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 #define FLAG_DSACK_TLP 0x20000 /* DSACK for tail loss probe */ 104 105 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) 106 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) 107 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK) 108 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) 109 110 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) 111 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH)) 112 113 #define REXMIT_NONE 0 /* no loss recovery to do */ 114 #define REXMIT_LOST 1 /* retransmit packets marked lost */ 115 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */ 116 117 #if IS_ENABLED(CONFIG_TLS_DEVICE) 118 static DEFINE_STATIC_KEY_DEFERRED_FALSE(clean_acked_data_enabled, HZ); 119 120 void clean_acked_data_enable(struct inet_connection_sock *icsk, 121 void (*cad)(struct sock *sk, u32 ack_seq)) 122 { 123 icsk->icsk_clean_acked = cad; 124 static_branch_deferred_inc(&clean_acked_data_enabled); 125 } 126 EXPORT_SYMBOL_GPL(clean_acked_data_enable); 127 128 void clean_acked_data_disable(struct inet_connection_sock *icsk) 129 { 130 static_branch_slow_dec_deferred(&clean_acked_data_enabled); 131 icsk->icsk_clean_acked = NULL; 132 } 133 EXPORT_SYMBOL_GPL(clean_acked_data_disable); 134 135 void clean_acked_data_flush(void) 136 { 137 static_key_deferred_flush(&clean_acked_data_enabled); 138 } 139 EXPORT_SYMBOL_GPL(clean_acked_data_flush); 140 #endif 141 142 #ifdef CONFIG_CGROUP_BPF 143 static void bpf_skops_parse_hdr(struct sock *sk, struct sk_buff *skb) 144 { 145 bool unknown_opt = tcp_sk(sk)->rx_opt.saw_unknown && 146 BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), 147 BPF_SOCK_OPS_PARSE_UNKNOWN_HDR_OPT_CB_FLAG); 148 bool parse_all_opt = BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), 149 BPF_SOCK_OPS_PARSE_ALL_HDR_OPT_CB_FLAG); 150 struct bpf_sock_ops_kern sock_ops; 151 152 if (likely(!unknown_opt && !parse_all_opt)) 153 return; 154 155 /* The skb will be handled in the 156 * bpf_skops_established() or 157 * bpf_skops_write_hdr_opt(). 158 */ 159 switch (sk->sk_state) { 160 case TCP_SYN_RECV: 161 case TCP_SYN_SENT: 162 case TCP_LISTEN: 163 return; 164 } 165 166 sock_owned_by_me(sk); 167 168 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp)); 169 sock_ops.op = BPF_SOCK_OPS_PARSE_HDR_OPT_CB; 170 sock_ops.is_fullsock = 1; 171 sock_ops.sk = sk; 172 bpf_skops_init_skb(&sock_ops, skb, tcp_hdrlen(skb)); 173 174 BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops); 175 } 176 177 static void bpf_skops_established(struct sock *sk, int bpf_op, 178 struct sk_buff *skb) 179 { 180 struct bpf_sock_ops_kern sock_ops; 181 182 sock_owned_by_me(sk); 183 184 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp)); 185 sock_ops.op = bpf_op; 186 sock_ops.is_fullsock = 1; 187 sock_ops.sk = sk; 188 /* sk with TCP_REPAIR_ON does not have skb in tcp_finish_connect */ 189 if (skb) 190 bpf_skops_init_skb(&sock_ops, skb, tcp_hdrlen(skb)); 191 192 BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops); 193 } 194 #else 195 static void bpf_skops_parse_hdr(struct sock *sk, struct sk_buff *skb) 196 { 197 } 198 199 static void bpf_skops_established(struct sock *sk, int bpf_op, 200 struct sk_buff *skb) 201 { 202 } 203 #endif 204 205 static void tcp_gro_dev_warn(struct sock *sk, const struct sk_buff *skb, 206 unsigned int len) 207 { 208 static bool __once __read_mostly; 209 210 if (!__once) { 211 struct net_device *dev; 212 213 __once = true; 214 215 rcu_read_lock(); 216 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif); 217 if (!dev || len >= dev->mtu) 218 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n", 219 dev ? dev->name : "Unknown driver"); 220 rcu_read_unlock(); 221 } 222 } 223 224 /* Adapt the MSS value used to make delayed ack decision to the 225 * real world. 226 */ 227 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb) 228 { 229 struct inet_connection_sock *icsk = inet_csk(sk); 230 const unsigned int lss = icsk->icsk_ack.last_seg_size; 231 unsigned int len; 232 233 icsk->icsk_ack.last_seg_size = 0; 234 235 /* skb->len may jitter because of SACKs, even if peer 236 * sends good full-sized frames. 237 */ 238 len = skb_shinfo(skb)->gso_size ? : skb->len; 239 if (len >= icsk->icsk_ack.rcv_mss) { 240 /* Note: divides are still a bit expensive. 241 * For the moment, only adjust scaling_ratio 242 * when we update icsk_ack.rcv_mss. 243 */ 244 if (unlikely(len != icsk->icsk_ack.rcv_mss)) { 245 u64 val = (u64)skb->len << TCP_RMEM_TO_WIN_SCALE; 246 247 do_div(val, skb->truesize); 248 tcp_sk(sk)->scaling_ratio = val ? val : 1; 249 } 250 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len, 251 tcp_sk(sk)->advmss); 252 /* Account for possibly-removed options */ 253 if (unlikely(len > icsk->icsk_ack.rcv_mss + 254 MAX_TCP_OPTION_SPACE)) 255 tcp_gro_dev_warn(sk, skb, len); 256 /* If the skb has a len of exactly 1*MSS and has the PSH bit 257 * set then it is likely the end of an application write. So 258 * more data may not be arriving soon, and yet the data sender 259 * may be waiting for an ACK if cwnd-bound or using TX zero 260 * copy. So we set ICSK_ACK_PUSHED here so that 261 * tcp_cleanup_rbuf() will send an ACK immediately if the app 262 * reads all of the data and is not ping-pong. If len > MSS 263 * then this logic does not matter (and does not hurt) because 264 * tcp_cleanup_rbuf() will always ACK immediately if the app 265 * reads data and there is more than an MSS of unACKed data. 266 */ 267 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_PSH) 268 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 269 } else { 270 /* Otherwise, we make more careful check taking into account, 271 * that SACKs block is variable. 272 * 273 * "len" is invariant segment length, including TCP header. 274 */ 275 len += skb->data - skb_transport_header(skb); 276 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) || 277 /* If PSH is not set, packet should be 278 * full sized, provided peer TCP is not badly broken. 279 * This observation (if it is correct 8)) allows 280 * to handle super-low mtu links fairly. 281 */ 282 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && 283 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) { 284 /* Subtract also invariant (if peer is RFC compliant), 285 * tcp header plus fixed timestamp option length. 286 * Resulting "len" is MSS free of SACK jitter. 287 */ 288 len -= tcp_sk(sk)->tcp_header_len; 289 icsk->icsk_ack.last_seg_size = len; 290 if (len == lss) { 291 icsk->icsk_ack.rcv_mss = len; 292 return; 293 } 294 } 295 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) 296 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; 297 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 298 } 299 } 300 301 static void tcp_incr_quickack(struct sock *sk, unsigned int max_quickacks) 302 { 303 struct inet_connection_sock *icsk = inet_csk(sk); 304 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); 305 306 if (quickacks == 0) 307 quickacks = 2; 308 quickacks = min(quickacks, max_quickacks); 309 if (quickacks > icsk->icsk_ack.quick) 310 icsk->icsk_ack.quick = quickacks; 311 } 312 313 static void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks) 314 { 315 struct inet_connection_sock *icsk = inet_csk(sk); 316 317 tcp_incr_quickack(sk, max_quickacks); 318 inet_csk_exit_pingpong_mode(sk); 319 icsk->icsk_ack.ato = TCP_ATO_MIN; 320 } 321 322 /* Send ACKs quickly, if "quick" count is not exhausted 323 * and the session is not interactive. 324 */ 325 326 static bool tcp_in_quickack_mode(struct sock *sk) 327 { 328 const struct inet_connection_sock *icsk = inet_csk(sk); 329 const struct dst_entry *dst = __sk_dst_get(sk); 330 331 return (dst && dst_metric(dst, RTAX_QUICKACK)) || 332 (icsk->icsk_ack.quick && !inet_csk_in_pingpong_mode(sk)); 333 } 334 335 static void tcp_ecn_queue_cwr(struct tcp_sock *tp) 336 { 337 if (tp->ecn_flags & TCP_ECN_OK) 338 tp->ecn_flags |= TCP_ECN_QUEUE_CWR; 339 } 340 341 static void tcp_ecn_accept_cwr(struct sock *sk, const struct sk_buff *skb) 342 { 343 if (tcp_hdr(skb)->cwr) { 344 tcp_sk(sk)->ecn_flags &= ~TCP_ECN_DEMAND_CWR; 345 346 /* If the sender is telling us it has entered CWR, then its 347 * cwnd may be very low (even just 1 packet), so we should ACK 348 * immediately. 349 */ 350 if (TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) 351 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW; 352 } 353 } 354 355 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp) 356 { 357 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 358 } 359 360 static void __tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb) 361 { 362 struct tcp_sock *tp = tcp_sk(sk); 363 364 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) { 365 case INET_ECN_NOT_ECT: 366 /* Funny extension: if ECT is not set on a segment, 367 * and we already seen ECT on a previous segment, 368 * it is probably a retransmit. 369 */ 370 if (tp->ecn_flags & TCP_ECN_SEEN) 371 tcp_enter_quickack_mode(sk, 2); 372 break; 373 case INET_ECN_CE: 374 if (tcp_ca_needs_ecn(sk)) 375 tcp_ca_event(sk, CA_EVENT_ECN_IS_CE); 376 377 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) { 378 /* Better not delay acks, sender can have a very low cwnd */ 379 tcp_enter_quickack_mode(sk, 2); 380 tp->ecn_flags |= TCP_ECN_DEMAND_CWR; 381 } 382 tp->ecn_flags |= TCP_ECN_SEEN; 383 break; 384 default: 385 if (tcp_ca_needs_ecn(sk)) 386 tcp_ca_event(sk, CA_EVENT_ECN_NO_CE); 387 tp->ecn_flags |= TCP_ECN_SEEN; 388 break; 389 } 390 } 391 392 static void tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb) 393 { 394 if (tcp_sk(sk)->ecn_flags & TCP_ECN_OK) 395 __tcp_ecn_check_ce(sk, skb); 396 } 397 398 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th) 399 { 400 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr)) 401 tp->ecn_flags &= ~TCP_ECN_OK; 402 } 403 404 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th) 405 { 406 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr)) 407 tp->ecn_flags &= ~TCP_ECN_OK; 408 } 409 410 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th) 411 { 412 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK)) 413 return true; 414 return false; 415 } 416 417 /* Buffer size and advertised window tuning. 418 * 419 * 1. Tuning sk->sk_sndbuf, when connection enters established state. 420 */ 421 422 static void tcp_sndbuf_expand(struct sock *sk) 423 { 424 const struct tcp_sock *tp = tcp_sk(sk); 425 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 426 int sndmem, per_mss; 427 u32 nr_segs; 428 429 /* Worst case is non GSO/TSO : each frame consumes one skb 430 * and skb->head is kmalloced using power of two area of memory 431 */ 432 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) + 433 MAX_TCP_HEADER + 434 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 435 436 per_mss = roundup_pow_of_two(per_mss) + 437 SKB_DATA_ALIGN(sizeof(struct sk_buff)); 438 439 nr_segs = max_t(u32, TCP_INIT_CWND, tcp_snd_cwnd(tp)); 440 nr_segs = max_t(u32, nr_segs, tp->reordering + 1); 441 442 /* Fast Recovery (RFC 5681 3.2) : 443 * Cubic needs 1.7 factor, rounded to 2 to include 444 * extra cushion (application might react slowly to EPOLLOUT) 445 */ 446 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2; 447 sndmem *= nr_segs * per_mss; 448 449 if (sk->sk_sndbuf < sndmem) 450 WRITE_ONCE(sk->sk_sndbuf, 451 min(sndmem, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[2]))); 452 } 453 454 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh) 455 * 456 * All tcp_full_space() is split to two parts: "network" buffer, allocated 457 * forward and advertised in receiver window (tp->rcv_wnd) and 458 * "application buffer", required to isolate scheduling/application 459 * latencies from network. 460 * window_clamp is maximal advertised window. It can be less than 461 * tcp_full_space(), in this case tcp_full_space() - window_clamp 462 * is reserved for "application" buffer. The less window_clamp is 463 * the smoother our behaviour from viewpoint of network, but the lower 464 * throughput and the higher sensitivity of the connection to losses. 8) 465 * 466 * rcv_ssthresh is more strict window_clamp used at "slow start" 467 * phase to predict further behaviour of this connection. 468 * It is used for two goals: 469 * - to enforce header prediction at sender, even when application 470 * requires some significant "application buffer". It is check #1. 471 * - to prevent pruning of receive queue because of misprediction 472 * of receiver window. Check #2. 473 * 474 * The scheme does not work when sender sends good segments opening 475 * window and then starts to feed us spaghetti. But it should work 476 * in common situations. Otherwise, we have to rely on queue collapsing. 477 */ 478 479 /* Slow part of check#2. */ 480 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb, 481 unsigned int skbtruesize) 482 { 483 const struct tcp_sock *tp = tcp_sk(sk); 484 /* Optimize this! */ 485 int truesize = tcp_win_from_space(sk, skbtruesize) >> 1; 486 int window = tcp_win_from_space(sk, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2])) >> 1; 487 488 while (tp->rcv_ssthresh <= window) { 489 if (truesize <= skb->len) 490 return 2 * inet_csk(sk)->icsk_ack.rcv_mss; 491 492 truesize >>= 1; 493 window >>= 1; 494 } 495 return 0; 496 } 497 498 /* Even if skb appears to have a bad len/truesize ratio, TCP coalescing 499 * can play nice with us, as sk_buff and skb->head might be either 500 * freed or shared with up to MAX_SKB_FRAGS segments. 501 * Only give a boost to drivers using page frag(s) to hold the frame(s), 502 * and if no payload was pulled in skb->head before reaching us. 503 */ 504 static u32 truesize_adjust(bool adjust, const struct sk_buff *skb) 505 { 506 u32 truesize = skb->truesize; 507 508 if (adjust && !skb_headlen(skb)) { 509 truesize -= SKB_TRUESIZE(skb_end_offset(skb)); 510 /* paranoid check, some drivers might be buggy */ 511 if (unlikely((int)truesize < (int)skb->len)) 512 truesize = skb->truesize; 513 } 514 return truesize; 515 } 516 517 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb, 518 bool adjust) 519 { 520 struct tcp_sock *tp = tcp_sk(sk); 521 int room; 522 523 room = min_t(int, tp->window_clamp, tcp_space(sk)) - tp->rcv_ssthresh; 524 525 if (room <= 0) 526 return; 527 528 /* Check #1 */ 529 if (!tcp_under_memory_pressure(sk)) { 530 unsigned int truesize = truesize_adjust(adjust, skb); 531 int incr; 532 533 /* Check #2. Increase window, if skb with such overhead 534 * will fit to rcvbuf in future. 535 */ 536 if (tcp_win_from_space(sk, truesize) <= skb->len) 537 incr = 2 * tp->advmss; 538 else 539 incr = __tcp_grow_window(sk, skb, truesize); 540 541 if (incr) { 542 incr = max_t(int, incr, 2 * skb->len); 543 tp->rcv_ssthresh += min(room, incr); 544 inet_csk(sk)->icsk_ack.quick |= 1; 545 } 546 } else { 547 /* Under pressure: 548 * Adjust rcv_ssthresh according to reserved mem 549 */ 550 tcp_adjust_rcv_ssthresh(sk); 551 } 552 } 553 554 /* 3. Try to fixup all. It is made immediately after connection enters 555 * established state. 556 */ 557 static void tcp_init_buffer_space(struct sock *sk) 558 { 559 int tcp_app_win = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_app_win); 560 struct tcp_sock *tp = tcp_sk(sk); 561 int maxwin; 562 563 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) 564 tcp_sndbuf_expand(sk); 565 566 tcp_mstamp_refresh(tp); 567 tp->rcvq_space.time = tp->tcp_mstamp; 568 tp->rcvq_space.seq = tp->copied_seq; 569 570 maxwin = tcp_full_space(sk); 571 572 if (tp->window_clamp >= maxwin) { 573 tp->window_clamp = maxwin; 574 575 if (tcp_app_win && maxwin > 4 * tp->advmss) 576 tp->window_clamp = max(maxwin - 577 (maxwin >> tcp_app_win), 578 4 * tp->advmss); 579 } 580 581 /* Force reservation of one segment. */ 582 if (tcp_app_win && 583 tp->window_clamp > 2 * tp->advmss && 584 tp->window_clamp + tp->advmss > maxwin) 585 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss); 586 587 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); 588 tp->snd_cwnd_stamp = tcp_jiffies32; 589 tp->rcvq_space.space = min3(tp->rcv_ssthresh, tp->rcv_wnd, 590 (u32)TCP_INIT_CWND * tp->advmss); 591 } 592 593 /* 4. Recalculate window clamp after socket hit its memory bounds. */ 594 static void tcp_clamp_window(struct sock *sk) 595 { 596 struct tcp_sock *tp = tcp_sk(sk); 597 struct inet_connection_sock *icsk = inet_csk(sk); 598 struct net *net = sock_net(sk); 599 int rmem2; 600 601 icsk->icsk_ack.quick = 0; 602 rmem2 = READ_ONCE(net->ipv4.sysctl_tcp_rmem[2]); 603 604 if (sk->sk_rcvbuf < rmem2 && 605 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && 606 !tcp_under_memory_pressure(sk) && 607 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) { 608 WRITE_ONCE(sk->sk_rcvbuf, 609 min(atomic_read(&sk->sk_rmem_alloc), rmem2)); 610 } 611 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) 612 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss); 613 } 614 615 /* Initialize RCV_MSS value. 616 * RCV_MSS is an our guess about MSS used by the peer. 617 * We haven't any direct information about the MSS. 618 * It's better to underestimate the RCV_MSS rather than overestimate. 619 * Overestimations make us ACKing less frequently than needed. 620 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). 621 */ 622 void tcp_initialize_rcv_mss(struct sock *sk) 623 { 624 const struct tcp_sock *tp = tcp_sk(sk); 625 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); 626 627 hint = min(hint, tp->rcv_wnd / 2); 628 hint = min(hint, TCP_MSS_DEFAULT); 629 hint = max(hint, TCP_MIN_MSS); 630 631 inet_csk(sk)->icsk_ack.rcv_mss = hint; 632 } 633 EXPORT_SYMBOL(tcp_initialize_rcv_mss); 634 635 /* Receiver "autotuning" code. 636 * 637 * The algorithm for RTT estimation w/o timestamps is based on 638 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. 639 * <https://public.lanl.gov/radiant/pubs.html#DRS> 640 * 641 * More detail on this code can be found at 642 * <http://staff.psc.edu/jheffner/>, 643 * though this reference is out of date. A new paper 644 * is pending. 645 */ 646 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) 647 { 648 u32 new_sample = tp->rcv_rtt_est.rtt_us; 649 long m = sample; 650 651 if (new_sample != 0) { 652 /* If we sample in larger samples in the non-timestamp 653 * case, we could grossly overestimate the RTT especially 654 * with chatty applications or bulk transfer apps which 655 * are stalled on filesystem I/O. 656 * 657 * Also, since we are only going for a minimum in the 658 * non-timestamp case, we do not smooth things out 659 * else with timestamps disabled convergence takes too 660 * long. 661 */ 662 if (!win_dep) { 663 m -= (new_sample >> 3); 664 new_sample += m; 665 } else { 666 m <<= 3; 667 if (m < new_sample) 668 new_sample = m; 669 } 670 } else { 671 /* No previous measure. */ 672 new_sample = m << 3; 673 } 674 675 tp->rcv_rtt_est.rtt_us = new_sample; 676 } 677 678 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) 679 { 680 u32 delta_us; 681 682 if (tp->rcv_rtt_est.time == 0) 683 goto new_measure; 684 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) 685 return; 686 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time); 687 if (!delta_us) 688 delta_us = 1; 689 tcp_rcv_rtt_update(tp, delta_us, 1); 690 691 new_measure: 692 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; 693 tp->rcv_rtt_est.time = tp->tcp_mstamp; 694 } 695 696 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, 697 const struct sk_buff *skb) 698 { 699 struct tcp_sock *tp = tcp_sk(sk); 700 701 if (tp->rx_opt.rcv_tsecr == tp->rcv_rtt_last_tsecr) 702 return; 703 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr; 704 705 if (TCP_SKB_CB(skb)->end_seq - 706 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) { 707 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr; 708 u32 delta_us; 709 710 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) { 711 if (!delta) 712 delta = 1; 713 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ); 714 tcp_rcv_rtt_update(tp, delta_us, 0); 715 } 716 } 717 } 718 719 /* 720 * This function should be called every time data is copied to user space. 721 * It calculates the appropriate TCP receive buffer space. 722 */ 723 void tcp_rcv_space_adjust(struct sock *sk) 724 { 725 struct tcp_sock *tp = tcp_sk(sk); 726 u32 copied; 727 int time; 728 729 trace_tcp_rcv_space_adjust(sk); 730 731 tcp_mstamp_refresh(tp); 732 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time); 733 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0) 734 return; 735 736 /* Number of bytes copied to user in last RTT */ 737 copied = tp->copied_seq - tp->rcvq_space.seq; 738 if (copied <= tp->rcvq_space.space) 739 goto new_measure; 740 741 /* A bit of theory : 742 * copied = bytes received in previous RTT, our base window 743 * To cope with packet losses, we need a 2x factor 744 * To cope with slow start, and sender growing its cwin by 100 % 745 * every RTT, we need a 4x factor, because the ACK we are sending 746 * now is for the next RTT, not the current one : 747 * <prev RTT . ><current RTT .. ><next RTT .... > 748 */ 749 750 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) && 751 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { 752 u64 rcvwin, grow; 753 int rcvbuf; 754 755 /* minimal window to cope with packet losses, assuming 756 * steady state. Add some cushion because of small variations. 757 */ 758 rcvwin = ((u64)copied << 1) + 16 * tp->advmss; 759 760 /* Accommodate for sender rate increase (eg. slow start) */ 761 grow = rcvwin * (copied - tp->rcvq_space.space); 762 do_div(grow, tp->rcvq_space.space); 763 rcvwin += (grow << 1); 764 765 rcvbuf = min_t(u64, tcp_space_from_win(sk, rcvwin), 766 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2])); 767 if (rcvbuf > sk->sk_rcvbuf) { 768 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf); 769 770 /* Make the window clamp follow along. */ 771 tp->window_clamp = tcp_win_from_space(sk, rcvbuf); 772 } 773 } 774 tp->rcvq_space.space = copied; 775 776 new_measure: 777 tp->rcvq_space.seq = tp->copied_seq; 778 tp->rcvq_space.time = tp->tcp_mstamp; 779 } 780 781 /* There is something which you must keep in mind when you analyze the 782 * behavior of the tp->ato delayed ack timeout interval. When a 783 * connection starts up, we want to ack as quickly as possible. The 784 * problem is that "good" TCP's do slow start at the beginning of data 785 * transmission. The means that until we send the first few ACK's the 786 * sender will sit on his end and only queue most of his data, because 787 * he can only send snd_cwnd unacked packets at any given time. For 788 * each ACK we send, he increments snd_cwnd and transmits more of his 789 * queue. -DaveM 790 */ 791 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb) 792 { 793 struct tcp_sock *tp = tcp_sk(sk); 794 struct inet_connection_sock *icsk = inet_csk(sk); 795 u32 now; 796 797 inet_csk_schedule_ack(sk); 798 799 tcp_measure_rcv_mss(sk, skb); 800 801 tcp_rcv_rtt_measure(tp); 802 803 now = tcp_jiffies32; 804 805 if (!icsk->icsk_ack.ato) { 806 /* The _first_ data packet received, initialize 807 * delayed ACK engine. 808 */ 809 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS); 810 icsk->icsk_ack.ato = TCP_ATO_MIN; 811 } else { 812 int m = now - icsk->icsk_ack.lrcvtime; 813 814 if (m <= TCP_ATO_MIN / 2) { 815 /* The fastest case is the first. */ 816 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; 817 } else if (m < icsk->icsk_ack.ato) { 818 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; 819 if (icsk->icsk_ack.ato > icsk->icsk_rto) 820 icsk->icsk_ack.ato = icsk->icsk_rto; 821 } else if (m > icsk->icsk_rto) { 822 /* Too long gap. Apparently sender failed to 823 * restart window, so that we send ACKs quickly. 824 */ 825 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS); 826 } 827 } 828 icsk->icsk_ack.lrcvtime = now; 829 830 tcp_ecn_check_ce(sk, skb); 831 832 if (skb->len >= 128) 833 tcp_grow_window(sk, skb, true); 834 } 835 836 /* Called to compute a smoothed rtt estimate. The data fed to this 837 * routine either comes from timestamps, or from segments that were 838 * known _not_ to have been retransmitted [see Karn/Partridge 839 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 840 * piece by Van Jacobson. 841 * NOTE: the next three routines used to be one big routine. 842 * To save cycles in the RFC 1323 implementation it was better to break 843 * it up into three procedures. -- erics 844 */ 845 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us) 846 { 847 struct tcp_sock *tp = tcp_sk(sk); 848 long m = mrtt_us; /* RTT */ 849 u32 srtt = tp->srtt_us; 850 851 /* The following amusing code comes from Jacobson's 852 * article in SIGCOMM '88. Note that rtt and mdev 853 * are scaled versions of rtt and mean deviation. 854 * This is designed to be as fast as possible 855 * m stands for "measurement". 856 * 857 * On a 1990 paper the rto value is changed to: 858 * RTO = rtt + 4 * mdev 859 * 860 * Funny. This algorithm seems to be very broken. 861 * These formulae increase RTO, when it should be decreased, increase 862 * too slowly, when it should be increased quickly, decrease too quickly 863 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely 864 * does not matter how to _calculate_ it. Seems, it was trap 865 * that VJ failed to avoid. 8) 866 */ 867 if (srtt != 0) { 868 m -= (srtt >> 3); /* m is now error in rtt est */ 869 srtt += m; /* rtt = 7/8 rtt + 1/8 new */ 870 if (m < 0) { 871 m = -m; /* m is now abs(error) */ 872 m -= (tp->mdev_us >> 2); /* similar update on mdev */ 873 /* This is similar to one of Eifel findings. 874 * Eifel blocks mdev updates when rtt decreases. 875 * This solution is a bit different: we use finer gain 876 * for mdev in this case (alpha*beta). 