_cwnd_reduction(sk, true); 2843 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO); 2844 tcp_try_keep_open(sk); 2845 return true; 2846 } 2847 return false; 2848 } 2849 2850 static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag) 2851 { 2852 struct tcp_sock *tp = tcp_sk(sk); 2853 2854 if (tcp_rtx_queue_empty(sk)) 2855 return; 2856 2857 if (unlikely(tcp_is_reno(tp))) { 2858 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED); 2859 } else if (tcp_is_rack(sk)) { 2860 u32 prior_retrans = tp->retrans_out; 2861 2862 tcp_rack_mark_lost(sk); 2863 if (prior_retrans > tp->retrans_out) 2864 *ack_flag |= FLAG_LOST_RETRANS; 2865 } 2866 } 2867 2868 static bool tcp_force_fast_retransmit(struct sock *sk) 2869 { 2870 struct tcp_sock *tp = tcp_sk(sk); 2871 2872 return after(tcp_highest_sack_seq(tp), 2873 tp->snd_una + tp->reordering * tp->mss_cache); 2874 } 2875 2876 /* Process an event, which can update packets-in-flight not trivially. 2877 * Main goal of this function is to calculate new estimate for left_out, 2878 * taking into account both packets sitting in receiver's buffer and 2879 * packets lost by network. 2880 * 2881 * Besides that it updates the congestion state when packet loss or ECN 2882 * is detected. But it does not reduce the cwnd, it is done by the 2883 * congestion control later. 2884 * 2885 * It does _not_ decide what to send, it is made in function 2886 * tcp_xmit_retransmit_queue(). 2887 */ 2888 static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una, 2889 int num_dupack, int *ack_flag, int *rexmit) 2890 { 2891 struct inet_connection_sock *icsk = inet_csk(sk); 2892 struct tcp_sock *tp = tcp_sk(sk); 2893 int fast_rexmit = 0, flag = *ack_flag; 2894 bool ece_ack = flag & FLAG_ECE; 2895 bool do_lost = num_dupack || ((flag & FLAG_DATA_SACKED) && 2896 tcp_force_fast_retransmit(sk)); 2897 2898 if (!tp->packets_out && tp->sacked_out) 2899 tp->sacked_out = 0; 2900 2901 /* Now state machine starts. 2902 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */ 2903 if (ece_ack) 2904 tp->prior_ssthresh = 0; 2905 2906 /* B. In all the states check for reneging SACKs. */ 2907 if (tcp_check_sack_reneging(sk, flag)) 2908 return; 2909 2910 /* C. Check consistency of the current state. */ 2911 tcp_verify_left_out(tp); 2912 2913 /* D. Check state exit conditions. State can be terminated 2914 * when high_seq is ACKed. */ 2915 if (icsk->icsk_ca_state == TCP_CA_Open) { 2916 WARN_ON(tp->retrans_out != 0); 2917 tp->retrans_stamp = 0; 2918 } else if (!before(tp->snd_una, tp->high_seq)) { 2919 switch (icsk->icsk_ca_state) { 2920 case TCP_CA_CWR: 2921 /* CWR is to be held something *above* high_seq 2922 * is ACKed for CWR bit to reach receiver. */ 2923 if (tp->snd_una != tp->high_seq) { 2924 tcp_end_cwnd_reduction(sk); 2925 tcp_set_ca_state(sk, TCP_CA_Open); 2926 } 2927 break; 2928 2929 case TCP_CA_Recovery: 2930 if (tcp_is_reno(tp)) 2931 tcp_reset_reno_sack(tp); 2932 if (tcp_try_undo_recovery(sk)) 2933 return; 2934 tcp_end_cwnd_reduction(sk); 2935 break; 2936 } 2937 } 2938 2939 /* E. Process state. */ 2940 switch (icsk->icsk_ca_state) { 2941 case TCP_CA_Recovery: 2942 if (!(flag & FLAG_SND_UNA_ADVANCED)) { 2943 if (tcp_is_reno(tp)) 2944 tcp_add_reno_sack(sk, num_dupack, ece_ack); 2945 } else { 2946 if (tcp_try_undo_partial(sk, prior_snd_una)) 2947 return; 2948 /* Partial ACK arrived. Force fast retransmit. */ 2949 do_lost = tcp_force_fast_retransmit(sk); 2950 } 2951 if (tcp_try_undo_dsack(sk)) { 2952 tcp_try_keep_open(sk); 2953 return; 2954 } 2955 tcp_identify_packet_loss(sk, ack_flag); 2956 break; 2957 case TCP_CA_Loss: 2958 tcp_process_loss(sk, flag, num_dupack, rexmit); 2959 tcp_identify_packet_loss(sk, ack_flag); 2960 if (!(icsk->icsk_ca_state == TCP_CA_Open || 2961 (*ack_flag & FLAG_LOST_RETRANS))) 2962 return; 2963 /* Change state if cwnd is undone or retransmits are lost */ 2964 fallthrough; 2965 default: 2966 if (tcp_is_reno(tp)) { 2967 if (flag & FLAG_SND_UNA_ADVANCED) 2968 tcp_reset_reno_sack(tp); 2969 tcp_add_reno_sack(sk, num_dupack, ece_ack); 2970 } 2971 2972 if (icsk->icsk_ca_state <= TCP_CA_Disorder) 2973 tcp_try_undo_dsack(sk); 2974 2975 tcp_identify_packet_loss(sk, ack_flag); 2976 if (!tcp_time_to_recover(sk, flag)) { 2977 tcp_try_to_open(sk, flag); 2978 return; 2979 } 2980 2981 /* MTU probe failure: don't reduce cwnd */ 2982 if (icsk->icsk_ca_state < TCP_CA_CWR && 2983 icsk->icsk_mtup.probe_size && 2984 tp->snd_una == tp->mtu_probe.probe_seq_start) { 2985 tcp_mtup_probe_failed(sk); 2986 /* Restores the reduction we did in tcp_mtup_probe() */ 2987 tp->snd_cwnd++; 2988 tcp_simple_retransmit(sk); 2989 return; 2990 } 2991 2992 /* Otherwise enter Recovery state */ 2993 tcp_enter_recovery(sk, ece_ack); 2994 fast_rexmit = 1; 2995 } 2996 2997 if (!tcp_is_rack(sk) && do_lost) 2998 tcp_update_scoreboard(sk, fast_rexmit); 2999 *rexmit = REXMIT_LOST; 3000 } 3001 3002 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag) 3003 { 3004 u32 wlen = sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen * HZ; 3005 struct tcp_sock *tp = tcp_sk(sk); 3006 3007 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) { 3008 /* If the remote keeps returning delayed ACKs, eventually 3009 * the min filter would pick it up and overestimate the 3010 * prop. delay when it expires. Skip suspected delayed ACKs. 3011 */ 3012 return; 3013 } 3014 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32, 3015 rtt_us ? : jiffies_to_usecs(1)); 3016 } 3017 3018 static bool tcp_ack_update_rtt(struct sock *sk, const int flag, 3019 long seq_rtt_us, long sack_rtt_us, 3020 long ca_rtt_us, struct rate_sample *rs) 3021 { 3022 const struct tcp_sock *tp = tcp_sk(sk); 3023 3024 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because 3025 * broken middle-boxes or peers may corrupt TS-ECR fields. But 3026 * Karn's algorithm forbids taking RTT if some retransmitted data 3027 * is acked (RFC6298). 3028 */ 3029 if (seq_rtt_us < 0) 3030 seq_rtt_us = sack_rtt_us; 3031 3032 /* RTTM Rule: A TSecr value received in a segment is used to 3033 * update the averaged RTT measurement only if the segment 3034 * acknowledges some new data, i.e., only if it advances the 3035 * left edge of the send window. 3036 * See draft-ietf-tcplw-high-performance-00, section 3.3. 3037 */ 3038 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 3039 flag & FLAG_ACKED) { 3040 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr; 3041 3042 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) { 3043 if (!delta) 3044 delta = 1; 3045 seq_rtt_us = delta * (USEC_PER_SEC / TCP_TS_HZ); 3046 ca_rtt_us = seq_rtt_us; 3047 } 3048 } 3049 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */ 3050 if (seq_rtt_us < 0) 3051 return false; 3052 3053 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is 3054 * always taken together with ACK, SACK, or TS-opts. Any negative 3055 * values will be skipped with the seq_rtt_us < 0 check above. 3056 */ 3057 tcp_update_rtt_min(sk, ca_rtt_us, flag); 3058 tcp_rtt_estimator(sk, seq_rtt_us); 3059 tcp_set_rto(sk); 3060 3061 /* RFC6298: only reset backoff on valid RTT measurement. */ 3062 inet_csk(sk)->icsk_backoff = 0; 3063 return true; 3064 } 3065 3066 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */ 3067 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req) 3068 { 3069 struct rate_sample rs; 3070 long rtt_us = -1L; 3071 3072 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack) 3073 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack); 3074 3075 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs); 3076 } 3077 3078 3079 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked) 3080 { 3081 const struct inet_connection_sock *icsk = inet_csk(sk); 3082 3083 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked); 3084 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32; 3085 } 3086 3087 /* Restart timer after forward progress on connection. 3088 * RFC2988 recommends to restart timer to now+rto. 3089 */ 3090 void tcp_rearm_rto(struct sock *sk) 3091 { 3092 const struct inet_connection_sock *icsk = inet_csk(sk); 3093 struct tcp_sock *tp = tcp_sk(sk); 3094 3095 /* If the retrans timer is currently being used by Fast Open 3096 * for SYN-ACK retrans purpose, stay put. 3097 */ 3098 if (rcu_access_pointer(tp->fastopen_rsk)) 3099 return; 3100 3101 if (!tp->packets_out) { 3102 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS); 3103 } else { 3104 u32 rto = inet_csk(sk)->icsk_rto; 3105 /* Offset the time elapsed after installing regular RTO */ 3106 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT || 3107 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) { 3108 s64 delta_us = tcp_rto_delta_us(sk); 3109 /* delta_us may not be positive if the socket is locked 3110 * when the retrans timer fires and is rescheduled. 3111 */ 3112 rto = usecs_to_jiffies(max_t(int, delta_us, 1)); 3113 } 3114 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto, 3115 TCP_RTO_MAX); 3116 } 3117 } 3118 3119 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */ 3120 static void tcp_set_xmit_timer(struct sock *sk) 3121 { 3122 if (!tcp_schedule_loss_probe(sk, true)) 3123 tcp_rearm_rto(sk); 3124 } 3125 3126 /* If we get here, the whole TSO packet has not been acked. */ 3127 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb) 3128 { 3129 struct tcp_sock *tp = tcp_sk(sk); 3130 u32 packets_acked; 3131 3132 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)); 3133 3134 packets_acked = tcp_skb_pcount(skb); 3135 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 3136 return 0; 3137 packets_acked -= tcp_skb_pcount(skb); 3138 3139 if (packets_acked) { 3140 BUG_ON(tcp_skb_pcount(skb) == 0); 3141 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)); 3142 } 3143 3144 return packets_acked; 3145 } 3146 3147 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb, 3148 u32 prior_snd_una) 3149 { 3150 const struct skb_shared_info *shinfo; 3151 3152 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */ 3153 if (likely(!TCP_SKB_CB(skb)->txstamp_ack)) 3154 return; 3155 3156 shinfo = skb_shinfo(skb); 3157 if (!before(shinfo->tskey, prior_snd_una) && 3158 before(shinfo->tskey, tcp_sk(sk)->snd_una)) { 3159 tcp_skb_tsorted_save(skb) { 3160 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK); 3161 } tcp_skb_tsorted_restore(skb); 3162 } 3163 } 3164 3165 /* Remove acknowledged frames from the retransmission queue. If our packet 3166 * is before the ack sequence we can discard it as it's confirmed to have 3167 * arrived at the other end. 3168 */ 3169 static int tcp_clean_rtx_queue(struct sock *sk, u32 prior_fack, 3170 u32 prior_snd_una, 3171 struct tcp_sacktag_state *sack, bool ece_ack) 3172 { 3173 const struct inet_connection_sock *icsk = inet_csk(sk); 3174 u64 first_ackt, last_ackt; 3175 struct tcp_sock *tp = tcp_sk(sk); 3176 u32 prior_sacked = tp->sacked_out; 3177 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */ 3178 struct sk_buff *skb, *next; 3179 bool fully_acked = true; 3180 long sack_rtt_us = -1L; 3181 long seq_rtt_us = -1L; 3182 long ca_rtt_us = -1L; 3183 u32 pkts_acked = 0; 3184 u32 last_in_flight = 0; 3185 bool rtt_update; 3186 int flag = 0; 3187 3188 first_ackt = 0; 3189 3190 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) { 3191 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 3192 const u32 start_seq = scb->seq; 3193 u8 sacked = scb->sacked; 3194 u32 acked_pcount; 3195 3196 /* Determine how many packets and what bytes were acked, tso and else */ 3197 if (after(scb->end_seq, tp->snd_una)) { 3198 if (tcp_skb_pcount(skb) == 1 || 3199 !after(tp->snd_una, scb->seq)) 3200 break; 3201 3202 acked_pcount = tcp_tso_acked(sk, skb); 3203 if (!acked_pcount) 3204 break; 3205 fully_acked = false; 3206 } else { 3207 acked_pcount = tcp_skb_pcount(skb); 3208 } 3209 3210 if (unlikely(sacked & TCPCB_RETRANS)) { 3211 if (sacked & TCPCB_SACKED_RETRANS) 3212 tp->retrans_out -= acked_pcount; 3213 flag |= FLAG_RETRANS_DATA_ACKED; 3214 } else if (!(sacked & TCPCB_SACKED_ACKED)) { 3215 last_ackt = tcp_skb_timestamp_us(skb); 3216 WARN_ON_ONCE(last_ackt == 0); 3217 if (!first_ackt) 3218 first_ackt = last_ackt; 3219 3220 last_in_flight = TCP_SKB_CB(skb)->tx.in_flight; 3221 if (before(start_seq, reord)) 3222 reord = start_seq; 3223 if (!after(scb->end_seq, tp->high_seq)) 3224 flag |= FLAG_ORIG_SACK_ACKED; 3225 } 3226 3227 if (sacked & TCPCB_SACKED_ACKED) { 3228 tp->sacked_out -= acked_pcount; 3229 } else if (tcp_is_sack(tp)) { 3230 tcp_count_delivered(tp, acked_pcount, ece_ack); 3231 if (!tcp_skb_spurious_retrans(tp, skb)) 3232 tcp_rack_advance(tp, sacked, scb->end_seq, 3233 tcp_skb_timestamp_us(skb)); 3234 } 3235 if (sacked & TCPCB_LOST) 3236 tp->lost_out -= acked_pcount; 3237 3238 tp->packets_out -= acked_pcount; 3239 pkts_acked += acked_pcount; 3240 tcp_rate_skb_delivered(sk, skb, sack->rate); 3241 3242 /* Initial outgoing SYN's get put onto the write_queue 3243 * just like anything else we transmit. It is not 3244 * true data, and if we misinform our callers that 3245 * this ACK acks real data, we will erroneously exit 3246 * connection startup slow start one packet too 3247 * quickly. This is severely frowned upon behavior. 3248 */ 3249 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) { 3250 flag |= FLAG_DATA_ACKED; 3251 } else { 3252 flag |= FLAG_SYN_ACKED; 3253 tp->retrans_stamp = 0; 3254 } 3255 3256 if (!fully_acked) 3257 break; 3258 3259 tcp_ack_tstamp(sk, skb, prior_snd_una); 3260 3261 next = skb_rb_next(skb); 3262 if (unlikely(skb == tp->retransmit_skb_hint)) 3263 tp->retransmit_skb_hint = NULL; 3264 if (unlikely(skb == tp->lost_skb_hint)) 3265 tp->lost_skb_hint = NULL; 3266 tcp_highest_sack_replace(sk, skb, next); 3267 tcp_rtx_queue_unlink_and_free(skb, sk); 3268 } 3269 3270 if (!skb) 3271 tcp_chrono_stop(sk, TCP_CHRONO_BUSY); 3272 3273 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una))) 3274 tp->snd_up = tp->snd_una; 3275 3276 if (skb) { 3277 tcp_ack_tstamp(sk, skb, prior_snd_una); 3278 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 3279 flag |= FLAG_SACK_RENEGING; 3280 } 3281 3282 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) { 3283 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt); 3284 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt); 3285 3286 if (pkts_acked == 1 && last_in_flight < tp->mss_cache && 3287 last_in_flight && !prior_sacked && fully_acked && 3288 sack->rate->prior_delivered + 1 == tp->delivered && 3289 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) { 3290 /* Conservatively mark a delayed ACK. It's typically 3291 * from a lone runt packet over the round trip to 3292 * a receiver w/o out-of-order or CE events. 3293 */ 3294 flag |= FLAG_ACK_MAYBE_DELAYED; 3295 } 3296 } 3297 if (sack->first_sackt) { 3298 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt); 3299 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt); 3300 } 3301 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us, 3302 ca_rtt_us, sack->rate); 3303 3304 if (flag & FLAG_ACKED) { 3305 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */ 3306 if (unlikely(icsk->icsk_mtup.probe_size && 3307 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) { 3308 tcp_mtup_probe_success(sk); 3309 } 3310 3311 if (tcp_is_reno(tp)) { 3312 tcp_remove_reno_sacks(sk, pkts_acked, ece_ack); 3313 3314 /* If any of the cumulatively ACKed segments was 3315 * retransmitted, non-SACK case cannot confirm that 3316 * progress was due to original transmission due to 3317 * lack of TCPCB_SACKED_ACKED bits even if some of 3318 * the packets may have been never retransmitted. 3319 */ 3320 if (flag & FLAG_RETRANS_DATA_ACKED) 3321 flag &= ~FLAG_ORIG_SACK_ACKED; 3322 } else { 3323 int delta; 3324 3325 /* Non-retransmitted hole got filled? That's reordering */ 3326 if (before(reord, prior_fack)) 3327 tcp_check_sack_reordering(sk, reord, 0); 3328 3329 delta = prior_sacked - tp->sacked_out; 3330 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta); 3331 } 3332 } else if (skb && rtt_update && sack_rtt_us >= 0 && 3333 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp, 3334 tcp_skb_timestamp_us(skb))) { 3335 /* Do not re-arm RTO if the sack RTT is measured from data sent 3336 * after when the head was last (re)transmitted. Otherwise the 3337 * timeout may continue to extend in loss recovery. 3338 */ 3339 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */ 3340 } 3341 3342 if (icsk->icsk_ca_ops->pkts_acked) { 3343 struct ack_sample sample = { .pkts_acked = pkts_acked, 3344 .rtt_us = sack->rate->rtt_us, 3345 .in_flight = last_in_flight }; 3346 3347 icsk->icsk_ca_ops->pkts_acked(sk, &sample); 3348 } 3349 3350 #if FASTRETRANS_DEBUG > 0 3351 WARN_ON((int)tp->sacked_out < 0); 3352 WARN_ON((int)tp->lost_out < 0); 3353 WARN_ON((int)tp->retrans_out < 0); 3354 if (!tp->packets_out && tcp_is_sack(tp)) { 3355 icsk = inet_csk(sk); 3356 if (tp->lost_out) { 3357 pr_debug("Leak l=%u %d\n", 3358 tp->lost_out, icsk->icsk_ca_state); 3359 tp->lost_out = 0; 3360 } 3361 if (tp->sacked_out) { 3362 pr_debug("Leak s=%u %d\n", 3363 tp->sacked_out, icsk->icsk_ca_state); 3364 tp->sacked_out = 0; 3365 } 3366 if (tp->retrans_out) { 3367 pr_debug("Leak r=%u %d\n", 3368 tp->retrans_out, icsk->icsk_ca_state); 3369 tp->retrans_out = 0; 3370 } 3371 } 3372 #endif 3373 return flag; 3374 } 3375 3376 static void tcp_ack_probe(struct sock *sk) 3377 { 3378 struct inet_connection_sock *icsk = inet_csk(sk); 3379 struct sk_buff *head = tcp_send_head(sk); 3380 const struct tcp_sock *tp = tcp_sk(sk); 3381 3382 /* Was it a usable window open? */ 3383 if (!head) 3384 return; 3385 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) { 3386 icsk->icsk_backoff = 0; 3387 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0); 3388 /* Socket must be waked up by subsequent tcp_data_snd_check(). 3389 * This function is not for random using! 3390 */ 3391 } else { 3392 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX); 3393 3394 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 3395 when, TCP_RTO_MAX); 3396 } 3397 } 3398 3399 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag) 3400 { 3401 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) || 3402 inet_csk(sk)->icsk_ca_state != TCP_CA_Open; 3403 } 3404 3405 /* Decide wheather to run the increase function of congestion control. */ 3406 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag) 3407 { 3408 /* If reordering is high then always grow cwnd whenever data is 3409 * delivered regardless of its ordering. Otherwise stay conservative 3410 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/ 3411 * new SACK or ECE mark may first advance cwnd here and later reduce 3412 * cwnd in tcp_fastretrans_alert() based on more states. 3413 */ 3414 if (tcp_sk(sk)->reordering > sock_net(sk)->ipv4.sysctl_tcp_reordering) 3415 return flag & FLAG_FORWARD_PROGRESS; 3416 3417 return flag & FLAG_DATA_ACKED; 3418 } 3419 3420 /* The "ultimate" congestion control function that aims to replace the rigid 3421 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction). 3422 * It's called toward the end of processing an ACK with precise rate 3423 * information. All transmission or retransmission are delayed afterwards. 3424 */ 3425 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked, 3426 int flag, const struct rate_sample *rs) 3427 { 3428 const struct inet_connection_sock *icsk = inet_csk(sk); 3429 3430 if (icsk->icsk_ca_ops->cong_control) { 3431 icsk->icsk_ca_ops->cong_control(sk, rs); 3432 return; 3433 } 3434 3435 if (tcp_in_cwnd_reduction(sk)) { 3436 /* Reduce cwnd if state mandates */ 3437 tcp_cwnd_reduction(sk, acked_sacked, rs->losses, flag); 3438 } else if (tcp_may_raise_cwnd(sk, flag)) { 3439 /* Advance cwnd if state allows */ 3440 tcp_cong_avoid(sk, ack, acked_sacked); 3441 } 3442 tcp_update_pacing_rate(sk); 3443 } 3444 3445 /* Check that window update is acceptable. 3446 * The function assumes that snd_una<=ack<=snd_next. 3447 */ 3448 static inline bool tcp_may_update_window(const struct tcp_sock *tp, 3449 const u32 ack, const u32 ack_seq, 3450 const u32 nwin) 3451 { 3452 return after(ack, tp->snd_una) || 3453 after(ack_seq, tp->snd_wl1) || 3454 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd); 3455 } 3456 3457 /* If we update tp->snd_una, also update tp->bytes_acked */ 3458 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack) 3459 { 3460 u32 delta = ack - tp->snd_una; 3461 3462 sock_owned_by_me((struct sock *)tp); 3463 tp->bytes_acked += delta; 3464 tp->snd_una = ack; 3465 } 3466 3467 /* If we update tp->rcv_nxt, also update tp->bytes_received */ 3468 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq) 3469 { 3470 u32 delta = seq - tp->rcv_nxt; 3471 3472 sock_owned_by_me((struct sock *)tp); 3473 tp->bytes_received += delta; 3474 WRITE_ONCE(tp->rcv_nxt, seq); 3475 } 3476 3477 /* Update our send window. 3478 * 3479 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2 3480 * and in FreeBSD. NetBSD's one is even worse.) is wrong. 3481 */ 3482 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack, 3483 u32 ack_seq) 3484 { 3485 struct tcp_sock *tp = tcp_sk(sk); 3486 int flag = 0; 3487 u32 nwin = ntohs(tcp_hdr(skb)->window); 3488 3489 if (likely(!tcp_hdr(skb)->syn)) 3490 nwin <<= tp->rx_opt.snd_wscale; 3491 3492 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) { 3493 flag |= FLAG_WIN_UPDATE; 3494 tcp_update_wl(tp, ack_seq); 3495 3496 if (tp->snd_wnd != nwin) { 3497 tp->snd_wnd = nwin; 3498 3499 /* Note, it is the only place, where 3500 * fast path is recovered for sending TCP. 3501 */ 3502 tp->pred_flags = 0; 3503 tcp_fast_path_check(sk); 3504 3505 if (!tcp_write_queue_empty(sk)) 3506 tcp_slow_start_after_idle_check(sk); 3507 3508 if (nwin > tp->max_window) { 3509 tp->max_window = nwin; 3510 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie); 3511 } 3512 } 3513 } 3514 3515 tcp_snd_una_update(tp, ack); 3516 3517 return flag; 3518 } 3519 3520 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx, 3521 u32 *last_oow_ack_time) 3522 { 3523 if (*last_oow_ack_time) { 3524 s32 elapsed = (s32)(tcp_jiffies32 - *last_oow_ack_time); 3525 3526 if (0 <= elapsed && elapsed < net->ipv4.sysctl_tcp_invalid_ratelimit) { 3527 NET_INC_STATS(net, mib_idx); 3528 return true; /* rate-limited: don't send yet! */ 3529 } 3530 } 3531 3532 *last_oow_ack_time = tcp_jiffies32; 3533 3534 return false; /* not rate-limited: go ahead, send dupack now! */ 3535 } 3536 3537 /* Return true if we're currently rate-limiting out-of-window ACKs and 3538 * thus shouldn't send a dupack right now. We rate-limit dupacks in 3539 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS 3540 * attacks that send repeated SYNs or ACKs for the same connection. To 3541 * do this, we do not send a duplicate SYNACK or ACK if the remote 3542 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate. 3543 */ 3544 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb, 3545 int mib_idx, u32 *last_oow_ack_time) 3546 { 3547 /* Data packets without SYNs are not likely part of an ACK loop. */ 3548 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) && 3549 !tcp_hdr(skb)->syn) 3550 return false; 3551 3552 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time); 3553 } 3554 3555 /* RFC 5961 7 [ACK Throttling] */ 3556 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb) 3557 { 3558 /* unprotected vars, we dont care of overwrites */ 3559 static u32 challenge_timestamp; 3560 static unsigned int challenge_count; 3561 struct tcp_sock *tp = tcp_sk(sk); 3562 struct net *net = sock_net(sk); 3563 u32 count, now; 3564 3565 /* First check our per-socket dupack rate limit. */ 3566 if (__tcp_oow_rate_limited(net, 3567 LINUX_MIB_TCPACKSKIPPEDCHALLENGE, 3568 &tp->last_oow_ack_time)) 3569 return; 3570 3571 /* Then check host-wide RFC 5961 rate limit. */ 3572 now = jiffies / HZ; 3573 if (now != challenge_timestamp) { 3574 u32 ack_limit = net->ipv4.sysctl_tcp_challenge_ack_limit; 3575 u32 half = (ack_limit + 1) >> 1; 3576 3577 challenge_timestamp = now; 3578 WRITE_ONCE(challenge_count, half + prandom_u32_max(ack_limit)); 3579 } 3580 count = READ_ONCE(challenge_count); 3581 if (count > 0) { 3582 WRITE_ONCE(challenge_count, count - 1); 3583 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK); 3584 tcp_send_ack(sk); 3585 } 3586 } 3587 3588 static void tcp_store_ts_recent(struct tcp_sock *tp) 3589 { 3590 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval; 3591 tp->rx_opt.ts_recent_stamp = ktime_get_seconds(); 3592 } 3593 3594 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq) 3595 { 3596 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) { 3597 /* PAWS bug workaround wrt. ACK frames, the PAWS discard 3598 * extra check below makes sure this can only happen 3599 * for pure ACK frames. -DaveM 3600 * 3601 * Not only, also it occurs for expired timestamps. 3602 */ 3603 3604 if (tcp_paws_check(&tp->rx_opt, 0)) 3605 tcp_store_ts_recent(tp); 3606 } 3607 } 3608 3609 /* This routine deals with acks during a TLP episode and ends an episode by 3610 * resetting tlp_high_seq. Ref: TLP algorithm in draft-ietf-tcpm-rack 3611 */ 3612 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag) 3613 { 3614 struct tcp_sock *tp = tcp_sk(sk); 3615 3616 if (before(ack, tp->tlp_high_seq)) 3617 return; 3618 3619 if (!tp->tlp_retrans) { 3620 /* TLP of new data has been acknowledged */ 3621 tp->tlp_high_seq = 0; 3622 } else if (flag & FLAG_DSACKING_ACK) { 3623 /* This DSACK means original and TLP probe arrived; no loss */ 3624 tp->tlp_high_seq = 0; 3625 } else if (after(ack, tp->tlp_high_seq)) { 3626 /* ACK advances: there was a loss, so reduce cwnd. Reset 3627 * tlp_high_seq in tcp_init_cwnd_reduction() 3628 */ 3629 tcp_init_cwnd_reduction(sk); 3630 tcp_set_ca_state(sk, TCP_CA_CWR); 3631 tcp_end_cwnd_reduction(sk); 3632 tcp_try_keep_open(sk); 3633 NET_INC_STATS(sock_net(sk), 3634 LINUX_MIB_TCPLOSSPROBERECOVERY); 3635 } else if (!(flag & (FLAG_SND_UNA_ADVANCED | 3636 FLAG_NOT_DUP | FLAG_DATA_SACKED))) { 3637 /* Pure dupack: original and TLP probe arrived; no loss */ 3638 tp->tlp_high_seq = 0; 3639 } 3640 } 3641 3642 static inline void tcp_in_ack_event(struct sock *sk, u32 flags) 3643 { 3644 const struct inet_connection_sock *icsk = inet_csk(sk); 3645 3646 if (icsk->icsk_ca_ops->in_ack_event) 3647 icsk->icsk_ca_ops->in_ack_event(sk, flags); 3648 } 3649 3650 /* Congestion control has updated the cwnd already. So if we're in 3651 * loss recovery then now we do any new sends (for FRTO) or 3652 * retransmits (for CA_Loss or CA_recovery) that make sense. 3653 */ 3654 static void tcp_xmit_recovery(struct sock *sk, int rexmit) 3655 { 3656 struct tcp_sock *tp = tcp_sk(sk); 3657 3658 if (rexmit == REXMIT_NONE || sk->sk_state == TCP_SYN_SENT) 3659 return; 3660 3661 if (unlikely(rexmit == REXMIT_NEW)) { 3662 __tcp_push_pending_frames(sk, tcp_current_mss(sk), 3663 TCP_NAGLE_OFF); 3664 if (after(tp->snd_nxt, tp->high_seq)) 3665 return; 3666 tp->frto = 0; 3667 } 3668 tcp_xmit_retransmit_queue(sk); 3669 } 3670 3671 /* Returns the number of packets newly acked or sacked by the current ACK */ 3672 static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, int flag) 3673 { 3674 const struct net *net = sock_net(sk); 3675 struct tcp_sock *tp = tcp_sk(sk); 3676 u32 delivered; 3677 3678 delivered = tp->delivered - prior_delivered; 3679 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered); 3680 if (flag & FLAG_ECE) 3681 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, delivered); 3682 3683 return delivered; 3684 } 3685 3686 /* This routine deals with incoming acks, but not outgoing ones. */ 3687 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag) 3688 { 3689 struct inet_connection_sock *icsk = inet_csk(sk); 3690 struct tcp_sock *tp = tcp_sk(sk); 3691 struct tcp_sacktag_state sack_state; 3692 struct rate_sample rs = { .prior_delivered = 0 }; 3693 u32 prior_snd_una = tp->snd_una; 3694 bool is_sack_reneg = tp->is_sack_reneg; 3695 u32 ack_seq = TCP_SKB_CB(skb)->seq; 3696 u32 ack = TCP_SKB_CB(skb)->ack_seq; 3697 int num_dupack = 0; 3698 int prior_packets = tp->packets_out; 3699 u32 delivered = tp->delivered; 3700 u32 lost = tp->lost; 3701 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */ 3702 u32 prior_fack; 3703 3704 sack_state.first_sackt = 0; 3705 sack_state.rate = &rs; 3706 sack_state.sack_delivered = 0; 3707 3708 /* We very likely will need to access rtx queue. */ 3709 prefetch(sk->tcp_rtx_queue.rb_node); 3710 3711 /* If the ack is older than previous acks 3712 * then we can probably ignore it. 3713 */ 3714 if (before(ack, prior_snd_una)) { 3715 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */ 3716 if (before(ack, prior_snd_una - tp->max_window)) { 3717 if (!(flag & FLAG_NO_CHALLENGE_ACK)) 3718 tcp_send_challenge_ack(sk, skb); 3719 return -1; 3720 } 3721 goto old_ack; 3722 } 3723 3724 /* If the ack includes data we haven't sent yet, discard 3725 * this segment (RFC793 Section 3.9). 3726 */ 3727 if (after(ack, tp->snd_nxt)) 3728 return -1; 3729 3730 if (after(ack, prior_snd_una)) { 3731 flag |= FLAG_SND_UNA_ADVANCED; 3732 icsk->icsk_retransmits = 0; 3733 3734 #if IS_ENABLED(CONFIG_TLS_DEVICE) 3735 if (static_branch_unlikely(&clean_acked_data_enabled.key)) 3736 if (icsk->icsk_clean_acked) 3737 icsk->icsk_clean_acked(sk, ack); 3738 #endif 3739 } 3740 3741 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una; 3742 rs.prior_in_flight = tcp_packets_in_flight(tp); 3743 3744 /* ts_recent update must be made after we are sure that the packet 3745 * is in window. 3746 */ 3747 if (flag & FLAG_UPDATE_TS_RECENT) 3748 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); 3749 3750 if ((flag & (FLAG_SLOWPATH | FLAG_SND_UNA_ADVANCED)) == 3751 FLAG_SND_UNA_ADVANCED) { 3752 /* Window is constant, pure forward advance. 3753 * No more checks are required. 3754 * Note, we use the fact that SND.UNA>=SND.WL2. 3755 */ 3756 tcp_update_wl(tp, ack_seq); 3757 tcp_snd_una_update(tp, ack); 3758 flag |= FLAG_WIN_UPDATE; 3759 3760 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE); 3761 3762 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS); 3763 } else { 3764 u32 ack_ev_flags = CA_ACK_SLOWPATH; 3765 3766 if (ack_seq != TCP_SKB_CB(skb)->end_seq) 3767 flag |= FLAG_DATA; 3768 else 3769 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS); 3770 3771 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq); 3772 3773 if (TCP_SKB_CB(skb)->sacked) 3774 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una, 3775 &sack_state); 3776 3777 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) { 3778 flag |= FLAG_ECE; 3779 ack_ev_flags |= CA_ACK_ECE; 3780 } 3781 3782 if (sack_state.sack_delivered) 3783 tcp_count_delivered(tp, sack_state.sack_delivered, 3784 flag & FLAG_ECE); 3785 3786 if (flag & FLAG_WIN_UPDATE) 3787 ack_ev_flags |= CA_ACK_WIN_UPDATE; 3788 3789 tcp_in_ack_event(sk, ack_ev_flags); 3790 } 3791 3792 /* This is a deviation from RFC3168 since it states that: 3793 * "When the TCP data sender is ready to set the CWR bit after reducing 3794 * the congestion window, it SHOULD set the CWR bit only on the first 3795 * new data packet that it transmits." 3796 * We accept CWR on pure ACKs to be more robust 3797 * with widely-deployed TCP implementations that do this. 3798 */ 3799 tcp_ecn_accept_cwr(sk, skb); 3800 3801 /* We passed data and got it acked, remove any soft error 3802 * log. Something worked... 