877 * Like Eifel it also prevents growth of rto, 878 * but also it limits too fast rto decreases, 879 * happening in pure Eifel. 880 */ 881 if (m > 0) 882 m >>= 3; 883 } else { 884 m -= (tp->mdev_us >> 2); /* similar update on mdev */ 885 } 886 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */ 887 if (tp->mdev_us > tp->mdev_max_us) { 888 tp->mdev_max_us = tp->mdev_us; 889 if (tp->mdev_max_us > tp->rttvar_us) 890 tp->rttvar_us = tp->mdev_max_us; 891 } 892 if (after(tp->snd_una, tp->rtt_seq)) { 893 if (tp->mdev_max_us < tp->rttvar_us) 894 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2; 895 tp->rtt_seq = tp->snd_nxt; 896 tp->mdev_max_us = tcp_rto_min_us(sk); 897 898 tcp_bpf_rtt(sk); 899 } 900 } else { 901 /* no previous measure. */ 902 srtt = m << 3; /* take the measured time to be rtt */ 903 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */ 904 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk)); 905 tp->mdev_max_us = tp->rttvar_us; 906 tp->rtt_seq = tp->snd_nxt; 907 908 tcp_bpf_rtt(sk); 909 } 910 tp->srtt_us = max(1U, srtt); 911 } 912 913 static void tcp_update_pacing_rate(struct sock *sk) 914 { 915 const struct tcp_sock *tp = tcp_sk(sk); 916 u64 rate; 917 918 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */ 919 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3); 920 921 /* current rate is (cwnd * mss) / srtt 922 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate. 923 * In Congestion Avoidance phase, set it to 120 % the current rate. 924 * 925 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh) 926 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching 927 * end of slow start and should slow down. 928 */ 929 if (tcp_snd_cwnd(tp) < tp->snd_ssthresh / 2) 930 rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio); 931 else 932 rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio); 933 934 rate *= max(tcp_snd_cwnd(tp), tp->packets_out); 935 936 if (likely(tp->srtt_us)) 937 do_div(rate, tp->srtt_us); 938 939 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate 940 * without any lock. We want to make sure compiler wont store 941 * intermediate values in this location. 942 */ 943 WRITE_ONCE(sk->sk_pacing_rate, min_t(u64, rate, 944 sk->sk_max_pacing_rate)); 945 } 946 947 /* Calculate rto without backoff. This is the second half of Van Jacobson's 948 * routine referred to above. 949 */ 950 static void tcp_set_rto(struct sock *sk) 951 { 952 const struct tcp_sock *tp = tcp_sk(sk); 953 /* Old crap is replaced with new one. 8) 954 * 955 * More seriously: 956 * 1. If rtt variance happened to be less 50msec, it is hallucination. 957 * It cannot be less due to utterly erratic ACK generation made 958 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ 959 * to do with delayed acks, because at cwnd>2 true delack timeout 960 * is invisible. Actually, Linux-2.4 also generates erratic 961 * ACKs in some circumstances. 962 */ 963 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp); 964 965 /* 2. Fixups made earlier cannot be right. 966 * If we do not estimate RTO correctly without them, 967 * all the algo is pure shit and should be replaced 968 * with correct one. It is exactly, which we pretend to do. 969 */ 970 971 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo 972 * guarantees that rto is higher. 973 */ 974 tcp_bound_rto(sk); 975 } 976 977 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst) 978 { 979 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); 980 981 if (!cwnd) 982 cwnd = TCP_INIT_CWND; 983 return min_t(__u32, cwnd, tp->snd_cwnd_clamp); 984 } 985 986 struct tcp_sacktag_state { 987 /* Timestamps for earliest and latest never-retransmitted segment 988 * that was SACKed. RTO needs the earliest RTT to stay conservative, 989 * but congestion control should still get an accurate delay signal. 990 */ 991 u64 first_sackt; 992 u64 last_sackt; 993 u32 reord; 994 u32 sack_delivered; 995 int flag; 996 unsigned int mss_now; 997 struct rate_sample *rate; 998 }; 999 1000 /* Take a notice that peer is sending D-SACKs. Skip update of data delivery 1001 * and spurious retransmission information if this DSACK is unlikely caused by 1002 * sender's action: 1003 * - DSACKed sequence range is larger than maximum receiver's window. 1004 * - Total no. of DSACKed segments exceed the total no. of retransmitted segs. 1005 */ 1006 static u32 tcp_dsack_seen(struct tcp_sock *tp, u32 start_seq, 1007 u32 end_seq, struct tcp_sacktag_state *state) 1008 { 1009 u32 seq_len, dup_segs = 1; 1010 1011 if (!before(start_seq, end_seq)) 1012 return 0; 1013 1014 seq_len = end_seq - start_seq; 1015 /* Dubious DSACK: DSACKed range greater than maximum advertised rwnd */ 1016 if (seq_len > tp->max_window) 1017 return 0; 1018 if (seq_len > tp->mss_cache) 1019 dup_segs = DIV_ROUND_UP(seq_len, tp->mss_cache); 1020 else if (tp->tlp_high_seq && tp->tlp_high_seq == end_seq) 1021 state->flag |= FLAG_DSACK_TLP; 1022 1023 tp->dsack_dups += dup_segs; 1024 /* Skip the DSACK if dup segs weren't retransmitted by sender */ 1025 if (tp->dsack_dups > tp->total_retrans) 1026 return 0; 1027 1028 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN; 1029 /* We increase the RACK ordering window in rounds where we receive 1030 * DSACKs that may have been due to reordering causing RACK to trigger 1031 * a spurious fast recovery. Thus RACK ignores DSACKs that happen 1032 * without having seen reordering, or that match TLP probes (TLP 1033 * is timer-driven, not triggered by RACK). 1034 */ 1035 if (tp->reord_seen && !(state->flag & FLAG_DSACK_TLP)) 1036 tp->rack.dsack_seen = 1; 1037 1038 state->flag |= FLAG_DSACKING_ACK; 1039 /* A spurious retransmission is delivered */ 1040 state->sack_delivered += dup_segs; 1041 1042 return dup_segs; 1043 } 1044 1045 /* It's reordering when higher sequence was delivered (i.e. sacked) before 1046 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering 1047 * distance is approximated in full-mss packet distance ("reordering"). 1048 */ 1049 static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq, 1050 const int ts) 1051 { 1052 struct tcp_sock *tp = tcp_sk(sk); 1053 const u32 mss = tp->mss_cache; 1054 u32 fack, metric; 1055 1056 fack = tcp_highest_sack_seq(tp); 1057 if (!before(low_seq, fack)) 1058 return; 1059 1060 metric = fack - low_seq; 1061 if ((metric > tp->reordering * mss) && mss) { 1062 #if FASTRETRANS_DEBUG > 1 1063 pr_debug("Disorder%d %d %u f%u s%u rr%d\n", 1064 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, 1065 tp->reordering, 1066 0, 1067 tp->sacked_out, 1068 tp->undo_marker ? tp->undo_retrans : 0); 1069 #endif 1070 tp->reordering = min_t(u32, (metric + mss - 1) / mss, 1071 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering)); 1072 } 1073 1074 /* This exciting event is worth to be remembered. 8) */ 1075 tp->reord_seen++; 1076 NET_INC_STATS(sock_net(sk), 1077 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER); 1078 } 1079 1080 /* This must be called before lost_out or retrans_out are updated 1081 * on a new loss, because we want to know if all skbs previously 1082 * known to be lost have already been retransmitted, indicating 1083 * that this newly lost skb is our next skb to retransmit. 1084 */ 1085 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb) 1086 { 1087 if ((!tp->retransmit_skb_hint && tp->retrans_out >= tp->lost_out) || 1088 (tp->retransmit_skb_hint && 1089 before(TCP_SKB_CB(skb)->seq, 1090 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))) 1091 tp->retransmit_skb_hint = skb; 1092 } 1093 1094 /* Sum the number of packets on the wire we have marked as lost, and 1095 * notify the congestion control module that the given skb was marked lost. 1096 */ 1097 static void tcp_notify_skb_loss_event(struct tcp_sock *tp, const struct sk_buff *skb) 1098 { 1099 tp->lost += tcp_skb_pcount(skb); 1100 } 1101 1102 void tcp_mark_skb_lost(struct sock *sk, struct sk_buff *skb) 1103 { 1104 __u8 sacked = TCP_SKB_CB(skb)->sacked; 1105 struct tcp_sock *tp = tcp_sk(sk); 1106 1107 if (sacked & TCPCB_SACKED_ACKED) 1108 return; 1109 1110 tcp_verify_retransmit_hint(tp, skb); 1111 if (sacked & TCPCB_LOST) { 1112 if (sacked & TCPCB_SACKED_RETRANS) { 1113 /* Account for retransmits that are lost again */ 1114 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1115 tp->retrans_out -= tcp_skb_pcount(skb); 1116 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT, 1117 tcp_skb_pcount(skb)); 1118 tcp_notify_skb_loss_event(tp, skb); 1119 } 1120 } else { 1121 tp->lost_out += tcp_skb_pcount(skb); 1122 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1123 tcp_notify_skb_loss_event(tp, skb); 1124 } 1125 } 1126 1127 /* Updates the delivered and delivered_ce counts */ 1128 static void tcp_count_delivered(struct tcp_sock *tp, u32 delivered, 1129 bool ece_ack) 1130 { 1131 tp->delivered += delivered; 1132 if (ece_ack) 1133 tp->delivered_ce += delivered; 1134 } 1135 1136 /* This procedure tags the retransmission queue when SACKs arrive. 1137 * 1138 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). 1139 * Packets in queue with these bits set are counted in variables 1140 * sacked_out, retrans_out and lost_out, correspondingly. 1141 * 1142 * Valid combinations are: 1143 * Tag InFlight Description 1144 * 0 1 - orig segment is in flight. 1145 * S 0 - nothing flies, orig reached receiver. 1146 * L 0 - nothing flies, orig lost by net. 1147 * R 2 - both orig and retransmit are in flight. 1148 * L|R 1 - orig is lost, retransmit is in flight. 1149 * S|R 1 - orig reached receiver, retrans is still in flight. 1150 * (L|S|R is logically valid, it could occur when L|R is sacked, 1151 * but it is equivalent to plain S and code short-curcuits it to S. 1152 * L|S is logically invalid, it would mean -1 packet in flight 8)) 1153 * 1154 * These 6 states form finite state machine, controlled by the following events: 1155 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) 1156 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) 1157 * 3. Loss detection event of two flavors: 1158 * A. Scoreboard estimator decided the packet is lost. 1159 * A'. Reno "three dupacks" marks head of queue lost. 1160 * B. SACK arrives sacking SND.NXT at the moment, when the 1161 * segment was retransmitted. 1162 * 4. D-SACK added new rule: D-SACK changes any tag to S. 1163 * 1164 * It is pleasant to note, that state diagram turns out to be commutative, 1165 * so that we are allowed not to be bothered by order of our actions, 1166 * when multiple events arrive simultaneously. (see the function below). 1167 * 1168 * Reordering detection. 1169 * -------------------- 1170 * Reordering metric is maximal distance, which a packet can be displaced 1171 * in packet stream. With SACKs we can estimate it: 1172 * 1173 * 1. SACK fills old hole and the corresponding segment was not 1174 * ever retransmitted -> reordering. Alas, we cannot use it 1175 * when segment was retransmitted. 1176 * 2. The last flaw is solved with D-SACK. D-SACK arrives 1177 * for retransmitted and already SACKed segment -> reordering.. 1178 * Both of these heuristics are not used in Loss state, when we cannot 1179 * account for retransmits accurately. 1180 * 1181 * SACK block validation. 1182 * ---------------------- 1183 * 1184 * SACK block range validation checks that the received SACK block fits to 1185 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT. 1186 * Note that SND.UNA is not included to the range though being valid because 1187 * it means that the receiver is rather inconsistent with itself reporting 1188 * SACK reneging when it should advance SND.UNA. Such SACK block this is 1189 * perfectly valid, however, in light of RFC2018 which explicitly states 1190 * that "SACK block MUST reflect the newest segment. Even if the newest 1191 * segment is going to be discarded ...", not that it looks very clever 1192 * in case of head skb. Due to potentional receiver driven attacks, we 1193 * choose to avoid immediate execution of a walk in write queue due to 1194 * reneging and defer head skb's loss recovery to standard loss recovery 1195 * procedure that will eventually trigger (nothing forbids us doing this). 1196 * 1197 * Implements also blockage to start_seq wrap-around. Problem lies in the 1198 * fact that though start_seq (s) is before end_seq (i.e., not reversed), 1199 * there's no guarantee that it will be before snd_nxt (n). The problem 1200 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt 1201 * wrap (s_w): 1202 * 1203 * <- outs wnd -> <- wrapzone -> 1204 * u e n u_w e_w s n_w 1205 * | | | | | | | 1206 * |<------------+------+----- TCP seqno space --------------+---------->| 1207 * ...-- <2^31 ->| |<--------... 1208 * ...---- >2^31 ------>| |<--------... 1209 * 1210 * Current code wouldn't be vulnerable but it's better still to discard such 1211 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat 1212 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in 1213 * snd_nxt wrap -> snd_una region will then become "well defined", i.e., 1214 * equal to the ideal case (infinite seqno space without wrap caused issues). 1215 * 1216 * With D-SACK the lower bound is extended to cover sequence space below 1217 * SND.UNA down to undo_marker, which is the last point of interest. Yet 1218 * again, D-SACK block must not to go across snd_una (for the same reason as 1219 * for the normal SACK blocks, explained above). But there all simplicity 1220 * ends, TCP might receive valid D-SACKs below that. As long as they reside 1221 * fully below undo_marker they do not affect behavior in anyway and can 1222 * therefore be safely ignored. In rare cases (which are more or less 1223 * theoretical ones), the D-SACK will nicely cross that boundary due to skb 1224 * fragmentation and packet reordering past skb's retransmission. To consider 1225 * them correctly, the acceptable range must be extended even more though 1226 * the exact amount is rather hard to quantify. However, tp->max_window can 1227 * be used as an exaggerated estimate. 1228 */ 1229 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack, 1230 u32 start_seq, u32 end_seq) 1231 { 1232 /* Too far in future, or reversed (interpretation is ambiguous) */ 1233 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq)) 1234 return false; 1235 1236 /* Nasty start_seq wrap-around check (see comments above) */ 1237 if (!before(start_seq, tp->snd_nxt)) 1238 return false; 1239 1240 /* In outstanding window? ...This is valid exit for D-SACKs too. 1241 * start_seq == snd_una is non-sensical (see comments above) 1242 */ 1243 if (after(start_seq, tp->snd_una)) 1244 return true; 1245 1246 if (!is_dsack || !tp->undo_marker) 1247 return false; 1248 1249 /* ...Then it's D-SACK, and must reside below snd_una completely */ 1250 if (after(end_seq, tp->snd_una)) 1251 return false; 1252 1253 if (!before(start_seq, tp->undo_marker)) 1254 return true; 1255 1256 /* Too old */ 1257 if (!after(end_seq, tp->undo_marker)) 1258 return false; 1259 1260 /* Undo_marker boundary crossing (overestimates a lot). Known already: 1261 * start_seq < undo_marker and end_seq >= undo_marker. 1262 */ 1263 return !before(start_seq, end_seq - tp->max_window); 1264 } 1265 1266 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb, 1267 struct tcp_sack_block_wire *sp, int num_sacks, 1268 u32 prior_snd_una, struct tcp_sacktag_state *state) 1269 { 1270 struct tcp_sock *tp = tcp_sk(sk); 1271 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq); 1272 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq); 1273 u32 dup_segs; 1274 1275 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) { 1276 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV); 1277 } else if (num_sacks > 1) { 1278 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq); 1279 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq); 1280 1281 if (after(end_seq_0, end_seq_1) || before(start_seq_0, start_seq_1)) 1282 return false; 1283 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKOFORECV); 1284 } else { 1285 return false; 1286 } 1287 1288 dup_segs = tcp_dsack_seen(tp, start_seq_0, end_seq_0, state); 1289 if (!dup_segs) { /* Skip dubious DSACK */ 1290 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKIGNOREDDUBIOUS); 1291 return false; 1292 } 1293 1294 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECVSEGS, dup_segs); 1295 1296 /* D-SACK for already forgotten data... Do dumb counting. */ 1297 if (tp->undo_marker && tp->undo_retrans > 0 && 1298 !after(end_seq_0, prior_snd_una) && 1299 after(end_seq_0, tp->undo_marker)) 1300 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - dup_segs); 1301 1302 return true; 1303 } 1304 1305 /* Check if skb is fully within the SACK block. In presence of GSO skbs, 1306 * the incoming SACK may not exactly match but we can find smaller MSS 1307 * aligned portion of it that matches. Therefore we might need to fragment 1308 * which may fail and creates some hassle (caller must handle error case 1309 * returns). 1310 * 1311 * FIXME: this could be merged to shift decision code 1312 */ 1313 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb, 1314 u32 start_seq, u32 end_seq) 1315 { 1316 int err; 1317 bool in_sack; 1318 unsigned int pkt_len; 1319 unsigned int mss; 1320 1321 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1322 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1323 1324 if (tcp_skb_pcount(skb) > 1 && !in_sack && 1325 after(TCP_SKB_CB(skb)->end_seq, start_seq)) { 1326 mss = tcp_skb_mss(skb); 1327 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); 1328 1329 if (!in_sack) { 1330 pkt_len = start_seq - TCP_SKB_CB(skb)->seq; 1331 if (pkt_len < mss) 1332 pkt_len = mss; 1333 } else { 1334 pkt_len = end_seq - TCP_SKB_CB(skb)->seq; 1335 if (pkt_len < mss) 1336 return -EINVAL; 1337 } 1338 1339 /* Round if necessary so that SACKs cover only full MSSes 1340 * and/or the remaining small portion (if present) 1341 */ 1342 if (pkt_len > mss) { 1343 unsigned int new_len = (pkt_len / mss) * mss; 1344 if (!in_sack && new_len < pkt_len) 1345 new_len += mss; 1346 pkt_len = new_len; 1347 } 1348 1349 if (pkt_len >= skb->len && !in_sack) 1350 return 0; 1351 1352 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, 1353 pkt_len, mss, GFP_ATOMIC); 1354 if (err < 0) 1355 return err; 1356 } 1357 1358 return in_sack; 1359 } 1360 1361 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */ 1362 static u8 tcp_sacktag_one(struct sock *sk, 1363 struct tcp_sacktag_state *state, u8 sacked, 1364 u32 start_seq, u32 end_seq, 1365 int dup_sack, int pcount, 1366 u64 xmit_time) 1367 { 1368 struct tcp_sock *tp = tcp_sk(sk); 1369 1370 /* Account D-SACK for retransmitted packet. */ 1371 if (dup_sack && (sacked & TCPCB_RETRANS)) { 1372 if (tp->undo_marker && tp->undo_retrans > 0 && 1373 after(end_seq, tp->undo_marker)) 1374 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - pcount); 1375 if ((sacked & TCPCB_SACKED_ACKED) && 1376 before(start_seq, state->reord)) 1377 state->reord = start_seq; 1378 } 1379 1380 /* Nothing to do; acked frame is about to be dropped (was ACKed). */ 1381 if (!after(end_seq, tp->snd_una)) 1382 return sacked; 1383 1384 if (!(sacked & TCPCB_SACKED_ACKED)) { 1385 tcp_rack_advance(tp, sacked, end_seq, xmit_time); 1386 1387 if (sacked & TCPCB_SACKED_RETRANS) { 1388 /* If the segment is not tagged as lost, 1389 * we do not clear RETRANS, believing 1390 * that retransmission is still in flight. 1391 */ 1392 if (sacked & TCPCB_LOST) { 1393 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); 1394 tp->lost_out -= pcount; 1395 tp->retrans_out -= pcount; 1396 } 1397 } else { 1398 if (!(sacked & TCPCB_RETRANS)) { 1399 /* New sack for not retransmitted frame, 1400 * which was in hole. It is reordering. 1401 */ 1402 if (before(start_seq, 1403 tcp_highest_sack_seq(tp)) && 1404 before(start_seq, state->reord)) 1405 state->reord = start_seq; 1406 1407 if (!after(end_seq, tp->high_seq)) 1408 state->flag |= FLAG_ORIG_SACK_ACKED; 1409 if (state->first_sackt == 0) 1410 state->first_sackt = xmit_time; 1411 state->last_sackt = xmit_time; 1412 } 1413 1414 if (sacked & TCPCB_LOST) { 1415 sacked &= ~TCPCB_LOST; 1416 tp->lost_out -= pcount; 1417 } 1418 } 1419 1420 sacked |= TCPCB_SACKED_ACKED; 1421 state->flag |= FLAG_DATA_SACKED; 1422 tp->sacked_out += pcount; 1423 /* Out-of-order packets delivered */ 1424 state->sack_delivered += pcount; 1425 1426 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */ 1427 if (tp->lost_skb_hint && 1428 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq)) 1429 tp->lost_cnt_hint += pcount; 1430 } 1431 1432 /* D-SACK. We can detect redundant retransmission in S|R and plain R 1433 * frames and clear it. undo_retrans is decreased above, L|R frames 1434 * are accounted above as well. 1435 */ 1436 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) { 1437 sacked &= ~TCPCB_SACKED_RETRANS; 1438 tp->retrans_out -= pcount; 1439 } 1440 1441 return sacked; 1442 } 1443 1444 /* Shift newly-SACKed bytes from this skb to the immediately previous 1445 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such. 1446 */ 1447 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev, 1448 struct sk_buff *skb, 1449 struct tcp_sacktag_state *state, 1450 unsigned int pcount, int shifted, int mss, 1451 bool dup_sack) 1452 { 1453 struct tcp_sock *tp = tcp_sk(sk); 1454 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */ 1455 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */ 1456 1457 BUG_ON(!pcount); 1458 1459 /* Adjust counters and hints for the newly sacked sequence 1460 * range but discard the return value since prev is already 1461 * marked. We must tag the range first because the seq 1462 * advancement below implicitly advances 1463 * tcp_highest_sack_seq() when skb is highest_sack. 1464 */ 1465 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked, 1466 start_seq, end_seq, dup_sack, pcount, 1467 tcp_skb_timestamp_us(skb)); 1468 tcp_rate_skb_delivered(sk, skb, state->rate); 1469 1470 if (skb == tp->lost_skb_hint) 1471 tp->lost_cnt_hint += pcount; 1472 1473 TCP_SKB_CB(prev)->end_seq += shifted; 1474 TCP_SKB_CB(skb)->seq += shifted; 1475 1476 tcp_skb_pcount_add(prev, pcount); 1477 WARN_ON_ONCE(tcp_skb_pcount(skb) < pcount); 1478 tcp_skb_pcount_add(skb, -pcount); 1479 1480 /* When we're adding to gso_segs == 1, gso_size will be zero, 1481 * in theory this shouldn't be necessary but as long as DSACK 1482 * code can come after this skb later on it's better to keep 1483 * setting gso_size to something. 1484 */ 1485 if (!TCP_SKB_CB(prev)->tcp_gso_size) 1486 TCP_SKB_CB(prev)->tcp_gso_size = mss; 1487 1488 /* CHECKME: To clear or not to clear? Mimics normal skb currently */ 1489 if (tcp_skb_pcount(skb) <= 1) 1490 TCP_SKB_CB(skb)->tcp_gso_size = 0; 1491 1492 /* Difference in this won't matter, both ACKed by the same cumul. ACK */ 1493 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS); 1494 1495 if (skb->len > 0) { 1496 BUG_ON(!tcp_skb_pcount(skb)); 1497 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED); 1498 return false; 1499 } 1500 1501 /* Whole SKB was eaten :-) */ 1502 1503 if (skb == tp->retransmit_skb_hint) 1504 tp->retransmit_skb_hint = prev; 1505 if (skb == tp->lost_skb_hint) { 1506 tp->lost_skb_hint = prev; 1507 tp->lost_cnt_hint -= tcp_skb_pcount(prev); 1508 } 1509 1510 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 1511 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor; 1512 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1513 TCP_SKB_CB(prev)->end_seq++; 1514 1515 if (skb == tcp_highest_sack(sk)) 1516 tcp_advance_highest_sack(sk, skb); 1517 1518 tcp_skb_collapse_tstamp(prev, skb); 1519 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp)) 1520 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0; 1521 1522 tcp_rtx_queue_unlink_and_free(skb, sk); 1523 1524 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED); 1525 1526 return true; 1527 } 1528 1529 /* I wish gso_size would have a bit more sane initialization than 1530 * something-or-zero which complicates things 1531 */ 1532 static int tcp_skb_seglen(const struct sk_buff *skb) 1533 { 1534 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb); 1535 } 1536 1537 /* Shifting pages past head area doesn't work */ 1538 static int skb_can_shift(const struct sk_buff *skb) 1539 { 1540 return !skb_headlen(skb) && skb_is_nonlinear(skb); 1541 } 1542 1543 int tcp_skb_shift(struct sk_buff *to, struct sk_buff *from, 1544 int pcount, int shiftlen) 1545 { 1546 /* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE) 1547 * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need 1548 * to make sure not storing more than 65535 * 8 bytes per skb, 1549 * even if current MSS is bigger. 1550 */ 1551 if (unlikely(to->len + shiftlen >= 65535 * TCP_MIN_GSO_SIZE)) 1552 return 0; 1553 if (unlikely(tcp_skb_pcount(to) + pcount > 65535)) 1554 return 0; 1555 return skb_shift(to, from, shiftlen); 1556 } 1557 1558 /* Try collapsing SACK blocks spanning across multiple skbs to a single 1559 * skb. 1560 */ 1561 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb, 1562 struct tcp_sacktag_state *state, 1563 u32 start_seq, u32 end_seq, 1564 bool dup_sack) 1565 { 1566 struct tcp_sock *tp = tcp_sk(sk); 1567 struct sk_buff *prev; 1568 int mss; 1569 int pcount = 0; 1570 int len; 1571 int in_sack; 1572 1573 /* Normally R but no L won't result in plain S */ 1574 if (!dup_sack && 1575 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS) 1576 goto fallback; 1577 if (!skb_can_shift(skb)) 1578 goto fallback; 1579 /* This frame is about to be dropped (was ACKed). */ 1580 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1581 goto fallback; 1582 1583 /* Can only happen with delayed DSACK + discard craziness */ 1584 prev = skb_rb_prev(skb); 1585 if (!prev) 1586 goto fallback; 1587 1588 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) 1589 goto fallback; 1590 1591 if (!tcp_skb_can_collapse(prev, skb)) 1592 goto fallback; 1593 1594 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1595 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1596 1597 if (in_sack) { 1598 len = skb->len; 1599 pcount = tcp_skb_pcount(skb); 1600 mss = tcp_skb_seglen(skb); 1601 1602 /* TODO: Fix DSACKs to not fragment already SACKed and we can 1603 * drop this restriction as unnecessary 1604 */ 1605 if (mss != tcp_skb_seglen(prev)) 1606 goto fallback; 1607 } else { 1608 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq)) 1609 goto noop; 1610 /* CHECKME: This is non-MSS split case only?, this will 1611 * cause skipped skbs due to advancing loop btw, original 1612 * has that feature too 1613 */ 1614 if (tcp_skb_pcount(skb) <= 1) 1615 goto noop; 1616 1617 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); 1618 if (!in_sack) { 1619 /* TODO: head merge to next could be attempted here 1620 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)), 1621 * though it might not be worth of the additional hassle 1622 * 1623 * ...we can probably just fallback to what was done 1624 * previously. We could try merging non-SACKed ones 1625 * as well but it probably isn't going to buy off 1626 * because later SACKs might again split them, and 1627 * it would make skb timestamp tracking considerably 1628 * harder problem. 1629 */ 1630 goto fallback; 1631 } 1632 1633 len = end_seq - TCP_SKB_CB(skb)->seq; 1634 BUG_ON(len < 0); 1635 BUG_ON(len > skb->len); 1636 1637 /* MSS boundaries should be honoured or else pcount will 1638 * severely break even though it makes things bit trickier. 1639 * Optimize common case to avoid most of the divides 1640 */ 1641 mss = tcp_skb_mss(skb); 1642 1643 /* TODO: Fix DSACKs to not fragment already SACKed and we can 1644 * drop this restriction as unnecessary 1645 */ 1646 if (mss != tcp_skb_seglen(prev)) 1647 goto fallback; 1648 1649 if (len == mss) { 1650 pcount = 1; 1651 } else if (len < mss) { 1652 goto noop; 1653 } else { 1654 pcount = len / mss; 1655 len = pcount * mss; 1656 } 1657 } 1658 1659 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */ 1660 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una)) 1661 goto fallback; 1662 1663 if (!tcp_skb_shift(prev, skb, pcount, len)) 1664 goto fallback; 1665 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack)) 1666 goto out; 1667 1668 /* Hole filled allows collapsing with the next as well, this is very 1669 * useful when hole on every nth skb pattern happens 1670 */ 1671 skb = skb_rb_next(prev); 1672 if (!skb) 1673 goto out; 1674 1675 if (!skb_can_shift(skb) || 1676 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) || 1677 (mss != tcp_skb_seglen(skb))) 1678 goto out; 1679 1680 if (!tcp_skb_can_collapse(prev, skb)) 1681 goto out; 1682 len = skb->len; 1683 pcount = tcp_skb_pcount(skb); 1684 if (tcp_skb_shift(prev, skb, pcount, len)) 1685 tcp_shifted_skb(sk, prev, skb, state, pcount, 1686 len, mss, 0); 1687 1688 out: 1689 return prev; 1690 1691 noop: 1692 return skb; 1693 1694 fallback: 1695 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK); 1696 return NULL; 1697 } 1698 1699 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk, 1700 struct tcp_sack_block *next_dup, 1701 struct tcp_sacktag_state *state, 1702 u32 start_seq, u32 end_seq, 1703 bool dup_sack_in) 1704 { 1705 struct tcp_sock *tp = tcp_sk(sk); 1706 struct sk_buff *tmp; 1707 1708 skb_rbtree_walk_from(skb) { 1709 int in_sack = 0; 1710 bool dup_sack = dup_sack_in; 1711 1712 /* queue is in-order => we can short-circuit the walk early */ 1713 if (!before(TCP_SKB_CB(skb)->seq, end_seq)) 1714 break; 1715 1716 if (next_dup && 1717 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) { 1718 in_sack = tcp_match_skb_to_sack(sk, skb, 1719 next_dup->start_seq, 1720 next_dup->end_seq); 1721 if (in_sack > 0) 1722 dup_sack = true; 1723 } 1724 1725 /* skb reference here is a bit tricky to get right, since 1726 * shifting can eat and free both this skb and the next, 1727 * so not even _safe variant of the loop is enough. 1728 */ 1729 if (in_sack <= 0) { 1730 tmp = tcp_shift_skb_data(sk, skb, state, 1731 start_seq, end_seq, dup_sack); 1732 if (tmp) { 1733 if (tmp != skb) { 1734 skb = tmp; 1735 continue; 1736 } 1737 1738 in_sack = 0; 1739 } else { 1740 in_sack = tcp_match_skb_to_sack(sk, skb, 1741 start_seq, 1742 end_seq); 1743 } 1744 } 1745 1746 if (unlikely(in_sack < 0)) 1747 break; 1748 1749 if (in_sack) { 1750 TCP_SKB_CB(skb)->sacked = 1751 tcp_sacktag_one(sk, 1752 state, 1753 TCP_SKB_CB(skb)->sacked, 1754 TCP_SKB_CB(skb)->seq, 1755 TCP_SKB_CB(skb)->end_seq, 1756 dup_sack, 1757 tcp_skb_pcount(skb), 1758 tcp_skb_timestamp_us(skb)); 1759 tcp_rate_skb_delivered(sk, skb, state->rate); 1760 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1761 list_del_init(&skb->tcp_tsorted_anchor); 1762 1763 if (!before(TCP_SKB_CB(skb)->seq, 1764 tcp_highest_sack_seq(tp))) 1765 tcp_advance_highest_sack(sk, skb); 1766 } 1767 } 1768 return skb; 1769 } 1770 1771 static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk, u32 seq) 1772 { 1773 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node; 1774 struct sk_buff *skb; 1775 1776 while (*p) { 1777 parent = *p; 1778 skb = rb_to_skb(parent); 1779 if (before(seq, TCP_SKB_CB(skb)->seq)) { 1780 p = &parent->rb_left; 1781 continue; 1782 } 1783 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) { 1784 p = &parent->rb_right; 1785 continue; 1786 } 1787 return skb; 1788 } 1789 return NULL; 1790 } 1791 1792 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk, 1793 u32 skip_to_seq) 1794 { 1795 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq)) 1796 return skb; 1797 1798 return tcp_sacktag_bsearch(sk, skip_to_seq); 1799 } 1800 1801 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb, 1802 struct sock *sk, 1803 struct tcp_sack_block *next_dup, 1804 struct tcp_sacktag_state *state, 1805 u32 skip_to_seq) 1806 { 1807 if (!next_dup) 1808 return skb; 1809 1810 if (before(next_dup->start_seq, skip_to_seq)) { 1811 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq); 1812 skb = tcp_sacktag_walk(skb, sk, NULL, state, 1813 next_dup->start_seq, next_dup->end_seq, 1814 1); 1815 } 1816 1817 return skb; 1818 } 1819 1820 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache) 1821 { 1822 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); 1823 } 1824 1825 static int 1826 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb, 1827 u32 prior_snd_una, struct tcp_sacktag_state *state) 1828 { 1829 struct tcp_sock *tp = tcp_sk(sk); 1830 const unsigned char *ptr = (skb_transport_header(ack_skb) + 1831 TCP_SKB_CB(ack_skb)->sacked); 1832 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2); 1833 struct tcp_sack_block sp[TCP_NUM_SACKS]; 1834 struct tcp_sack_block *cache; 1835 struct sk_buff *skb; 1836 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3); 1837 int used_sacks; 1838 bool found_dup_sack = false; 1839 int i, j; 1840 int first_sack_index; 1841 1842 state->flag = 0; 1843 state->reord = tp->snd_nxt; 1844 1845 if (!tp->sacked_out) 1846 tcp_highest_sack_reset(sk); 1847 1848 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire, 1849 num_sacks, prior_snd_una, state); 1850 1851 /* Eliminate too old ACKs, but take into 1852 * account more or less fresh ones, they can 1853 * contain valid SACK info. 1854 */ 1855 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) 1856 return 0; 1857 1858 if (!tp->packets_out) 1859 goto out; 1860 1861 used_sacks = 0; 1862 first_sack_index = 0; 1863 for (i = 0; i < num_sacks; i++) { 1864 bool dup_sack = !i && found_dup_sack; 1865 1866 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq); 1867 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq); 1868 1869 if (!tcp_is_sackblock_valid(tp, dup_sack, 1870 sp[used_sacks].start_seq, 1871 sp[used_sacks].end_seq)) { 1872 int mib_idx; 1873 1874 if (dup_sack) { 1875 if (!tp->undo_marker) 1876 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO; 1877 else 1878 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD; 1879 } else { 1880 /* Don't count olds caused by ACK reordering */ 1881 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) && 1882 !after(sp[used_sacks].end_seq, tp->snd_una)) 1883 continue; 1884 mib_idx = LINUX_MIB_TCPSACKDISCARD; 1885 } 1886 1887 NET_INC_STATS(sock_net(sk), mib_idx); 1888 if (i == 0) 1889 first_sack_index = -1; 1890 continue; 1891 } 1892 1893 /* Ignore very old stuff early */ 1894 if (!after(sp[used_sacks].end_seq, prior_snd_una)) { 1895 if (i == 0) 1896 first_sack_index = -1; 1897 continue; 1898 } 1899 1900 used_sacks++; 1901 } 1902 1903 /* order SACK blocks to allow in order walk of the retrans queue */ 1904 for (i = used_sacks - 1; i > 0; i--) { 1905 for (j = 0; j < i; j++) { 1906 if (after(sp[j].start_seq, sp[j + 1].start_seq)) { 1907 swap(sp[j], sp[j + 1]); 1908 1909 /* Track where the first SACK block goes to */ 1910 if (j == first_sack_index) 1911 first_sack_index = j + 1; 1912 } 1913 } 1914 } 1915 1916 state->mss_now = tcp_current_mss(sk); 1917 skb = NULL; 1918 i = 0; 1919 1920 if (!tp->sacked_out) { 1921 /* It's already past, so skip checking against it */ 1922 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); 1923 } else { 1924 cache = tp->recv_sack_cache; 1925 /* Skip empty blocks in at head of the cache */ 1926 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq && 1927 !cache->end_seq) 1928 cache++; 1929 } 1930 1931 while (i < used_sacks) { 1932 u32 start_seq = sp[i].start_seq; 1933 u32 end_seq = sp[i].end_seq; 1934 bool dup_sack = (found_dup_sack && (i == first_sack_index)); 1935 struct tcp_sack_block *next_dup = NULL; 1936 1937 if (found_dup_sack && ((i + 1) == first_sack_index)) 1938 next_dup = &sp[i + 1]; 1939 1940 /* Skip too early cached blocks */ 1941 while (tcp_sack_cache_ok(tp, cache) && 1942 !before(start_seq, cache->end_seq)) 1943 cache++; 1944 1945 /* Can skip some work by looking recv_sack_cache? */ 1946 if (tcp_sack_cache_ok(tp, cache) && !dup_sack && 1947 after(end_seq, cache->start_seq)) { 1948 1949 /* Head todo? */ 1950 if (before(start_seq, cache->start_seq)) { 1951 skb = tcp_sacktag_skip(skb, sk, start_seq); 1952 skb = tcp_sacktag_walk(skb, sk, next_dup, 1953 state, 1954 start_seq, 1955 cache->start_seq, 1956 dup_sack); 1957 } 1958 1959 /* Rest of the block already fully processed? */ 1960 if (!after(end_seq, cache->end_seq)) 1961 goto advance_sp; 1962 1963 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup, 1964 state, 1965 cache->end_seq); 1966 1967 /* ...tail remains todo... */ 1968 if (tcp_highest_sack_seq(tp) == cache->end_seq) { 1969 /* ...but better entrypoint exists! */ 1970 skb = tcp_highest_sack(sk); 1971 if (!skb) 1972 break; 1973 cache++; 1974 goto walk; 1975 } 1976 1977 skb = tcp_sacktag_skip(skb, sk, cache->end_seq); 1978 /* Check overlap against next cached too (past this one already) */ 1979 cache++; 1980 continue; 1981 } 1982 1983 if (!before(start_seq, tcp_highest_sack_seq(tp))) { 1984 skb = tcp_highest_sack(sk); 1985 if (!skb) 1986 break; 1987 } 1988 skb = tcp_sacktag_skip(skb, sk, start_seq); 1989 1990 walk: 1991 skb = tcp_sacktag_walk(skb, sk, next_dup, state, 1992 start_seq, end_seq, dup_sack); 1993 1994 advance_sp: 1995 i++; 1996 } 1997 1998 /* Clear the head of the cache sack blocks so we can skip it next time */ 1999 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) { 2000 tp->recv_sack_cache[i].start_seq = 0; 2001 tp->recv_sack_cache[i].end_seq = 0; 2002 } 2003 for (j = 0; j < used_sacks; j++) 2004 tp->recv_sack_cache[i++] = sp[j]; 2005 2006 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker) 2007 tcp_check_sack_reordering(sk, state->reord, 0); 2008 2009 tcp_verify_left_out(tp); 2010 out: 2011 2012 #if FASTRETRANS_DEBUG > 0 2013 WARN_ON((int)tp->sacked_out < 0); 2014 WARN_ON((int)tp->lost_out < 0); 2015 WARN_ON((int)tp->retrans_out < 0); 2016 WARN_ON((int)tcp_packets_in_flight(tp) < 0); 2017 #endif 2018 return state->flag; 2019 } 2020 2021 /* Limits sacked_out so that sum with lost_out isn't ever larger than 2022 * packets_out. Returns false if sacked_out adjustement wasn't necessary. 2023 */ 2024 static bool tcp_limit_reno_sacked(struct tcp_sock *tp) 2025 { 2026 u32 holes; 2027 2028 holes = max(tp->lost_out, 1U); 2029 holes = min(holes, tp->packets_out); 2030 2031 if ((tp->sacked_out + holes) > tp->packets_out) { 2032 tp->sacked_out = tp->packets_out - holes; 2033 return true; 2034 } 2035 return false; 2036 } 2037 2038 /* If we receive more dupacks than we expected counting segments 2039 * in assumption of absent reordering, interpret this as reordering. 2040 * The only another reason could be bug in receiver TCP. 2041 */ 2042 static void tcp_check_reno_reordering(struct sock *sk, const int addend) 2043 { 2044 struct tcp_sock *tp = tcp_sk(sk); 2045 2046 if (!tcp_limit_reno_sacked(tp)) 2047 return; 2048 2049 tp->reordering = min_t(u32, tp->packets_out + addend, 2050 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering)); 2051 tp->reord_seen++; 2052 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER); 2053 } 2054 2055 /* Emulate SACKs for SACKless connection: account for a new dupack. */ 2056 2057 static void tcp_add_reno_sack(struct sock *sk, int num_dupack, bool ece_ack) 2058 { 2059 if (num_dupack) { 2060 struct tcp_sock *tp = tcp_sk(sk); 2061 u32 prior_sacked = tp->sacked_out; 2062 s32 delivered; 2063 2064 tp->sacked_out += num_dupack; 2065 tcp_check_reno_reordering(sk, 0); 2066 delivered = tp->sacked_out - prior_sacked; 2067 if (delivered > 0) 2068 tcp_count_delivered(tp, delivered, ece_ack); 2069 tcp_verify_left_out(tp); 2070 } 2071 } 2072 2073 /* Account for ACK, ACKing some data in Reno Recovery phase. */ 2074 2075 static void tcp_remove_reno_sacks(struct sock *sk, int acked, bool ece_ack) 2076 { 2077 struct tcp_sock *tp = tcp_sk(sk); 2078 2079 if (acked > 0) { 2080 /* One ACK acked hole. The rest eat duplicate ACKs. */ 2081 tcp_count_delivered(tp, max_t(int, acked - tp->sacked_out, 1), 2082 ece_ack); 2083 if (acked - 1 >= tp->sacked_out) 2084 tp->sacked_out = 0; 2085 else 2086 tp->sacked_out -= acked - 1; 2087 } 2088 tcp_check_reno_reordering(sk, acked); 2089 tcp_verify_left_out(tp); 2090 } 2091 2092 static inline void tcp_reset_reno_sack(struct tcp_sock *tp) 2093 { 2094 tp->sacked_out = 0; 2095 } 2096 2097 void tcp_clear_retrans(struct tcp_sock *tp) 2098 { 2099 tp->retrans_out = 0; 2100 tp->lost_out = 0; 2101 tp->undo_marker = 0; 2102 tp->undo_retrans = -1; 2103 tp->sacked_out = 0; 2104 } 2105 2106 static inline void tcp_init_undo(struct tcp_sock *tp) 2107 { 2108 tp->undo_marker = tp->snd_una; 2109 2110 /* Retransmission still in flight may cause DSACKs later. */ 2111 /* First, account for regular retransmits in flight: */ 2112 tp->undo_retrans = tp->retrans_out; 2113 /* Next, account for TLP retransmits in flight: */ 2114 if (tp->tlp_high_seq && tp->tlp_retrans) 2115 tp->undo_retrans++; 2116 /* Finally, avoid 0, because undo_retrans==0 means "can undo now": */ 2117 if (!tp->undo_retrans) 2118 tp->undo_retrans = -1; 2119 } 2120 2121 static bool tcp_is_rack(const struct sock *sk) 2122 { 2123 return READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_recovery) & 2124 TCP_RACK_LOSS_DETECTION; 2125 } 2126 2127 /* If we detect SACK reneging, forget all SACK information 2128 * and reset tags completely, otherwise preserve SACKs. If receiver 2129 * dropped its ofo queue, we will know this due to reneging detection. 2130 */ 2131 static void tcp_timeout_mark_lost(struct sock *sk) 2132 { 2133 struct tcp_sock *tp = tcp_sk(sk); 2134 struct sk_buff *skb, *head; 2135 bool is_reneg; /* is receiver reneging on SACKs? */ 2136 2137 head = tcp_rtx_queue_head(sk); 2138 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED); 2139 if (is_reneg) { 2140 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING); 2141 tp->sacked_out = 0; 2142 /* Mark SACK reneging until we recover from this loss event. */ 2143 tp->is_sack_reneg = 1; 2144 } else if (tcp_is_reno(tp)) { 2145 tcp_reset_reno_sack(tp); 2146 } 2147 2148 skb = head; 2149 skb_rbtree_walk_from(skb) { 2150 if (is_reneg) 2151 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED; 2152 else if (tcp_is_rack(sk) && skb != head && 2153 tcp_rack_skb_timeout(tp, skb, 0) > 0) 2154 continue; /* Don't mark recently sent ones lost yet */ 2155 tcp_mark_skb_lost(sk, skb); 2156 } 2157 tcp_verify_left_out(tp); 2158 tcp_clear_all_retrans_hints(tp); 2159 } 2160 2161 /* Enter Loss state. */ 2162 void tcp_enter_loss(struct sock *sk) 2163 { 2164 const struct inet_connection_sock *icsk = inet_csk(sk); 2165 struct tcp_sock *tp = tcp_sk(sk); 2166 struct net *net = sock_net(sk); 2167 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery; 2168 u8 reordering; 2169 2170 tcp_timeout_mark_lost(sk); 2171 2172 /* Reduce ssthresh if it has not yet been made inside this window. */ 2173 if (icsk->icsk_ca_state <= TCP_CA_Disorder || 2174 !after(tp->high_seq, tp->snd_una) || 2175 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) { 2176 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2177 tp->prior_cwnd = tcp_snd_cwnd(tp); 2178 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 2179 tcp_ca_event(sk, CA_EVENT_LOSS); 2180 tcp_init_undo(tp); 2181 } 2182 tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + 1); 2183 tp->snd_cwnd_cnt = 0; 2184 tp->snd_cwnd_stamp = tcp_jiffies32; 2185 2186 /* Timeout in disordered state after receiving substantial DUPACKs 2187 * suggests that the degree of reordering is over-estimated. 2188 */ 2189 reordering = READ_ONCE(net->ipv4.sysctl_tcp_reordering); 2190 if (icsk->icsk_ca_state <= TCP_CA_Disorder && 2191 tp->sacked_out >= reordering) 2192 tp->reordering = min_t(unsigned int, tp->reordering, 2193 reordering); 2194 2195 tcp_set_ca_state(sk, TCP_CA_Loss); 2196 tp->high_seq = tp->snd_nxt; 2197 tp->tlp_high_seq = 0; 2198 tcp_ecn_queue_cwr(tp); 2199 2200 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous 2201 * loss recovery is underway except recurring timeout(s) on 2202 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing 2203 */ 2204 tp->frto = READ_ONCE(net->ipv4.sysctl_tcp_frto) && 2205 (new_recovery || icsk->icsk_retransmits) && 2206 !inet_csk(sk)->icsk_mtup.probe_size; 2207 } 2208 2209 /* If ACK arrived pointing to a remembered SACK, it means that our 2210 * remembered SACKs do not reflect real state of receiver i.e. 2211 * receiver _host_ is heavily congested (or buggy). 2212 * 2213 * To avoid big spurious retransmission bursts due to transient SACK 2214 * scoreboard oddities that look like reneging, we give the receiver a 2215 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will 2216 * restore sanity to the SACK scoreboard. If the apparent reneging 2217 * persists until this RTO then we'll clear the SACK scoreboard. 2218 */ 2219 static bool tcp_check_sack_reneging(struct sock *sk, int *ack_flag) 2220 { 2221 if (*ack_flag & FLAG_SACK_RENEGING && 2222 *ack_flag & FLAG_SND_UNA_ADVANCED) { 2223 struct tcp_sock *tp = tcp_sk(sk); 2224 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4), 2225 msecs_to_jiffies(10)); 2226 2227 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 2228 delay, TCP_RTO_MAX); 2229 *ack_flag &= ~FLAG_SET_XMIT_TIMER; 2230 return true; 2231 } 2232 return false; 2233 } 2234 2235 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs 2236 * counter when SACK is enabled (without SACK, sacked_out is used for 2237 * that purpose). 2238 * 2239 * With reordering, holes may still be in flight, so RFC3517 recovery 2240 * uses pure sacked_out (total number of SACKed segments) even though 2241 * it violates the RFC that uses duplicate ACKs, often these are equal 2242 * but when e.g. out-of-window ACKs or packet duplication occurs, 2243 * they differ. Since neither occurs due to loss, TCP should really 2244 * ignore them. 2245 */ 2246 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp) 2247 { 2248 return tp->sacked_out + 1; 2249 } 2250 2251 /* Linux NewReno/SACK/ECN state machine. 2252 * -------------------------------------- 2253 * 2254 * "Open" Normal state, no dubious events, fast path. 2255 * "Disorder" In all the respects it is "Open", 2256 * but requires a bit more attention. It is entered when 2257 * we see some SACKs or dupacks. It is split of "Open" 2258 * mainly to move some processing from fast path to slow one. 2259 * "CWR" CWND was reduced due to some Congestion Notification event. 2260 * It can be ECN, ICMP source quench, local device congestion. 2261 * "Recovery" CWND was reduced, we are fast-retransmitting. 2262 * "Loss" CWND was reduced due to RTO timeout or SACK reneging. 2263 * 2264 * tcp_fastretrans_alert() is entered: 2265 * - each incoming ACK, if state is not "Open" 2266 * - when arrived ACK is unusual, namely: 2267 * * SACK 2268 * * Duplicate ACK. 2269 * * ECN ECE. 2270 * 2271 * Counting packets in flight is pretty simple. 2272 * 2273 * in_flight = packets_out - left_out + retrans_out 2274 * 2275 * packets_out is SND.NXT-SND.UNA counted in packets. 2276 * 2277 * retrans_out is number of retransmitted segments. 2278 * 2279 * left_out is number of segments left network, but not ACKed yet. 2280 * 2281 * left_out = sacked_out + lost_out 2282 * 2283 * sacked_out: Packets, which arrived to receiver out of order 2284 * and hence not ACKed. With SACKs this number is simply 2285 * amount of SACKed data. Even without SACKs 2286 * it is easy to give pretty reliable estimate of this number, 2287 * counting duplicate ACKs. 2288 * 2289 * lost_out: Packets lost by network. TCP has no explicit 2290 * "loss notification" feedback from network (for now). 2291 * It means that this number can be only _guessed_. 2292 * Actually, it is the heuristics to predict lossage that 2293 * distinguishes different algorithms. 2294 * 2295 * F.e. after RTO, when all the queue is considered as lost, 2296 * lost_out = packets_out and in_flight = retrans_out. 2297 * 2298 * Essentially, we have now a few algorithms detecting 2299 * lost packets. 2300 * 2301 * If the receiver supports SACK: 2302 * 2303 * RFC6675/3517: It is the conventional algorithm. A packet is 2304 * considered lost if the number of higher sequence packets 2305 * SACKed is greater than or equal the DUPACK thoreshold 2306 * (reordering). This is implemented in tcp_mark_head_lost and 2307 * tcp_update_scoreboard. 2308 * 2309 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm 2310 * (2017-) that checks timing instead of counting DUPACKs. 2311 * Essentially a packet is considered lost if it's not S/ACKed 2312 * after RTT + reordering_window, where both metrics are 2313 * dynamically measured and adjusted. This is implemented in 2314 * tcp_rack_mark_lost. 2315 * 2316 * If the receiver does not support SACK: 2317 * 2318 * NewReno (RFC6582): in Recovery we assume that one segment 2319 * is lost (classic Reno). While we are in Recovery and 2320 * a partial ACK arrives, we assume that one more packet 2321 * is lost (NewReno). This heuristics are the same in NewReno 2322 * and SACK. 2323 * 2324 * Really tricky (and requiring careful tuning) part of algorithm 2325 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue(). 2326 * The first determines the moment _when_ we should reduce CWND and, 2327 * hence, slow down forward transmission. In fact, it determines the moment 2328 * when we decide that hole is caused by loss, rather than by a reorder. 2329 * 2330 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill 2331 * holes, caused by lost packets. 2332 * 2333 * And the most logically complicated part of algorithm is undo 2334 * heuristics. We detect false retransmits due to both too early 2335 * fast retransmit (reordering) and underestimated RTO, analyzing 2336 * timestamps and D-SACKs. When we detect that some segments were 2337 * retransmitted by mistake and CWND reduction was wrong, we undo 2338 * window reduction and abort recovery phase. This logic is hidden 2339 * inside several functions named tcp_try_undo_<something>. 2340 */ 2341 2342 /* This function decides, when we should leave Disordered state 2343 * and enter Recovery phase, reducing congestion window. 2344 * 2345 * Main question: may we further continue forward transmission 2346 * with the same cwnd? 2347 */ 2348 static bool tcp_time_to_recover(struct sock *sk, int flag) 2349 { 2350 struct tcp_sock *tp = tcp_sk(sk); 2351 2352 /* Trick#1: The loss is proven. */ 2353 if (tp->lost_out) 2354 return true; 2355 2356 /* Not-A-Trick#2 : Classic rule... */ 2357 if (!tcp_is_rack(sk) && tcp_dupack_heuristics(tp) > tp->reordering) 2358 return true; 2359 2360 return false; 2361 } 2362 2363 /* Detect loss in event "A" above by marking head of queue up as lost. 2364 * For RFC3517 SACK, a segment is considered lost if it 2365 * has at least tp->reordering SACKed seqments above it; "packets" refers to 2366 * the maximum SACKed segments to pass before reaching this limit. 