3803 */ 3804 sk->sk_err_soft = 0; 3805 icsk->icsk_probes_out = 0; 3806 tp->rcv_tstamp = tcp_jiffies32; 3807 if (!prior_packets) 3808 goto no_queue; 3809 3810 /* See if we can take anything off of the retransmit queue. */ 3811 flag |= tcp_clean_rtx_queue(sk, prior_fack, prior_snd_una, &sack_state, 3812 flag & FLAG_ECE); 3813 3814 tcp_rack_update_reo_wnd(sk, &rs); 3815 3816 if (tp->tlp_high_seq) 3817 tcp_process_tlp_ack(sk, ack, flag); 3818 /* If needed, reset TLP/RTO timer; RACK may later override this. */ 3819 if (flag & FLAG_SET_XMIT_TIMER) 3820 tcp_set_xmit_timer(sk); 3821 3822 if (tcp_ack_is_dubious(sk, flag)) { 3823 if (!(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP))) { 3824 num_dupack = 1; 3825 /* Consider if pure acks were aggregated in tcp_add_backlog() */ 3826 if (!(flag & FLAG_DATA)) 3827 num_dupack = max_t(u16, 1, skb_shinfo(skb)->gso_segs); 3828 } 3829 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag, 3830 &rexmit); 3831 } 3832 3833 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) 3834 sk_dst_confirm(sk); 3835 3836 delivered = tcp_newly_delivered(sk, delivered, flag); 3837 lost = tp->lost - lost; /* freshly marked lost */ 3838 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED); 3839 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate); 3840 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate); 3841 tcp_xmit_recovery(sk, rexmit); 3842 return 1; 3843 3844 no_queue: 3845 /* If data was DSACKed, see if we can undo a cwnd reduction. */ 3846 if (flag & FLAG_DSACKING_ACK) { 3847 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag, 3848 &rexmit); 3849 tcp_newly_delivered(sk, delivered, flag); 3850 } 3851 /* If this ack opens up a zero window, clear backoff. It was 3852 * being used to time the probes, and is probably far higher than 3853 * it needs to be for normal retransmission. 3854 */ 3855 tcp_ack_probe(sk); 3856 3857 if (tp->tlp_high_seq) 3858 tcp_process_tlp_ack(sk, ack, flag); 3859 return 1; 3860 3861 old_ack: 3862 /* If data was SACKed, tag it and see if we should send more data. 3863 * If data was DSACKed, see if we can undo a cwnd reduction. 3864 */ 3865 if (TCP_SKB_CB(skb)->sacked) { 3866 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una, 3867 &sack_state); 3868 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag, 3869 &rexmit); 3870 tcp_newly_delivered(sk, delivered, flag); 3871 tcp_xmit_recovery(sk, rexmit); 3872 } 3873 3874 return 0; 3875 } 3876 3877 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie, 3878 bool syn, struct tcp_fastopen_cookie *foc, 3879 bool exp_opt) 3880 { 3881 /* Valid only in SYN or SYN-ACK with an even length. */ 3882 if (!foc || !syn || len < 0 || (len & 1)) 3883 return; 3884 3885 if (len >= TCP_FASTOPEN_COOKIE_MIN && 3886 len <= TCP_FASTOPEN_COOKIE_MAX) 3887 memcpy(foc->val, cookie, len); 3888 else if (len != 0) 3889 len = -1; 3890 foc->len = len; 3891 foc->exp = exp_opt; 3892 } 3893 3894 static bool smc_parse_options(const struct tcphdr *th, 3895 struct tcp_options_received *opt_rx, 3896 const unsigned char *ptr, 3897 int opsize) 3898 { 3899 #if IS_ENABLED(CONFIG_SMC) 3900 if (static_branch_unlikely(&tcp_have_smc)) { 3901 if (th->syn && !(opsize & 1) && 3902 opsize >= TCPOLEN_EXP_SMC_BASE && 3903 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC) { 3904 opt_rx->smc_ok = 1; 3905 return true; 3906 } 3907 } 3908 #endif 3909 return false; 3910 } 3911 3912 /* Try to parse the MSS option from the TCP header. Return 0 on failure, clamped 3913 * value on success. 3914 */ 3915 static u16 tcp_parse_mss_option(const struct tcphdr *th, u16 user_mss) 3916 { 3917 const unsigned char *ptr = (const unsigned char *)(th + 1); 3918 int length = (th->doff * 4) - sizeof(struct tcphdr); 3919 u16 mss = 0; 3920 3921 while (length > 0) { 3922 int opcode = *ptr++; 3923 int opsize; 3924 3925 switch (opcode) { 3926 case TCPOPT_EOL: 3927 return mss; 3928 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ 3929 length--; 3930 continue; 3931 default: 3932 if (length < 2) 3933 return mss; 3934 opsize = *ptr++; 3935 if (opsize < 2) /* "silly options" */ 3936 return mss; 3937 if (opsize > length) 3938 return mss; /* fail on partial options */ 3939 if (opcode == TCPOPT_MSS && opsize == TCPOLEN_MSS) { 3940 u16 in_mss = get_unaligned_be16(ptr); 3941 3942 if (in_mss) { 3943 if (user_mss && user_mss < in_mss) 3944 in_mss = user_mss; 3945 mss = in_mss; 3946 } 3947 } 3948 ptr += opsize - 2; 3949 length -= opsize; 3950 } 3951 } 3952 return mss; 3953 } 3954 3955 /* Look for tcp options. Normally only called on SYN and SYNACK packets. 3956 * But, this can also be called on packets in the established flow when 3957 * the fast version below fails. 3958 */ 3959 void tcp_parse_options(const struct net *net, 3960 const struct sk_buff *skb, 3961 struct tcp_options_received *opt_rx, int estab, 3962 struct tcp_fastopen_cookie *foc) 3963 { 3964 const unsigned char *ptr; 3965 const struct tcphdr *th = tcp_hdr(skb); 3966 int length = (th->doff * 4) - sizeof(struct tcphdr); 3967 3968 ptr = (const unsigned char *)(th + 1); 3969 opt_rx->saw_tstamp = 0; 3970 opt_rx->saw_unknown = 0; 3971 3972 while (length > 0) { 3973 int opcode = *ptr++; 3974 int opsize; 3975 3976 switch (opcode) { 3977 case TCPOPT_EOL: 3978 return; 3979 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ 3980 length--; 3981 continue; 3982 default: 3983 if (length < 2) 3984 return; 3985 opsize = *ptr++; 3986 if (opsize < 2) /* "silly options" */ 3987 return; 3988 if (opsize > length) 3989 return; /* don't parse partial options */ 3990 switch (opcode) { 3991 case TCPOPT_MSS: 3992 if (opsize == TCPOLEN_MSS && th->syn && !estab) { 3993 u16 in_mss = get_unaligned_be16(ptr); 3994 if (in_mss) { 3995 if (opt_rx->user_mss && 3996 opt_rx->user_mss < in_mss) 3997 in_mss = opt_rx->user_mss; 3998 opt_rx->mss_clamp = in_mss; 3999 } 4000 } 4001 break; 4002 case TCPOPT_WINDOW: 4003 if (opsize == TCPOLEN_WINDOW && th->syn && 4004 !estab && net->ipv4.sysctl_tcp_window_scaling) { 4005 __u8 snd_wscale = *(__u8 *)ptr; 4006 opt_rx->wscale_ok = 1; 4007 if (snd_wscale > TCP_MAX_WSCALE) { 4008 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n", 4009 __func__, 4010 snd_wscale, 4011 TCP_MAX_WSCALE); 4012 snd_wscale = TCP_MAX_WSCALE; 4013 } 4014 opt_rx->snd_wscale = snd_wscale; 4015 } 4016 break; 4017 case TCPOPT_TIMESTAMP: 4018 if ((opsize == TCPOLEN_TIMESTAMP) && 4019 ((estab && opt_rx->tstamp_ok) || 4020 (!estab && net->ipv4.sysctl_tcp_timestamps))) { 4021 opt_rx->saw_tstamp = 1; 4022 opt_rx->rcv_tsval = get_unaligned_be32(ptr); 4023 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4); 4024 } 4025 break; 4026 case TCPOPT_SACK_PERM: 4027 if (opsize == TCPOLEN_SACK_PERM && th->syn && 4028 !estab && net->ipv4.sysctl_tcp_sack) { 4029 opt_rx->sack_ok = TCP_SACK_SEEN; 4030 tcp_sack_reset(opt_rx); 4031 } 4032 break; 4033 4034 case TCPOPT_SACK: 4035 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) && 4036 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) && 4037 opt_rx->sack_ok) { 4038 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th; 4039 } 4040 break; 4041 #ifdef CONFIG_TCP_MD5SIG 4042 case TCPOPT_MD5SIG: 4043 /* 4044 * The MD5 Hash has already been 4045 * checked (see tcp_v{4,6}_do_rcv()). 4046 */ 4047 break; 4048 #endif 4049 case TCPOPT_FASTOPEN: 4050 tcp_parse_fastopen_option( 4051 opsize - TCPOLEN_FASTOPEN_BASE, 4052 ptr, th->syn, foc, false); 4053 break; 4054 4055 case TCPOPT_EXP: 4056 /* Fast Open option shares code 254 using a 4057 * 16 bits magic number. 4058 */ 4059 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE && 4060 get_unaligned_be16(ptr) == 4061 TCPOPT_FASTOPEN_MAGIC) { 4062 tcp_parse_fastopen_option(opsize - 4063 TCPOLEN_EXP_FASTOPEN_BASE, 4064 ptr + 2, th->syn, foc, true); 4065 break; 4066 } 4067 4068 if (smc_parse_options(th, opt_rx, ptr, opsize)) 4069 break; 4070 4071 opt_rx->saw_unknown = 1; 4072 break; 4073 4074 default: 4075 opt_rx->saw_unknown = 1; 4076 } 4077 ptr += opsize-2; 4078 length -= opsize; 4079 } 4080 } 4081 } 4082 EXPORT_SYMBOL(tcp_parse_options); 4083 4084 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th) 4085 { 4086 const __be32 *ptr = (const __be32 *)(th + 1); 4087 4088 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) 4089 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) { 4090 tp->rx_opt.saw_tstamp = 1; 4091 ++ptr; 4092 tp->rx_opt.rcv_tsval = ntohl(*ptr); 4093 ++ptr; 4094 if (*ptr) 4095 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset; 4096 else 4097 tp->rx_opt.rcv_tsecr = 0; 4098 return true; 4099 } 4100 return false; 4101 } 4102 4103 /* Fast parse options. This hopes to only see timestamps. 4104 * If it is wrong it falls back on tcp_parse_options(). 4105 */ 4106 static bool tcp_fast_parse_options(const struct net *net, 4107 const struct sk_buff *skb, 4108 const struct tcphdr *th, struct tcp_sock *tp) 4109 { 4110 /* In the spirit of fast parsing, compare doff directly to constant 4111 * values. Because equality is used, short doff can be ignored here. 4112 */ 4113 if (th->doff == (sizeof(*th) / 4)) { 4114 tp->rx_opt.saw_tstamp = 0; 4115 return false; 4116 } else if (tp->rx_opt.tstamp_ok && 4117 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) { 4118 if (tcp_parse_aligned_timestamp(tp, th)) 4119 return true; 4120 } 4121 4122 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL); 4123 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) 4124 tp->rx_opt.rcv_tsecr -= tp->tsoffset; 4125 4126 return true; 4127 } 4128 4129 #ifdef CONFIG_TCP_MD5SIG 4130 /* 4131 * Parse MD5 Signature option 4132 */ 4133 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th) 4134 { 4135 int length = (th->doff << 2) - sizeof(*th); 4136 const u8 *ptr = (const u8 *)(th + 1); 4137 4138 /* If not enough data remaining, we can short cut */ 4139 while (length >= TCPOLEN_MD5SIG) { 4140 int opcode = *ptr++; 4141 int opsize; 4142 4143 switch (opcode) { 4144 case TCPOPT_EOL: 4145 return NULL; 4146 case TCPOPT_NOP: 4147 length--; 4148 continue; 4149 default: 4150 opsize = *ptr++; 4151 if (opsize < 2 || opsize > length) 4152 return NULL; 4153 if (opcode == TCPOPT_MD5SIG) 4154 return opsize == TCPOLEN_MD5SIG ? ptr : NULL; 4155 } 4156 ptr += opsize - 2; 4157 length -= opsize; 4158 } 4159 return NULL; 4160 } 4161 EXPORT_SYMBOL(tcp_parse_md5sig_option); 4162 #endif 4163 4164 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM 4165 * 4166 * It is not fatal. If this ACK does _not_ change critical state (seqs, window) 4167 * it can pass through stack. So, the following predicate verifies that 4168 * this segment is not used for anything but congestion avoidance or 4169 * fast retransmit. Moreover, we even are able to eliminate most of such 4170 * second order effects, if we apply some small "replay" window (~RTO) 4171 * to timestamp space. 4172 * 4173 * All these measures still do not guarantee that we reject wrapped ACKs 4174 * on networks with high bandwidth, when sequence space is recycled fastly, 4175 * but it guarantees that such events will be very rare and do not affect 4176 * connection seriously. This doesn't look nice, but alas, PAWS is really 4177 * buggy extension. 4178 * 4179 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC 4180 * states that events when retransmit arrives after original data are rare. 4181 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is 4182 * the biggest problem on large power networks even with minor reordering. 4183 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe 4184 * up to bandwidth of 18Gigabit/sec. 8) ] 4185 */ 4186 4187 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb) 4188 { 4189 const struct tcp_sock *tp = tcp_sk(sk); 4190 const struct tcphdr *th = tcp_hdr(skb); 4191 u32 seq = TCP_SKB_CB(skb)->seq; 4192 u32 ack = TCP_SKB_CB(skb)->ack_seq; 4193 4194 return (/* 1. Pure ACK with correct sequence number. */ 4195 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) && 4196 4197 /* 2. ... and duplicate ACK. */ 4198 ack == tp->snd_una && 4199 4200 /* 3. ... and does not update window. */ 4201 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) && 4202 4203 /* 4. ... and sits in replay window. */ 4204 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ); 4205 } 4206 4207 static inline bool tcp_paws_discard(const struct sock *sk, 4208 const struct sk_buff *skb) 4209 { 4210 const struct tcp_sock *tp = tcp_sk(sk); 4211 4212 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) && 4213 !tcp_disordered_ack(sk, skb); 4214 } 4215 4216 /* Check segment sequence number for validity. 4217 * 4218 * Segment controls are considered valid, if the segment 4219 * fits to the window after truncation to the window. Acceptability 4220 * of data (and SYN, FIN, of course) is checked separately. 4221 * See tcp_data_queue(), for example. 4222 * 4223 * Also, controls (RST is main one) are accepted using RCV.WUP instead 4224 * of RCV.NXT. Peer still did not advance his SND.UNA when we 4225 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP. 4226 * (borrowed from freebsd) 4227 */ 4228 4229 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq) 4230 { 4231 return !before(end_seq, tp->rcv_wup) && 4232 !after(seq, tp->rcv_nxt + tcp_receive_window(tp)); 4233 } 4234 4235 /* When we get a reset we do this. */ 4236 void tcp_reset(struct sock *sk, struct sk_buff *skb) 4237 { 4238 trace_tcp_receive_reset(sk); 4239 4240 if (sk_is_mptcp(sk)) 4241 mptcp_incoming_options(sk, skb); 4242 4243 /* We want the right error as BSD sees it (and indeed as we do). */ 4244 switch (sk->sk_state) { 4245 case TCP_SYN_SENT: 4246 sk->sk_err = ECONNREFUSED; 4247 break; 4248 case TCP_CLOSE_WAIT: 4249 sk->sk_err = EPIPE; 4250 break; 4251 case TCP_CLOSE: 4252 return; 4253 default: 4254 sk->sk_err = ECONNRESET; 4255 } 4256 /* This barrier is coupled with smp_rmb() in tcp_poll() */ 4257 smp_wmb(); 4258 4259 tcp_write_queue_purge(sk); 4260 tcp_done(sk); 4261 4262 if (!sock_flag(sk, SOCK_DEAD)) 4263 sk->sk_error_report(sk); 4264 } 4265 4266 /* 4267 * Process the FIN bit. This now behaves as it is supposed to work 4268 * and the FIN takes effect when it is validly part of sequence 4269 * space. Not before when we get holes. 4270 * 4271 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT 4272 * (and thence onto LAST-ACK and finally, CLOSE, we never enter 4273 * TIME-WAIT) 4274 * 4275 * If we are in FINWAIT-1, a received FIN indicates simultaneous 4276 * close and we go into CLOSING (and later onto TIME-WAIT) 4277 * 4278 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT. 4279 */ 4280 void tcp_fin(struct sock *sk) 4281 { 4282 struct tcp_sock *tp = tcp_sk(sk); 4283 4284 inet_csk_schedule_ack(sk); 4285 4286 sk->sk_shutdown |= RCV_SHUTDOWN; 4287 sock_set_flag(sk, SOCK_DONE); 4288 4289 switch (sk->sk_state) { 4290 case TCP_SYN_RECV: 4291 case TCP_ESTABLISHED: 4292 /* Move to CLOSE_WAIT */ 4293 tcp_set_state(sk, TCP_CLOSE_WAIT); 4294 inet_csk_enter_pingpong_mode(sk); 4295 break; 4296 4297 case TCP_CLOSE_WAIT: 4298 case TCP_CLOSING: 4299 /* Received a retransmission of the FIN, do 4300 * nothing. 4301 */ 4302 break; 4303 case TCP_LAST_ACK: 4304 /* RFC793: Remain in the LAST-ACK state. */ 4305 break; 4306 4307 case TCP_FIN_WAIT1: 4308 /* This case occurs when a simultaneous close 4309 * happens, we must ack the received FIN and 4310 * enter the CLOSING state. 4311 */ 4312 tcp_send_ack(sk); 4313 tcp_set_state(sk, TCP_CLOSING); 4314 break; 4315 case TCP_FIN_WAIT2: 4316 /* Received a FIN -- send ACK and enter TIME_WAIT. */ 4317 tcp_send_ack(sk); 4318 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 4319 break; 4320 default: 4321 /* Only TCP_LISTEN and TCP_CLOSE are left, in these 4322 * cases we should never reach this piece of code. 4323 */ 4324 pr_err("%s: Impossible, sk->sk_state=%d\n", 4325 __func__, sk->sk_state); 4326 break; 4327 } 4328 4329 /* It _is_ possible, that we have something out-of-order _after_ FIN. 4330 * Probably, we should reset in this case. For now drop them. 4331 */ 4332 skb_rbtree_purge(&tp->out_of_order_queue); 4333 if (tcp_is_sack(tp)) 4334 tcp_sack_reset(&tp->rx_opt); 4335 sk_mem_reclaim(sk); 4336 4337 if (!sock_flag(sk, SOCK_DEAD)) { 4338 sk->sk_state_change(sk); 4339 4340 /* Do not send POLL_HUP for half duplex close. */ 4341 if (sk->sk_shutdown == SHUTDOWN_MASK || 4342 sk->sk_state == TCP_CLOSE) 4343 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP); 4344 else 4345 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN); 4346 } 4347 } 4348 4349 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, 4350 u32 end_seq) 4351 { 4352 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) { 4353 if (before(seq, sp->start_seq)) 4354 sp->start_seq = seq; 4355 if (after(end_seq, sp->end_seq)) 4356 sp->end_seq = end_seq; 4357 return true; 4358 } 4359 return false; 4360 } 4361 4362 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq) 4363 { 4364 struct tcp_sock *tp = tcp_sk(sk); 4365 4366 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) { 4367 int mib_idx; 4368 4369 if (before(seq, tp->rcv_nxt)) 4370 mib_idx = LINUX_MIB_TCPDSACKOLDSENT; 4371 else 4372 mib_idx = LINUX_MIB_TCPDSACKOFOSENT; 4373 4374 NET_INC_STATS(sock_net(sk), mib_idx); 4375 4376 tp->rx_opt.dsack = 1; 4377 tp->duplicate_sack[0].start_seq = seq; 4378 tp->duplicate_sack[0].end_seq = end_seq; 4379 } 4380 } 4381 4382 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq) 4383 { 4384 struct tcp_sock *tp = tcp_sk(sk); 4385 4386 if (!tp->rx_opt.dsack) 4387 tcp_dsack_set(sk, seq, end_seq); 4388 else 4389 tcp_sack_extend(tp->duplicate_sack, seq, end_seq); 4390 } 4391 4392 static void tcp_rcv_spurious_retrans(struct sock *sk, const struct sk_buff *skb) 4393 { 4394 /* When the ACK path fails or drops most ACKs, the sender would 4395 * timeout and spuriously retransmit the same segment repeatedly. 4396 * The receiver remembers and reflects via DSACKs. Leverage the 4397 * DSACK state and change the txhash to re-route speculatively. 4398 */ 4399 if (TCP_SKB_CB(skb)->seq == tcp_sk(sk)->duplicate_sack[0].start_seq) { 4400 sk_rethink_txhash(sk); 4401 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDUPLICATEDATAREHASH); 4402 } 4403 } 4404 4405 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb) 4406 { 4407 struct tcp_sock *tp = tcp_sk(sk); 4408 4409 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 4410 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4411 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); 4412 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS); 4413 4414 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) { 4415 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 4416 4417 tcp_rcv_spurious_retrans(sk, skb); 4418 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) 4419 end_seq = tp->rcv_nxt; 4420 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq); 4421 } 4422 } 4423 4424 tcp_send_ack(sk); 4425 } 4426 4427 /* These routines update the SACK block as out-of-order packets arrive or 4428 * in-order packets close up the sequence space. 4429 */ 4430 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp) 4431 { 4432 int this_sack; 4433 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4434 struct tcp_sack_block *swalk = sp + 1; 4435 4436 /* See if the recent change to the first SACK eats into 4437 * or hits the sequence space of other SACK blocks, if so coalesce. 4438 */ 4439 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) { 4440 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) { 4441 int i; 4442 4443 /* Zap SWALK, by moving every further SACK up by one slot. 4444 * Decrease num_sacks. 4445 */ 4446 tp->rx_opt.num_sacks--; 4447 for (i = this_sack; i < tp->rx_opt.num_sacks; i++) 4448 sp[i] = sp[i + 1]; 4449 continue; 4450 } 4451 this_sack++; 4452 swalk++; 4453 } 4454 } 4455 4456 static void tcp_sack_compress_send_ack(struct sock *sk) 4457 { 4458 struct tcp_sock *tp = tcp_sk(sk); 4459 4460 if (!tp->compressed_ack) 4461 return; 4462 4463 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1) 4464 __sock_put(sk); 4465 4466 /* Since we have to send one ack finally, 4467 * substract one from tp->compressed_ack to keep 4468 * LINUX_MIB_TCPACKCOMPRESSED accurate. 4469 */ 4470 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED, 4471 tp->compressed_ack - 1); 4472 4473 tp->compressed_ack = 0; 4474 tcp_send_ack(sk); 4475 } 4476 4477 /* Reasonable amount of sack blocks included in TCP SACK option 4478 * The max is 4, but this becomes 3 if TCP timestamps are there. 4479 * Given that SACK packets might be lost, be conservative and use 2. 4480 */ 4481 #define TCP_SACK_BLOCKS_EXPECTED 2 4482 4483 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq) 4484 { 4485 struct tcp_sock *tp = tcp_sk(sk); 4486 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4487 int cur_sacks = tp->rx_opt.num_sacks; 4488 int this_sack; 4489 4490 if (!cur_sacks) 4491 goto new_sack; 4492 4493 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) { 4494 if (tcp_sack_extend(sp, seq, end_seq)) { 4495 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED) 4496 tcp_sack_compress_send_ack(sk); 4497 /* Rotate this_sack to the first one. */ 4498 for (; this_sack > 0; this_sack--, sp--) 4499 swap(*sp, *(sp - 1)); 4500 if (cur_sacks > 1) 4501 tcp_sack_maybe_coalesce(tp); 4502 return; 4503 } 4504 } 4505 4506 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED) 4507 tcp_sack_compress_send_ack(sk); 4508 4509 /* Could not find an adjacent existing SACK, build a new one, 4510 * put it at the front, and shift everyone else down. We 4511 * always know there is at least one SACK present already here. 4512 * 4513 * If the sack array is full, forget about the last one. 4514 */ 4515 if (this_sack >= TCP_NUM_SACKS) { 4516 this_sack--; 4517 tp->rx_opt.num_sacks--; 4518 sp--; 4519 } 4520 for (; this_sack > 0; this_sack--, sp--) 4521 *sp = *(sp - 1); 4522 4523 new_sack: 4524 /* Build the new head SACK, and we're done. */ 4525 sp->start_seq = seq; 4526 sp->end_seq = end_seq; 4527 tp->rx_opt.num_sacks++; 4528 } 4529 4530 /* RCV.NXT advances, some SACKs should be eaten. */ 4531 4532 static void tcp_sack_remove(struct tcp_sock *tp) 4533 { 4534 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4535 int num_sacks = tp->rx_opt.num_sacks; 4536 int this_sack; 4537 4538 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */ 4539 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) { 4540 tp->rx_opt.num_sacks = 0; 4541 return; 4542 } 4543 4544 for (this_sack = 0; this_sack < num_sacks;) { 4545 /* Check if the start of the sack is covered by RCV.NXT. */ 4546 if (!before(tp->rcv_nxt, sp->start_seq)) { 4547 int i; 4548 4549 /* RCV.NXT must cover all the block! */ 4550 WARN_ON(before(tp->rcv_nxt, sp->end_seq)); 4551 4552 /* Zap this SACK, by moving forward any other SACKS. */ 4553 for (i = this_sack+1; i < num_sacks; i++) 4554 tp->selective_acks[i-1] = tp->selective_acks[i]; 4555 num_sacks--; 4556 continue; 4557 } 4558 this_sack++; 4559 sp++; 4560 } 4561 tp->rx_opt.num_sacks = num_sacks; 4562 } 4563 4564 /** 4565 * tcp_try_coalesce - try to merge skb to prior one 4566 * @sk: socket 4567 * @to: prior buffer 4568 * @from: buffer to add in queue 4569 * @fragstolen: pointer to boolean 4570 * 4571 * Before queueing skb @from after @to, try to merge them 4572 * to reduce overall memory use and queue lengths, if cost is small. 4573 * Packets in ofo or receive queues can stay a long time. 4574 * Better try to coalesce them right now to avoid future collapses. 4575 * Returns true if caller should free @from instead of queueing it 4576 */ 4577 static bool tcp_try_coalesce(struct sock *sk, 4578 struct sk_buff *to, 4579 struct sk_buff *from, 4580 bool *fragstolen) 4581 { 4582 int delta; 4583 4584 *fragstolen = false; 4585 4586 /* Its possible this segment overlaps with prior segment in queue */ 4587 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq) 4588 return false; 4589 4590 if (!mptcp_skb_can_collapse(to, from)) 4591 return false; 4592 4593 #ifdef CONFIG_TLS_DEVICE 4594 if (from->decrypted != to->decrypted) 4595 return false; 4596 #endif 4597 4598 if (!skb_try_coalesce(to, from, fragstolen, &delta)) 4599 return false; 4600 4601 atomic_add(delta, &sk->sk_rmem_alloc); 4602 sk_mem_charge(sk, delta); 4603 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE); 4604 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq; 4605 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq; 4606 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags; 4607 4608 if (TCP_SKB_CB(from)->has_rxtstamp) { 4609 TCP_SKB_CB(to)->has_rxtstamp = true; 4610 to->tstamp = from->tstamp; 4611 skb_hwtstamps(to)->hwtstamp = skb_hwtstamps(from)->hwtstamp; 4612 } 4613 4614 return true; 4615 } 4616 4617 static bool tcp_ooo_try_coalesce(struct sock *sk, 4618 struct sk_buff *to, 4619 struct sk_buff *from, 4620 bool *fragstolen) 4621 { 4622 bool res = tcp_try_coalesce(sk, to, from, fragstolen); 4623 4624 /* In case tcp_drop() is called later, update to->gso_segs */ 4625 if (res) { 4626 u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) + 4627 max_t(u16, 1, skb_shinfo(from)->gso_segs); 4628 4629 skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF); 4630 } 4631 return res; 4632 } 4633 4634 static void tcp_drop(struct sock *sk, struct sk_buff *skb) 4635 { 4636 sk_drops_add(sk, skb); 4637 __kfree_skb(skb); 4638 } 4639 4640 /* This one checks to see if we can put data from the 4641 * out_of_order queue into the receive_queue. 4642 */ 4643 static void tcp_ofo_queue(struct sock *sk) 4644 { 4645 struct tcp_sock *tp = tcp_sk(sk); 4646 __u32 dsack_high = tp->rcv_nxt; 4647 bool fin, fragstolen, eaten; 4648 struct sk_buff *skb, *tail; 4649 struct rb_node *p; 4650 4651 p = rb_first(&tp->out_of_order_queue); 4652 while (p) { 4653 skb = rb_to_skb(p); 4654 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 4655 break; 4656 4657 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) { 4658 __u32 dsack = dsack_high; 4659 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high)) 4660 dsack_high = TCP_SKB_CB(skb)->end_seq; 4661 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack); 4662 } 4663 p = rb_next(p); 4664 rb_erase(&skb->rbnode, &tp->out_of_order_queue); 4665 4666 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) { 4667 tcp_drop(sk, skb); 4668 continue; 4669 } 4670 4671 tail = skb_peek_tail(&sk->sk_receive_queue); 4672 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen); 4673 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq); 4674 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN; 4675 if (!eaten) 4676 __skb_queue_tail(&sk->sk_receive_queue, skb); 4677 else 4678 kfree_skb_partial(skb, fragstolen); 4679 4680 if (unlikely(fin)) { 4681 tcp_fin(sk); 4682 /* tcp_fin() purges tp->out_of_order_queue, 4683 * so we must end this loop right now. 4684 */ 4685 break; 4686 } 4687 } 4688 } 4689 4690 static bool tcp_prune_ofo_queue(struct sock *sk); 4691 static int tcp_prune_queue(struct sock *sk); 4692 4693 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb, 4694 unsigned int size) 4695 { 4696 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || 4697 !sk_rmem_schedule(sk, skb, size)) { 4698 4699 if (tcp_prune_queue(sk) < 0) 4700 return -1; 4701 4702 while (!sk_rmem_schedule(sk, skb, size)) { 4703 if (!tcp_prune_ofo_queue(sk)) 4704 return -1; 4705 } 4706 } 4707 return 0; 4708 } 4709 4710 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb) 4711 { 4712 struct tcp_sock *tp = tcp_sk(sk); 4713 struct rb_node **p, *parent; 4714 struct sk_buff *skb1; 4715 u32 seq, end_seq; 4716 bool fragstolen; 4717 4718 tcp_ecn_check_ce(sk, skb); 4719 4720 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) { 4721 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP); 4722 sk->sk_data_ready(sk); 4723 tcp_drop(sk, skb); 4724 return; 4725 } 4726 4727 /* Disable header prediction. */ 4728 tp->pred_flags = 0; 4729 inet_csk_schedule_ack(sk); 4730 4731 tp->rcv_ooopack += max_t(u16, 1, skb_shinfo(skb)->gso_segs); 4732 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE); 4733 seq = TCP_SKB_CB(skb)->seq; 4734 end_seq = TCP_SKB_CB(skb)->end_seq; 4735 4736 p = &tp->out_of_order_queue.rb_node; 4737 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) { 4738 /* Initial out of order segment, build 1 SACK. */ 4739 if (tcp_is_sack(tp)) { 4740 tp->rx_opt.num_sacks = 1; 4741 tp->selective_acks[0].start_seq = seq; 4742 tp->selective_acks[0].end_seq = end_seq; 4743 } 4744 rb_link_node(&skb->rbnode, NULL, p); 4745 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue); 4746 tp->ooo_last_skb = skb; 4747 goto end; 4748 } 4749 4750 /* In the typical case, we are adding an skb to the end of the list. 4751 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup. 4752 */ 4753 if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb, 4754 skb, &fragstolen)) { 4755 coalesce_done: 4756 /* For non sack flows, do not grow window to force DUPACK 4757 * and trigger fast retransmit. 4758 */ 4759 if (tcp_is_sack(tp)) 4760 tcp_grow_window(sk, skb); 4761 kfree_skb_partial(skb, fragstolen); 4762 skb = NULL; 4763 goto add_sack; 4764 } 4765 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */ 4766 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) { 4767 parent = &tp->ooo_last_skb->rbnode; 4768 p = &parent->rb_right; 4769 goto insert; 4770 } 4771 4772 /* Find place to insert this segment. Handle overlaps on the way. */ 4773 parent = NULL; 4774 while (*p) { 4775 parent = *p; 4776 skb1 = rb_to_skb(parent); 4777 if (before(seq, TCP_SKB_CB(skb1)->seq)) { 4778 p = &parent->rb_left; 4779 continue; 4780 } 4781 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) { 4782 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 4783 /* All the bits are present. Drop. */ 4784 NET_INC_STATS(sock_net(sk), 4785 LINUX_MIB_TCPOFOMERGE); 4786 tcp_drop(sk, skb); 4787 skb = NULL; 4788 tcp_dsack_set(sk, seq, end_seq); 4789 goto add_sack; 4790 } 4791 if (after(seq, TCP_SKB_CB(skb1)->seq)) { 4792 /* Partial overlap. */ 4793 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq); 4794 } else { 4795 /* skb's seq == skb1's seq and skb covers skb1. 4796 * Replace skb1 with skb. 4797 */ 4798 rb_replace_node(&skb1->rbnode, &skb->rbnode, 4799 &tp->out_of_order_queue); 4800 tcp_dsack_extend(sk, 4801 TCP_SKB_CB(skb1)->seq, 4802 TCP_SKB_CB(skb1)->end_seq); 4803 NET_INC_STATS(sock_net(sk), 4804 LINUX_MIB_TCPOFOMERGE); 4805 tcp_drop(sk, skb1); 4806 goto merge_right; 4807 } 4808 } else if (tcp_ooo_try_coalesce(sk, skb1, 4809 skb, &fragstolen)) { 4810 goto coalesce_done; 4811 } 4812 p = &parent->rb_right; 4813 } 4814 insert: 4815 /* Insert segment into RB tree. */ 4816 rb_link_node(&skb->rbnode, parent, p); 4817 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue); 4818 4819 merge_right: 4820 /* Remove other segments covered by skb. */ 4821 while ((skb1 = skb_rb_next(skb)) != NULL) { 4822 if (!after(end_seq, TCP_SKB_CB(skb1)->seq)) 4823 break; 4824 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 4825 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, 4826 end_seq); 4827 break; 4828 } 4829 rb_erase(&skb1->rbnode, &tp->out_of_order_queue); 4830 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, 4831 TCP_SKB_CB(skb1)->end_seq); 4832 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE); 4833 tcp_drop(sk, skb1); 4834 } 4835 /* If there is no skb after us, we are the last_skb ! */ 4836 if (!skb1) 4837 tp->ooo_last_skb = skb; 4838 4839 add_sack: 4840 if (tcp_is_sack(tp)) 4841 tcp_sack_new_ofo_skb(sk, seq, end_seq); 4842 end: 4843 if (skb) { 4844 /* For non sack flows, do not grow window to force DUPACK 4845 * and trigger fast retransmit. 