2367 */ 2368 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head) 2369 { 2370 struct tcp_sock *tp = tcp_sk(sk); 2371 struct sk_buff *skb; 2372 int cnt; 2373 /* Use SACK to deduce losses of new sequences sent during recovery */ 2374 const u32 loss_high = tp->snd_nxt; 2375 2376 WARN_ON(packets > tp->packets_out); 2377 skb = tp->lost_skb_hint; 2378 if (skb) { 2379 /* Head already handled? */ 2380 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una)) 2381 return; 2382 cnt = tp->lost_cnt_hint; 2383 } else { 2384 skb = tcp_rtx_queue_head(sk); 2385 cnt = 0; 2386 } 2387 2388 skb_rbtree_walk_from(skb) { 2389 /* TODO: do this better */ 2390 /* this is not the most efficient way to do this... */ 2391 tp->lost_skb_hint = skb; 2392 tp->lost_cnt_hint = cnt; 2393 2394 if (after(TCP_SKB_CB(skb)->end_seq, loss_high)) 2395 break; 2396 2397 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2398 cnt += tcp_skb_pcount(skb); 2399 2400 if (cnt > packets) 2401 break; 2402 2403 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_LOST)) 2404 tcp_mark_skb_lost(sk, skb); 2405 2406 if (mark_head) 2407 break; 2408 } 2409 tcp_verify_left_out(tp); 2410 } 2411 2412 /* Account newly detected lost packet(s) */ 2413 2414 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit) 2415 { 2416 struct tcp_sock *tp = tcp_sk(sk); 2417 2418 if (tcp_is_sack(tp)) { 2419 int sacked_upto = tp->sacked_out - tp->reordering; 2420 if (sacked_upto >= 0) 2421 tcp_mark_head_lost(sk, sacked_upto, 0); 2422 else if (fast_rexmit) 2423 tcp_mark_head_lost(sk, 1, 1); 2424 } 2425 } 2426 2427 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when) 2428 { 2429 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 2430 before(tp->rx_opt.rcv_tsecr, when); 2431 } 2432 2433 /* skb is spurious retransmitted if the returned timestamp echo 2434 * reply is prior to the skb transmission time 2435 */ 2436 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp, 2437 const struct sk_buff *skb) 2438 { 2439 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) && 2440 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb)); 2441 } 2442 2443 /* Nothing was retransmitted or returned timestamp is less 2444 * than timestamp of the first retransmission. 2445 */ 2446 static inline bool tcp_packet_delayed(const struct tcp_sock *tp) 2447 { 2448 return tp->retrans_stamp && 2449 tcp_tsopt_ecr_before(tp, tp->retrans_stamp); 2450 } 2451 2452 /* Undo procedures. */ 2453 2454 /* We can clear retrans_stamp when there are no retransmissions in the 2455 * window. It would seem that it is trivially available for us in 2456 * tp->retrans_out, however, that kind of assumptions doesn't consider 2457 * what will happen if errors occur when sending retransmission for the 2458 * second time. ...It could the that such segment has only 2459 * TCPCB_EVER_RETRANS set at the present time. It seems that checking 2460 * the head skb is enough except for some reneging corner cases that 2461 * are not worth the effort. 2462 * 2463 * Main reason for all this complexity is the fact that connection dying 2464 * time now depends on the validity of the retrans_stamp, in particular, 2465 * that successive retransmissions of a segment must not advance 2466 * retrans_stamp under any conditions. 2467 */ 2468 static bool tcp_any_retrans_done(const struct sock *sk) 2469 { 2470 const struct tcp_sock *tp = tcp_sk(sk); 2471 struct sk_buff *skb; 2472 2473 if (tp->retrans_out) 2474 return true; 2475 2476 skb = tcp_rtx_queue_head(sk); 2477 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS)) 2478 return true; 2479 2480 return false; 2481 } 2482 2483 static void DBGUNDO(struct sock *sk, const char *msg) 2484 { 2485 #if FASTRETRANS_DEBUG > 1 2486 struct tcp_sock *tp = tcp_sk(sk); 2487 struct inet_sock *inet = inet_sk(sk); 2488 2489 if (sk->sk_family == AF_INET) { 2490 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n", 2491 msg, 2492 &inet->inet_daddr, ntohs(inet->inet_dport), 2493 tcp_snd_cwnd(tp), tcp_left_out(tp), 2494 tp->snd_ssthresh, tp->prior_ssthresh, 2495 tp->packets_out); 2496 } 2497 #if IS_ENABLED(CONFIG_IPV6) 2498 else if (sk->sk_family == AF_INET6) { 2499 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n", 2500 msg, 2501 &sk->sk_v6_daddr, ntohs(inet->inet_dport), 2502 tcp_snd_cwnd(tp), tcp_left_out(tp), 2503 tp->snd_ssthresh, tp->prior_ssthresh, 2504 tp->packets_out); 2505 } 2506 #endif 2507 #endif 2508 } 2509 2510 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss) 2511 { 2512 struct tcp_sock *tp = tcp_sk(sk); 2513 2514 if (unmark_loss) { 2515 struct sk_buff *skb; 2516 2517 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) { 2518 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 2519 } 2520 tp->lost_out = 0; 2521 tcp_clear_all_retrans_hints(tp); 2522 } 2523 2524 if (tp->prior_ssthresh) { 2525 const struct inet_connection_sock *icsk = inet_csk(sk); 2526 2527 tcp_snd_cwnd_set(tp, icsk->icsk_ca_ops->undo_cwnd(sk)); 2528 2529 if (tp->prior_ssthresh > tp->snd_ssthresh) { 2530 tp->snd_ssthresh = tp->prior_ssthresh; 2531 tcp_ecn_withdraw_cwr(tp); 2532 } 2533 } 2534 tp->snd_cwnd_stamp = tcp_jiffies32; 2535 tp->undo_marker = 0; 2536 tp->rack.advanced = 1; /* Force RACK to re-exam losses */ 2537 } 2538 2539 static inline bool tcp_may_undo(const struct tcp_sock *tp) 2540 { 2541 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp)); 2542 } 2543 2544 static bool tcp_is_non_sack_preventing_reopen(struct sock *sk) 2545 { 2546 struct tcp_sock *tp = tcp_sk(sk); 2547 2548 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) { 2549 /* Hold old state until something *above* high_seq 2550 * is ACKed. For Reno it is MUST to prevent false 2551 * fast retransmits (RFC2582). SACK TCP is safe. */ 2552 if (!tcp_any_retrans_done(sk)) 2553 tp->retrans_stamp = 0; 2554 return true; 2555 } 2556 return false; 2557 } 2558 2559 /* People celebrate: "We love our President!" */ 2560 static bool tcp_try_undo_recovery(struct sock *sk) 2561 { 2562 struct tcp_sock *tp = tcp_sk(sk); 2563 2564 if (tcp_may_undo(tp)) { 2565 int mib_idx; 2566 2567 /* Happy end! We did not retransmit anything 2568 * or our original transmission succeeded. 2569 */ 2570 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans"); 2571 tcp_undo_cwnd_reduction(sk, false); 2572 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss) 2573 mib_idx = LINUX_MIB_TCPLOSSUNDO; 2574 else 2575 mib_idx = LINUX_MIB_TCPFULLUNDO; 2576 2577 NET_INC_STATS(sock_net(sk), mib_idx); 2578 } else if (tp->rack.reo_wnd_persist) { 2579 tp->rack.reo_wnd_persist--; 2580 } 2581 if (tcp_is_non_sack_preventing_reopen(sk)) 2582 return true; 2583 tcp_set_ca_state(sk, TCP_CA_Open); 2584 tp->is_sack_reneg = 0; 2585 return false; 2586 } 2587 2588 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */ 2589 static bool tcp_try_undo_dsack(struct sock *sk) 2590 { 2591 struct tcp_sock *tp = tcp_sk(sk); 2592 2593 if (tp->undo_marker && !tp->undo_retrans) { 2594 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH, 2595 tp->rack.reo_wnd_persist + 1); 2596 DBGUNDO(sk, "D-SACK"); 2597 tcp_undo_cwnd_reduction(sk, false); 2598 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO); 2599 return true; 2600 } 2601 return false; 2602 } 2603 2604 /* Undo during loss recovery after partial ACK or using F-RTO. */ 2605 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo) 2606 { 2607 struct tcp_sock *tp = tcp_sk(sk); 2608 2609 if (frto_undo || tcp_may_undo(tp)) { 2610 tcp_undo_cwnd_reduction(sk, true); 2611 2612 DBGUNDO(sk, "partial loss"); 2613 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO); 2614 if (frto_undo) 2615 NET_INC_STATS(sock_net(sk), 2616 LINUX_MIB_TCPSPURIOUSRTOS); 2617 inet_csk(sk)->icsk_retransmits = 0; 2618 if (tcp_is_non_sack_preventing_reopen(sk)) 2619 return true; 2620 if (frto_undo || tcp_is_sack(tp)) { 2621 tcp_set_ca_state(sk, TCP_CA_Open); 2622 tp->is_sack_reneg = 0; 2623 } 2624 return true; 2625 } 2626 return false; 2627 } 2628 2629 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937. 2630 * It computes the number of packets to send (sndcnt) based on packets newly 2631 * delivered: 2632 * 1) If the packets in flight is larger than ssthresh, PRR spreads the 2633 * cwnd reductions across a full RTT. 2634 * 2) Otherwise PRR uses packet conservation to send as much as delivered. 2635 * But when SND_UNA is acked without further losses, 2636 * slow starts cwnd up to ssthresh to speed up the recovery. 2637 */ 2638 static void tcp_init_cwnd_reduction(struct sock *sk) 2639 { 2640 struct tcp_sock *tp = tcp_sk(sk); 2641 2642 tp->high_seq = tp->snd_nxt; 2643 tp->tlp_high_seq = 0; 2644 tp->snd_cwnd_cnt = 0; 2645 tp->prior_cwnd = tcp_snd_cwnd(tp); 2646 tp->prr_delivered = 0; 2647 tp->prr_out = 0; 2648 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk); 2649 tcp_ecn_queue_cwr(tp); 2650 } 2651 2652 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int newly_lost, int flag) 2653 { 2654 struct tcp_sock *tp = tcp_sk(sk); 2655 int sndcnt = 0; 2656 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp); 2657 2658 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd)) 2659 return; 2660 2661 tp->prr_delivered += newly_acked_sacked; 2662 if (delta < 0) { 2663 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered + 2664 tp->prior_cwnd - 1; 2665 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out; 2666 } else { 2667 sndcnt = max_t(int, tp->prr_delivered - tp->prr_out, 2668 newly_acked_sacked); 2669 if (flag & FLAG_SND_UNA_ADVANCED && !newly_lost) 2670 sndcnt++; 2671 sndcnt = min(delta, sndcnt); 2672 } 2673 /* Force a fast retransmit upon entering fast recovery */ 2674 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1)); 2675 tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + sndcnt); 2676 } 2677 2678 static inline void tcp_end_cwnd_reduction(struct sock *sk) 2679 { 2680 struct tcp_sock *tp = tcp_sk(sk); 2681 2682 if (inet_csk(sk)->icsk_ca_ops->cong_control) 2683 return; 2684 2685 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */ 2686 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH && 2687 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) { 2688 tcp_snd_cwnd_set(tp, tp->snd_ssthresh); 2689 tp->snd_cwnd_stamp = tcp_jiffies32; 2690 } 2691 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR); 2692 } 2693 2694 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */ 2695 void tcp_enter_cwr(struct sock *sk) 2696 { 2697 struct tcp_sock *tp = tcp_sk(sk); 2698 2699 tp->prior_ssthresh = 0; 2700 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) { 2701 tp->undo_marker = 0; 2702 tcp_init_cwnd_reduction(sk); 2703 tcp_set_ca_state(sk, TCP_CA_CWR); 2704 } 2705 } 2706 EXPORT_SYMBOL(tcp_enter_cwr); 2707 2708 static void tcp_try_keep_open(struct sock *sk) 2709 { 2710 struct tcp_sock *tp = tcp_sk(sk); 2711 int state = TCP_CA_Open; 2712 2713 if (tcp_left_out(tp) || tcp_any_retrans_done(sk)) 2714 state = TCP_CA_Disorder; 2715 2716 if (inet_csk(sk)->icsk_ca_state != state) { 2717 tcp_set_ca_state(sk, state); 2718 tp->high_seq = tp->snd_nxt; 2719 } 2720 } 2721 2722 static void tcp_try_to_open(struct sock *sk, int flag) 2723 { 2724 struct tcp_sock *tp = tcp_sk(sk); 2725 2726 tcp_verify_left_out(tp); 2727 2728 if (!tcp_any_retrans_done(sk)) 2729 tp->retrans_stamp = 0; 2730 2731 if (flag & FLAG_ECE) 2732 tcp_enter_cwr(sk); 2733 2734 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) { 2735 tcp_try_keep_open(sk); 2736 } 2737 } 2738 2739 static void tcp_mtup_probe_failed(struct sock *sk) 2740 { 2741 struct inet_connection_sock *icsk = inet_csk(sk); 2742 2743 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1; 2744 icsk->icsk_mtup.probe_size = 0; 2745 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL); 2746 } 2747 2748 static void tcp_mtup_probe_success(struct sock *sk) 2749 { 2750 struct tcp_sock *tp = tcp_sk(sk); 2751 struct inet_connection_sock *icsk = inet_csk(sk); 2752 u64 val; 2753 2754 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2755 2756 val = (u64)tcp_snd_cwnd(tp) * tcp_mss_to_mtu(sk, tp->mss_cache); 2757 do_div(val, icsk->icsk_mtup.probe_size); 2758 DEBUG_NET_WARN_ON_ONCE((u32)val != val); 2759 tcp_snd_cwnd_set(tp, max_t(u32, 1U, val)); 2760 2761 tp->snd_cwnd_cnt = 0; 2762 tp->snd_cwnd_stamp = tcp_jiffies32; 2763 tp->snd_ssthresh = tcp_current_ssthresh(sk); 2764 2765 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size; 2766 icsk->icsk_mtup.probe_size = 0; 2767 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 2768 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS); 2769 } 2770 2771 /* Sometimes we deduce that packets have been dropped due to reasons other than 2772 * congestion, like path MTU reductions or failed client TFO attempts. In these 2773 * cases we call this function to retransmit as many packets as cwnd allows, 2774 * without reducing cwnd. Given that retransmits will set retrans_stamp to a 2775 * non-zero value (and may do so in a later calling context due to TSQ), we 2776 * also enter CA_Loss so that we track when all retransmitted packets are ACKed 2777 * and clear retrans_stamp when that happens (to ensure later recurring RTOs 2778 * are using the correct retrans_stamp and don't declare ETIMEDOUT 2779 * prematurely). 2780 */ 2781 static void tcp_non_congestion_loss_retransmit(struct sock *sk) 2782 { 2783 const struct inet_connection_sock *icsk = inet_csk(sk); 2784 struct tcp_sock *tp = tcp_sk(sk); 2785 2786 if (icsk->icsk_ca_state != TCP_CA_Loss) { 2787 tp->high_seq = tp->snd_nxt; 2788 tp->snd_ssthresh = tcp_current_ssthresh(sk); 2789 tp->prior_ssthresh = 0; 2790 tp->undo_marker = 0; 2791 tcp_set_ca_state(sk, TCP_CA_Loss); 2792 } 2793 tcp_xmit_retransmit_queue(sk); 2794 } 2795 2796 /* Do a simple retransmit without using the backoff mechanisms in 2797 * tcp_timer. This is used for path mtu discovery. 2798 * The socket is already locked here. 2799 */ 2800 void tcp_simple_retransmit(struct sock *sk) 2801 { 2802 struct tcp_sock *tp = tcp_sk(sk); 2803 struct sk_buff *skb; 2804 int mss; 2805 2806 /* A fastopen SYN request is stored as two separate packets within 2807 * the retransmit queue, this is done by tcp_send_syn_data(). 2808 * As a result simply checking the MSS of the frames in the queue 2809 * will not work for the SYN packet. 2810 * 2811 * Us being here is an indication of a path MTU issue so we can 2812 * assume that the fastopen SYN was lost and just mark all the 2813 * frames in the retransmit queue as lost. We will use an MSS of 2814 * -1 to mark all frames as lost, otherwise compute the current MSS. 2815 */ 2816 if (tp->syn_data && sk->sk_state == TCP_SYN_SENT) 2817 mss = -1; 2818 else 2819 mss = tcp_current_mss(sk); 2820 2821 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) { 2822 if (tcp_skb_seglen(skb) > mss) 2823 tcp_mark_skb_lost(sk, skb); 2824 } 2825 2826 tcp_clear_retrans_hints_partial(tp); 2827 2828 if (!tp->lost_out) 2829 return; 2830 2831 if (tcp_is_reno(tp)) 2832 tcp_limit_reno_sacked(tp); 2833 2834 tcp_verify_left_out(tp); 2835 2836 /* Don't muck with the congestion window here. 2837 * Reason is that we do not increase amount of _data_ 2838 * in network, but units changed and effective 2839 * cwnd/ssthresh really reduced now. 2840 */ 2841 tcp_non_congestion_loss_retransmit(sk); 2842 } 2843 EXPORT_SYMBOL(tcp_simple_retransmit); 2844 2845 void tcp_enter_recovery(struct sock *sk, bool ece_ack) 2846 { 2847 struct tcp_sock *tp = tcp_sk(sk); 2848 int mib_idx; 2849 2850 if (tcp_is_reno(tp)) 2851 mib_idx = LINUX_MIB_TCPRENORECOVERY; 2852 else 2853 mib_idx = LINUX_MIB_TCPSACKRECOVERY; 2854 2855 NET_INC_STATS(sock_net(sk), mib_idx); 2856 2857 tp->prior_ssthresh = 0; 2858 tcp_init_undo(tp); 2859 2860 if (!tcp_in_cwnd_reduction(sk)) { 2861 if (!ece_ack) 2862 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2863 tcp_init_cwnd_reduction(sk); 2864 } 2865 tcp_set_ca_state(sk, TCP_CA_Recovery); 2866 } 2867 2868 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are 2869 * recovered or spurious. Otherwise retransmits more on partial ACKs. 2870 */ 2871 static void tcp_process_loss(struct sock *sk, int flag, int num_dupack, 2872 int *rexmit) 2873 { 2874 struct tcp_sock *tp = tcp_sk(sk); 2875 bool recovered = !before(tp->snd_una, tp->high_seq); 2876 2877 if ((flag & FLAG_SND_UNA_ADVANCED || rcu_access_pointer(tp->fastopen_rsk)) && 2878 tcp_try_undo_loss(sk, false)) 2879 return; 2880 2881 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */ 2882 /* Step 3.b. A timeout is spurious if not all data are 2883 * lost, i.e., never-retransmitted data are (s)acked. 2884 */ 2885 if ((flag & FLAG_ORIG_SACK_ACKED) && 2886 tcp_try_undo_loss(sk, true)) 2887 return; 2888 2889 if (after(tp->snd_nxt, tp->high_seq)) { 2890 if (flag & FLAG_DATA_SACKED || num_dupack) 2891 tp->frto = 0; /* Step 3.a. loss was real */ 2892 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) { 2893 tp->high_seq = tp->snd_nxt; 2894 /* Step 2.b. Try send new data (but deferred until cwnd 2895 * is updated in tcp_ack()). Otherwise fall back to 2896 * the conventional recovery. 2897 */ 2898 if (!tcp_write_queue_empty(sk) && 2899 after(tcp_wnd_end(tp), tp->snd_nxt)) { 2900 *rexmit = REXMIT_NEW; 2901 return; 2902 } 2903 tp->frto = 0; 2904 } 2905 } 2906 2907 if (recovered) { 2908 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */ 2909 tcp_try_undo_recovery(sk); 2910 return; 2911 } 2912 if (tcp_is_reno(tp)) { 2913 /* A Reno DUPACK means new data in F-RTO step 2.b above are 2914 * delivered. Lower inflight to clock out (re)transmissions. 2915 */ 2916 if (after(tp->snd_nxt, tp->high_seq) && num_dupack) 2917 tcp_add_reno_sack(sk, num_dupack, flag & FLAG_ECE); 2918 else if (flag & FLAG_SND_UNA_ADVANCED) 2919 tcp_reset_reno_sack(tp); 2920 } 2921 *rexmit = REXMIT_LOST; 2922 } 2923 2924 static bool tcp_force_fast_retransmit(struct sock *sk) 2925 { 2926 struct tcp_sock *tp = tcp_sk(sk); 2927 2928 return after(tcp_highest_sack_seq(tp), 2929 tp->snd_una + tp->reordering * tp->mss_cache); 2930 } 2931 2932 /* Undo during fast recovery after partial ACK. */ 2933 static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una, 2934 bool *do_lost) 2935 { 2936 struct tcp_sock *tp = tcp_sk(sk); 2937 2938 if (tp->undo_marker && tcp_packet_delayed(tp)) { 2939 /* Plain luck! Hole if filled with delayed 2940 * packet, rather than with a retransmit. Check reordering. 2941 */ 2942 tcp_check_sack_reordering(sk, prior_snd_una, 1); 2943 2944 /* We are getting evidence that the reordering degree is higher 2945 * than we realized. If there are no retransmits out then we 2946 * can undo. Otherwise we clock out new packets but do not 2947 * mark more packets lost or retransmit more. 2948 */ 2949 if (tp->retrans_out) 2950 return true; 2951 2952 if (!tcp_any_retrans_done(sk)) 2953 tp->retrans_stamp = 0; 2954 2955 DBGUNDO(sk, "partial recovery"); 2956 tcp_undo_cwnd_reduction(sk, true); 2957 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO); 2958 tcp_try_keep_open(sk); 2959 } else { 2960 /* Partial ACK arrived. Force fast retransmit. */ 2961 *do_lost = tcp_force_fast_retransmit(sk); 2962 } 2963 return false; 2964 } 2965 2966 static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag) 2967 { 2968 struct tcp_sock *tp = tcp_sk(sk); 2969 2970 if (tcp_rtx_queue_empty(sk)) 2971 return; 2972 2973 if (unlikely(tcp_is_reno(tp))) { 2974 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED); 2975 } else if (tcp_is_rack(sk)) { 2976 u32 prior_retrans = tp->retrans_out; 2977 2978 if (tcp_rack_mark_lost(sk)) 2979 *ack_flag &= ~FLAG_SET_XMIT_TIMER; 2980 if (prior_retrans > tp->retrans_out) 2981 *ack_flag |= FLAG_LOST_RETRANS; 2982 } 2983 } 2984 2985 /* Process an event, which can update packets-in-flight not trivially. 2986 * Main goal of this function is to calculate new estimate for left_out, 2987 * taking into account both packets sitting in receiver's buffer and 2988 * packets lost by network. 2989 * 2990 * Besides that it updates the congestion state when packet loss or ECN 2991 * is detected. But it does not reduce the cwnd, it is done by the 2992 * congestion control later. 2993 * 2994 * It does _not_ decide what to send, it is made in function 2995 * tcp_xmit_retransmit_queue(). 2996 */ 2997 static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una, 2998 int num_dupack, int *ack_flag, int *rexmit) 2999 { 3000 struct inet_connection_sock *icsk = inet_csk(sk); 3001 struct tcp_sock *tp = tcp_sk(sk); 3002 int fast_rexmit = 0, flag = *ack_flag; 3003 bool ece_ack = flag & FLAG_ECE; 3004 bool do_lost = num_dupack || ((flag & FLAG_DATA_SACKED) && 3005 tcp_force_fast_retransmit(sk)); 3006 3007 if (!tp->packets_out && tp->sacked_out) 3008 tp->sacked_out = 0; 3009 3010 /* Now state machine starts. 3011 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */ 3012 if (ece_ack) 3013 tp->prior_ssthresh = 0; 3014 3015 /* B. In all the states check for reneging SACKs. */ 3016 if (tcp_check_sack_reneging(sk, ack_flag)) 3017 return; 3018 3019 /* C. Check consistency of the current state. */ 3020 tcp_verify_left_out(tp); 3021 3022 /* D. Check state exit conditions. State can be terminated 3023 * when high_seq is ACKed. */ 3024 if (icsk->icsk_ca_state == TCP_CA_Open) { 3025 WARN_ON(tp->retrans_out != 0 && !tp->syn_data); 3026 tp->retrans_stamp = 0; 3027 } else if (!before(tp->snd_una, tp->high_seq)) { 3028 switch (icsk->icsk_ca_state) { 3029 case TCP_CA_CWR: 3030 /* CWR is to be held something *above* high_seq 3031 * is ACKed for CWR bit to reach receiver. */ 3032 if (tp->snd_una != tp->high_seq) { 3033 tcp_end_cwnd_reduction(sk); 3034 tcp_set_ca_state(sk, TCP_CA_Open); 3035 } 3036 break; 3037 3038 case TCP_CA_Recovery: 3039 if (tcp_is_reno(tp)) 3040 tcp_reset_reno_sack(tp); 3041 if (tcp_try_undo_recovery(sk)) 3042 return; 3043 tcp_end_cwnd_reduction(sk); 3044 break; 3045 } 3046 } 3047 3048 /* E. Process state. */ 3049 switch (icsk->icsk_ca_state) { 3050 case TCP_CA_Recovery: 3051 if (!(flag & FLAG_SND_UNA_ADVANCED)) { 3052 if (tcp_is_reno(tp)) 3053 tcp_add_reno_sack(sk, num_dupack, ece_ack); 3054 } else if (tcp_try_undo_partial(sk, prior_snd_una, &do_lost)) 3055 return; 3056 3057 if (tcp_try_undo_dsack(sk)) 3058 tcp_try_to_open(sk, flag); 3059 3060 tcp_identify_packet_loss(sk, ack_flag); 3061 if (icsk->icsk_ca_state != TCP_CA_Recovery) { 3062 if (!tcp_time_to_recover(sk, flag)) 3063 return; 3064 /* Undo reverts the recovery state. If loss is evident, 3065 * starts a new recovery (e.g. reordering then loss); 3066 */ 3067 tcp_enter_recovery(sk, ece_ack); 3068 } 3069 break; 3070 case TCP_CA_Loss: 3071 tcp_process_loss(sk, flag, num_dupack, rexmit); 3072 tcp_identify_packet_loss(sk, ack_flag); 3073 if (!(icsk->icsk_ca_state == TCP_CA_Open || 3074 (*ack_flag & FLAG_LOST_RETRANS))) 3075 return; 3076 /* Change state if cwnd is undone or retransmits are lost */ 3077 fallthrough; 3078 default: 3079 if (tcp_is_reno(tp)) { 3080 if (flag & FLAG_SND_UNA_ADVANCED) 3081 tcp_reset_reno_sack(tp); 3082 tcp_add_reno_sack(sk, num_dupack, ece_ack); 3083 } 3084 3085 if (icsk->icsk_ca_state <= TCP_CA_Disorder) 3086 tcp_try_undo_dsack(sk); 3087 3088 tcp_identify_packet_loss(sk, ack_flag); 3089 if (!tcp_time_to_recover(sk, flag)) { 3090 tcp_try_to_open(sk, flag); 3091 return; 3092 } 3093 3094 /* MTU probe failure: don't reduce cwnd */ 3095 if (icsk->icsk_ca_state < TCP_CA_CWR && 3096 icsk->icsk_mtup.probe_size && 3097 tp->snd_una == tp->mtu_probe.probe_seq_start) { 3098 tcp_mtup_probe_failed(sk); 3099 /* Restores the reduction we did in tcp_mtup_probe() */ 3100 tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) + 1); 3101 tcp_simple_retransmit(sk); 3102 return; 3103 } 3104 3105 /* Otherwise enter Recovery state */ 3106 tcp_enter_recovery(sk, ece_ack); 3107 fast_rexmit = 1; 3108 } 3109 3110 if (!tcp_is_rack(sk) && do_lost) 3111 tcp_update_scoreboard(sk, fast_rexmit); 3112 *rexmit = REXMIT_LOST; 3113 } 3114 3115 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag) 3116 { 3117 u32 wlen = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen) * HZ; 3118 struct tcp_sock *tp = tcp_sk(sk); 3119 3120 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) { 3121 /* If the remote keeps returning delayed ACKs, eventually 3122 * the min filter would pick it up and overestimate the 3123 * prop. delay when it expires. Skip suspected delayed ACKs. 3124 */ 3125 return; 3126 } 3127 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32, 3128 rtt_us ? : jiffies_to_usecs(1)); 3129 } 3130 3131 static bool tcp_ack_update_rtt(struct sock *sk, const int flag, 3132 long seq_rtt_us, long sack_rtt_us, 3133 long ca_rtt_us, struct rate_sample *rs) 3134 { 3135 const struct tcp_sock *tp = tcp_sk(sk); 3136 3137 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because 3138 * broken middle-boxes or peers may corrupt TS-ECR fields. But 3139 * Karn's algorithm forbids taking RTT if some retransmitted data 3140 * is acked (RFC6298). 3141 */ 3142 if (seq_rtt_us < 0) 3143 seq_rtt_us = sack_rtt_us; 3144 3145 /* RTTM Rule: A TSecr value received in a segment is used to 3146 * update the averaged RTT measurement only if the segment 3147 * acknowledges some new data, i.e., only if it advances the 3148 * left edge of the send window. 3149 * See draft-ietf-tcplw-high-performance-00, section 3.3. 3150 */ 3151 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 3152 flag & FLAG_ACKED) { 3153 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr; 3154 3155 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) { 3156 if (!delta) 3157 delta = 1; 3158 seq_rtt_us = delta * (USEC_PER_SEC / TCP_TS_HZ); 3159 ca_rtt_us = seq_rtt_us; 3160 } 3161 } 3162 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */ 3163 if (seq_rtt_us < 0) 3164 return false; 3165 3166 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is 3167 * always taken together with ACK, SACK, or TS-opts. Any negative 3168 * values will be skipped with the seq_rtt_us < 0 check above. 3169 */ 3170 tcp_update_rtt_min(sk, ca_rtt_us, flag); 3171 tcp_rtt_estimator(sk, seq_rtt_us); 3172 tcp_set_rto(sk); 3173 3174 /* RFC6298: only reset backoff on valid RTT measurement. */ 3175 inet_csk(sk)->icsk_backoff = 0; 3176 return true; 3177 } 3178 3179 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */ 3180 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req) 3181 { 3182 struct rate_sample rs; 3183 long rtt_us = -1L; 3184 3185 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack) 3186 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack); 3187 3188 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs); 3189 } 3190 3191 3192 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked) 3193 { 3194 const struct inet_connection_sock *icsk = inet_csk(sk); 3195 3196 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked); 3197 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32; 3198 } 3199 3200 /* Restart timer after forward progress on connection. 