4846 */ 4847 if (tcp_is_sack(tp)) 4848 tcp_grow_window(sk, skb); 4849 skb_condense(skb); 4850 skb_set_owner_r(skb, sk); 4851 } 4852 } 4853 4854 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, 4855 bool *fragstolen) 4856 { 4857 int eaten; 4858 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue); 4859 4860 eaten = (tail && 4861 tcp_try_coalesce(sk, tail, 4862 skb, fragstolen)) ? 1 : 0; 4863 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq); 4864 if (!eaten) { 4865 __skb_queue_tail(&sk->sk_receive_queue, skb); 4866 skb_set_owner_r(skb, sk); 4867 } 4868 return eaten; 4869 } 4870 4871 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size) 4872 { 4873 struct sk_buff *skb; 4874 int err = -ENOMEM; 4875 int data_len = 0; 4876 bool fragstolen; 4877 4878 if (size == 0) 4879 return 0; 4880 4881 if (size > PAGE_SIZE) { 4882 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS); 4883 4884 data_len = npages << PAGE_SHIFT; 4885 size = data_len + (size & ~PAGE_MASK); 4886 } 4887 skb = alloc_skb_with_frags(size - data_len, data_len, 4888 PAGE_ALLOC_COSTLY_ORDER, 4889 &err, sk->sk_allocation); 4890 if (!skb) 4891 goto err; 4892 4893 skb_put(skb, size - data_len); 4894 skb->data_len = data_len; 4895 skb->len = size; 4896 4897 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) { 4898 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP); 4899 goto err_free; 4900 } 4901 4902 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size); 4903 if (err) 4904 goto err_free; 4905 4906 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt; 4907 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size; 4908 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1; 4909 4910 if (tcp_queue_rcv(sk, skb, &fragstolen)) { 4911 WARN_ON_ONCE(fragstolen); /* should not happen */ 4912 __kfree_skb(skb); 4913 } 4914 return size; 4915 4916 err_free: 4917 kfree_skb(skb); 4918 err: 4919 return err; 4920 4921 } 4922 4923 void tcp_data_ready(struct sock *sk) 4924 { 4925 const struct tcp_sock *tp = tcp_sk(sk); 4926 int avail = tp->rcv_nxt - tp->copied_seq; 4927 4928 if (avail < sk->sk_rcvlowat && !tcp_rmem_pressure(sk) && 4929 !sock_flag(sk, SOCK_DONE) && 4930 tcp_receive_window(tp) > inet_csk(sk)->icsk_ack.rcv_mss) 4931 return; 4932 4933 sk->sk_data_ready(sk); 4934 } 4935 4936 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb) 4937 { 4938 struct tcp_sock *tp = tcp_sk(sk); 4939 bool fragstolen; 4940 int eaten; 4941 4942 if (sk_is_mptcp(sk)) 4943 mptcp_incoming_options(sk, skb); 4944 4945 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) { 4946 __kfree_skb(skb); 4947 return; 4948 } 4949 skb_dst_drop(skb); 4950 __skb_pull(skb, tcp_hdr(skb)->doff * 4); 4951 4952 tp->rx_opt.dsack = 0; 4953 4954 /* Queue data for delivery to the user. 4955 * Packets in sequence go to the receive queue. 4956 * Out of sequence packets to the out_of_order_queue. 4957 */ 4958 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { 4959 if (tcp_receive_window(tp) == 0) { 4960 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP); 4961 goto out_of_window; 4962 } 4963 4964 /* Ok. In sequence. In window. */ 4965 queue_and_out: 4966 if (skb_queue_len(&sk->sk_receive_queue) == 0) 4967 sk_forced_mem_schedule(sk, skb->truesize); 4968 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) { 4969 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP); 4970 sk->sk_data_ready(sk); 4971 goto drop; 4972 } 4973 4974 eaten = tcp_queue_rcv(sk, skb, &fragstolen); 4975 if (skb->len) 4976 tcp_event_data_recv(sk, skb); 4977 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 4978 tcp_fin(sk); 4979 4980 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) { 4981 tcp_ofo_queue(sk); 4982 4983 /* RFC5681. 4.2. SHOULD send immediate ACK, when 4984 * gap in queue is filled. 4985 */ 4986 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) 4987 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW; 4988 } 4989 4990 if (tp->rx_opt.num_sacks) 4991 tcp_sack_remove(tp); 4992 4993 tcp_fast_path_check(sk); 4994 4995 if (eaten > 0) 4996 kfree_skb_partial(skb, fragstolen); 4997 if (!sock_flag(sk, SOCK_DEAD)) 4998 tcp_data_ready(sk); 4999 return; 5000 } 5001 5002 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 5003 tcp_rcv_spurious_retrans(sk, skb); 5004 /* A retransmit, 2nd most common case. Force an immediate ack. */ 5005 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); 5006 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 5007 5008 out_of_window: 5009 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS); 5010 inet_csk_schedule_ack(sk); 5011 drop: 5012 tcp_drop(sk, skb); 5013 return; 5014 } 5015 5016 /* Out of window. F.e. zero window probe. */ 5017 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp))) 5018 goto out_of_window; 5019 5020 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 5021 /* Partial packet, seq < rcv_next < end_seq */ 5022 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt); 5023 5024 /* If window is closed, drop tail of packet. But after 5025 * remembering D-SACK for its head made in previous line. 5026 */ 5027 if (!tcp_receive_window(tp)) { 5028 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP); 5029 goto out_of_window; 5030 } 5031 goto queue_and_out; 5032 } 5033 5034 tcp_data_queue_ofo(sk, skb); 5035 } 5036 5037 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list) 5038 { 5039 if (list) 5040 return !skb_queue_is_last(list, skb) ? skb->next : NULL; 5041 5042 return skb_rb_next(skb); 5043 } 5044 5045 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb, 5046 struct sk_buff_head *list, 5047 struct rb_root *root) 5048 { 5049 struct sk_buff *next = tcp_skb_next(skb, list); 5050 5051 if (list) 5052 __skb_unlink(skb, list); 5053 else 5054 rb_erase(&skb->rbnode, root); 5055 5056 __kfree_skb(skb); 5057 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED); 5058 5059 return next; 5060 } 5061 5062 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */ 5063 void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb) 5064 { 5065 struct rb_node **p = &root->rb_node; 5066 struct rb_node *parent = NULL; 5067 struct sk_buff *skb1; 5068 5069 while (*p) { 5070 parent = *p; 5071 skb1 = rb_to_skb(parent); 5072 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq)) 5073 p = &parent->rb_left; 5074 else 5075 p = &parent->rb_right; 5076 } 5077 rb_link_node(&skb->rbnode, parent, p); 5078 rb_insert_color(&skb->rbnode, root); 5079 } 5080 5081 /* Collapse contiguous sequence of skbs head..tail with 5082 * sequence numbers start..end. 5083 * 5084 * If tail is NULL, this means until the end of the queue. 5085 * 5086 * Segments with FIN/SYN are not collapsed (only because this 5087 * simplifies code) 5088 */ 5089 static void 5090 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root, 5091 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end) 5092 { 5093 struct sk_buff *skb = head, *n; 5094 struct sk_buff_head tmp; 5095 bool end_of_skbs; 5096 5097 /* First, check that queue is collapsible and find 5098 * the point where collapsing can be useful. 5099 */ 5100 restart: 5101 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) { 5102 n = tcp_skb_next(skb, list); 5103 5104 /* No new bits? It is possible on ofo queue. */ 5105 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 5106 skb = tcp_collapse_one(sk, skb, list, root); 5107 if (!skb) 5108 break; 5109 goto restart; 5110 } 5111 5112 /* The first skb to collapse is: 5113 * - not SYN/FIN and 5114 * - bloated or contains data before "start" or 5115 * overlaps to the next one and mptcp allow collapsing. 5116 */ 5117 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) && 5118 (tcp_win_from_space(sk, skb->truesize) > skb->len || 5119 before(TCP_SKB_CB(skb)->seq, start))) { 5120 end_of_skbs = false; 5121 break; 5122 } 5123 5124 if (n && n != tail && mptcp_skb_can_collapse(skb, n) && 5125 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) { 5126 end_of_skbs = false; 5127 break; 5128 } 5129 5130 /* Decided to skip this, advance start seq. */ 5131 start = TCP_SKB_CB(skb)->end_seq; 5132 } 5133 if (end_of_skbs || 5134 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN))) 5135 return; 5136 5137 __skb_queue_head_init(&tmp); 5138 5139 while (before(start, end)) { 5140 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start); 5141 struct sk_buff *nskb; 5142 5143 nskb = alloc_skb(copy, GFP_ATOMIC); 5144 if (!nskb) 5145 break; 5146 5147 memcpy(nskb->cb, skb->cb, sizeof(skb->cb)); 5148 #ifdef CONFIG_TLS_DEVICE 5149 nskb->decrypted = skb->decrypted; 5150 #endif 5151 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start; 5152 if (list) 5153 __skb_queue_before(list, skb, nskb); 5154 else 5155 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */ 5156 skb_set_owner_r(nskb, sk); 5157 mptcp_skb_ext_move(nskb, skb); 5158 5159 /* Copy data, releasing collapsed skbs. */ 5160 while (copy > 0) { 5161 int offset = start - TCP_SKB_CB(skb)->seq; 5162 int size = TCP_SKB_CB(skb)->end_seq - start; 5163 5164 BUG_ON(offset < 0); 5165 if (size > 0) { 5166 size = min(copy, size); 5167 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size)) 5168 BUG(); 5169 TCP_SKB_CB(nskb)->end_seq += size; 5170 copy -= size; 5171 start += size; 5172 } 5173 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 5174 skb = tcp_collapse_one(sk, skb, list, root); 5175 if (!skb || 5176 skb == tail || 5177 !mptcp_skb_can_collapse(nskb, skb) || 5178 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN))) 5179 goto end; 5180 #ifdef CONFIG_TLS_DEVICE 5181 if (skb->decrypted != nskb->decrypted) 5182 goto end; 5183 #endif 5184 } 5185 } 5186 } 5187 end: 5188 skb_queue_walk_safe(&tmp, skb, n) 5189 tcp_rbtree_insert(root, skb); 5190 } 5191 5192 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs 5193 * and tcp_collapse() them until all the queue is collapsed. 5194 */ 5195 static void tcp_collapse_ofo_queue(struct sock *sk) 5196 { 5197 struct tcp_sock *tp = tcp_sk(sk); 5198 u32 range_truesize, sum_tiny = 0; 5199 struct sk_buff *skb, *head; 5200 u32 start, end; 5201 5202 skb = skb_rb_first(&tp->out_of_order_queue); 5203 new_range: 5204 if (!skb) { 5205 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue); 5206 return; 5207 } 5208 start = TCP_SKB_CB(skb)->seq; 5209 end = TCP_SKB_CB(skb)->end_seq; 5210 range_truesize = skb->truesize; 5211 5212 for (head = skb;;) { 5213 skb = skb_rb_next(skb); 5214 5215 /* Range is terminated when we see a gap or when 5216 * we are at the queue end. 5217 */ 5218 if (!skb || 5219 after(TCP_SKB_CB(skb)->seq, end) || 5220 before(TCP_SKB_CB(skb)->end_seq, start)) { 5221 /* Do not attempt collapsing tiny skbs */ 5222 if (range_truesize != head->truesize || 5223 end - start >= SKB_WITH_OVERHEAD(SK_MEM_QUANTUM)) { 5224 tcp_collapse(sk, NULL, &tp->out_of_order_queue, 5225 head, skb, start, end); 5226 } else { 5227 sum_tiny += range_truesize; 5228 if (sum_tiny > sk->sk_rcvbuf >> 3) 5229 return; 5230 } 5231 goto new_range; 5232 } 5233 5234 range_truesize += skb->truesize; 5235 if (unlikely(before(TCP_SKB_CB(skb)->seq, start))) 5236 start = TCP_SKB_CB(skb)->seq; 5237 if (after(TCP_SKB_CB(skb)->end_seq, end)) 5238 end = TCP_SKB_CB(skb)->end_seq; 5239 } 5240 } 5241 5242 /* 5243 * Clean the out-of-order queue to make room. 5244 * We drop high sequences packets to : 5245 * 1) Let a chance for holes to be filled. 5246 * 2) not add too big latencies if thousands of packets sit there. 5247 * (But if application shrinks SO_RCVBUF, we could still end up 5248 * freeing whole queue here) 5249 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks. 5250 * 5251 * Return true if queue has shrunk. 5252 */ 5253 static bool tcp_prune_ofo_queue(struct sock *sk) 5254 { 5255 struct tcp_sock *tp = tcp_sk(sk); 5256 struct rb_node *node, *prev; 5257 int goal; 5258 5259 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) 5260 return false; 5261 5262 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED); 5263 goal = sk->sk_rcvbuf >> 3; 5264 node = &tp->ooo_last_skb->rbnode; 5265 do { 5266 prev = rb_prev(node); 5267 rb_erase(node, &tp->out_of_order_queue); 5268 goal -= rb_to_skb(node)->truesize; 5269 tcp_drop(sk, rb_to_skb(node)); 5270 if (!prev || goal <= 0) { 5271 sk_mem_reclaim(sk); 5272 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf && 5273 !tcp_under_memory_pressure(sk)) 5274 break; 5275 goal = sk->sk_rcvbuf >> 3; 5276 } 5277 node = prev; 5278 } while (node); 5279 tp->ooo_last_skb = rb_to_skb(prev); 5280 5281 /* Reset SACK state. A conforming SACK implementation will 5282 * do the same at a timeout based retransmit. When a connection 5283 * is in a sad state like this, we care only about integrity 5284 * of the connection not performance. 5285 */ 5286 if (tp->rx_opt.sack_ok) 5287 tcp_sack_reset(&tp->rx_opt); 5288 return true; 5289 } 5290 5291 /* Reduce allocated memory if we can, trying to get 5292 * the socket within its memory limits again. 5293 * 5294 * Return less than zero if we should start dropping frames 5295 * until the socket owning process reads some of the data 5296 * to stabilize the situation. 5297 */ 5298 static int tcp_prune_queue(struct sock *sk) 5299 { 5300 struct tcp_sock *tp = tcp_sk(sk); 5301 5302 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED); 5303 5304 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf) 5305 tcp_clamp_window(sk); 5306 else if (tcp_under_memory_pressure(sk)) 5307 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss); 5308 5309 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 5310 return 0; 5311 5312 tcp_collapse_ofo_queue(sk); 5313 if (!skb_queue_empty(&sk->sk_receive_queue)) 5314 tcp_collapse(sk, &sk->sk_receive_queue, NULL, 5315 skb_peek(&sk->sk_receive_queue), 5316 NULL, 5317 tp->copied_seq, tp->rcv_nxt); 5318 sk_mem_reclaim(sk); 5319 5320 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 5321 return 0; 5322 5323 /* Collapsing did not help, destructive actions follow. 5324 * This must not ever occur. */ 5325 5326 tcp_prune_ofo_queue(sk); 5327 5328 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 5329 return 0; 5330 5331 /* If we are really being abused, tell the caller to silently 5332 * drop receive data on the floor. It will get retransmitted 5333 * and hopefully then we'll have sufficient space. 5334 */ 5335 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED); 5336 5337 /* Massive buffer overcommit. */ 5338 tp->pred_flags = 0; 5339 return -1; 5340 } 5341 5342 static bool tcp_should_expand_sndbuf(const struct sock *sk) 5343 { 5344 const struct tcp_sock *tp = tcp_sk(sk); 5345 5346 /* If the user specified a specific send buffer setting, do 5347 * not modify it. 5348 */ 5349 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK) 5350 return false; 5351 5352 /* If we are under global TCP memory pressure, do not expand. */ 5353 if (tcp_under_memory_pressure(sk)) 5354 return false; 5355 5356 /* If we are under soft global TCP memory pressure, do not expand. */ 5357 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0)) 5358 return false; 5359 5360 /* If we filled the congestion window, do not expand. */ 5361 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 5362 return false; 5363 5364 return true; 5365 } 5366 5367 static void tcp_new_space(struct sock *sk) 5368 { 5369 struct tcp_sock *tp = tcp_sk(sk); 5370 5371 if (tcp_should_expand_sndbuf(sk)) { 5372 tcp_sndbuf_expand(sk); 5373 tp->snd_cwnd_stamp = tcp_jiffies32; 5374 } 5375 5376 sk->sk_write_space(sk); 5377 } 5378 5379 static void tcp_check_space(struct sock *sk) 5380 { 5381 /* pairs with tcp_poll() */ 5382 smp_mb(); 5383 if (sk->sk_socket && 5384 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 5385 tcp_new_space(sk); 5386 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) 5387 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED); 5388 } 5389 } 5390 5391 static inline void tcp_data_snd_check(struct sock *sk) 5392 { 5393 tcp_push_pending_frames(sk); 5394 tcp_check_space(sk); 5395 } 5396 5397 /* 5398 * Check if sending an ack is needed. 5399 */ 5400 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible) 5401 { 5402 struct tcp_sock *tp = tcp_sk(sk); 5403 unsigned long rtt, delay; 5404 5405 /* More than one full frame received... */ 5406 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss && 5407 /* ... and right edge of window advances far enough. 5408 * (tcp_recvmsg() will send ACK otherwise). 5409 * If application uses SO_RCVLOWAT, we want send ack now if 5410 * we have not received enough bytes to satisfy the condition. 5411 */ 5412 (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat || 5413 __tcp_select_window(sk) >= tp->rcv_wnd)) || 5414 /* We ACK each frame or... */ 5415 tcp_in_quickack_mode(sk) || 5416 /* Protocol state mandates a one-time immediate ACK */ 5417 inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) { 5418 send_now: 5419 tcp_send_ack(sk); 5420 return; 5421 } 5422 5423 if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) { 5424 tcp_send_delayed_ack(sk); 5425 return; 5426 } 5427 5428 if (!tcp_is_sack(tp) || 5429 tp->compressed_ack >= sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr) 5430 goto send_now; 5431 5432 if (tp->compressed_ack_rcv_nxt != tp->rcv_nxt) { 5433 tp->compressed_ack_rcv_nxt = tp->rcv_nxt; 5434 tp->dup_ack_counter = 0; 5435 } 5436 if (tp->dup_ack_counter < TCP_FASTRETRANS_THRESH) { 5437 tp->dup_ack_counter++; 5438 goto send_now; 5439 } 5440 tp->compressed_ack++; 5441 if (hrtimer_is_queued(&tp->compressed_ack_timer)) 5442 return; 5443 5444 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */ 5445 5446 rtt = tp->rcv_rtt_est.rtt_us; 5447 if (tp->srtt_us && tp->srtt_us < rtt) 5448 rtt = tp->srtt_us; 5449 5450 delay = min_t(unsigned long, sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns, 5451 rtt * (NSEC_PER_USEC >> 3)/20); 5452 sock_hold(sk); 5453 hrtimer_start_range_ns(&tp->compressed_ack_timer, ns_to_ktime(delay), 5454 sock_net(sk)->ipv4.sysctl_tcp_comp_sack_slack_ns, 5455 HRTIMER_MODE_REL_PINNED_SOFT); 5456 } 5457 5458 static inline void tcp_ack_snd_check(struct sock *sk) 5459 { 5460 if (!inet_csk_ack_scheduled(sk)) { 5461 /* We sent a data segment already. */ 5462 return; 5463 } 5464 __tcp_ack_snd_check(sk, 1); 5465 } 5466 5467 /* 5468 * This routine is only called when we have urgent data 5469 * signaled. Its the 'slow' part of tcp_urg. It could be 5470 * moved inline now as tcp_urg is only called from one 5471 * place. We handle URGent data wrong. We have to - as 5472 * BSD still doesn't use the correction from RFC961. 5473 * For 1003.1g we should support a new option TCP_STDURG to permit 5474 * either form (or just set the sysctl tcp_stdurg). 5475 */ 5476 5477 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th) 5478 { 5479 struct tcp_sock *tp = tcp_sk(sk); 5480 u32 ptr = ntohs(th->urg_ptr); 5481 5482 if (ptr && !sock_net(sk)->ipv4.sysctl_tcp_stdurg) 5483 ptr--; 5484 ptr += ntohl(th->seq); 5485 5486 /* Ignore urgent data that we've already seen and read. */ 5487 if (after(tp->copied_seq, ptr)) 5488 return; 5489 5490 /* Do not replay urg ptr. 5491 * 5492 * NOTE: interesting situation not covered by specs. 5493 * Misbehaving sender may send urg ptr, pointing to segment, 5494 * which we already have in ofo queue. We are not able to fetch 5495 * such data and will stay in TCP_URG_NOTYET until will be eaten 5496 * by recvmsg(). Seems, we are not obliged to handle such wicked 5497 * situations. But it is worth to think about possibility of some 5498 * DoSes using some hypothetical application level deadlock. 5499 */ 5500 if (before(ptr, tp->rcv_nxt)) 5501 return; 5502 5503 /* Do we already have a newer (or duplicate) urgent pointer? */ 5504 if (tp->urg_data && !after(ptr, tp->urg_seq)) 5505 return; 5506 5507 /* Tell the world about our new urgent pointer. */ 5508 sk_send_sigurg(sk); 5509 5510 /* We may be adding urgent data when the last byte read was 5511 * urgent. To do this requires some care. We cannot just ignore 5512 * tp->copied_seq since we would read the last urgent byte again 5513 * as data, nor can we alter copied_seq until this data arrives 5514 * or we break the semantics of SIOCATMARK (and thus sockatmark()) 5515 * 5516 * NOTE. Double Dutch. Rendering to plain English: author of comment 5517 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB); 5518 * and expect that both A and B disappear from stream. This is _wrong_. 5519 * Though this happens in BSD with high probability, this is occasional. 5520 * Any application relying on this is buggy. Note also, that fix "works" 5521 * only in this artificial test. Insert some normal data between A and B and we will 5522 * decline of BSD again. Verdict: it is better to remove to trap 5523 * buggy users. 5524 */ 5525 if (tp->urg_seq == tp->copied_seq && tp->urg_data && 5526 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) { 5527 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 5528 tp->copied_seq++; 5529 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) { 5530 __skb_unlink(skb, &sk->sk_receive_queue); 5531 __kfree_skb(skb); 5532 } 5533 } 5534 5535 tp->urg_data = TCP_URG_NOTYET; 5536 WRITE_ONCE(tp->urg_seq, ptr); 5537 5538 /* Disable header prediction. */ 5539 tp->pred_flags = 0; 5540 } 5541 5542 /* This is the 'fast' part of urgent handling. */ 5543 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th) 5544 { 5545 struct tcp_sock *tp = tcp_sk(sk); 5546 5547 /* Check if we get a new urgent pointer - normally not. */ 5548 if (th->urg) 5549 tcp_check_urg(sk, th); 5550 5551 /* Do we wait for any urgent data? - normally not... */ 5552 if (tp->urg_data == TCP_URG_NOTYET) { 5553 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) - 5554 th->syn; 5555 5556 /* Is the urgent pointer pointing into this packet? */ 5557 if (ptr < skb->len) { 5558 u8 tmp; 5559 if (skb_copy_bits(skb, ptr, &tmp, 1)) 5560 BUG(); 5561 tp->urg_data = TCP_URG_VALID | tmp; 5562 if (!sock_flag(sk, SOCK_DEAD)) 5563 sk->sk_data_ready(sk); 5564 } 5565 } 5566 } 5567 5568 /* Accept RST for rcv_nxt - 1 after a FIN. 5569 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a 5570 * FIN is sent followed by a RST packet. The RST is sent with the same 5571 * sequence number as the FIN, and thus according to RFC 5961 a challenge 5572 * ACK should be sent. However, Mac OSX rate limits replies to challenge 5573 * ACKs on the closed socket. In addition middleboxes can drop either the 5574 * challenge ACK or a subsequent RST. 5575 */ 5576 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb) 5577 { 5578 struct tcp_sock *tp = tcp_sk(sk); 5579 5580 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) && 5581 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK | 5582 TCPF_CLOSING)); 5583 } 5584 5585 /* Does PAWS and seqno based validation of an incoming segment, flags will 5586 * play significant role here. 5587 */ 5588 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb, 5589 const struct tcphdr *th, int syn_inerr) 5590 { 5591 struct tcp_sock *tp = tcp_sk(sk); 5592 bool rst_seq_match = false; 5593 5594 /* RFC1323: H1. Apply PAWS check first. */ 5595 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) && 5596 tp->rx_opt.saw_tstamp && 5597 tcp_paws_discard(sk, skb)) { 5598 if (!th->rst) { 5599 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED); 5600 if (!tcp_oow_rate_limited(sock_net(sk), skb, 5601 LINUX_MIB_TCPACKSKIPPEDPAWS, 5602 &tp->last_oow_ack_time)) 5603 tcp_send_dupack(sk, skb); 5604 goto discard; 5605 } 5606 /* Reset is accepted even if it did not pass PAWS. */ 5607 } 5608 5609 /* Step 1: check sequence number */ 5610 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { 5611 /* RFC793, page 37: "In all states except SYN-SENT, all reset 5612 * (RST) segments are validated by checking their SEQ-fields." 5613 * And page 69: "If an incoming segment is not acceptable, 5614 * an acknowledgment should be sent in reply (unless the RST 5615 * bit is set, if so drop the segment and return)". 5616 */ 5617 if (!th->rst) { 5618 if (th->syn) 5619 goto syn_challenge; 5620 if (!tcp_oow_rate_limited(sock_net(sk), skb, 5621 LINUX_MIB_TCPACKSKIPPEDSEQ, 5622 &tp->last_oow_ack_time)) 5623 tcp_send_dupack(sk, skb); 5624 } else if (tcp_reset_check(sk, skb)) { 5625 tcp_reset(sk, skb); 5626 } 5627 goto discard; 5628 } 5629 5630 /* Step 2: check RST bit */ 5631 if (th->rst) { 5632 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a 5633 * FIN and SACK too if available): 5634 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or 5635 * the right-most SACK block, 5636 * then 5637 * RESET the connection 5638 * else 5639 * Send a challenge ACK 5640 */ 5641 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt || 5642 tcp_reset_check(sk, skb)) { 5643 rst_seq_match = true; 5644 } else if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) { 5645 struct tcp_sack_block *sp = &tp->selective_acks[0]; 5646 int max_sack = sp[0].end_seq; 5647 int this_sack; 5648 5649 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; 5650 ++this_sack) { 5651 max_sack = after(sp[this_sack].end_seq, 5652 max_sack) ? 5653 sp[this_sack].end_seq : max_sack; 5654 } 5655 5656 if (TCP_SKB_CB(skb)->seq == max_sack) 5657 rst_seq_match = true; 5658 } 5659 5660 if (rst_seq_match) 5661 tcp_reset(sk, skb); 5662 else { 5663 /* Disable TFO if RST is out-of-order 5664 * and no data has been received 5665 * for current active TFO socket 5666 */ 5667 if (tp->syn_fastopen && !tp->data_segs_in && 5668 sk->sk_state == TCP_ESTABLISHED) 5669 tcp_fastopen_active_disable(sk); 5670 tcp_send_challenge_ack(sk, skb); 5671 } 5672 goto discard; 5673 } 5674 5675 /* step 3: check security and precedence [ignored] */ 5676 5677 /* step 4: Check for a SYN 5678 * RFC 5961 4.2 : Send a challenge ack 5679 */ 5680 if (th->syn) { 5681 syn_challenge: 5682 if (syn_inerr) 5683 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS); 5684 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE); 5685 tcp_send_challenge_ack(sk, skb); 5686 goto discard; 5687 } 5688 5689 bpf_skops_parse_hdr(sk, skb); 5690 5691 return true; 5692 5693 discard: 5694 tcp_drop(sk, skb); 5695 return false; 5696 } 5697 5698 /* 5699 * TCP receive function for the ESTABLISHED state. 5700 * 5701 * It is split into a fast path and a slow path. The fast path is 5702 * disabled when: 5703 * - A zero window was announced from us - zero window probing 5704 * is only handled properly in the slow path. 5705 * - Out of order segments arrived. 5706 * - Urgent data is expected. 5707 * - There is no buffer space left 5708 * - Unexpected TCP flags/window values/header lengths are received 5709 * (detected by checking the TCP header against pred_flags) 5710 * - Data is sent in both directions. Fast path only supports pure senders 5711 * or pure receivers (this means either the sequence number or the ack 5712 * value must stay constant) 5713 * - Unexpected TCP option. 5714 * 5715 * When these conditions are not satisfied it drops into a standard 5716 * receive procedure patterned after RFC793 to handle all cases. 5717 * The first three cases are guaranteed by proper pred_flags setting, 5718 * the rest is checked inline. Fast processing is turned on in 5719 * tcp_data_queue when everything is OK. 5720 */ 5721 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb) 5722 { 5723 const struct tcphdr *th = (const struct tcphdr *)skb->data; 5724 struct tcp_sock *tp = tcp_sk(sk); 5725 unsigned int len = skb->len; 5726 5727 /* TCP congestion window tracking */ 5728 trace_tcp_probe(sk, skb); 5729 5730 tcp_mstamp_refresh(tp); 5731 if (unlikely(!sk->sk_rx_dst)) 5732 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb); 5733 /* 5734 * Header prediction. 5735 * The code loosely follows the one in the famous 5736 * "30 instruction TCP receive" Van Jacobson mail. 5737 * 5738 * Van's trick is to deposit buffers into socket queue 5739 * on a device interrupt, to call tcp_recv function 5740 * on the receive process context and checksum and copy 5741 * the buffer to user space. smart... 5742 * 5743 * Our current scheme is not silly either but we take the 5744 * extra cost of the net_bh soft interrupt processing... 5745 * We do checksum and copy also but from device to kernel. 5746 */ 5747 5748 tp->rx_opt.saw_tstamp = 0; 5749 5750 /* pred_flags is 0xS?10 << 16 + snd_wnd 5751 * if header_prediction is to be made 5752 * 'S' will always be tp->tcp_header_len >> 2 5753 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to 5754 * turn it off (when there are holes in the receive 5755 * space for instance) 5756 * PSH flag is ignored. 5757 */ 5758 5759 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags && 5760 TCP_SKB_CB(skb)->seq == tp->rcv_nxt && 5761 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) { 5762 int tcp_header_len = tp->tcp_header_len; 5763 5764 /* Timestamp header prediction: tcp_header_len 5765 * is automatically equal to th->doff*4 due to pred_flags 5766 * match. 5767 */ 5768 5769 /* Check timestamp */ 5770 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) { 5771 /* No? Slow path! */ 5772 if (!tcp_parse_aligned_timestamp(tp, th)) 5773 goto slow_path; 5774 5775 /* If PAWS failed, check it more carefully in slow path */ 5776 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0) 5777 goto slow_path; 5778 5779 /* DO NOT update ts_recent here, if checksum fails 5780 * and timestamp was corrupted part, it will result 5781 * in a hung connection since we will drop all 5782 * future packets due to the PAWS test. 5783 */ 5784 } 5785 5786 if (len <= tcp_header_len) { 5787 /* Bulk data transfer: sender */ 5788 if (len == tcp_header_len) { 5789 /* Predicted packet is in window by definition. 5790 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 5791 * Hence, check seq<=rcv_wup reduces to: 5792 */ 5793 if (tcp_header_len == 5794 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 5795 tp->rcv_nxt == tp->rcv_wup) 5796 tcp_store_ts_recent(tp); 5797 5798 /* We know that such packets are checksummed 5799 * on entry. 5800 */ 5801 tcp_ack(sk, skb, 0); 5802 __kfree_skb(skb); 5803 tcp_data_snd_check(sk); 5804 /* When receiving pure ack in fast path, update 5805 * last ts ecr directly instead of calling 5806 * tcp_rcv_rtt_measure_ts() 5807 */ 5808 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr; 5809 return; 5810 } else { /* Header too small */ 5811 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS); 5812 goto discard; 5813 } 5814 } else { 5815 int eaten = 0; 5816 bool fragstolen = false; 5817 5818 if (tcp_checksum_complete(skb)) 5819 goto csum_error; 5820 5821 if ((int)skb->truesize > sk->sk_forward_alloc) 5822 goto step5; 5823 5824 /* Predicted packet is in window by definition. 