3201 * RFC2988 recommends to restart timer to now+rto. 3202 */ 3203 void tcp_rearm_rto(struct sock *sk) 3204 { 3205 const struct inet_connection_sock *icsk = inet_csk(sk); 3206 struct tcp_sock *tp = tcp_sk(sk); 3207 3208 /* If the retrans timer is currently being used by Fast Open 3209 * for SYN-ACK retrans purpose, stay put. 3210 */ 3211 if (rcu_access_pointer(tp->fastopen_rsk)) 3212 return; 3213 3214 if (!tp->packets_out) { 3215 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS); 3216 } else { 3217 u32 rto = inet_csk(sk)->icsk_rto; 3218 /* Offset the time elapsed after installing regular RTO */ 3219 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT || 3220 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) { 3221 s64 delta_us = tcp_rto_delta_us(sk); 3222 /* delta_us may not be positive if the socket is locked 3223 * when the retrans timer fires and is rescheduled. 3224 */ 3225 rto = usecs_to_jiffies(max_t(int, delta_us, 1)); 3226 } 3227 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto, 3228 TCP_RTO_MAX); 3229 } 3230 } 3231 3232 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */ 3233 static void tcp_set_xmit_timer(struct sock *sk) 3234 { 3235 if (!tcp_schedule_loss_probe(sk, true)) 3236 tcp_rearm_rto(sk); 3237 } 3238 3239 /* If we get here, the whole TSO packet has not been acked. */ 3240 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb) 3241 { 3242 struct tcp_sock *tp = tcp_sk(sk); 3243 u32 packets_acked; 3244 3245 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)); 3246 3247 packets_acked = tcp_skb_pcount(skb); 3248 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 3249 return 0; 3250 packets_acked -= tcp_skb_pcount(skb); 3251 3252 if (packets_acked) { 3253 BUG_ON(tcp_skb_pcount(skb) == 0); 3254 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)); 3255 } 3256 3257 return packets_acked; 3258 } 3259 3260 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb, 3261 const struct sk_buff *ack_skb, u32 prior_snd_una) 3262 { 3263 const struct skb_shared_info *shinfo; 3264 3265 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */ 3266 if (likely(!TCP_SKB_CB(skb)->txstamp_ack)) 3267 return; 3268 3269 shinfo = skb_shinfo(skb); 3270 if (!before(shinfo->tskey, prior_snd_una) && 3271 before(shinfo->tskey, tcp_sk(sk)->snd_una)) { 3272 tcp_skb_tsorted_save(skb) { 3273 __skb_tstamp_tx(skb, ack_skb, NULL, sk, SCM_TSTAMP_ACK); 3274 } tcp_skb_tsorted_restore(skb); 3275 } 3276 } 3277 3278 /* Remove acknowledged frames from the retransmission queue. If our packet 3279 * is before the ack sequence we can discard it as it's confirmed to have 3280 * arrived at the other end. 3281 */ 3282 static int tcp_clean_rtx_queue(struct sock *sk, const struct sk_buff *ack_skb, 3283 u32 prior_fack, u32 prior_snd_una, 3284 struct tcp_sacktag_state *sack, bool ece_ack) 3285 { 3286 const struct inet_connection_sock *icsk = inet_csk(sk); 3287 u64 first_ackt, last_ackt; 3288 struct tcp_sock *tp = tcp_sk(sk); 3289 u32 prior_sacked = tp->sacked_out; 3290 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */ 3291 struct sk_buff *skb, *next; 3292 bool fully_acked = true; 3293 long sack_rtt_us = -1L; 3294 long seq_rtt_us = -1L; 3295 long ca_rtt_us = -1L; 3296 u32 pkts_acked = 0; 3297 bool rtt_update; 3298 int flag = 0; 3299 3300 first_ackt = 0; 3301 3302 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) { 3303 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 3304 const u32 start_seq = scb->seq; 3305 u8 sacked = scb->sacked; 3306 u32 acked_pcount; 3307 3308 /* Determine how many packets and what bytes were acked, tso and else */ 3309 if (after(scb->end_seq, tp->snd_una)) { 3310 if (tcp_skb_pcount(skb) == 1 || 3311 !after(tp->snd_una, scb->seq)) 3312 break; 3313 3314 acked_pcount = tcp_tso_acked(sk, skb); 3315 if (!acked_pcount) 3316 break; 3317 fully_acked = false; 3318 } else { 3319 acked_pcount = tcp_skb_pcount(skb); 3320 } 3321 3322 if (unlikely(sacked & TCPCB_RETRANS)) { 3323 if (sacked & TCPCB_SACKED_RETRANS) 3324 tp->retrans_out -= acked_pcount; 3325 flag |= FLAG_RETRANS_DATA_ACKED; 3326 } else if (!(sacked & TCPCB_SACKED_ACKED)) { 3327 last_ackt = tcp_skb_timestamp_us(skb); 3328 WARN_ON_ONCE(last_ackt == 0); 3329 if (!first_ackt) 3330 first_ackt = last_ackt; 3331 3332 if (before(start_seq, reord)) 3333 reord = start_seq; 3334 if (!after(scb->end_seq, tp->high_seq)) 3335 flag |= FLAG_ORIG_SACK_ACKED; 3336 } 3337 3338 if (sacked & TCPCB_SACKED_ACKED) { 3339 tp->sacked_out -= acked_pcount; 3340 } else if (tcp_is_sack(tp)) { 3341 tcp_count_delivered(tp, acked_pcount, ece_ack); 3342 if (!tcp_skb_spurious_retrans(tp, skb)) 3343 tcp_rack_advance(tp, sacked, scb->end_seq, 3344 tcp_skb_timestamp_us(skb)); 3345 } 3346 if (sacked & TCPCB_LOST) 3347 tp->lost_out -= acked_pcount; 3348 3349 tp->packets_out -= acked_pcount; 3350 pkts_acked += acked_pcount; 3351 tcp_rate_skb_delivered(sk, skb, sack->rate); 3352 3353 /* Initial outgoing SYN's get put onto the write_queue 3354 * just like anything else we transmit. It is not 3355 * true data, and if we misinform our callers that 3356 * this ACK acks real data, we will erroneously exit 3357 * connection startup slow start one packet too 3358 * quickly. This is severely frowned upon behavior. 3359 */ 3360 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) { 3361 flag |= FLAG_DATA_ACKED; 3362 } else { 3363 flag |= FLAG_SYN_ACKED; 3364 tp->retrans_stamp = 0; 3365 } 3366 3367 if (!fully_acked) 3368 break; 3369 3370 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una); 3371 3372 next = skb_rb_next(skb); 3373 if (unlikely(skb == tp->retransmit_skb_hint)) 3374 tp->retransmit_skb_hint = NULL; 3375 if (unlikely(skb == tp->lost_skb_hint)) 3376 tp->lost_skb_hint = NULL; 3377 tcp_highest_sack_replace(sk, skb, next); 3378 tcp_rtx_queue_unlink_and_free(skb, sk); 3379 } 3380 3381 if (!skb) 3382 tcp_chrono_stop(sk, TCP_CHRONO_BUSY); 3383 3384 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una))) 3385 tp->snd_up = tp->snd_una; 3386 3387 if (skb) { 3388 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una); 3389 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 3390 flag |= FLAG_SACK_RENEGING; 3391 } 3392 3393 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) { 3394 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt); 3395 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt); 3396 3397 if (pkts_acked == 1 && fully_acked && !prior_sacked && 3398 (tp->snd_una - prior_snd_una) < tp->mss_cache && 3399 sack->rate->prior_delivered + 1 == tp->delivered && 3400 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) { 3401 /* Conservatively mark a delayed ACK. It's typically 3402 * from a lone runt packet over the round trip to 3403 * a receiver w/o out-of-order or CE events. 3404 */ 3405 flag |= FLAG_ACK_MAYBE_DELAYED; 3406 } 3407 } 3408 if (sack->first_sackt) { 3409 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt); 3410 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt); 3411 } 3412 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us, 3413 ca_rtt_us, sack->rate); 3414 3415 if (flag & FLAG_ACKED) { 3416 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */ 3417 if (unlikely(icsk->icsk_mtup.probe_size && 3418 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) { 3419 tcp_mtup_probe_success(sk); 3420 } 3421 3422 if (tcp_is_reno(tp)) { 3423 tcp_remove_reno_sacks(sk, pkts_acked, ece_ack); 3424 3425 /* If any of the cumulatively ACKed segments was 3426 * retransmitted, non-SACK case cannot confirm that 3427 * progress was due to original transmission due to 3428 * lack of TCPCB_SACKED_ACKED bits even if some of 3429 * the packets may have been never retransmitted. 3430 */ 3431 if (flag & FLAG_RETRANS_DATA_ACKED) 3432 flag &= ~FLAG_ORIG_SACK_ACKED; 3433 } else { 3434 int delta; 3435 3436 /* Non-retransmitted hole got filled? That's reordering */ 3437 if (before(reord, prior_fack)) 3438 tcp_check_sack_reordering(sk, reord, 0); 3439 3440 delta = prior_sacked - tp->sacked_out; 3441 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta); 3442 } 3443 } else if (skb && rtt_update && sack_rtt_us >= 0 && 3444 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp, 3445 tcp_skb_timestamp_us(skb))) { 3446 /* Do not re-arm RTO if the sack RTT is measured from data sent 3447 * after when the head was last (re)transmitted. Otherwise the 3448 * timeout may continue to extend in loss recovery. 3449 */ 3450 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */ 3451 } 3452 3453 if (icsk->icsk_ca_ops->pkts_acked) { 3454 struct ack_sample sample = { .pkts_acked = pkts_acked, 3455 .rtt_us = sack->rate->rtt_us }; 3456 3457 sample.in_flight = tp->mss_cache * 3458 (tp->delivered - sack->rate->prior_delivered); 3459 icsk->icsk_ca_ops->pkts_acked(sk, &sample); 3460 } 3461 3462 #if FASTRETRANS_DEBUG > 0 3463 WARN_ON((int)tp->sacked_out < 0); 3464 WARN_ON((int)tp->lost_out < 0); 3465 WARN_ON((int)tp->retrans_out < 0); 3466 if (!tp->packets_out && tcp_is_sack(tp)) { 3467 icsk = inet_csk(sk); 3468 if (tp->lost_out) { 3469 pr_debug("Leak l=%u %d\n", 3470 tp->lost_out, icsk->icsk_ca_state); 3471 tp->lost_out = 0; 3472 } 3473 if (tp->sacked_out) { 3474 pr_debug("Leak s=%u %d\n", 3475 tp->sacked_out, icsk->icsk_ca_state); 3476 tp->sacked_out = 0; 3477 } 3478 if (tp->retrans_out) { 3479 pr_debug("Leak r=%u %d\n", 3480 tp->retrans_out, icsk->icsk_ca_state); 3481 tp->retrans_out = 0; 3482 } 3483 } 3484 #endif 3485 return flag; 3486 } 3487 3488 static void tcp_ack_probe(struct sock *sk) 3489 { 3490 struct inet_connection_sock *icsk = inet_csk(sk); 3491 struct sk_buff *head = tcp_send_head(sk); 3492 const struct tcp_sock *tp = tcp_sk(sk); 3493 3494 /* Was it a usable window open? */ 3495 if (!head) 3496 return; 3497 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) { 3498 icsk->icsk_backoff = 0; 3499 icsk->icsk_probes_tstamp = 0; 3500 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0); 3501 /* Socket must be waked up by subsequent tcp_data_snd_check(). 3502 * This function is not for random using! 3503 */ 3504 } else { 3505 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX); 3506 3507 when = tcp_clamp_probe0_to_user_timeout(sk, when); 3508 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, when, TCP_RTO_MAX); 3509 } 3510 } 3511 3512 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag) 3513 { 3514 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) || 3515 inet_csk(sk)->icsk_ca_state != TCP_CA_Open; 3516 } 3517 3518 /* Decide wheather to run the increase function of congestion control. */ 3519 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag) 3520 { 3521 /* If reordering is high then always grow cwnd whenever data is 3522 * delivered regardless of its ordering. Otherwise stay conservative 3523 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/ 3524 * new SACK or ECE mark may first advance cwnd here and later reduce 3525 * cwnd in tcp_fastretrans_alert() based on more states. 3526 */ 3527 if (tcp_sk(sk)->reordering > 3528 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reordering)) 3529 return flag & FLAG_FORWARD_PROGRESS; 3530 3531 return flag & FLAG_DATA_ACKED; 3532 } 3533 3534 /* The "ultimate" congestion control function that aims to replace the rigid 3535 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction). 3536 * It's called toward the end of processing an ACK with precise rate 3537 * information. All transmission or retransmission are delayed afterwards. 3538 */ 3539 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked, 3540 int flag, const struct rate_sample *rs) 3541 { 3542 const struct inet_connection_sock *icsk = inet_csk(sk); 3543 3544 if (icsk->icsk_ca_ops->cong_control) { 3545 icsk->icsk_ca_ops->cong_control(sk, rs); 3546 return; 3547 } 3548 3549 if (tcp_in_cwnd_reduction(sk)) { 3550 /* Reduce cwnd if state mandates */ 3551 tcp_cwnd_reduction(sk, acked_sacked, rs->losses, flag); 3552 } else if (tcp_may_raise_cwnd(sk, flag)) { 3553 /* Advance cwnd if state allows */ 3554 tcp_cong_avoid(sk, ack, acked_sacked); 3555 } 3556 tcp_update_pacing_rate(sk); 3557 } 3558 3559 /* Check that window update is acceptable. 3560 * The function assumes that snd_una<=ack<=snd_next. 3561 */ 3562 static inline bool tcp_may_update_window(const struct tcp_sock *tp, 3563 const u32 ack, const u32 ack_seq, 3564 const u32 nwin) 3565 { 3566 return after(ack, tp->snd_una) || 3567 after(ack_seq, tp->snd_wl1) || 3568 (ack_seq == tp->snd_wl1 && (nwin > tp->snd_wnd || !nwin)); 3569 } 3570 3571 /* If we update tp->snd_una, also update tp->bytes_acked */ 3572 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack) 3573 { 3574 u32 delta = ack - tp->snd_una; 3575 3576 sock_owned_by_me((struct sock *)tp); 3577 tp->bytes_acked += delta; 3578 tp->snd_una = ack; 3579 } 3580 3581 /* If we update tp->rcv_nxt, also update tp->bytes_received */ 3582 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq) 3583 { 3584 u32 delta = seq - tp->rcv_nxt; 3585 3586 sock_owned_by_me((struct sock *)tp); 3587 tp->bytes_received += delta; 3588 WRITE_ONCE(tp->rcv_nxt, seq); 3589 } 3590 3591 /* Update our send window. 3592 * 3593 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2 3594 * and in FreeBSD. NetBSD's one is even worse.) is wrong. 3595 */ 3596 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack, 3597 u32 ack_seq) 3598 { 3599 struct tcp_sock *tp = tcp_sk(sk); 3600 int flag = 0; 3601 u32 nwin = ntohs(tcp_hdr(skb)->window); 3602 3603 if (likely(!tcp_hdr(skb)->syn)) 3604 nwin <<= tp->rx_opt.snd_wscale; 3605 3606 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) { 3607 flag |= FLAG_WIN_UPDATE; 3608 tcp_update_wl(tp, ack_seq); 3609 3610 if (tp->snd_wnd != nwin) { 3611 tp->snd_wnd = nwin; 3612 3613 /* Note, it is the only place, where 3614 * fast path is recovered for sending TCP. 3615 */ 3616 tp->pred_flags = 0; 3617 tcp_fast_path_check(sk); 3618 3619 if (!tcp_write_queue_empty(sk)) 3620 tcp_slow_start_after_idle_check(sk); 3621 3622 if (nwin > tp->max_window) { 3623 tp->max_window = nwin; 3624 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie); 3625 } 3626 } 3627 } 3628 3629 tcp_snd_una_update(tp, ack); 3630 3631 return flag; 3632 } 3633 3634 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx, 3635 u32 *last_oow_ack_time) 3636 { 3637 /* Paired with the WRITE_ONCE() in this function. */ 3638 u32 val = READ_ONCE(*last_oow_ack_time); 3639 3640 if (val) { 3641 s32 elapsed = (s32)(tcp_jiffies32 - val); 3642 3643 if (0 <= elapsed && 3644 elapsed < READ_ONCE(net->ipv4.sysctl_tcp_invalid_ratelimit)) { 3645 NET_INC_STATS(net, mib_idx); 3646 return true; /* rate-limited: don't send yet! */ 3647 } 3648 } 3649 3650 /* Paired with the prior READ_ONCE() and with itself, 3651 * as we might be lockless. 3652 */ 3653 WRITE_ONCE(*last_oow_ack_time, tcp_jiffies32); 3654 3655 return false; /* not rate-limited: go ahead, send dupack now! */ 3656 } 3657 3658 /* Return true if we're currently rate-limiting out-of-window ACKs and 3659 * thus shouldn't send a dupack right now. We rate-limit dupacks in 3660 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS 3661 * attacks that send repeated SYNs or ACKs for the same connection. To 3662 * do this, we do not send a duplicate SYNACK or ACK if the remote 3663 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate. 3664 */ 3665 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb, 3666 int mib_idx, u32 *last_oow_ack_time) 3667 { 3668 /* Data packets without SYNs are not likely part of an ACK loop. */ 3669 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) && 3670 !tcp_hdr(skb)->syn) 3671 return false; 3672 3673 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time); 3674 } 3675 3676 /* RFC 5961 7 [ACK Throttling] */ 3677 static void tcp_send_challenge_ack(struct sock *sk) 3678 { 3679 struct tcp_sock *tp = tcp_sk(sk); 3680 struct net *net = sock_net(sk); 3681 u32 count, now, ack_limit; 3682 3683 /* First check our per-socket dupack rate limit. */ 3684 if (__tcp_oow_rate_limited(net, 3685 LINUX_MIB_TCPACKSKIPPEDCHALLENGE, 3686 &tp->last_oow_ack_time)) 3687 return; 3688 3689 ack_limit = READ_ONCE(net->ipv4.sysctl_tcp_challenge_ack_limit); 3690 if (ack_limit == INT_MAX) 3691 goto send_ack; 3692 3693 /* Then check host-wide RFC 5961 rate limit. */ 3694 now = jiffies / HZ; 3695 if (now != READ_ONCE(net->ipv4.tcp_challenge_timestamp)) { 3696 u32 half = (ack_limit + 1) >> 1; 3697 3698 WRITE_ONCE(net->ipv4.tcp_challenge_timestamp, now); 3699 WRITE_ONCE(net->ipv4.tcp_challenge_count, 3700 get_random_u32_inclusive(half, ack_limit + half - 1)); 3701 } 3702 count = READ_ONCE(net->ipv4.tcp_challenge_count); 3703 if (count > 0) { 3704 WRITE_ONCE(net->ipv4.tcp_challenge_count, count - 1); 3705 send_ack: 3706 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK); 3707 tcp_send_ack(sk); 3708 } 3709 } 3710 3711 static void tcp_store_ts_recent(struct tcp_sock *tp) 3712 { 3713 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval; 3714 tp->rx_opt.ts_recent_stamp = ktime_get_seconds(); 3715 } 3716 3717 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq) 3718 { 3719 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) { 3720 /* PAWS bug workaround wrt. ACK frames, the PAWS discard 3721 * extra check below makes sure this can only happen 3722 * for pure ACK frames. -DaveM 3723 * 3724 * Not only, also it occurs for expired timestamps. 3725 */ 3726 3727 if (tcp_paws_check(&tp->rx_opt, 0)) 3728 tcp_store_ts_recent(tp); 3729 } 3730 } 3731 3732 /* This routine deals with acks during a TLP episode and ends an episode by 3733 * resetting tlp_high_seq. Ref: TLP algorithm in draft-ietf-tcpm-rack 3734 */ 3735 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag) 3736 { 3737 struct tcp_sock *tp = tcp_sk(sk); 3738 3739 if (before(ack, tp->tlp_high_seq)) 3740 return; 3741 3742 if (!tp->tlp_retrans) { 3743 /* TLP of new data has been acknowledged */ 3744 tp->tlp_high_seq = 0; 3745 } else if (flag & FLAG_DSACK_TLP) { 3746 /* This DSACK means original and TLP probe arrived; no loss */ 3747 tp->tlp_high_seq = 0; 3748 } else if (after(ack, tp->tlp_high_seq)) { 3749 /* ACK advances: there was a loss, so reduce cwnd. Reset 3750 * tlp_high_seq in tcp_init_cwnd_reduction() 3751 */ 3752 tcp_init_cwnd_reduction(sk); 3753 tcp_set_ca_state(sk, TCP_CA_CWR); 3754 tcp_end_cwnd_reduction(sk); 3755 tcp_try_keep_open(sk); 3756 NET_INC_STATS(sock_net(sk), 3757 LINUX_MIB_TCPLOSSPROBERECOVERY); 3758 } else if (!(flag & (FLAG_SND_UNA_ADVANCED | 3759 FLAG_NOT_DUP | FLAG_DATA_SACKED))) { 3760 /* Pure dupack: original and TLP probe arrived; no loss */ 3761 tp->tlp_high_seq = 0; 3762 } 3763 } 3764 3765 static inline void tcp_in_ack_event(struct sock *sk, u32 flags) 3766 { 3767 const struct inet_connection_sock *icsk = inet_csk(sk); 3768 3769 if (icsk->icsk_ca_ops->in_ack_event) 3770 icsk->icsk_ca_ops->in_ack_event(sk, flags); 3771 } 3772 3773 /* Congestion control has updated the cwnd already. So if we're in 3774 * loss recovery then now we do any new sends (for FRTO) or 3775 * retransmits (for CA_Loss or CA_recovery) that make sense. 3776 */ 3777 static void tcp_xmit_recovery(struct sock *sk, int rexmit) 3778 { 3779 struct tcp_sock *tp = tcp_sk(sk); 3780 3781 if (rexmit == REXMIT_NONE || sk->sk_state == TCP_SYN_SENT) 3782 return; 3783 3784 if (unlikely(rexmit == REXMIT_NEW)) { 3785 __tcp_push_pending_frames(sk, tcp_current_mss(sk), 3786 TCP_NAGLE_OFF); 3787 if (after(tp->snd_nxt, tp->high_seq)) 3788 return; 3789 tp->frto = 0; 3790 } 3791 tcp_xmit_retransmit_queue(sk); 3792 } 3793 3794 /* Returns the number of packets newly acked or sacked by the current ACK */ 3795 static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, int flag) 3796 { 3797 const struct net *net = sock_net(sk); 3798 struct tcp_sock *tp = tcp_sk(sk); 3799 u32 delivered; 3800 3801 delivered = tp->delivered - prior_delivered; 3802 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered); 3803 if (flag & FLAG_ECE) 3804 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, delivered); 3805 3806 return delivered; 3807 } 3808 3809 /* This routine deals with incoming acks, but not outgoing ones. */ 3810 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag) 3811 { 3812 struct inet_connection_sock *icsk = inet_csk(sk); 3813 struct tcp_sock *tp = tcp_sk(sk); 3814 struct tcp_sacktag_state sack_state; 3815 struct rate_sample rs = { .prior_delivered = 0 }; 3816 u32 prior_snd_una = tp->snd_una; 3817 bool is_sack_reneg = tp->is_sack_reneg; 3818 u32 ack_seq = TCP_SKB_CB(skb)->seq; 3819 u32 ack = TCP_SKB_CB(skb)->ack_seq; 3820 int num_dupack = 0; 3821 int prior_packets = tp->packets_out; 3822 u32 delivered = tp->delivered; 3823 u32 lost = tp->lost; 3824 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */ 3825 u32 prior_fack; 3826 3827 sack_state.first_sackt = 0; 3828 sack_state.rate = &rs; 3829 sack_state.sack_delivered = 0; 3830 3831 /* We very likely will need to access rtx queue. */ 3832 prefetch(sk->tcp_rtx_queue.rb_node); 3833 3834 /* If the ack is older than previous acks 3835 * then we can probably ignore it. 3836 */ 3837 if (before(ack, prior_snd_una)) { 3838 u32 max_window; 3839 3840 /* do not accept ACK for bytes we never sent. */ 3841 max_window = min_t(u64, tp->max_window, tp->bytes_acked); 3842 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */ 3843 if (before(ack, prior_snd_una - max_window)) { 3844 if (!(flag & FLAG_NO_CHALLENGE_ACK)) 3845 tcp_send_challenge_ack(sk); 3846 return -SKB_DROP_REASON_TCP_TOO_OLD_ACK; 3847 } 3848 goto old_ack; 3849 } 3850 3851 /* If the ack includes data we haven't sent yet, discard 3852 * this segment (RFC793 Section 3.9). 3853 */ 3854 if (after(ack, tp->snd_nxt)) 3855 return -SKB_DROP_REASON_TCP_ACK_UNSENT_DATA; 3856 3857 if (after(ack, prior_snd_una)) { 3858 flag |= FLAG_SND_UNA_ADVANCED; 3859 icsk->icsk_retransmits = 0; 3860 3861 #if IS_ENABLED(CONFIG_TLS_DEVICE) 3862 if (static_branch_unlikely(&clean_acked_data_enabled.key)) 3863 if (icsk->icsk_clean_acked) 3864 icsk->icsk_clean_acked(sk, ack); 3865 #endif 3866 } 3867 3868 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una; 3869 rs.prior_in_flight = tcp_packets_in_flight(tp); 3870 3871 /* ts_recent update must be made after we are sure that the packet 3872 * is in window. 3873 */ 3874 if (flag & FLAG_UPDATE_TS_RECENT) 3875 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); 3876 3877 if ((flag & (FLAG_SLOWPATH | FLAG_SND_UNA_ADVANCED)) == 3878 FLAG_SND_UNA_ADVANCED) { 3879 /* Window is constant, pure forward advance. 3880 * No more checks are required. 3881 * Note, we use the fact that SND.UNA>=SND.WL2. 3882 */ 3883 tcp_update_wl(tp, ack_seq); 3884 tcp_snd_una_update(tp, ack); 3885 flag |= FLAG_WIN_UPDATE; 3886 3887 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE); 3888 3889 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS); 3890 } else { 3891 u32 ack_ev_flags = CA_ACK_SLOWPATH; 3892 3893 if (ack_seq != TCP_SKB_CB(skb)->end_seq) 3894 flag |= FLAG_DATA; 3895 else 3896 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS); 3897 3898 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq); 3899 3900 if (TCP_SKB_CB(skb)->sacked) 3901 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una, 3902 &sack_state); 3903 3904 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) { 3905 flag |= FLAG_ECE; 3906 ack_ev_flags |= CA_ACK_ECE; 3907 } 3908 3909 if (sack_state.sack_delivered) 3910 tcp_count_delivered(tp, sack_state.sack_delivered, 3911 flag & FLAG_ECE); 3912 3913 if (flag & FLAG_WIN_UPDATE) 3914 ack_ev_flags |= CA_ACK_WIN_UPDATE; 3915 3916 tcp_in_ack_event(sk, ack_ev_flags); 3917 } 3918 3919 /* This is a deviation from RFC3168 since it states that: 3920 * "When the TCP data sender is ready to set the CWR bit after reducing 3921 * the congestion window, it SHOULD set the CWR bit only on the first 3922 * new data packet that it transmits." 3923 * We accept CWR on pure ACKs to be more robust 3924 * with widely-deployed TCP implementations that do this. 3925 */ 3926 tcp_ecn_accept_cwr(sk, skb); 3927 3928 /* We passed data and got it acked, remove any soft error 3929 * log. Something worked... 3930 */ 3931 WRITE_ONCE(sk->sk_err_soft, 0); 3932 icsk->icsk_probes_out = 0; 3933 tp->rcv_tstamp = tcp_jiffies32; 3934 if (!prior_packets) 3935 goto no_queue; 3936 3937 /* See if we can take anything off of the retransmit queue. */ 3938 flag |= tcp_clean_rtx_queue(sk, skb, prior_fack, prior_snd_una, 3939 &sack_state, flag & FLAG_ECE); 3940 3941 tcp_rack_update_reo_wnd(sk, &rs); 3942 3943 if (tp->tlp_high_seq) 3944 tcp_process_tlp_ack(sk, ack, flag); 3945 3946 if (tcp_ack_is_dubious(sk, flag)) { 3947 if (!(flag & (FLAG_SND_UNA_ADVANCED | 3948 FLAG_NOT_DUP | FLAG_DSACKING_ACK))) { 3949 num_dupack = 1; 3950 /* Consider if pure acks were aggregated in tcp_add_backlog() */ 3951 if (!(flag & FLAG_DATA)) 3952 num_dupack = max_t(u16, 1, skb_shinfo(skb)->gso_segs); 3953 } 3954 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag, 3955 &rexmit); 3956 } 3957 3958 /* If needed, reset TLP/RTO timer when RACK doesn't set. */ 3959 if (flag & FLAG_SET_XMIT_TIMER) 3960 tcp_set_xmit_timer(sk); 3961 3962 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) 3963 sk_dst_confirm(sk); 3964 3965 delivered = tcp_newly_delivered(sk, delivered, flag); 3966 lost = tp->lost - lost; /* freshly marked lost */ 3967 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED); 3968 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate); 3969 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate); 3970 tcp_xmit_recovery(sk, rexmit); 3971 return 1; 3972 3973 no_queue: 3974 /* If data was DSACKed, see if we can undo a cwnd reduction. */ 3975 if (flag & FLAG_DSACKING_ACK) { 3976 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag, 3977 &rexmit); 3978 tcp_newly_delivered(sk, delivered, flag); 3979 } 3980 /* If this ack opens up a zero window, clear backoff. It was 3981 * being used to time the probes, and is probably far higher than 3982 * it needs to be for normal retransmission. 