5825 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 5826 * Hence, check seq<=rcv_wup reduces to: 5827 */ 5828 if (tcp_header_len == 5829 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 5830 tp->rcv_nxt == tp->rcv_wup) 5831 tcp_store_ts_recent(tp); 5832 5833 tcp_rcv_rtt_measure_ts(sk, skb); 5834 5835 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS); 5836 5837 /* Bulk data transfer: receiver */ 5838 __skb_pull(skb, tcp_header_len); 5839 eaten = tcp_queue_rcv(sk, skb, &fragstolen); 5840 5841 tcp_event_data_recv(sk, skb); 5842 5843 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) { 5844 /* Well, only one small jumplet in fast path... */ 5845 tcp_ack(sk, skb, FLAG_DATA); 5846 tcp_data_snd_check(sk); 5847 if (!inet_csk_ack_scheduled(sk)) 5848 goto no_ack; 5849 } else { 5850 tcp_update_wl(tp, TCP_SKB_CB(skb)->seq); 5851 } 5852 5853 __tcp_ack_snd_check(sk, 0); 5854 no_ack: 5855 if (eaten) 5856 kfree_skb_partial(skb, fragstolen); 5857 tcp_data_ready(sk); 5858 return; 5859 } 5860 } 5861 5862 slow_path: 5863 if (len < (th->doff << 2) || tcp_checksum_complete(skb)) 5864 goto csum_error; 5865 5866 if (!th->ack && !th->rst && !th->syn) 5867 goto discard; 5868 5869 /* 5870 * Standard slow path. 5871 */ 5872 5873 if (!tcp_validate_incoming(sk, skb, th, 1)) 5874 return; 5875 5876 step5: 5877 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0) 5878 goto discard; 5879 5880 tcp_rcv_rtt_measure_ts(sk, skb); 5881 5882 /* Process urgent data. */ 5883 tcp_urg(sk, skb, th); 5884 5885 /* step 7: process the segment text */ 5886 tcp_data_queue(sk, skb); 5887 5888 tcp_data_snd_check(sk); 5889 tcp_ack_snd_check(sk); 5890 return; 5891 5892 csum_error: 5893 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS); 5894 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS); 5895 5896 discard: 5897 tcp_drop(sk, skb); 5898 } 5899 EXPORT_SYMBOL(tcp_rcv_established); 5900 5901 void tcp_init_transfer(struct sock *sk, int bpf_op, struct sk_buff *skb) 5902 { 5903 struct inet_connection_sock *icsk = inet_csk(sk); 5904 struct tcp_sock *tp = tcp_sk(sk); 5905 5906 tcp_mtup_init(sk); 5907 icsk->icsk_af_ops->rebuild_header(sk); 5908 tcp_init_metrics(sk); 5909 5910 /* Initialize the congestion window to start the transfer. 5911 * Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been 5912 * retransmitted. In light of RFC6298 more aggressive 1sec 5913 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK 5914 * retransmission has occurred. 5915 */ 5916 if (tp->total_retrans > 1 && tp->undo_marker) 5917 tp->snd_cwnd = 1; 5918 else 5919 tp->snd_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk)); 5920 tp->snd_cwnd_stamp = tcp_jiffies32; 5921 5922 icsk->icsk_ca_initialized = 0; 5923 bpf_skops_established(sk, bpf_op, skb); 5924 if (!icsk->icsk_ca_initialized) 5925 tcp_init_congestion_control(sk); 5926 tcp_init_buffer_space(sk); 5927 } 5928 5929 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb) 5930 { 5931 struct tcp_sock *tp = tcp_sk(sk); 5932 struct inet_connection_sock *icsk = inet_csk(sk); 5933 5934 tcp_set_state(sk, TCP_ESTABLISHED); 5935 icsk->icsk_ack.lrcvtime = tcp_jiffies32; 5936 5937 if (skb) { 5938 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb); 5939 security_inet_conn_established(sk, skb); 5940 sk_mark_napi_id(sk, skb); 5941 } 5942 5943 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB, skb); 5944 5945 /* Prevent spurious tcp_cwnd_restart() on first data 5946 * packet. 5947 */ 5948 tp->lsndtime = tcp_jiffies32; 5949 5950 if (sock_flag(sk, SOCK_KEEPOPEN)) 5951 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp)); 5952 5953 if (!tp->rx_opt.snd_wscale) 5954 __tcp_fast_path_on(tp, tp->snd_wnd); 5955 else 5956 tp->pred_flags = 0; 5957 } 5958 5959 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack, 5960 struct tcp_fastopen_cookie *cookie) 5961 { 5962 struct tcp_sock *tp = tcp_sk(sk); 5963 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL; 5964 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0; 5965 bool syn_drop = false; 5966 5967 if (mss == tp->rx_opt.user_mss) { 5968 struct tcp_options_received opt; 5969 5970 /* Get original SYNACK MSS value if user MSS sets mss_clamp */ 5971 tcp_clear_options(&opt); 5972 opt.user_mss = opt.mss_clamp = 0; 5973 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL); 5974 mss = opt.mss_clamp; 5975 } 5976 5977 if (!tp->syn_fastopen) { 5978 /* Ignore an unsolicited cookie */ 5979 cookie->len = -1; 5980 } else if (tp->total_retrans) { 5981 /* SYN timed out and the SYN-ACK neither has a cookie nor 5982 * acknowledges data. Presumably the remote received only 5983 * the retransmitted (regular) SYNs: either the original 5984 * SYN-data or the corresponding SYN-ACK was dropped. 5985 */ 5986 syn_drop = (cookie->len < 0 && data); 5987 } else if (cookie->len < 0 && !tp->syn_data) { 5988 /* We requested a cookie but didn't get it. If we did not use 5989 * the (old) exp opt format then try so next time (try_exp=1). 5990 * Otherwise we go back to use the RFC7413 opt (try_exp=2). 5991 */ 5992 try_exp = tp->syn_fastopen_exp ? 2 : 1; 5993 } 5994 5995 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp); 5996 5997 if (data) { /* Retransmit unacked data in SYN */ 5998 if (tp->total_retrans) 5999 tp->fastopen_client_fail = TFO_SYN_RETRANSMITTED; 6000 else 6001 tp->fastopen_client_fail = TFO_DATA_NOT_ACKED; 6002 skb_rbtree_walk_from(data) { 6003 if (__tcp_retransmit_skb(sk, data, 1)) 6004 break; 6005 } 6006 tcp_rearm_rto(sk); 6007 NET_INC_STATS(sock_net(sk), 6008 LINUX_MIB_TCPFASTOPENACTIVEFAIL); 6009 return true; 6010 } 6011 tp->syn_data_acked = tp->syn_data; 6012 if (tp->syn_data_acked) { 6013 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE); 6014 /* SYN-data is counted as two separate packets in tcp_ack() */ 6015 if (tp->delivered > 1) 6016 --tp->delivered; 6017 } 6018 6019 tcp_fastopen_add_skb(sk, synack); 6020 6021 return false; 6022 } 6023 6024 static void smc_check_reset_syn(struct tcp_sock *tp) 6025 { 6026 #if IS_ENABLED(CONFIG_SMC) 6027 if (static_branch_unlikely(&tcp_have_smc)) { 6028 if (tp->syn_smc && !tp->rx_opt.smc_ok) 6029 tp->syn_smc = 0; 6030 } 6031 #endif 6032 } 6033 6034 static void tcp_try_undo_spurious_syn(struct sock *sk) 6035 { 6036 struct tcp_sock *tp = tcp_sk(sk); 6037 u32 syn_stamp; 6038 6039 /* undo_marker is set when SYN or SYNACK times out. The timeout is 6040 * spurious if the ACK's timestamp option echo value matches the 6041 * original SYN timestamp. 6042 */ 6043 syn_stamp = tp->retrans_stamp; 6044 if (tp->undo_marker && syn_stamp && tp->rx_opt.saw_tstamp && 6045 syn_stamp == tp->rx_opt.rcv_tsecr) 6046 tp->undo_marker = 0; 6047 } 6048 6049 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb, 6050 const struct tcphdr *th) 6051 { 6052 struct inet_connection_sock *icsk = inet_csk(sk); 6053 struct tcp_sock *tp = tcp_sk(sk); 6054 struct tcp_fastopen_cookie foc = { .len = -1 }; 6055 int saved_clamp = tp->rx_opt.mss_clamp; 6056 bool fastopen_fail; 6057 6058 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc); 6059 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) 6060 tp->rx_opt.rcv_tsecr -= tp->tsoffset; 6061 6062 if (th->ack) { 6063 /* rfc793: 6064 * "If the state is SYN-SENT then 6065 * first check the ACK bit 6066 * If the ACK bit is set 6067 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send 6068 * a reset (unless the RST bit is set, if so drop 6069 * the segment and return)" 6070 */ 6071 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) || 6072 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) { 6073 /* Previous FIN/ACK or RST/ACK might be ignored. */ 6074 if (icsk->icsk_retransmits == 0) 6075 inet_csk_reset_xmit_timer(sk, 6076 ICSK_TIME_RETRANS, 6077 TCP_TIMEOUT_MIN, TCP_RTO_MAX); 6078 goto reset_and_undo; 6079 } 6080 6081 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 6082 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp, 6083 tcp_time_stamp(tp))) { 6084 NET_INC_STATS(sock_net(sk), 6085 LINUX_MIB_PAWSACTIVEREJECTED); 6086 goto reset_and_undo; 6087 } 6088 6089 /* Now ACK is acceptable. 6090 * 6091 * "If the RST bit is set 6092 * If the ACK was acceptable then signal the user "error: 6093 * connection reset", drop the segment, enter CLOSED state, 6094 * delete TCB, and return." 6095 */ 6096 6097 if (th->rst) { 6098 tcp_reset(sk, skb); 6099 goto discard; 6100 } 6101 6102 /* rfc793: 6103 * "fifth, if neither of the SYN or RST bits is set then 6104 * drop the segment and return." 6105 * 6106 * See note below! 6107 * --ANK(990513) 6108 */ 6109 if (!th->syn) 6110 goto discard_and_undo; 6111 6112 /* rfc793: 6113 * "If the SYN bit is on ... 6114 * are acceptable then ... 6115 * (our SYN has been ACKed), change the connection 6116 * state to ESTABLISHED..." 6117 */ 6118 6119 tcp_ecn_rcv_synack(tp, th); 6120 6121 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); 6122 tcp_try_undo_spurious_syn(sk); 6123 tcp_ack(sk, skb, FLAG_SLOWPATH); 6124 6125 /* Ok.. it's good. Set up sequence numbers and 6126 * move to established. 6127 */ 6128 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1); 6129 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 6130 6131 /* RFC1323: The window in SYN & SYN/ACK segments is 6132 * never scaled. 6133 */ 6134 tp->snd_wnd = ntohs(th->window); 6135 6136 if (!tp->rx_opt.wscale_ok) { 6137 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0; 6138 tp->window_clamp = min(tp->window_clamp, 65535U); 6139 } 6140 6141 if (tp->rx_opt.saw_tstamp) { 6142 tp->rx_opt.tstamp_ok = 1; 6143 tp->tcp_header_len = 6144 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 6145 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 6146 tcp_store_ts_recent(tp); 6147 } else { 6148 tp->tcp_header_len = sizeof(struct tcphdr); 6149 } 6150 6151 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 6152 tcp_initialize_rcv_mss(sk); 6153 6154 /* Remember, tcp_poll() does not lock socket! 6155 * Change state from SYN-SENT only after copied_seq 6156 * is initialized. */ 6157 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 6158 6159 smc_check_reset_syn(tp); 6160 6161 smp_mb(); 6162 6163 tcp_finish_connect(sk, skb); 6164 6165 fastopen_fail = (tp->syn_fastopen || tp->syn_data) && 6166 tcp_rcv_fastopen_synack(sk, skb, &foc); 6167 6168 if (!sock_flag(sk, SOCK_DEAD)) { 6169 sk->sk_state_change(sk); 6170 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT); 6171 } 6172 if (fastopen_fail) 6173 return -1; 6174 if (sk->sk_write_pending || 6175 icsk->icsk_accept_queue.rskq_defer_accept || 6176 inet_csk_in_pingpong_mode(sk)) { 6177 /* Save one ACK. Data will be ready after 6178 * several ticks, if write_pending is set. 6179 * 6180 * It may be deleted, but with this feature tcpdumps 6181 * look so _wonderfully_ clever, that I was not able 6182 * to stand against the temptation 8) --ANK 6183 */ 6184 inet_csk_schedule_ack(sk); 6185 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS); 6186 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 6187 TCP_DELACK_MAX, TCP_RTO_MAX); 6188 6189 discard: 6190 tcp_drop(sk, skb); 6191 return 0; 6192 } else { 6193 tcp_send_ack(sk); 6194 } 6195 return -1; 6196 } 6197 6198 /* No ACK in the segment */ 6199 6200 if (th->rst) { 6201 /* rfc793: 6202 * "If the RST bit is set 6203 * 6204 * Otherwise (no ACK) drop the segment and return." 6205 */ 6206 6207 goto discard_and_undo; 6208 } 6209 6210 /* PAWS check. */ 6211 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && 6212 tcp_paws_reject(&tp->rx_opt, 0)) 6213 goto discard_and_undo; 6214 6215 if (th->syn) { 6216 /* We see SYN without ACK. It is attempt of 6217 * simultaneous connect with crossed SYNs. 6218 * Particularly, it can be connect to self. 6219 */ 6220 tcp_set_state(sk, TCP_SYN_RECV); 6221 6222 if (tp->rx_opt.saw_tstamp) { 6223 tp->rx_opt.tstamp_ok = 1; 6224 tcp_store_ts_recent(tp); 6225 tp->tcp_header_len = 6226 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 6227 } else { 6228 tp->tcp_header_len = sizeof(struct tcphdr); 6229 } 6230 6231 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1); 6232 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 6233 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 6234 6235 /* RFC1323: The window in SYN & SYN/ACK segments is 6236 * never scaled. 6237 */ 6238 tp->snd_wnd = ntohs(th->window); 6239 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 6240 tp->max_window = tp->snd_wnd; 6241 6242 tcp_ecn_rcv_syn(tp, th); 6243 6244 tcp_mtup_init(sk); 6245 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 6246 tcp_initialize_rcv_mss(sk); 6247 6248 tcp_send_synack(sk); 6249 #if 0 6250 /* Note, we could accept data and URG from this segment. 6251 * There are no obstacles to make this (except that we must 6252 * either change tcp_recvmsg() to prevent it from returning data 6253 * before 3WHS completes per RFC793, or employ TCP Fast Open). 6254 * 6255 * However, if we ignore data in ACKless segments sometimes, 6256 * we have no reasons to accept it sometimes. 6257 * Also, seems the code doing it in step6 of tcp_rcv_state_process 6258 * is not flawless. So, discard packet for sanity. 6259 * Uncomment this return to process the data. 6260 */ 6261 return -1; 6262 #else 6263 goto discard; 6264 #endif 6265 } 6266 /* "fifth, if neither of the SYN or RST bits is set then 6267 * drop the segment and return." 6268 */ 6269 6270 discard_and_undo: 6271 tcp_clear_options(&tp->rx_opt); 6272 tp->rx_opt.mss_clamp = saved_clamp; 6273 goto discard; 6274 6275 reset_and_undo: 6276 tcp_clear_options(&tp->rx_opt); 6277 tp->rx_opt.mss_clamp = saved_clamp; 6278 return 1; 6279 } 6280 6281 static void tcp_rcv_synrecv_state_fastopen(struct sock *sk) 6282 { 6283 struct request_sock *req; 6284 6285 /* If we are still handling the SYNACK RTO, see if timestamp ECR allows 6286 * undo. If peer SACKs triggered fast recovery, we can't undo here. 6287 */ 6288 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss) 6289 tcp_try_undo_loss(sk, false); 6290 6291 /* Reset rtx states to prevent spurious retransmits_timed_out() */ 6292 tcp_sk(sk)->retrans_stamp = 0; 6293 inet_csk(sk)->icsk_retransmits = 0; 6294 6295 /* Once we leave TCP_SYN_RECV or TCP_FIN_WAIT_1, 6296 * we no longer need req so release it. 6297 */ 6298 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk, 6299 lockdep_sock_is_held(sk)); 6300 reqsk_fastopen_remove(sk, req, false); 6301 6302 /* Re-arm the timer because data may have been sent out. 6303 * This is similar to the regular data transmission case 6304 * when new data has just been ack'ed. 6305 * 6306 * (TFO) - we could try to be more aggressive and 6307 * retransmitting any data sooner based on when they 6308 * are sent out. 6309 */ 6310 tcp_rearm_rto(sk); 6311 } 6312 6313 /* 6314 * This function implements the receiving procedure of RFC 793 for 6315 * all states except ESTABLISHED and TIME_WAIT. 6316 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be 6317 * address independent. 6318 */ 6319 6320 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb) 6321 { 6322 struct tcp_sock *tp = tcp_sk(sk); 6323 struct inet_connection_sock *icsk = inet_csk(sk); 6324 const struct tcphdr *th = tcp_hdr(skb); 6325 struct request_sock *req; 6326 int queued = 0; 6327 bool acceptable; 6328 6329 switch (sk->sk_state) { 6330 case TCP_CLOSE: 6331 goto discard; 6332 6333 case TCP_LISTEN: 6334 if (th->ack) 6335 return 1; 6336 6337 if (th->rst) 6338 goto discard; 6339 6340 if (th->syn) { 6341 if (th->fin) 6342 goto discard; 6343 /* It is possible that we process SYN packets from backlog, 6344 * so we need to make sure to disable BH and RCU right there. 6345 */ 6346 rcu_read_lock(); 6347 local_bh_disable(); 6348 acceptable = icsk->icsk_af_ops->conn_request(sk, skb) >= 0; 6349 local_bh_enable(); 6350 rcu_read_unlock(); 6351 6352 if (!acceptable) 6353 return 1; 6354 consume_skb(skb); 6355 return 0; 6356 } 6357 goto discard; 6358 6359 case TCP_SYN_SENT: 6360 tp->rx_opt.saw_tstamp = 0; 6361 tcp_mstamp_refresh(tp); 6362 queued = tcp_rcv_synsent_state_process(sk, skb, th); 6363 if (queued >= 0) 6364 return queued; 6365 6366 /* Do step6 onward by hand. */ 6367 tcp_urg(sk, skb, th); 6368 __kfree_skb(skb); 6369 tcp_data_snd_check(sk); 6370 return 0; 6371 } 6372 6373 tcp_mstamp_refresh(tp); 6374 tp->rx_opt.saw_tstamp = 0; 6375 req = rcu_dereference_protected(tp->fastopen_rsk, 6376 lockdep_sock_is_held(sk)); 6377 if (req) { 6378 bool req_stolen; 6379 6380 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV && 6381 sk->sk_state != TCP_FIN_WAIT1); 6382 6383 if (!tcp_check_req(sk, skb, req, true, &req_stolen)) 6384 goto discard; 6385 } 6386 6387 if (!th->ack && !th->rst && !th->syn) 6388 goto discard; 6389 6390 if (!tcp_validate_incoming(sk, skb, th, 0)) 6391 return 0; 6392 6393 /* step 5: check the ACK field */ 6394 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH | 6395 FLAG_UPDATE_TS_RECENT | 6396 FLAG_NO_CHALLENGE_ACK) > 0; 6397 6398 if (!acceptable) { 6399 if (sk->sk_state == TCP_SYN_RECV) 6400 return 1; /* send one RST */ 6401 tcp_send_challenge_ack(sk, skb); 6402 goto discard; 6403 } 6404 switch (sk->sk_state) { 6405 case TCP_SYN_RECV: 6406 tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */ 6407 if (!tp->srtt_us) 6408 tcp_synack_rtt_meas(sk, req); 6409 6410 if (req) { 6411 tcp_rcv_synrecv_state_fastopen(sk); 6412 } else { 6413 tcp_try_undo_spurious_syn(sk); 6414 tp->retrans_stamp = 0; 6415 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB, 6416 skb); 6417 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 6418 } 6419 smp_mb(); 6420 tcp_set_state(sk, TCP_ESTABLISHED); 6421 sk->sk_state_change(sk); 6422 6423 /* Note, that this wakeup is only for marginal crossed SYN case. 6424 * Passively open sockets are not waked up, because 6425 * sk->sk_sleep == NULL and sk->sk_socket == NULL. 6426 */ 6427 if (sk->sk_socket) 6428 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT); 6429 6430 tp->snd_una = TCP_SKB_CB(skb)->ack_seq; 6431 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale; 6432 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); 6433 6434 if (tp->rx_opt.tstamp_ok) 6435 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 6436 6437 if (!inet_csk(sk)->icsk_ca_ops->cong_control) 6438 tcp_update_pacing_rate(sk); 6439 6440 /* Prevent spurious tcp_cwnd_restart() on first data packet */ 6441 tp->lsndtime = tcp_jiffies32; 6442 6443 tcp_initialize_rcv_mss(sk); 6444 tcp_fast_path_on(tp); 6445 break; 6446 6447 case TCP_FIN_WAIT1: { 6448 int tmo; 6449 6450 if (req) 6451 tcp_rcv_synrecv_state_fastopen(sk); 6452 6453 if (tp->snd_una != tp->write_seq) 6454 break; 6455 6456 tcp_set_state(sk, TCP_FIN_WAIT2); 6457 sk->sk_shutdown |= SEND_SHUTDOWN; 6458 6459 sk_dst_confirm(sk); 6460 6461 if (!sock_flag(sk, SOCK_DEAD)) { 6462 /* Wake up lingering close() */ 6463 sk->sk_state_change(sk); 6464 break; 6465 } 6466 6467 if (tp->linger2 < 0) { 6468 tcp_done(sk); 6469 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 6470 return 1; 6471 } 6472 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 6473 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { 6474 /* Receive out of order FIN after close() */ 6475 if (tp->syn_fastopen && th->fin) 6476 tcp_fastopen_active_disable(sk); 6477 tcp_done(sk); 6478 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 6479 return 1; 6480 } 6481 6482 tmo = tcp_fin_time(sk); 6483 if (tmo > TCP_TIMEWAIT_LEN) { 6484 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN); 6485 } else if (th->fin || sock_owned_by_user(sk)) { 6486 /* Bad case. We could lose such FIN otherwise. 6487 * It is not a big problem, but it looks confusing 6488 * and not so rare event. We still can lose it now, 6489 * if it spins in bh_lock_sock(), but it is really 6490 * marginal case. 6491 */ 6492 inet_csk_reset_keepalive_timer(sk, tmo); 6493 } else { 6494 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 6495 goto discard; 6496 } 6497 break; 6498 } 6499 6500 case TCP_CLOSING: 6501 if (tp->snd_una == tp->write_seq) { 6502 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 6503 goto discard; 6504 } 6505 break; 6506 6507 case TCP_LAST_ACK: 6508 if (tp->snd_una == tp->write_seq) { 6509 tcp_update_metrics(sk); 6510 tcp_done(sk); 6511 goto discard; 6512 } 6513 break; 6514 } 6515 6516 /* step 6: check the URG bit */ 6517 tcp_urg(sk, skb, th); 6518 6519 /* step 7: process the segment text */ 6520 switch (sk->sk_state) { 6521 case TCP_CLOSE_WAIT: 6522 case TCP_CLOSING: 6523 case TCP_LAST_ACK: 6524 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 6525 if (sk_is_mptcp(sk)) 6526 mptcp_incoming_options(sk, skb); 6527 break; 6528 } 6529 fallthrough; 6530 case TCP_FIN_WAIT1: 6531 case TCP_FIN_WAIT2: 6532 /* RFC 793 says to queue data in these states, 6533 * RFC 1122 says we MUST send a reset. 6534 * BSD 4.4 also does reset. 6535 */ 6536 if (sk->sk_shutdown & RCV_SHUTDOWN) { 6537 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 6538 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { 6539 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 6540 tcp_reset(sk, skb); 6541 return 1; 6542 } 6543 } 6544 fallthrough; 6545 case TCP_ESTABLISHED: 6546 tcp_data_queue(sk, skb); 6547 queued = 1; 6548 break; 6549 } 6550 6551 /* tcp_data could move socket to TIME-WAIT */ 6552 if (sk->sk_state != TCP_CLOSE) { 6553 tcp_data_snd_check(sk); 6554 tcp_ack_snd_check(sk); 6555 } 6556 6557 if (!queued) { 6558 discard: 6559 tcp_drop(sk, skb); 6560 } 6561 return 0; 6562 } 6563 EXPORT_SYMBOL(tcp_rcv_state_process); 6564 6565 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family) 6566 { 6567 struct inet_request_sock *ireq = inet_rsk(req); 6568 6569 if (family == AF_INET) 6570 net_dbg_ratelimited("drop open request from %pI4/%u\n", 6571 &ireq->ir_rmt_addr, port); 6572 #if IS_ENABLED(CONFIG_IPV6) 6573 else if (family == AF_INET6) 6574 net_dbg_ratelimited("drop open request from %pI6/%u\n", 6575 &ireq->ir_v6_rmt_addr, port); 6576 #endif 6577 } 6578 6579 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set 6580 * 6581 * If we receive a SYN packet with these bits set, it means a 6582 * network is playing bad games with TOS bits. In order to 6583 * avoid possible false congestion notifications, we disable 6584 * TCP ECN negotiation. 6585 * 6586 * Exception: tcp_ca wants ECN. This is required for DCTCP 6587 * congestion control: Linux DCTCP asserts ECT on all packets, 6588 * including SYN, which is most optimal solution; however, 6589 * others, such as FreeBSD do not. 6590 * 6591 * Exception: At least one of the reserved bits of the TCP header (th->res1) is 6592 * set, indicating the use of a future TCP extension (such as AccECN). See 6593 * RFC8311 §4.3 which updates RFC3168 to allow the development of such 6594 * extensions. 6595 */ 6596 static void tcp_ecn_create_request(struct request_sock *req, 6597 const struct sk_buff *skb, 6598 const struct sock *listen_sk, 6599 const struct dst_entry *dst) 6600 { 6601 const struct tcphdr *th = tcp_hdr(skb); 6602 const struct net *net = sock_net(listen_sk); 6603 bool th_ecn = th->ece && th->cwr; 6604 bool ect, ecn_ok; 6605 u32 ecn_ok_dst; 6606 6607 if (!th_ecn) 6608 return; 6609 6610 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield); 6611 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK); 6612 ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst; 6613 6614 if (((!ect || th->res1) && ecn_ok) || tcp_ca_needs_ecn(listen_sk) || 6615 (ecn_ok_dst & DST_FEATURE_ECN_CA) || 6616 tcp_bpf_ca_needs_ecn((struct sock *)req)) 6617 inet_rsk(req)->ecn_ok = 1; 6618 } 6619 6620 static void tcp_openreq_init(struct request_sock *req, 6621 const struct tcp_options_received *rx_opt, 6622 struct sk_buff *skb, const struct sock *sk) 6623 { 6624 struct inet_request_sock *ireq = inet_rsk(req); 6625 6626 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */ 6627 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq; 6628 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 6629 tcp_rsk(req)->snt_synack = 0; 6630 tcp_rsk(req)->last_oow_ack_time = 0; 6631 req->mss = rx_opt->mss_clamp; 6632 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0; 6633 ireq->tstamp_ok = rx_opt->tstamp_ok; 6634 ireq->sack_ok = rx_opt->sack_ok; 6635 ireq->snd_wscale = rx_opt->snd_wscale; 6636 ireq->wscale_ok = rx_opt->wscale_ok; 6637 ireq->acked = 0; 6638 ireq->ecn_ok = 0; 6639 ireq->ir_rmt_port = tcp_hdr(skb)->source; 6640 ireq->ir_num = ntohs(tcp_hdr(skb)->dest); 6641 ireq->ir_mark = inet_request_mark(sk, skb); 6642 #if IS_ENABLED(CONFIG_SMC) 6643 ireq->smc_ok = rx_opt->smc_ok; 6644 #endif 6645 } 6646 6647 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops, 6648 struct sock *sk_listener, 6649 bool attach_listener) 6650 { 6651 struct request_sock *req = reqsk_alloc(ops, sk_listener, 6652 attach_listener); 6653 6654 if (req) { 6655 struct inet_request_sock *ireq = inet_rsk(req); 6656 6657 ireq->ireq_opt = NULL; 6658 #if IS_ENABLED(CONFIG_IPV6) 6659 ireq->pktopts = NULL; 6660 #endif 6661 atomic64_set(&ireq->ir_cookie, 0); 6662 ireq->ireq_state = TCP_NEW_SYN_RECV; 6663 write_pnet(&ireq->ireq_net, sock_net(sk_listener)); 6664 ireq->ireq_family = sk_listener->sk_family; 6665 } 6666 6667 return req; 6668 } 6669 EXPORT_SYMBOL(inet_reqsk_alloc); 6670 6671 /* 6672 * Return true if a syncookie should be sent 6673 */ 6674 static bool tcp_syn_flood_action(const struct sock *sk, const char *proto) 6675 { 6676 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue; 6677 const char *msg = "Dropping request"; 6678 bool want_cookie = false; 6679 struct net *net = sock_net(sk); 6680 6681 #ifdef CONFIG_SYN_COOKIES 6682 if (net->ipv4.sysctl_tcp_syncookies) { 6683 msg = "Sending cookies"; 6684 want_cookie = true; 6685 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES); 6686 } else 6687 #endif 6688 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP); 6689 6690 if (!queue->synflood_warned && 6691 net->ipv4.sysctl_tcp_syncookies != 2 && 6692 xchg(&queue->synflood_warned, 1) == 0) 6693 net_info_ratelimited("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n", 6694 proto, sk->sk_num, msg); 6695 6696 return want_cookie; 6697 } 6698 6699 static void tcp_reqsk_record_syn(const struct sock *sk, 6700 struct request_sock *req, 6701 const struct sk_buff *skb) 6702 { 6703 if (tcp_sk(sk)->save_syn) { 6704 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb); 6705 struct saved_syn *saved_syn; 6706 u32 mac_hdrlen; 6707 void *base; 6708 6709 if (tcp_sk(sk)->save_syn == 2) { /* Save full header. */ 6710 base = skb_mac_header(skb); 6711 mac_hdrlen = skb_mac_header_len(skb); 6712 len += mac_hdrlen; 6713 } else { 6714 base = skb_network_header(skb); 6715 mac_hdrlen = 0; 6716 } 6717 6718 saved_syn = kmalloc(struct_size(saved_syn, data, len), 6719 GFP_ATOMIC); 6720 if (saved_syn) { 6721 saved_syn->mac_hdrlen = mac_hdrlen; 6722 saved_syn->network_hdrlen = skb_network_header_len(skb); 6723 saved_syn->tcp_hdrlen = tcp_hdrlen(skb); 6724 memcpy(saved_syn->data, base, len); 6725 req->saved_syn = saved_syn; 6726 } 6727 } 6728 } 6729 6730 /* If a SYN cookie is required and supported, returns a clamped MSS value to be 6731 * used for SYN cookie generation. 6732 */ 6733 u16 tcp_get_syncookie_mss(struct request_sock_ops *rsk_ops, 6734 const struct tcp_request_sock_ops *af_ops, 6735 struct sock *sk, struct tcphdr *th) 6736 { 6737 struct tcp_sock *tp = tcp_sk(sk); 6738 u16 mss; 6739 6740 if (sock_net(sk)->ipv4.sysctl_tcp_syncookies != 2 && 6741 !inet_csk_reqsk_queue_is_full(sk)) 6742 return 0; 6743 6744 if (!tcp_syn_flood_action(sk, rsk_ops->slab_name)) 6745 return 0; 6746 6747 if (sk_acceptq_is_full(sk)) { 6748 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS); 6749 return 0; 6750 } 6751 6752 mss = tcp_parse_mss_option(th, tp->rx_opt.user_mss); 6753 if (!mss) 6754 mss = af_ops->mss_clamp; 6755 6756 return mss; 6757 } 6758 EXPORT_SYMBOL_GPL(tcp_get_syncookie_mss); 6759 6760 int tcp_conn_request(struct request_sock_ops *rsk_ops, 6761 const struct tcp_request_sock_ops *af_ops, 6762 struct sock *sk, struct sk_buff *skb) 6763 { 6764 struct tcp_fastopen_cookie foc = { .len = -1 }; 6765 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn; 6766 struct tcp_options_received tmp_opt; 6767 struct tcp_sock *tp = tcp_sk(sk); 6768 struct net *net = sock_net(sk); 6769 struct sock *fastopen_sk = NULL; 6770 struct request_sock *req; 6771 bool want_cookie = false; 6772 struct dst_entry *dst; 6773 struct flowi fl; 6774 6775 /* TW buckets are converted to open requests without 6776 * limitations, they conserve resources and peer is 6777 * evidently real one. 6778 */ 6779 if ((net->ipv4.sysctl_tcp_syncookies == 2 || 6780 inet_csk_reqsk_queue_is_full(sk)) && !isn) { 6781 want_cookie = tcp_syn_flood_action(sk, rsk_ops->slab_name); 6782 if (!want_cookie) 6783 goto drop; 6784 } 6785 6786 if (sk_acceptq_is_full(sk)) { 6787 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS); 6788 goto drop; 6789 } 6790 6791 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie); 6792 if (!req) 6793 goto drop; 6794 6795 req->syncookie = want_cookie; 6796 tcp_rsk(req)->af_specific = af_ops; 6797 tcp_rsk(req)->ts_off = 0; 6798 #if IS_ENABLED(CONFIG_MPTCP) 6799 tcp_rsk(req)->is_mptcp = 0; 6800 #endif 6801 6802 tcp_clear_options(&tmp_opt); 6803 tmp_opt.mss_clamp = af_ops->mss_clamp; 6804 tmp_opt.user_mss = tp->rx_opt.user_mss; 6805 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0, 6806 want_cookie ? NULL : &foc); 6807 6808 if (want_cookie && !tmp_opt.saw_tstamp) 6809 tcp_clear_options(&tmp_opt); 6810 6811 if (IS_ENABLED(CONFIG_SMC) && want_cookie) 6812 tmp_opt.smc_ok = 0; 6813 6814 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp; 6815 tcp_openreq_init(req, &tmp_opt, skb, sk); 6816 inet_rsk(req)->no_srccheck = inet_sk(sk)->transparent; 6817 6818 /* Note: tcp_v6_init_req() might override ir_iif for link locals */ 6819 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb); 6820 6821 dst = af_ops->route_req(sk, skb, &fl, req); 6822 if (!dst) 6823 goto drop_and_free; 6824 6825 if (tmp_opt.tstamp_ok) 6826 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb); 6827 6828 if (!want_cookie && !isn) { 6829 /* Kill the following clause, if you dislike this way. */ 6830 if (!net->ipv4.sysctl_tcp_syncookies && 6831 (net->ipv4.sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) < 6832 (net->ipv4.sysctl_max_syn_backlog >> 2)) && 6833 !tcp_peer_is_proven(req, dst)) { 6834 /* Without syncookies last quarter of 6835 * backlog is filled with destinations, 6836 * proven to be alive. 6837 * It means that we continue to communicate 6838 * to destinations, already remembered 6839 * to the moment of synflood. 6840 */ 6841 pr_drop_req(req, ntohs(tcp_hdr(skb)->source), 6842 rsk_ops->family); 6843 goto drop_and_release; 6844 } 6845 6846 isn = af_ops->init_seq(skb); 6847 } 6848 6849 tcp_ecn_create_request(req, skb, sk, dst); 6850 6851 if (want_cookie) { 6852 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss); 6853 if (!tmp_opt.tstamp_ok) 6854 inet_rsk(req)->ecn_ok = 0; 6855 } 6856 6857 tcp_rsk(req)->snt_isn = isn; 6858 tcp_rsk(req)->txhash = net_tx_rndhash(); 6859 tcp_rsk(req)->syn_tos = TCP_SKB_CB(skb)->ip_dsfield; 6860 tcp_openreq_init_rwin(req, sk, dst); 6861 sk_rx_queue_set(req_to_sk(req), skb); 6862 if (!want_cookie) { 6863 tcp_reqsk_record_syn(sk, req, skb); 6864 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst); 6865 } 6866 if (fastopen_sk) { 6867 af_ops->send_synack(fastopen_sk, dst, &fl, req, 6868 &foc, TCP_SYNACK_FASTOPEN, skb); 6869 /* Add the child socket directly into the accept queue */ 6870 if (!inet_csk_reqsk_queue_add(sk, req, fastopen_sk)) { 6871 reqsk_fastopen_remove(fastopen_sk, req, false); 6872 bh_unlock_sock(fastopen_sk); 6873 sock_put(fastopen_sk); 6874 goto drop_and_free; 6875 } 6876 sk->sk_data_ready(sk); 6877 bh_unlock_sock(fastopen_sk); 6878 sock_put(fastopen_sk); 6879 } else { 6880 tcp_rsk(req)->tfo_listener = false; 6881 if (!want_cookie) 6882 inet_csk_reqsk_queue_hash_add(sk, req, 6883 tcp_timeout_init((struct sock *)req)); 6884 af_ops->send_synack(sk, dst, &fl, req, &foc, 6885 !want_cookie ? TCP_SYNACK_NORMAL : 6886 TCP_SYNACK_COOKIE, 6887 skb); 6888 if (want_cookie) { 6889 reqsk_free(req); 6890 return 0; 6891 } 6892 } 6893 reqsk_put(req); 6894 return 0; 6895 6896 drop_and_release: 6897 dst_release(dst); 6898 drop_and_free: 6899 __reqsk_free(req); 6900 drop: 6901 tcp_listendrop(sk); 6902 return 0; 6903 } 6904 EXPORT_SYMBOL(tcp_conn_request); 6905