3983 */ 3984 tcp_ack_probe(sk); 3985 3986 if (tp->tlp_high_seq) 3987 tcp_process_tlp_ack(sk, ack, flag); 3988 return 1; 3989 3990 old_ack: 3991 /* If data was SACKed, tag it and see if we should send more data. 3992 * If data was DSACKed, see if we can undo a cwnd reduction. 3993 */ 3994 if (TCP_SKB_CB(skb)->sacked) { 3995 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una, 3996 &sack_state); 3997 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag, 3998 &rexmit); 3999 tcp_newly_delivered(sk, delivered, flag); 4000 tcp_xmit_recovery(sk, rexmit); 4001 } 4002 4003 return 0; 4004 } 4005 4006 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie, 4007 bool syn, struct tcp_fastopen_cookie *foc, 4008 bool exp_opt) 4009 { 4010 /* Valid only in SYN or SYN-ACK with an even length. */ 4011 if (!foc || !syn || len < 0 || (len & 1)) 4012 return; 4013 4014 if (len >= TCP_FASTOPEN_COOKIE_MIN && 4015 len <= TCP_FASTOPEN_COOKIE_MAX) 4016 memcpy(foc->val, cookie, len); 4017 else if (len != 0) 4018 len = -1; 4019 foc->len = len; 4020 foc->exp = exp_opt; 4021 } 4022 4023 static bool smc_parse_options(const struct tcphdr *th, 4024 struct tcp_options_received *opt_rx, 4025 const unsigned char *ptr, 4026 int opsize) 4027 { 4028 #if IS_ENABLED(CONFIG_SMC) 4029 if (static_branch_unlikely(&tcp_have_smc)) { 4030 if (th->syn && !(opsize & 1) && 4031 opsize >= TCPOLEN_EXP_SMC_BASE && 4032 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC) { 4033 opt_rx->smc_ok = 1; 4034 return true; 4035 } 4036 } 4037 #endif 4038 return false; 4039 } 4040 4041 /* Try to parse the MSS option from the TCP header. Return 0 on failure, clamped 4042 * value on success. 4043 */ 4044 u16 tcp_parse_mss_option(const struct tcphdr *th, u16 user_mss) 4045 { 4046 const unsigned char *ptr = (const unsigned char *)(th + 1); 4047 int length = (th->doff * 4) - sizeof(struct tcphdr); 4048 u16 mss = 0; 4049 4050 while (length > 0) { 4051 int opcode = *ptr++; 4052 int opsize; 4053 4054 switch (opcode) { 4055 case TCPOPT_EOL: 4056 return mss; 4057 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ 4058 length--; 4059 continue; 4060 default: 4061 if (length < 2) 4062 return mss; 4063 opsize = *ptr++; 4064 if (opsize < 2) /* "silly options" */ 4065 return mss; 4066 if (opsize > length) 4067 return mss; /* fail on partial options */ 4068 if (opcode == TCPOPT_MSS && opsize == TCPOLEN_MSS) { 4069 u16 in_mss = get_unaligned_be16(ptr); 4070 4071 if (in_mss) { 4072 if (user_mss && user_mss < in_mss) 4073 in_mss = user_mss; 4074 mss = in_mss; 4075 } 4076 } 4077 ptr += opsize - 2; 4078 length -= opsize; 4079 } 4080 } 4081 return mss; 4082 } 4083 EXPORT_SYMBOL_GPL(tcp_parse_mss_option); 4084 4085 /* Look for tcp options. Normally only called on SYN and SYNACK packets. 4086 * But, this can also be called on packets in the established flow when 4087 * the fast version below fails. 4088 */ 4089 void tcp_parse_options(const struct net *net, 4090 const struct sk_buff *skb, 4091 struct tcp_options_received *opt_rx, int estab, 4092 struct tcp_fastopen_cookie *foc) 4093 { 4094 const unsigned char *ptr; 4095 const struct tcphdr *th = tcp_hdr(skb); 4096 int length = (th->doff * 4) - sizeof(struct tcphdr); 4097 4098 ptr = (const unsigned char *)(th + 1); 4099 opt_rx->saw_tstamp = 0; 4100 opt_rx->saw_unknown = 0; 4101 4102 while (length > 0) { 4103 int opcode = *ptr++; 4104 int opsize; 4105 4106 switch (opcode) { 4107 case TCPOPT_EOL: 4108 return; 4109 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ 4110 length--; 4111 continue; 4112 default: 4113 if (length < 2) 4114 return; 4115 opsize = *ptr++; 4116 if (opsize < 2) /* "silly options" */ 4117 return; 4118 if (opsize > length) 4119 return; /* don't parse partial options */ 4120 switch (opcode) { 4121 case TCPOPT_MSS: 4122 if (opsize == TCPOLEN_MSS && th->syn && !estab) { 4123 u16 in_mss = get_unaligned_be16(ptr); 4124 if (in_mss) { 4125 if (opt_rx->user_mss && 4126 opt_rx->user_mss < in_mss) 4127 in_mss = opt_rx->user_mss; 4128 opt_rx->mss_clamp = in_mss; 4129 } 4130 } 4131 break; 4132 case TCPOPT_WINDOW: 4133 if (opsize == TCPOLEN_WINDOW && th->syn && 4134 !estab && READ_ONCE(net->ipv4.sysctl_tcp_window_scaling)) { 4135 __u8 snd_wscale = *(__u8 *)ptr; 4136 opt_rx->wscale_ok = 1; 4137 if (snd_wscale > TCP_MAX_WSCALE) { 4138 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n", 4139 __func__, 4140 snd_wscale, 4141 TCP_MAX_WSCALE); 4142 snd_wscale = TCP_MAX_WSCALE; 4143 } 4144 opt_rx->snd_wscale = snd_wscale; 4145 } 4146 break; 4147 case TCPOPT_TIMESTAMP: 4148 if ((opsize == TCPOLEN_TIMESTAMP) && 4149 ((estab && opt_rx->tstamp_ok) || 4150 (!estab && READ_ONCE(net->ipv4.sysctl_tcp_timestamps)))) { 4151 opt_rx->saw_tstamp = 1; 4152 opt_rx->rcv_tsval = get_unaligned_be32(ptr); 4153 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4); 4154 } 4155 break; 4156 case TCPOPT_SACK_PERM: 4157 if (opsize == TCPOLEN_SACK_PERM && th->syn && 4158 !estab && READ_ONCE(net->ipv4.sysctl_tcp_sack)) { 4159 opt_rx->sack_ok = TCP_SACK_SEEN; 4160 tcp_sack_reset(opt_rx); 4161 } 4162 break; 4163 4164 case TCPOPT_SACK: 4165 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) && 4166 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) && 4167 opt_rx->sack_ok) { 4168 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th; 4169 } 4170 break; 4171 #ifdef CONFIG_TCP_MD5SIG 4172 case TCPOPT_MD5SIG: 4173 /* The MD5 Hash has already been 4174 * checked (see tcp_v{4,6}_rcv()). 4175 */ 4176 break; 4177 #endif 4178 case TCPOPT_FASTOPEN: 4179 tcp_parse_fastopen_option( 4180 opsize - TCPOLEN_FASTOPEN_BASE, 4181 ptr, th->syn, foc, false); 4182 break; 4183 4184 case TCPOPT_EXP: 4185 /* Fast Open option shares code 254 using a 4186 * 16 bits magic number. 4187 */ 4188 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE && 4189 get_unaligned_be16(ptr) == 4190 TCPOPT_FASTOPEN_MAGIC) { 4191 tcp_parse_fastopen_option(opsize - 4192 TCPOLEN_EXP_FASTOPEN_BASE, 4193 ptr + 2, th->syn, foc, true); 4194 break; 4195 } 4196 4197 if (smc_parse_options(th, opt_rx, ptr, opsize)) 4198 break; 4199 4200 opt_rx->saw_unknown = 1; 4201 break; 4202 4203 default: 4204 opt_rx->saw_unknown = 1; 4205 } 4206 ptr += opsize-2; 4207 length -= opsize; 4208 } 4209 } 4210 } 4211 EXPORT_SYMBOL(tcp_parse_options); 4212 4213 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th) 4214 { 4215 const __be32 *ptr = (const __be32 *)(th + 1); 4216 4217 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) 4218 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) { 4219 tp->rx_opt.saw_tstamp = 1; 4220 ++ptr; 4221 tp->rx_opt.rcv_tsval = ntohl(*ptr); 4222 ++ptr; 4223 if (*ptr) 4224 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset; 4225 else 4226 tp->rx_opt.rcv_tsecr = 0; 4227 return true; 4228 } 4229 return false; 4230 } 4231 4232 /* Fast parse options. This hopes to only see timestamps. 4233 * If it is wrong it falls back on tcp_parse_options(). 4234 */ 4235 static bool tcp_fast_parse_options(const struct net *net, 4236 const struct sk_buff *skb, 4237 const struct tcphdr *th, struct tcp_sock *tp) 4238 { 4239 /* In the spirit of fast parsing, compare doff directly to constant 4240 * values. Because equality is used, short doff can be ignored here. 4241 */ 4242 if (th->doff == (sizeof(*th) / 4)) { 4243 tp->rx_opt.saw_tstamp = 0; 4244 return false; 4245 } else if (tp->rx_opt.tstamp_ok && 4246 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) { 4247 if (tcp_parse_aligned_timestamp(tp, th)) 4248 return true; 4249 } 4250 4251 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL); 4252 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) 4253 tp->rx_opt.rcv_tsecr -= tp->tsoffset; 4254 4255 return true; 4256 } 4257 4258 #ifdef CONFIG_TCP_MD5SIG 4259 /* 4260 * Parse MD5 Signature option 4261 */ 4262 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th) 4263 { 4264 int length = (th->doff << 2) - sizeof(*th); 4265 const u8 *ptr = (const u8 *)(th + 1); 4266 4267 /* If not enough data remaining, we can short cut */ 4268 while (length >= TCPOLEN_MD5SIG) { 4269 int opcode = *ptr++; 4270 int opsize; 4271 4272 switch (opcode) { 4273 case TCPOPT_EOL: 4274 return NULL; 4275 case TCPOPT_NOP: 4276 length--; 4277 continue; 4278 default: 4279 opsize = *ptr++; 4280 if (opsize < 2 || opsize > length) 4281 return NULL; 4282 if (opcode == TCPOPT_MD5SIG) 4283 return opsize == TCPOLEN_MD5SIG ? ptr : NULL; 4284 } 4285 ptr += opsize - 2; 4286 length -= opsize; 4287 } 4288 return NULL; 4289 } 4290 EXPORT_SYMBOL(tcp_parse_md5sig_option); 4291 #endif 4292 4293 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM 4294 * 4295 * It is not fatal. If this ACK does _not_ change critical state (seqs, window) 4296 * it can pass through stack. So, the following predicate verifies that 4297 * this segment is not used for anything but congestion avoidance or 4298 * fast retransmit. Moreover, we even are able to eliminate most of such 4299 * second order effects, if we apply some small "replay" window (~RTO) 4300 * to timestamp space. 4301 * 4302 * All these measures still do not guarantee that we reject wrapped ACKs 4303 * on networks with high bandwidth, when sequence space is recycled fastly, 4304 * but it guarantees that such events will be very rare and do not affect 4305 * connection seriously. This doesn't look nice, but alas, PAWS is really 4306 * buggy extension. 4307 * 4308 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC 4309 * states that events when retransmit arrives after original data are rare. 4310 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is 4311 * the biggest problem on large power networks even with minor reordering. 4312 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe 4313 * up to bandwidth of 18Gigabit/sec. 8) ] 4314 */ 4315 4316 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb) 4317 { 4318 const struct tcp_sock *tp = tcp_sk(sk); 4319 const struct tcphdr *th = tcp_hdr(skb); 4320 u32 seq = TCP_SKB_CB(skb)->seq; 4321 u32 ack = TCP_SKB_CB(skb)->ack_seq; 4322 4323 return (/* 1. Pure ACK with correct sequence number. */ 4324 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) && 4325 4326 /* 2. ... and duplicate ACK. */ 4327 ack == tp->snd_una && 4328 4329 /* 3. ... and does not update window. */ 4330 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) && 4331 4332 /* 4. ... and sits in replay window. */ 4333 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ); 4334 } 4335 4336 static inline bool tcp_paws_discard(const struct sock *sk, 4337 const struct sk_buff *skb) 4338 { 4339 const struct tcp_sock *tp = tcp_sk(sk); 4340 4341 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) && 4342 !tcp_disordered_ack(sk, skb); 4343 } 4344 4345 /* Check segment sequence number for validity. 4346 * 4347 * Segment controls are considered valid, if the segment 4348 * fits to the window after truncation to the window. Acceptability 4349 * of data (and SYN, FIN, of course) is checked separately. 4350 * See tcp_data_queue(), for example. 4351 * 4352 * Also, controls (RST is main one) are accepted using RCV.WUP instead 4353 * of RCV.NXT. Peer still did not advance his SND.UNA when we 4354 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP. 4355 * (borrowed from freebsd) 4356 */ 4357 4358 static enum skb_drop_reason tcp_sequence(const struct tcp_sock *tp, 4359 u32 seq, u32 end_seq) 4360 { 4361 if (before(end_seq, tp->rcv_wup)) 4362 return SKB_DROP_REASON_TCP_OLD_SEQUENCE; 4363 4364 if (after(seq, tp->rcv_nxt + tcp_receive_window(tp))) 4365 return SKB_DROP_REASON_TCP_INVALID_SEQUENCE; 4366 4367 return SKB_NOT_DROPPED_YET; 4368 } 4369 4370 /* When we get a reset we do this. */ 4371 void tcp_reset(struct sock *sk, struct sk_buff *skb) 4372 { 4373 trace_tcp_receive_reset(sk); 4374 4375 /* mptcp can't tell us to ignore reset pkts, 4376 * so just ignore the return value of mptcp_incoming_options(). 4377 */ 4378 if (sk_is_mptcp(sk)) 4379 mptcp_incoming_options(sk, skb); 4380 4381 /* We want the right error as BSD sees it (and indeed as we do). */ 4382 switch (sk->sk_state) { 4383 case TCP_SYN_SENT: 4384 WRITE_ONCE(sk->sk_err, ECONNREFUSED); 4385 break; 4386 case TCP_CLOSE_WAIT: 4387 WRITE_ONCE(sk->sk_err, EPIPE); 4388 break; 4389 case TCP_CLOSE: 4390 return; 4391 default: 4392 WRITE_ONCE(sk->sk_err, ECONNRESET); 4393 } 4394 /* This barrier is coupled with smp_rmb() in tcp_poll() */ 4395 smp_wmb(); 4396 4397 tcp_write_queue_purge(sk); 4398 tcp_done(sk); 4399 4400 if (!sock_flag(sk, SOCK_DEAD)) 4401 sk_error_report(sk); 4402 } 4403 4404 /* 4405 * Process the FIN bit. This now behaves as it is supposed to work 4406 * and the FIN takes effect when it is validly part of sequence 4407 * space. Not before when we get holes. 4408 * 4409 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT 4410 * (and thence onto LAST-ACK and finally, CLOSE, we never enter 4411 * TIME-WAIT) 4412 * 4413 * If we are in FINWAIT-1, a received FIN indicates simultaneous 4414 * close and we go into CLOSING (and later onto TIME-WAIT) 4415 * 4416 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT. 4417 */ 4418 void tcp_fin(struct sock *sk) 4419 { 4420 struct tcp_sock *tp = tcp_sk(sk); 4421 4422 inet_csk_schedule_ack(sk); 4423 4424 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN); 4425 sock_set_flag(sk, SOCK_DONE); 4426 4427 switch (sk->sk_state) { 4428 case TCP_SYN_RECV: 4429 case TCP_ESTABLISHED: 4430 /* Move to CLOSE_WAIT */ 4431 tcp_set_state(sk, TCP_CLOSE_WAIT); 4432 inet_csk_enter_pingpong_mode(sk); 4433 break; 4434 4435 case TCP_CLOSE_WAIT: 4436 case TCP_CLOSING: 4437 /* Received a retransmission of the FIN, do 4438 * nothing. 4439 */ 4440 break; 4441 case TCP_LAST_ACK: 4442 /* RFC793: Remain in the LAST-ACK state. */ 4443 break; 4444 4445 case TCP_FIN_WAIT1: 4446 /* This case occurs when a simultaneous close 4447 * happens, we must ack the received FIN and 4448 * enter the CLOSING state. 4449 */ 4450 tcp_send_ack(sk); 4451 tcp_set_state(sk, TCP_CLOSING); 4452 break; 4453 case TCP_FIN_WAIT2: 4454 /* Received a FIN -- send ACK and enter TIME_WAIT. */ 4455 tcp_send_ack(sk); 4456 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 4457 break; 4458 default: 4459 /* Only TCP_LISTEN and TCP_CLOSE are left, in these 4460 * cases we should never reach this piece of code. 4461 */ 4462 pr_err("%s: Impossible, sk->sk_state=%d\n", 4463 __func__, sk->sk_state); 4464 break; 4465 } 4466 4467 /* It _is_ possible, that we have something out-of-order _after_ FIN. 4468 * Probably, we should reset in this case. For now drop them. 4469 */ 4470 skb_rbtree_purge(&tp->out_of_order_queue); 4471 if (tcp_is_sack(tp)) 4472 tcp_sack_reset(&tp->rx_opt); 4473 4474 if (!sock_flag(sk, SOCK_DEAD)) { 4475 sk->sk_state_change(sk); 4476 4477 /* Do not send POLL_HUP for half duplex close. */ 4478 if (sk->sk_shutdown == SHUTDOWN_MASK || 4479 sk->sk_state == TCP_CLOSE) 4480 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP); 4481 else 4482 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN); 4483 } 4484 } 4485 4486 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, 4487 u32 end_seq) 4488 { 4489 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) { 4490 if (before(seq, sp->start_seq)) 4491 sp->start_seq = seq; 4492 if (after(end_seq, sp->end_seq)) 4493 sp->end_seq = end_seq; 4494 return true; 4495 } 4496 return false; 4497 } 4498 4499 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq) 4500 { 4501 struct tcp_sock *tp = tcp_sk(sk); 4502 4503 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) { 4504 int mib_idx; 4505 4506 if (before(seq, tp->rcv_nxt)) 4507 mib_idx = LINUX_MIB_TCPDSACKOLDSENT; 4508 else 4509 mib_idx = LINUX_MIB_TCPDSACKOFOSENT; 4510 4511 NET_INC_STATS(sock_net(sk), mib_idx); 4512 4513 tp->rx_opt.dsack = 1; 4514 tp->duplicate_sack[0].start_seq = seq; 4515 tp->duplicate_sack[0].end_seq = end_seq; 4516 } 4517 } 4518 4519 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq) 4520 { 4521 struct tcp_sock *tp = tcp_sk(sk); 4522 4523 if (!tp->rx_opt.dsack) 4524 tcp_dsack_set(sk, seq, end_seq); 4525 else 4526 tcp_sack_extend(tp->duplicate_sack, seq, end_seq); 4527 } 4528 4529 static void tcp_rcv_spurious_retrans(struct sock *sk, const struct sk_buff *skb) 4530 { 4531 /* When the ACK path fails or drops most ACKs, the sender would 4532 * timeout and spuriously retransmit the same segment repeatedly. 4533 * The receiver remembers and reflects via DSACKs. Leverage the 4534 * DSACK state and change the txhash to re-route speculatively. 4535 */ 4536 if (TCP_SKB_CB(skb)->seq == tcp_sk(sk)->duplicate_sack[0].start_seq && 4537 sk_rethink_txhash(sk)) 4538 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDUPLICATEDATAREHASH); 4539 } 4540 4541 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb) 4542 { 4543 struct tcp_sock *tp = tcp_sk(sk); 4544 4545 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 4546 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4547 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); 4548 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS); 4549 4550 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) { 4551 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 4552 4553 tcp_rcv_spurious_retrans(sk, skb); 4554 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) 4555 end_seq = tp->rcv_nxt; 4556 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq); 4557 } 4558 } 4559 4560 tcp_send_ack(sk); 4561 } 4562 4563 /* These routines update the SACK block as out-of-order packets arrive or 4564 * in-order packets close up the sequence space. 4565 */ 4566 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp) 4567 { 4568 int this_sack; 4569 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4570 struct tcp_sack_block *swalk = sp + 1; 4571 4572 /* See if the recent change to the first SACK eats into 4573 * or hits the sequence space of other SACK blocks, if so coalesce. 4574 */ 4575 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) { 4576 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) { 4577 int i; 4578 4579 /* Zap SWALK, by moving every further SACK up by one slot. 4580 * Decrease num_sacks. 4581 */ 4582 tp->rx_opt.num_sacks--; 4583 for (i = this_sack; i < tp->rx_opt.num_sacks; i++) 4584 sp[i] = sp[i + 1]; 4585 continue; 4586 } 4587 this_sack++; 4588 swalk++; 4589 } 4590 } 4591 4592 void tcp_sack_compress_send_ack(struct sock *sk) 4593 { 4594 struct tcp_sock *tp = tcp_sk(sk); 4595 4596 if (!tp->compressed_ack) 4597 return; 4598 4599 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1) 4600 __sock_put(sk); 4601 4602 /* Since we have to send one ack finally, 4603 * substract one from tp->compressed_ack to keep 4604 * LINUX_MIB_TCPACKCOMPRESSED accurate. 4605 */ 4606 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED, 4607 tp->compressed_ack - 1); 4608 4609 tp->compressed_ack = 0; 4610 tcp_send_ack(sk); 4611 } 4612 4613 /* Reasonable amount of sack blocks included in TCP SACK option 4614 * The max is 4, but this becomes 3 if TCP timestamps are there. 4615 * Given that SACK packets might be lost, be conservative and use 2. 4616 */ 4617 #define TCP_SACK_BLOCKS_EXPECTED 2 4618 4619 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq) 4620 { 4621 struct tcp_sock *tp = tcp_sk(sk); 4622 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4623 int cur_sacks = tp->rx_opt.num_sacks; 4624 int this_sack; 4625 4626 if (!cur_sacks) 4627 goto new_sack; 4628 4629 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) { 4630 if (tcp_sack_extend(sp, seq, end_seq)) { 4631 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED) 4632 tcp_sack_compress_send_ack(sk); 4633 /* Rotate this_sack to the first one. */ 4634 for (; this_sack > 0; this_sack--, sp--) 4635 swap(*sp, *(sp - 1)); 4636 if (cur_sacks > 1) 4637 tcp_sack_maybe_coalesce(tp); 4638 return; 4639 } 4640 } 4641 4642 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED) 4643 tcp_sack_compress_send_ack(sk); 4644 4645 /* Could not find an adjacent existing SACK, build a new one, 4646 * put it at the front, and shift everyone else down. We 4647 * always know there is at least one SACK present already here. 4648 * 4649 * If the sack array is full, forget about the last one. 4650 */ 4651 if (this_sack >= TCP_NUM_SACKS) { 4652 this_sack--; 4653 tp->rx_opt.num_sacks--; 4654 sp--; 4655 } 4656 for (; this_sack > 0; this_sack--, sp--) 4657 *sp = *(sp - 1); 4658 4659 new_sack: 4660 /* Build the new head SACK, and we're done. */ 4661 sp->start_seq = seq; 4662 sp->end_seq = end_seq; 4663 tp->rx_opt.num_sacks++; 4664 } 4665 4666 /* RCV.NXT advances, some SACKs should be eaten. */ 4667 4668 static void tcp_sack_remove(struct tcp_sock *tp) 4669 { 4670 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4671 int num_sacks = tp->rx_opt.num_sacks; 4672 int this_sack; 4673 4674 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */ 4675 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) { 4676 tp->rx_opt.num_sacks = 0; 4677 return; 4678 } 4679 4680 for (this_sack = 0; this_sack < num_sacks;) { 4681 /* Check if the start of the sack is covered by RCV.NXT. */ 4682 if (!before(tp->rcv_nxt, sp->start_seq)) { 4683 int i; 4684 4685 /* RCV.NXT must cover all the block! */ 4686 WARN_ON(before(tp->rcv_nxt, sp->end_seq)); 4687 4688 /* Zap this SACK, by moving forward any other SACKS. */ 4689 for (i = this_sack+1; i < num_sacks; i++) 4690 tp->selective_acks[i-1] = tp->selective_acks[i]; 4691 num_sacks--; 4692 continue; 4693 } 4694 this_sack++; 4695 sp++; 4696 } 4697 tp->rx_opt.num_sacks = num_sacks; 4698 } 4699 4700 /** 4701 * tcp_try_coalesce - try to merge skb to prior one 4702 * @sk: socket 4703 * @to: prior buffer 4704 * @from: buffer to add in queue 4705 * @fragstolen: pointer to boolean 4706 * 4707 * Before queueing skb @from after @to, try to merge them 4708 * to reduce overall memory use and queue lengths, if cost is small. 4709 * Packets in ofo or receive queues can stay a long time. 4710 * Better try to coalesce them right now to avoid future collapses. 4711 * Returns true if caller should free @from instead of queueing it 4712 */ 4713 static bool tcp_try_coalesce(struct sock *sk, 4714 struct sk_buff *to, 4715 struct sk_buff *from, 4716 bool *fragstolen) 4717 { 4718 int delta; 4719 4720 *fragstolen = false; 4721 4722 /* Its possible this segment overlaps with prior segment in queue */ 4723 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq) 4724 return false; 4725 4726 if (!mptcp_skb_can_collapse(to, from)) 4727 return false; 4728 4729 #ifdef CONFIG_TLS_DEVICE 4730 if (from->decrypted != to->decrypted) 4731 return false; 4732 #endif 4733 4734 if (!skb_try_coalesce(to, from, fragstolen, &delta)) 4735 return false; 4736 4737 atomic_add(delta, &sk->sk_rmem_alloc); 4738 sk_mem_charge(sk, delta); 4739 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE); 4740 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq; 4741 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq; 4742 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags; 4743 4744 if (TCP_SKB_CB(from)->has_rxtstamp) { 4745 TCP_SKB_CB(to)->has_rxtstamp = true; 4746 to->tstamp = from->tstamp; 4747 skb_hwtstamps(to)->hwtstamp = skb_hwtstamps(from)->hwtstamp; 4748 } 4749 4750 return true; 4751 } 4752 4753 static bool tcp_ooo_try_coalesce(struct sock *sk, 4754 struct sk_buff *to, 4755 struct sk_buff *from, 4756 bool *fragstolen) 4757 { 4758 bool res = tcp_try_coalesce(sk, to, from, fragstolen); 4759 4760 /* In case tcp_drop_reason() is called later, update to->gso_segs */ 4761 if (res) { 4762 u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) + 4763 max_t(u16, 1, skb_shinfo(from)->gso_segs); 4764 4765 skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF); 4766 } 4767 return res; 4768 } 4769 4770 static void tcp_drop_reason(struct sock *sk, struct sk_buff *skb, 4771 enum skb_drop_reason reason) 4772 { 4773 sk_drops_add(sk, skb); 4774 kfree_skb_reason(skb, reason); 4775 } 4776 4777 /* This one checks to see if we can put data from the 4778 * out_of_order queue into the receive_queue. 4779 */ 4780 static void tcp_ofo_queue(struct sock *sk) 4781 { 4782 struct tcp_sock *tp = tcp_sk(sk); 4783 __u32 dsack_high = tp->rcv_nxt; 4784 bool fin, fragstolen, eaten; 4785 struct sk_buff *skb, *tail; 4786 struct rb_node *p; 4787 4788 p = rb_first(&tp->out_of_order_queue); 4789 while (p) { 4790 skb = rb_to_skb(p); 4791 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 4792 break; 4793 4794 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) { 4795 __u32 dsack = dsack_high; 4796 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high)) 4797 dsack_high = TCP_SKB_CB(skb)->end_seq; 4798 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack); 4799 } 4800 p = rb_next(p); 4801 rb_erase(&skb->rbnode, &tp->out_of_order_queue); 4802 4803 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) { 4804 tcp_drop_reason(sk, skb, SKB_DROP_REASON_TCP_OFO_DROP); 4805 continue; 4806 } 4807 4808 tail = skb_peek_tail(&sk->sk_receive_queue); 4809 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen); 4810 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq); 4811 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN; 4812 if (!eaten) 4813 __skb_queue_tail(&sk->sk_receive_queue, skb); 4814 else 4815 kfree_skb_partial(skb, fragstolen); 4816 4817 if (unlikely(fin)) { 4818 tcp_fin(sk); 4819 /* tcp_fin() purges tp->out_of_order_queue, 4820 * so we must end this loop right now. 4821 */ 4822 break; 4823 } 4824 } 4825 } 4826 4827 static bool tcp_prune_ofo_queue(struct sock *sk, const struct sk_buff *in_skb); 4828 static int tcp_prune_queue(struct sock *sk, const struct sk_buff *in_skb); 4829 4830 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb, 4831 unsigned int size) 4832 { 4833 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || 4834 !sk_rmem_schedule(sk, skb, size)) { 4835 4836 if (tcp_prune_queue(sk, skb) < 0) 4837 return -1; 4838 4839 while (!sk_rmem_schedule(sk, skb, size)) { 4840 if (!tcp_prune_ofo_queue(sk, skb)) 4841 return -1; 4842 } 4843 } 4844 return 0; 4845 } 4846 4847 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb) 4848 { 4849 struct tcp_sock *tp = tcp_sk(sk); 4850 struct rb_node **p, *parent; 4851 struct sk_buff *skb1; 4852 u32 seq, end_seq; 4853 bool fragstolen; 4854 4855 tcp_ecn_check_ce(sk, skb); 4856 4857 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) { 4858 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP); 4859 sk->sk_data_ready(sk); 4860 tcp_drop_reason(sk, skb, SKB_DROP_REASON_PROTO_MEM); 4861 return; 4862 } 4863 4864 /* Disable header prediction. */ 4865 tp->pred_flags = 0; 4866 inet_csk_schedule_ack(sk); 4867 4868 tp->rcv_ooopack += max_t(u16, 1, skb_shinfo(skb)->gso_segs); 4869 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE); 4870 seq = TCP_SKB_CB(skb)->seq; 4871 end_seq = TCP_SKB_CB(skb)->end_seq; 4872 4873 p = &tp->out_of_order_queue.rb_node; 4874 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) { 4875 /* Initial out of order segment, build 1 SACK. */ 4876 if (tcp_is_sack(tp)) { 4877 tp->rx_opt.num_sacks = 1; 4878 tp->selective_acks[0].start_seq = seq; 4879 tp->selective_acks[0].end_seq = end_seq; 4880 } 4881 rb_link_node(&skb->rbnode, NULL, p); 4882 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue); 4883 tp->ooo_last_skb = skb; 4884 goto end; 4885 } 4886 4887 /* In the typical case, we are adding an skb to the end of the list. 4888 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup. 4889 */ 4890 if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb, 4891 skb, &fragstolen)) { 4892 coalesce_done: 4893 /* For non sack flows, do not grow window to force DUPACK 4894 * and trigger fast retransmit. 4895 */ 4896 if (tcp_is_sack(tp)) 4897 tcp_grow_window(sk, skb, true); 4898 kfree_skb_partial(skb, fragstolen); 4899 skb = NULL; 4900 goto add_sack; 4901 } 4902 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */ 4903 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) { 4904 parent = &tp->ooo_last_skb->rbnode; 4905 p = &parent->rb_right; 4906 goto insert; 4907 } 4908 4909 /* Find place to insert this segment. Handle overlaps on the way. */ 4910 parent = NULL; 4911 while (*p) { 4912 parent = *p; 4913 skb1 = rb_to_skb(parent); 4914 if (before(seq, TCP_SKB_CB(skb1)->seq)) { 4915 p = &parent->rb_left; 4916 continue; 4917 } 4918 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) { 4919 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 4920 /* All the bits are present. Drop. */ 4921 NET_INC_STATS(sock_net(sk), 4922 LINUX_MIB_TCPOFOMERGE); 4923 tcp_drop_reason(sk, skb, 4924 SKB_DROP_REASON_TCP_OFOMERGE); 4925 skb = NULL; 4926 tcp_dsack_set(sk, seq, end_seq); 4927 goto add_sack; 4928 } 4929 if (after(seq, TCP_SKB_CB(skb1)->seq)) { 4930 /* Partial overlap. */ 4931 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq); 4932 } else { 4933 /* skb's seq == skb1's seq and skb covers skb1. 4934 * Replace skb1 with skb. 4935 */ 4936 rb_replace_node(&skb1->rbnode, &skb->rbnode, 4937 &tp->out_of_order_queue); 4938 tcp_dsack_extend(sk, 4939 TCP_SKB_CB(skb1)->seq, 4940 TCP_SKB_CB(skb1)->end_seq); 4941 NET_INC_STATS(sock_net(sk), 4942 LINUX_MIB_TCPOFOMERGE); 4943 tcp_drop_reason(sk, skb1, 4944 SKB_DROP_REASON_TCP_OFOMERGE); 4945 goto merge_right; 4946 } 4947 } else if (tcp_ooo_try_coalesce(sk, skb1, 4948 skb, &fragstolen)) { 4949 goto coalesce_done; 4950 } 4951 p = &parent->rb_right; 4952 } 4953 insert: 4954 /* Insert segment into RB tree. */ 4955 rb_link_node(&skb->rbnode, parent, p); 4956 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue); 4957 4958 merge_right: 4959 /* Remove other segments covered by skb. */ 4960 while ((skb1 = skb_rb_next(skb)) != NULL) { 4961 if (!after(end_seq, TCP_SKB_CB(skb1)->seq)) 4962 break; 4963 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 4964 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, 4965 end_seq); 4966 break; 4967 } 4968 rb_erase(&skb1->rbnode, &tp->out_of_order_queue); 4969 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, 4970 TCP_SKB_CB(skb1)->end_seq); 4971 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE); 4972 tcp_drop_reason(sk, skb1, SKB_DROP_REASON_TCP_OFOMERGE); 4973 } 4974 /* If there is no skb after us, we are the last_skb ! */ 4975 if (!skb1) 4976 tp->ooo_last_skb = skb; 4977 4978 add_sack: 4979 if (tcp_is_sack(tp)) 4980 tcp_sack_new_ofo_skb(sk, seq, end_seq); 4981 end: 4982 if (skb) { 4983 /* For non sack flows, do not grow window to force DUPACK 4984 * and trigger fast retransmit. 4985 */ 4986 if (tcp_is_sack(tp)) 4987 tcp_grow_window(sk, skb, false); 4988 skb_condense(skb); 4989 skb_set_owner_r(skb, sk); 4990 } 4991 } 4992 4993 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, 4994 bool *fragstolen) 4995 { 4996 int eaten; 4997 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue); 4998 4999 eaten = (tail && 5000 tcp_try_coalesce(sk, tail, 5001 skb, fragstolen)) ? 1 : 0; 5002 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq); 5003 if (!eaten) { 5004 __skb_queue_tail(&sk->sk_receive_queue, skb); 5005 skb_set_owner_r(skb, sk); 5006 } 5007 return eaten; 5008 } 5009 5010 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size) 5011 { 5012 struct sk_buff *skb; 5013 int err = -ENOMEM; 5014 int data_len = 0; 5015 bool fragstolen; 5016 5017 if (size == 0) 5018 return 0; 5019 5020 if (size > PAGE_SIZE) { 5021 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS); 5022 5023 data_len = npages << PAGE_SHIFT; 5024 size = data_len + (size & ~PAGE_MASK); 5025 } 5026 skb = alloc_skb_with_frags(size - data_len, data_len, 5027 PAGE_ALLOC_COSTLY_ORDER, 5028 &err, sk->sk_allocation); 5029 if (!skb) 5030 goto err; 5031 5032 skb_put(skb, size - data_len); 5033 skb->data_len = data_len; 5034 skb->len = size; 5035 5036 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) { 5037 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP); 5038 goto err_free; 5039 } 5040 5041 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size); 5042 if (err) 5043 goto err_free; 5044 5045 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt; 5046 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size; 5047 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1; 5048 5049 if (tcp_queue_rcv(sk, skb, &fragstolen)) { 5050 WARN_ON_ONCE(fragstolen); /* should not happen */ 5051 __kfree_skb(skb); 5052 } 5053 return size; 5054 5055 err_free: 5056 kfree_skb(skb); 5057 err: 5058 return err; 5059 5060 } 5061 5062 void tcp_data_ready(struct sock *sk) 5063 { 5064 if (tcp_epollin_ready(sk, sk->sk_rcvlowat) || sock_flag(sk, SOCK_DONE)) 5065 sk->sk_data_ready(sk); 5066 } 5067 5068 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb) 5069 { 5070 struct tcp_sock *tp = tcp_sk(sk); 5071 enum skb_drop_reason reason; 5072 bool fragstolen; 5073 int eaten; 5074 5075 /* If a subflow has been reset, the packet should not continue 5076 * to be processed, drop the packet. 5077 */ 5078 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb)) { 5079 __kfree_skb(skb); 5080 return; 5081 } 5082 5083 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) { 5084 __kfree_skb(skb); 5085 return; 5086 } 5087 skb_dst_drop(skb); 5088 __skb_pull(skb, tcp_hdr(skb)->doff * 4); 5089 5090 reason = SKB_DROP_REASON_NOT_SPECIFIED; 5091 tp->rx_opt.dsack = 0; 5092 5093 /* Queue data for delivery to the user. 5094 * Packets in sequence go to the receive queue. 5095 * Out of sequence packets to the out_of_order_queue. 5096 */ 5097 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { 5098 if (tcp_receive_window(tp) == 0) { 5099 reason = SKB_DROP_REASON_TCP_ZEROWINDOW; 5100 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP); 5101 goto out_of_window; 5102 } 5103 5104 /* Ok. In sequence. In window. */ 5105 queue_and_out: 5106 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) { 5107 /* TODO: maybe ratelimit these WIN 0 ACK ? */ 5108 inet_csk(sk)->icsk_ack.pending |= 5109 (ICSK_ACK_NOMEM | ICSK_ACK_NOW); 5110 inet_csk_schedule_ack(sk); 5111 sk->sk_data_ready(sk); 5112 5113 if (skb_queue_len(&sk->sk_receive_queue)) { 5114 reason = SKB_DROP_REASON_PROTO_MEM; 5115 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP); 5116 goto drop; 5117 } 5118 sk_forced_mem_schedule(sk, skb->truesize); 5119 } 5120 5121 eaten = tcp_queue_rcv(sk, skb, &fragstolen); 5122 if (skb->len) 5123 tcp_event_data_recv(sk, skb); 5124 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 5125 tcp_fin(sk); 5126 5127 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) { 5128 tcp_ofo_queue(sk); 5129 5130 /* RFC5681. 4.2. SHOULD send immediate ACK, when 5131 * gap in queue is filled. 5132 */ 5133 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) 5134 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW; 5135 } 5136 5137 if (tp->rx_opt.num_sacks) 5138 tcp_sack_remove(tp); 5139 5140 tcp_fast_path_check(sk); 5141 5142 if (eaten > 0) 5143 kfree_skb_partial(skb, fragstolen); 5144 if (!sock_flag(sk, SOCK_DEAD)) 5145 tcp_data_ready(sk); 5146 return; 5147 } 5148 5149 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 5150 tcp_rcv_spurious_retrans(sk, skb); 5151 /* A retransmit, 2nd most common case. Force an immediate ack. */ 5152 reason = SKB_DROP_REASON_TCP_OLD_DATA; 5153 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); 5154 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 5155 5156 out_of_window: 5157 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS); 5158 inet_csk_schedule_ack(sk); 5159 drop: 5160 tcp_drop_reason(sk, skb, reason); 5161 return; 5162 } 5163 5164 /* Out of window. F.e. zero window probe. */ 5165 if (!before(TCP_SKB_CB(skb)->seq, 5166 tp->rcv_nxt + tcp_receive_window(tp))) { 5167 reason = SKB_DROP_REASON_TCP_OVERWINDOW; 5168 goto out_of_window; 5169 } 5170 5171 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 5172 /* Partial packet, seq < rcv_next < end_seq */ 5173 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt); 5174 5175 /* If window is closed, drop tail of packet. But after 5176 * remembering D-SACK for its head made in previous line. 5177 */ 5178 if (!tcp_receive_window(tp)) { 5179 reason = SKB_DROP_REASON_TCP_ZEROWINDOW; 5180 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP); 5181 goto out_of_window; 5182 } 5183 goto queue_and_out; 5184 } 5185 5186 tcp_data_queue_ofo(sk, skb); 5187 } 5188 5189 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list) 5190 { 5191 if (list) 5192 return !skb_queue_is_last(list, skb) ? skb->next : NULL; 5193 5194 return skb_rb_next(skb); 5195 } 5196 5197 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb, 5198 struct sk_buff_head *list, 5199 struct rb_root *root) 5200 { 5201 struct sk_buff *next = tcp_skb_next(skb, list); 5202 5203 if (list) 5204 __skb_unlink(skb, list); 5205 else 5206 rb_erase(&skb->rbnode, root); 5207 5208 __kfree_skb(skb); 5209 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED); 5210 5211 return next; 5212 } 5213 5214 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */ 5215 void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb) 5216 { 5217 struct rb_node **p = &root->rb_node; 5218 struct rb_node *parent = NULL; 5219 struct sk_buff *skb1; 5220 5221 while (*p) { 5222 parent = *p; 5223 skb1 = rb_to_skb(parent); 5224 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq)) 5225 p = &parent->rb_left; 5226 else 5227 p = &parent->rb_right; 5228 } 5229 rb_link_node(&skb->rbnode, parent, p); 5230 rb_insert_color(&skb->rbnode, root); 5231 } 5232 5233 /* Collapse contiguous sequence of skbs head..tail with 5234 * sequence numbers start..end. 5235 * 5236 * If tail is NULL, this means until the end of the queue. 5237 * 5238 * Segments with FIN/SYN are not collapsed (only because this 5239 * simplifies code) 5240 */ 5241 static void 5242 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root, 5243 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end) 5244 { 5245 struct sk_buff *skb = head, *n; 5246 struct sk_buff_head tmp; 5247 bool end_of_skbs; 5248 5249 /* First, check that queue is collapsible and find 5250 * the point where collapsing can be useful. 5251 */ 5252 restart: 5253 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) { 5254 n = tcp_skb_next(skb, list); 5255 5256 /* No new bits? It is possible on ofo queue. */ 5257 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 5258 skb = tcp_collapse_one(sk, skb, list, root); 5259 if (!skb) 5260 break; 5261 goto restart; 5262 } 5263 5264 /* The first skb to collapse is: 5265 * - not SYN/FIN and 5266 * - bloated or contains data before "start" or 5267 * overlaps to the next one and mptcp allow collapsing. 5268 */ 5269 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) && 5270 (tcp_win_from_space(sk, skb->truesize) > skb->len || 5271 before(TCP_SKB_CB(skb)->seq, start))) { 5272 end_of_skbs = false; 5273 break; 5274 } 5275 5276 if (n && n != tail && mptcp_skb_can_collapse(skb, n) && 5277 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) { 5278 end_of_skbs = false; 5279 break; 5280 } 5281 5282 /* Decided to skip this, advance start seq. */ 5283 start = TCP_SKB_CB(skb)->end_seq; 5284 } 5285 if (end_of_skbs || 5286 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN))) 5287 return; 5288 5289 __skb_queue_head_init(&tmp); 5290 5291 while (before(start, end)) { 5292 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start); 5293 struct sk_buff *nskb; 5294 5295 nskb = alloc_skb(copy, GFP_ATOMIC); 5296 if (!nskb) 5297 break; 5298 5299 memcpy(nskb->cb, skb->cb, sizeof(skb->cb)); 5300 #ifdef CONFIG_TLS_DEVICE 5301 nskb->decrypted = skb->decrypted; 5302 #endif 5303 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start; 5304 if (list) 5305 __skb_queue_before(list, skb, nskb); 5306 else 5307 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */ 5308 skb_set_owner_r(nskb, sk); 5309 mptcp_skb_ext_move(nskb, skb); 5310 5311 /* Copy data, releasing collapsed skbs. */ 5312 while (copy > 0) { 5313 int offset = start - TCP_SKB_CB(skb)->seq; 5314 int size = TCP_SKB_CB(skb)->end_seq - start; 5315 5316 BUG_ON(offset < 0); 5317 if (size > 0) { 5318 size = min(copy, size); 5319 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size)) 5320 BUG(); 5321 TCP_SKB_CB(nskb)->end_seq += size; 5322 copy -= size; 5323 start += size; 5324 } 5325 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 5326 skb = tcp_collapse_one(sk, skb, list, root); 5327 if (!skb || 5328 skb == tail || 5329 !mptcp_skb_can_collapse(nskb, skb) || 5330 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN))) 5331 goto end; 5332 #ifdef CONFIG_TLS_DEVICE 5333 if (skb->decrypted != nskb->decrypted) 5334 goto end; 5335 #endif 5336 } 5337 } 5338 } 5339 end: 5340 skb_queue_walk_safe(&tmp, skb, n) 5341 tcp_rbtree_insert(root, skb); 5342 } 5343 5344 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs 5345 * and tcp_collapse() them until all the queue is collapsed. 5346 */ 5347 static void tcp_collapse_ofo_queue(struct sock *sk) 5348 { 5349 struct tcp_sock *tp = tcp_sk(sk); 5350 u32 range_truesize, sum_tiny = 0; 5351 struct sk_buff *skb, *head; 5352 u32 start, end; 5353 5354 skb = skb_rb_first(&tp->out_of_order_queue); 5355 new_range: 5356 if (!skb) { 5357 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue); 5358 return; 5359 } 5360 start = TCP_SKB_CB(skb)->seq; 5361 end = TCP_SKB_CB(skb)->end_seq; 5362 range_truesize = skb->truesize; 5363 5364 for (head = skb;;) { 5365 skb = skb_rb_next(skb); 5366 5367 /* Range is terminated when we see a gap or when 5368 * we are at the queue end. 5369 */ 5370 if (!skb || 5371 after(TCP_SKB_CB(skb)->seq, end) || 5372 before(TCP_SKB_CB(skb)->end_seq, start)) { 5373 /* Do not attempt collapsing tiny skbs */ 5374 if (range_truesize != head->truesize || 5375 end - start >= SKB_WITH_OVERHEAD(PAGE_SIZE)) { 5376 tcp_collapse(sk, NULL, &tp->out_of_order_queue, 5377 head, skb, start, end); 5378 } else { 5379 sum_tiny += range_truesize; 5380 if (sum_tiny > sk->sk_rcvbuf >> 3) 5381 return; 5382 } 5383 goto new_range; 5384 } 5385 5386 range_truesize += skb->truesize; 5387 if (unlikely(before(TCP_SKB_CB(skb)->seq, start))) 5388 start = TCP_SKB_CB(skb)->seq; 5389 if (after(TCP_SKB_CB(skb)->end_seq, end)) 5390 end = TCP_SKB_CB(skb)->end_seq; 5391 } 5392 } 5393 5394 /* 5395 * Clean the out-of-order queue to make room. 5396 * We drop high sequences packets to : 5397 * 1) Let a chance for holes to be filled. 5398 * This means we do not drop packets from ooo queue if their sequence 5399 * is before incoming packet sequence. 5400 * 2) not add too big latencies if thousands of packets sit there. 5401 * (But if application shrinks SO_RCVBUF, we could still end up 5402 * freeing whole queue here) 5403 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks. 5404 * 5405 * Return true if queue has shrunk. 5406 */ 5407 static bool tcp_prune_ofo_queue(struct sock *sk, const struct sk_buff *in_skb) 5408 { 5409 struct tcp_sock *tp = tcp_sk(sk); 5410 struct rb_node *node, *prev; 5411 bool pruned = false; 5412 int goal; 5413 5414 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) 5415 return false; 5416 5417 goal = sk->sk_rcvbuf >> 3; 5418 node = &tp->ooo_last_skb->rbnode; 5419 5420 do { 5421 struct sk_buff *skb = rb_to_skb(node); 5422 5423 /* If incoming skb would land last in ofo queue, stop pruning. */ 5424 if (after(TCP_SKB_CB(in_skb)->seq, TCP_SKB_CB(skb)->seq)) 5425 break; 5426 pruned = true; 5427 prev = rb_prev(node); 5428 rb_erase(node, &tp->out_of_order_queue); 5429 goal -= skb->truesize; 5430 tcp_drop_reason(sk, skb, SKB_DROP_REASON_TCP_OFO_QUEUE_PRUNE); 5431 tp->ooo_last_skb = rb_to_skb(prev); 5432 if (!prev || goal <= 0) { 5433 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf && 5434 !tcp_under_memory_pressure(sk)) 5435 break; 5436 goal = sk->sk_rcvbuf >> 3; 5437 } 5438 node = prev; 5439 } while (node); 5440 5441 if (pruned) { 5442 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED); 5443 /* Reset SACK state. A conforming SACK implementation will 5444 * do the same at a timeout based retransmit. When a connection 5445 * is in a sad state like this, we care only about integrity 5446 * of the connection not performance. 5447 */ 5448 if (tp->rx_opt.sack_ok) 5449 tcp_sack_reset(&tp->rx_opt); 5450 } 5451 return pruned; 5452 } 5453 5454 /* Reduce allocated memory if we can, trying to get 5455 * the socket within its memory limits again. 5456 * 5457 * Return less than zero if we should start dropping frames 5458 * until the socket owning process reads some of the data 5459 * to stabilize the situation. 5460 */ 5461 static int tcp_prune_queue(struct sock *sk, const struct sk_buff *in_skb) 5462 { 5463 struct tcp_sock *tp = tcp_sk(sk); 5464 5465 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED); 5466 5467 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf) 5468 tcp_clamp_window(sk); 5469 else if (tcp_under_memory_pressure(sk)) 5470 tcp_adjust_rcv_ssthresh(sk); 5471 5472 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 5473 return 0; 5474 5475 tcp_collapse_ofo_queue(sk); 5476 if (!skb_queue_empty(&sk->sk_receive_queue)) 5477 tcp_collapse(sk, &sk->sk_receive_queue, NULL, 5478 skb_peek(&sk->sk_receive_queue), 5479 NULL, 5480 tp->copied_seq, tp->rcv_nxt); 5481 5482 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 5483 return 0; 5484 5485 /* Collapsing did not help, destructive actions follow. 5486 * This must not ever occur. */ 5487 5488 tcp_prune_ofo_queue(sk, in_skb); 5489 5490 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 5491 return 0; 5492 5493 /* If we are really being abused, tell the caller to silently 5494 * drop receive data on the floor. It will get retransmitted 5495 * and hopefully then we'll have sufficient space. 5496 */ 5497 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED); 5498 5499 /* Massive buffer overcommit. */ 5500 tp->pred_flags = 0; 5501 return -1; 5502 } 5503 5504 static bool tcp_should_expand_sndbuf(struct sock *sk) 5505 { 5506 const struct tcp_sock *tp = tcp_sk(sk); 5507 5508 /* If the user specified a specific send buffer setting, do 5509 * not modify it. 5510 */ 5511 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK) 5512 return false; 5513 5514 /* If we are under global TCP memory pressure, do not expand. */ 5515 if (tcp_under_memory_pressure(sk)) { 5516 int unused_mem = sk_unused_reserved_mem(sk); 5517 5518 /* Adjust sndbuf according to reserved mem. But make sure 5519 * it never goes below SOCK_MIN_SNDBUF. 5520 * See sk_stream_moderate_sndbuf() for more details. 5521 */ 5522 if (unused_mem > SOCK_MIN_SNDBUF) 5523 WRITE_ONCE(sk->sk_sndbuf, unused_mem); 5524 5525 return false; 5526 } 5527 5528 /* If we are under soft global TCP memory pressure, do not expand. */ 5529 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0)) 5530 return false; 5531 5532 /* If we filled the congestion window, do not expand. */ 5533 if (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp)) 5534 return false; 5535 5536 return true; 5537 } 5538 5539 static void tcp_new_space(struct sock *sk) 5540 { 5541 struct tcp_sock *tp = tcp_sk(sk); 5542 5543 if (tcp_should_expand_sndbuf(sk)) { 5544 tcp_sndbuf_expand(sk); 5545 tp->snd_cwnd_stamp = tcp_jiffies32; 5546 } 5547 5548 INDIRECT_CALL_1(sk->sk_write_space, sk_stream_write_space, sk); 5549 } 5550 5551 /* Caller made space either from: 5552 * 1) Freeing skbs in rtx queues (after tp->snd_una has advanced) 5553 * 2) Sent skbs from output queue (and thus advancing tp->snd_nxt) 5554 * 5555 * We might be able to generate EPOLLOUT to the application if: 5556 * 1) Space consumed in output/rtx queues is below sk->sk_sndbuf/2 5557 * 2) notsent amount (tp->write_seq - tp->snd_nxt) became 5558 * small enough that tcp_stream_memory_free() decides it 5559 * is time to generate EPOLLOUT. 5560 */ 5561 void tcp_check_space(struct sock *sk) 5562 { 5563 /* pairs with tcp_poll() */ 5564 smp_mb(); 5565 if (sk->sk_socket && 5566 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 5567 tcp_new_space(sk); 5568 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) 5569 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED); 5570 } 5571 } 5572 5573 static inline void tcp_data_snd_check(struct sock *sk) 5574 { 5575 tcp_push_pending_frames(sk); 5576 tcp_check_space(sk); 5577 } 5578 5579 /* 5580 * Check if sending an ack is needed. 5581 */ 5582 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible) 5583 { 5584 struct tcp_sock *tp = tcp_sk(sk); 5585 unsigned long rtt, delay; 5586 5587 /* More than one full frame received... */ 5588 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss && 5589 /* ... and right edge of window advances far enough. 5590 * (tcp_recvmsg() will send ACK otherwise). 5591 * If application uses SO_RCVLOWAT, we want send ack now if 5592 * we have not received enough bytes to satisfy the condition. 5593 */ 5594 (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat || 5595 __tcp_select_window(sk) >= tp->rcv_wnd)) || 5596 /* We ACK each frame or... */ 5597 tcp_in_quickack_mode(sk) || 5598 /* Protocol state mandates a one-time immediate ACK */ 5599 inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) { 5600 send_now: 5601 tcp_send_ack(sk); 5602 return; 5603 } 5604 5605 if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) { 5606 tcp_send_delayed_ack(sk); 5607 return; 5608 } 5609 5610 if (!tcp_is_sack(tp) || 5611 tp->compressed_ack >= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr)) 5612 goto send_now; 5613 5614 if (tp->compressed_ack_rcv_nxt != tp->rcv_nxt) { 5615 tp->compressed_ack_rcv_nxt = tp->rcv_nxt; 5616 tp->dup_ack_counter = 0; 5617 } 5618 if (tp->dup_ack_counter < TCP_FASTRETRANS_THRESH) { 5619 tp->dup_ack_counter++; 5620 goto send_now; 5621 } 5622 tp->compressed_ack++; 5623 if (hrtimer_is_queued(&tp->compressed_ack_timer)) 5624 return; 5625 5626 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */ 5627 5628 rtt = tp->rcv_rtt_est.rtt_us; 5629 if (tp->srtt_us && tp->srtt_us < rtt) 5630 rtt = tp->srtt_us; 5631 5632 delay = min_t(unsigned long, 5633 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns), 5634 rtt * (NSEC_PER_USEC >> 3)/20); 5635 sock_hold(sk); 5636 hrtimer_start_range_ns(&tp->compressed_ack_timer, ns_to_ktime(delay), 5637 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_slack_ns), 5638 HRTIMER_MODE_REL_PINNED_SOFT); 5639 } 5640 5641 static inline void tcp_ack_snd_check(struct sock *sk) 5642 { 5643 if (!inet_csk_ack_scheduled(sk)) { 5644 /* We sent a data segment already. */ 5645 return; 5646 } 5647 __tcp_ack_snd_check(sk, 1); 5648 } 5649 5650 /* 5651 * This routine is only called when we have urgent data 5652 * signaled. Its the 'slow' part of tcp_urg. It could be 5653 * moved inline now as tcp_urg is only called from one 5654 * place. We handle URGent data wrong. We have to - as 5655 * BSD still doesn't use the correction from RFC961. 5656 * For 1003.1g we should support a new option TCP_STDURG to permit 5657 * either form (or just set the sysctl tcp_stdurg). 5658 */ 5659 5660 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th) 5661 { 5662 struct tcp_sock *tp = tcp_sk(sk); 5663 u32 ptr = ntohs(th->urg_ptr); 5664 5665 if (ptr && !READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_stdurg)) 5666 ptr--; 5667 ptr += ntohl(th->seq); 5668 5669 /* Ignore urgent data that we've already seen and read. */ 5670 if (after(tp->copied_seq, ptr)) 5671 return; 5672 5673 /* Do not replay urg ptr. 5674 * 5675 * NOTE: interesting situation not covered by specs. 5676 * Misbehaving sender may send urg ptr, pointing to segment, 5677 * which we already have in ofo queue. We are not able to fetch 5678 * such data and will stay in TCP_URG_NOTYET until will be eaten 5679 * by recvmsg(). Seems, we are not obliged to handle such wicked 5680 * situations. But it is worth to think about possibility of some 5681 * DoSes using some hypothetical application level deadlock. 5682 */ 5683 if (before(ptr, tp->rcv_nxt)) 5684 return; 5685 5686 /* Do we already have a newer (or duplicate) urgent pointer? */ 5687 if (tp->urg_data && !after(ptr, tp->urg_seq)) 5688 return; 5689 5690 /* Tell the world about our new urgent pointer. */ 5691 sk_send_sigurg(sk); 5692 5693 /* We may be adding urgent data when the last byte read was 5694 * urgent. To do this requires some care. We cannot just ignore 5695 * tp->copied_seq since we would read the last urgent byte again 5696 * as data, nor can we alter copied_seq until this data arrives 5697 * or we break the semantics of SIOCATMARK (and thus sockatmark()) 5698 * 5699 * NOTE. Double Dutch. Rendering to plain English: author of comment 5700 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB); 5701 * and expect that both A and B disappear from stream. This is _wrong_. 5702 * Though this happens in BSD with high probability, this is occasional. 5703 * Any application relying on this is buggy. Note also, that fix "works" 5704 * only in this artificial test. Insert some normal data between A and B and we will 5705 * decline of BSD again. Verdict: it is better to remove to trap 5706 * buggy users. 5707 */ 5708 if (tp->urg_seq == tp->copied_seq && tp->urg_data && 5709 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) { 5710 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 5711 tp->copied_seq++; 5712 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) { 5713 __skb_unlink(skb, &sk->sk_receive_queue); 5714 __kfree_skb(skb); 5715 } 5716 } 5717 5718 WRITE_ONCE(tp->urg_data, TCP_URG_NOTYET); 5719 WRITE_ONCE(tp->urg_seq, ptr); 5720 5721 /* Disable header prediction. */ 5722 tp->pred_flags = 0; 5723 } 5724 5725 /* This is the 'fast' part of urgent handling. */ 5726 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th) 5727 { 5728 struct tcp_sock *tp = tcp_sk(sk); 5729 5730 /* Check if we get a new urgent pointer - normally not. */ 5731 if (unlikely(th->urg)) 5732 tcp_check_urg(sk, th); 5733 5734 /* Do we wait for any urgent data? - normally not... */ 5735 if (unlikely(tp->urg_data == TCP_URG_NOTYET)) { 5736 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) - 5737 th->syn; 5738 5739 /* Is the urgent pointer pointing into this packet? */ 5740 if (ptr < skb->len) { 5741 u8 tmp; 5742 if (skb_copy_bits(skb, ptr, &tmp, 1)) 5743 BUG(); 5744 WRITE_ONCE(tp->urg_data, TCP_URG_VALID | tmp); 5745 if (!sock_flag(sk, SOCK_DEAD)) 5746 sk->sk_data_ready(sk); 5747 } 5748 } 5749 } 5750 5751 /* Accept RST for rcv_nxt - 1 after a FIN. 5752 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a 5753 * FIN is sent followed by a RST packet. The RST is sent with the same 5754 * sequence number as the FIN, and thus according to RFC 5961 a challenge 5755 * ACK should be sent. However, Mac OSX rate limits replies to challenge 5756 * ACKs on the closed socket. In addition middleboxes can drop either the 5757 * challenge ACK or a subsequent RST. 5758 */ 5759 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb) 5760 { 5761 const struct tcp_sock *tp = tcp_sk(sk); 5762 5763 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) && 5764 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK | 5765 TCPF_CLOSING)); 5766 } 5767 5768 /* Does PAWS and seqno based validation of an incoming segment, flags will 5769 * play significant role here. 5770 */ 5771 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb, 5772 const struct tcphdr *th, int syn_inerr) 5773 { 5774 struct tcp_sock *tp = tcp_sk(sk); 5775 SKB_DR(reason); 5776 5777 /* RFC1323: H1. Apply PAWS check first. */ 5778 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) && 5779 tp->rx_opt.saw_tstamp && 5780 tcp_paws_discard(sk, skb)) { 5781 if (!th->rst) { 5782 if (unlikely(th->syn)) 5783 goto syn_challenge; 5784 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED); 5785 if (!tcp_oow_rate_limited(sock_net(sk), skb, 5786 LINUX_MIB_TCPACKSKIPPEDPAWS, 5787 &tp->last_oow_ack_time)) 5788 tcp_send_dupack(sk, skb); 5789 SKB_DR_SET(reason, TCP_RFC7323_PAWS); 5790 goto discard; 5791 } 5792 /* Reset is accepted even if it did not pass PAWS. */ 5793 } 5794 5795 /* Step 1: check sequence number */ 5796 reason = tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 5797 if (reason) { 5798 /* RFC793, page 37: "In all states except SYN-SENT, all reset 5799 * (RST) segments are validated by checking their SEQ-fields." 5800 * And page 69: "If an incoming segment is not acceptable, 5801 * an acknowledgment should be sent in reply (unless the RST 5802 * bit is set, if so drop the segment and return)". 5803 */ 5804 if (!th->rst) { 5805 if (th->syn) 5806 goto syn_challenge; 5807 if (!tcp_oow_rate_limited(sock_net(sk), skb, 5808 LINUX_MIB_TCPACKSKIPPEDSEQ, 5809 &tp->last_oow_ack_time)) 5810 tcp_send_dupack(sk, skb); 5811 } else if (tcp_reset_check(sk, skb)) { 5812 goto reset; 5813 } 5814 goto discard; 5815 } 5816 5817 /* Step 2: check RST bit */ 5818 if (th->rst) { 5819 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a 5820 * FIN and SACK too if available): 5821 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or 5822 * the right-most SACK block, 5823 * then 5824 * RESET the connection 5825 * else 5826 * Send a challenge ACK 5827 */ 5828 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt || 5829 tcp_reset_check(sk, skb)) 5830 goto reset; 5831 5832 if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) { 5833 struct tcp_sack_block *sp = &tp->selective_acks[0]; 5834 int max_sack = sp[0].end_seq; 5835 int this_sack; 5836 5837 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; 5838 ++this_sack) { 5839 max_sack = after(sp[this_sack].end_seq, 5840 max_sack) ? 5841 sp[this_sack].end_seq : max_sack; 5842 } 5843 5844 if (TCP_SKB_CB(skb)->seq == max_sack) 5845 goto reset; 5846 } 5847 5848 /* Disable TFO if RST is out-of-order 5849 * and no data has been received 5850 * for current active TFO socket 5851 */ 5852 if (tp->syn_fastopen && !tp->data_segs_in && 5853 sk->sk_state == TCP_ESTABLISHED) 5854 tcp_fastopen_active_disable(sk); 5855 tcp_send_challenge_ack(sk); 5856 SKB_DR_SET(reason, TCP_RESET); 5857 goto discard; 5858 } 5859 5860 /* step 3: check security and precedence [ignored] */ 5861 5862 /* step 4: Check for a SYN 5863 * RFC 5961 4.2 : Send a challenge ack 5864 */ 5865 if (th->syn) { 5866 syn_challenge: 5867 if (syn_inerr) 5868 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS); 5869 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE); 5870 tcp_send_challenge_ack(sk); 5871 SKB_DR_SET(reason, TCP_INVALID_SYN); 5872 goto discard; 5873 } 5874 5875 bpf_skops_parse_hdr(sk, skb); 5876 5877 return true; 5878 5879 discard: 5880 tcp_drop_reason(sk, skb, reason); 5881 return false; 5882 5883 reset: 5884 tcp_reset(sk, skb); 5885 __kfree_skb(skb); 5886 return false; 5887 } 5888 5889 /* 5890 * TCP receive function for the ESTABLISHED state. 5891 * 5892 * It is split into a fast path and a slow path. The fast path is 5893 * disabled when: 5894 * - A zero window was announced from us - zero window probing 5895 * is only handled properly in the slow path. 5896 * - Out of order segments arrived. 5897 * - Urgent data is expected. 5898 * - There is no buffer space left 5899 * - Unexpected TCP flags/window values/header lengths are received 5900 * (detected by checking the TCP header against pred_flags) 5901 * - Data is sent in both directions. Fast path only supports pure senders 5902 * or pure receivers (this means either the sequence number or the ack 5903 * value must stay constant) 5904 * - Unexpected TCP option. 5905 * 5906 * When these conditions are not satisfied it drops into a standard 5907 * receive procedure patterned after RFC793 to handle all cases. 5908 * The first three cases are guaranteed by proper pred_flags setting, 5909 * the rest is checked inline. Fast processing is turned on in 5910 * tcp_data_queue when everything is OK. 5911 */ 5912 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb) 5913 { 5914 enum skb_drop_reason reason = SKB_DROP_REASON_NOT_SPECIFIED; 5915 const struct tcphdr *th = (const struct tcphdr *)skb->data; 5916 struct tcp_sock *tp = tcp_sk(sk); 5917 unsigned int len = skb->len; 5918 5919 /* TCP congestion window tracking */ 5920 trace_tcp_probe(sk, skb); 5921 5922 tcp_mstamp_refresh(tp); 5923 if (unlikely(!rcu_access_pointer(sk->sk_rx_dst))) 5924 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb); 5925 /* 5926 * Header prediction. 5927 * The code loosely follows the one in the famous 5928 * "30 instruction TCP receive" Van Jacobson mail. 5929 * 5930 * Van's trick is to deposit buffers into socket queue 5931 * on a device interrupt, to call tcp_recv function 5932 * on the receive process context and checksum and copy 5933 * the buffer to user space. smart... 5934 * 5935 * Our current scheme is not silly either but we take the 5936 * extra cost of the net_bh soft interrupt processing... 5937 * We do checksum and copy also but from device to kernel. 5938 */ 5939 5940 tp->rx_opt.saw_tstamp = 0; 5941 5942 /* pred_flags is 0xS?10 << 16 + snd_wnd 5943 * if header_prediction is to be made 5944 * 'S' will always be tp->tcp_header_len >> 2 5945 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to 5946 * turn it off (when there are holes in the receive 5947 * space for instance) 5948 * PSH flag is ignored. 5949 */ 5950 5951 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags && 5952 TCP_SKB_CB(skb)->seq == tp->rcv_nxt && 5953 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) { 5954 int tcp_header_len = tp->tcp_header_len; 5955 5956 /* Timestamp header prediction: tcp_header_len 5957 * is automatically equal to th->doff*4 due to pred_flags 5958 * match. 5959 */ 5960 5961 /* Check timestamp */ 5962 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) { 5963 /* No? Slow path! */ 5964 if (!tcp_parse_aligned_timestamp(tp, th)) 5965 goto slow_path; 5966 5967 /* If PAWS failed, check it more carefully in slow path */ 5968 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0) 5969 goto slow_path; 5970 5971 /* DO NOT update ts_recent here, if checksum fails 5972 * and timestamp was corrupted part, it will result 5973 * in a hung connection since we will drop all 5974 * future packets due to the PAWS test. 5975 */ 5976 } 5977 5978 if (len <= tcp_header_len) { 5979 /* Bulk data transfer: sender */ 5980 if (len == tcp_header_len) { 5981 /* Predicted packet is in window by definition. 5982 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 5983 * Hence, check seq<=rcv_wup reduces to: 5984 */ 5985 if (tcp_header_len == 5986 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 5987 tp->rcv_nxt == tp->rcv_wup) 5988 tcp_store_ts_recent(tp); 5989 5990 /* We know that such packets are checksummed 5991 * on entry. 5992 */ 5993 tcp_ack(sk, skb, 0); 5994 __kfree_skb(skb); 5995 tcp_data_snd_check(sk); 5996 /* When receiving pure ack in fast path, update 5997 * last ts ecr directly instead of calling 5998 * tcp_rcv_rtt_measure_ts() 5999 */ 6000 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr; 6001 return; 6002 } else { /* Header too small */ 6003 reason = SKB_DROP_REASON_PKT_TOO_SMALL; 6004 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS); 6005 goto discard; 6006 } 6007 } else { 6008 int eaten = 0; 6009 bool fragstolen = false; 6010 6011 if (tcp_checksum_complete(skb)) 6012 goto csum_error; 6013 6014 if ((int)skb->truesize > sk->sk_forward_alloc) 6015 goto step5; 6016 6017 /* Predicted packet is in window by definition. 6018 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 6019 * Hence, check seq<=rcv_wup reduces to: 6020 */ 6021 if (tcp_header_len == 6022 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 6023 tp->rcv_nxt == tp->rcv_wup) 6024 tcp_store_ts_recent(tp); 6025 6026 tcp_rcv_rtt_measure_ts(sk, skb); 6027 6028 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS); 6029 6030 /* Bulk data transfer: receiver */ 6031 skb_dst_drop(skb); 6032 __skb_pull(skb, tcp_header_len); 6033 eaten = tcp_queue_rcv(sk, skb, &fragstolen); 6034 6035 tcp_event_data_recv(sk, skb); 6036 6037 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) { 6038 /* Well, only one small jumplet in fast path... */ 6039 tcp_ack(sk, skb, FLAG_DATA); 6040 tcp_data_snd_check(sk); 6041 if (!inet_csk_ack_scheduled(sk)) 6042 goto no_ack; 6043 } else { 6044 tcp_update_wl(tp, TCP_SKB_CB(skb)->seq); 6045 } 6046 6047 __tcp_ack_snd_check(sk, 0); 6048 no_ack: 6049 if (eaten) 6050 kfree_skb_partial(skb, fragstolen); 6051 tcp_data_ready(sk); 6052 return; 6053 } 6054 } 6055 6056 slow_path: 6057 if (len < (th->doff << 2) || tcp_checksum_complete(skb)) 6058 goto csum_error; 6059 6060 if (!th->ack && !th->rst && !th->syn) { 6061 reason = SKB_DROP_REASON_TCP_FLAGS; 6062 goto discard; 6063 } 6064 6065 /* 6066 * Standard slow path. 6067 */ 6068 6069 if (!tcp_validate_incoming(sk, skb, th, 1)) 6070 return; 6071 6072 step5: 6073 reason = tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT); 6074 if ((int)reason < 0) { 6075 reason = -reason; 6076 goto discard; 6077 } 6078 tcp_rcv_rtt_measure_ts(sk, skb); 6079 6080 /* Process urgent data. */ 6081 tcp_urg(sk, skb, th); 6082 6083 /* step 7: process the segment text */ 6084 tcp_data_queue(sk, skb); 6085 6086 tcp_data_snd_check(sk); 6087 tcp_ack_snd_check(sk); 6088 return; 6089 6090 csum_error: 6091 reason = SKB_DROP_REASON_TCP_CSUM; 6092 trace_tcp_bad_csum(skb); 6093 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS); 6094 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS); 6095 6096 discard: 6097 tcp_drop_reason(sk, skb, reason); 6098 } 6099 EXPORT_SYMBOL(tcp_rcv_established); 6100 6101 void tcp_init_transfer(struct sock *sk, int bpf_op, struct sk_buff *skb) 6102 { 6103 struct inet_connection_sock *icsk = inet_csk(sk); 6104 struct tcp_sock *tp = tcp_sk(sk); 6105 6106 tcp_mtup_init(sk); 6107 icsk->icsk_af_ops->rebuild_header(sk); 6108 tcp_init_metrics(sk); 6109 6110 /* Initialize the congestion window to start the transfer. 6111 * Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been 6112 * retransmitted. In light of RFC6298 more aggressive 1sec 6113 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK 6114 * retransmission has occurred. 6115 */ 6116 if (tp->total_retrans > 1 && tp->undo_marker) 6117 tcp_snd_cwnd_set(tp, 1); 6118 else 6119 tcp_snd_cwnd_set(tp, tcp_init_cwnd(tp, __sk_dst_get(sk))); 6120 tp->snd_cwnd_stamp = tcp_jiffies32; 6121 6122 bpf_skops_established(sk, bpf_op, skb); 6123 /* Initialize congestion control unless BPF initialized it already: */ 6124 if (!icsk->icsk_ca_initialized) 6125 tcp_init_congestion_control(sk); 6126 tcp_init_buffer_space(sk); 6127 } 6128 6129 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb) 6130 { 6131 struct tcp_sock *tp = tcp_sk(sk); 6132 struct inet_connection_sock *icsk = inet_csk(sk); 6133 6134 tcp_set_state(sk, TCP_ESTABLISHED); 6135 icsk->icsk_ack.lrcvtime = tcp_jiffies32; 6136 6137 if (skb) { 6138 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb); 6139 security_inet_conn_established(sk, skb); 6140 sk_mark_napi_id(sk, skb); 6141 } 6142 6143 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB, skb); 6144 6145 /* Prevent spurious tcp_cwnd_restart() on first data 6146 * packet. 6147 */ 6148 tp->lsndtime = tcp_jiffies32; 6149 6150 if (sock_flag(sk, SOCK_KEEPOPEN)) 6151 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp)); 6152 6153 if (!tp->rx_opt.snd_wscale) 6154 __tcp_fast_path_on(tp, tp->snd_wnd); 6155 else 6156 tp->pred_flags = 0; 6157 } 6158 6159 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack, 6160 struct tcp_fastopen_cookie *cookie) 6161 { 6162 struct tcp_sock *tp = tcp_sk(sk); 6163 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL; 6164 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0; 6165 bool syn_drop = false; 6166 6167 if (mss == tp->rx_opt.user_mss) { 6168 struct tcp_options_received opt; 6169 6170 /* Get original SYNACK MSS value if user MSS sets mss_clamp */ 6171 tcp_clear_options(&opt); 6172 opt.user_mss = opt.mss_clamp = 0; 6173 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL); 6174 mss = opt.mss_clamp; 6175 } 6176 6177 if (!tp->syn_fastopen) { 6178 /* Ignore an unsolicited cookie */ 6179 cookie->len = -1; 6180 } else if (tp->total_retrans) { 6181 /* SYN timed out and the SYN-ACK neither has a cookie nor 6182 * acknowledges data. Presumably the remote received only 6183 * the retransmitted (regular) SYNs: either the original 6184 * SYN-data or the corresponding SYN-ACK was dropped. 6185 */ 6186 syn_drop = (cookie->len < 0 && data); 6187 } else if (cookie->len < 0 && !tp->syn_data) { 6188 /* We requested a cookie but didn't get it. If we did not use 6189 * the (old) exp opt format then try so next time (try_exp=1). 6190 * Otherwise we go back to use the RFC7413 opt (try_exp=2). 6191 */ 6192 try_exp = tp->syn_fastopen_exp ? 2 : 1; 6193 } 6194 6195 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp); 6196 6197 if (data) { /* Retransmit unacked data in SYN */ 6198 if (tp->total_retrans) 6199 tp->fastopen_client_fail = TFO_SYN_RETRANSMITTED; 6200 else 6201 tp->fastopen_client_fail = TFO_DATA_NOT_ACKED; 6202 skb_rbtree_walk_from(data) 6203 tcp_mark_skb_lost(sk, data); 6204 tcp_non_congestion_loss_retransmit(sk); 6205 NET_INC_STATS(sock_net(sk), 6206 LINUX_MIB_TCPFASTOPENACTIVEFAIL); 6207 return true; 6208 } 6209 tp->syn_data_acked = tp->syn_data; 6210 if (tp->syn_data_acked) { 6211 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE); 6212 /* SYN-data is counted as two separate packets in tcp_ack() */ 6213 if (tp->delivered > 1) 6214 --tp->delivered; 6215 } 6216 6217 tcp_fastopen_add_skb(sk, synack); 6218 6219 return false; 6220 } 6221 6222 static void smc_check_reset_syn(struct tcp_sock *tp) 6223 { 6224 #if IS_ENABLED(CONFIG_SMC) 6225 if (static_branch_unlikely(&tcp_have_smc)) { 6226 if (tp->syn_smc && !tp->rx_opt.smc_ok) 6227 tp->syn_smc = 0; 6228 } 6229 #endif 6230 } 6231 6232 static void tcp_try_undo_spurious_syn(struct sock *sk) 6233 { 6234 struct tcp_sock *tp = tcp_sk(sk); 6235 u32 syn_stamp; 6236 6237 /* undo_marker is set when SYN or SYNACK times out. The timeout is 6238 * spurious if the ACK's timestamp option echo value matches the 6239 * original SYN timestamp. 6240 */ 6241 syn_stamp = tp->retrans_stamp; 6242 if (tp->undo_marker && syn_stamp && tp->rx_opt.saw_tstamp && 6243 syn_stamp == tp->rx_opt.rcv_tsecr) 6244 tp->undo_marker = 0; 6245 } 6246 6247 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb, 6248 const struct tcphdr *th) 6249 { 6250 struct inet_connection_sock *icsk = inet_csk(sk); 6251 struct tcp_sock *tp = tcp_sk(sk); 6252 struct tcp_fastopen_cookie foc = { .len = -1 }; 6253 int saved_clamp = tp->rx_opt.mss_clamp; 6254 bool fastopen_fail; 6255 SKB_DR(reason); 6256 6257 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc); 6258 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) 6259 tp->rx_opt.rcv_tsecr -= tp->tsoffset; 6260 6261 if (th->ack) { 6262 /* rfc793: 6263 * "If the state is SYN-SENT then 6264 * first check the ACK bit 6265 * If the ACK bit is set 6266 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send 6267 * a reset (unless the RST bit is set, if so drop 6268 * the segment and return)" 6269 */ 6270 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) || 6271 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) { 6272 /* Previous FIN/ACK or RST/ACK might be ignored. */ 6273 if (icsk->icsk_retransmits == 0) 6274 inet_csk_reset_xmit_timer(sk, 6275 ICSK_TIME_RETRANS, 6276 TCP_TIMEOUT_MIN, TCP_RTO_MAX); 6277 goto reset_and_undo; 6278 } 6279 6280 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 6281 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp, 6282 tcp_time_stamp(tp))) { 6283 NET_INC_STATS(sock_net(sk), 6284 LINUX_MIB_PAWSACTIVEREJECTED); 6285 goto reset_and_undo; 6286 } 6287 6288 /* Now ACK is acceptable. 6289 * 6290 * "If the RST bit is set 6291 * If the ACK was acceptable then signal the user "error: 6292 * connection reset", drop the segment, enter CLOSED state, 6293 * delete TCB, and return." 6294 */ 6295 6296 if (th->rst) { 6297 tcp_reset(sk, skb); 6298 consume: 6299 __kfree_skb(skb); 6300 return 0; 6301 } 6302 6303 /* rfc793: 6304 * "fifth, if neither of the SYN or RST bits is set then 6305 * drop the segment and return." 6306 * 6307 * See note below! 6308 * --ANK(990513) 6309 */ 6310 if (!th->syn) { 6311 SKB_DR_SET(reason, TCP_FLAGS); 6312 goto discard_and_undo; 6313 } 6314 /* rfc793: 6315 * "If the SYN bit is on ... 6316 * are acceptable then ... 6317 * (our SYN has been ACKed), change the connection 6318 * state to ESTABLISHED..." 6319 */ 6320 6321 tcp_ecn_rcv_synack(tp, th); 6322 6323 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); 6324 tcp_try_undo_spurious_syn(sk); 6325 tcp_ack(sk, skb, FLAG_SLOWPATH); 6326 6327 /* Ok.. it's good. Set up sequence numbers and 6328 * move to established. 6329 */ 6330 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1); 6331 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 6332 6333 /* RFC1323: The window in SYN & SYN/ACK segments is 6334 * never scaled. 6335 */ 6336 tp->snd_wnd = ntohs(th->window); 6337 6338 if (!tp->rx_opt.wscale_ok) { 6339 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0; 6340 tp->window_clamp = min(tp->window_clamp, 65535U); 6341 } 6342 6343 if (tp->rx_opt.saw_tstamp) { 6344 tp->rx_opt.tstamp_ok = 1; 6345 tp->tcp_header_len = 6346 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 6347 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 6348 tcp_store_ts_recent(tp); 6349 } else { 6350 tp->tcp_header_len = sizeof(struct tcphdr); 6351 } 6352 6353 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 6354 tcp_initialize_rcv_mss(sk); 6355 6356 /* Remember, tcp_poll() does not lock socket! 6357 * Change state from SYN-SENT only after copied_seq 6358 * is initialized. */ 6359 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 6360 6361 smc_check_reset_syn(tp); 6362 6363 smp_mb(); 6364 6365 tcp_finish_connect(sk, skb); 6366 6367 fastopen_fail = (tp->syn_fastopen || tp->syn_data) && 6368 tcp_rcv_fastopen_synack(sk, skb, &foc); 6369 6370 if (!sock_flag(sk, SOCK_DEAD)) { 6371 sk->sk_state_change(sk); 6372 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT); 6373 } 6374 if (fastopen_fail) 6375 return -1; 6376 if (sk->sk_write_pending || 6377 READ_ONCE(icsk->icsk_accept_queue.rskq_defer_accept) || 6378 inet_csk_in_pingpong_mode(sk)) { 6379 /* Save one ACK. Data will be ready after 6380 * several ticks, if write_pending is set. 6381 * 6382 * It may be deleted, but with this feature tcpdumps 6383 * look so _wonderfully_ clever, that I was not able 6384 * to stand against the temptation 8) --ANK 6385 */ 6386 inet_csk_schedule_ack(sk); 6387 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS); 6388 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 6389 TCP_DELACK_MAX, TCP_RTO_MAX); 6390 goto consume; 6391 } 6392 tcp_send_ack(sk); 6393 return -1; 6394 } 6395 6396 /* No ACK in the segment */ 6397 6398 if (th->rst) { 6399 /* rfc793: 6400 * "If the RST bit is set 6401 * 6402 * Otherwise (no ACK) drop the segment and return." 6403 */ 6404 SKB_DR_SET(reason, TCP_RESET); 6405 goto discard_and_undo; 6406 } 6407 6408 /* PAWS check. */ 6409 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && 6410 tcp_paws_reject(&tp->rx_opt, 0)) { 6411 SKB_DR_SET(reason, TCP_RFC7323_PAWS); 6412 goto discard_and_undo; 6413 } 6414 if (th->syn) { 6415 /* We see SYN without ACK. It is attempt of 6416 * simultaneous connect with crossed SYNs. 6417 * Particularly, it can be connect to self. 6418 */ 6419 tcp_set_state(sk, TCP_SYN_RECV); 6420 6421 if (tp->rx_opt.saw_tstamp) { 6422 tp->rx_opt.tstamp_ok = 1; 6423 tcp_store_ts_recent(tp); 6424 tp->tcp_header_len = 6425 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 6426 } else { 6427 tp->tcp_header_len = sizeof(struct tcphdr); 6428 } 6429 6430 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1); 6431 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 6432 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 6433 6434 /* RFC1323: The window in SYN & SYN/ACK segments is 6435 * never scaled. 6436 */ 6437 tp->snd_wnd = ntohs(th->window); 6438 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 6439 tp->max_window = tp->snd_wnd; 6440 6441 tcp_ecn_rcv_syn(tp, th); 6442 6443 tcp_mtup_init(sk); 6444 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 6445 tcp_initialize_rcv_mss(sk); 6446 6447 tcp_send_synack(sk); 6448 #if 0 6449 /* Note, we could accept data and URG from this segment. 6450 * There are no obstacles to make this (except that we must 6451 * either change tcp_recvmsg() to prevent it from returning data 6452 * before 3WHS completes per RFC793, or employ TCP Fast Open). 6453 * 6454 * However, if we ignore data in ACKless segments sometimes, 6455 * we have no reasons to accept it sometimes. 6456 * Also, seems the code doing it in step6 of tcp_rcv_state_process 6457 * is not flawless. So, discard packet for sanity. 6458 * Uncomment this return to process the data. 6459 */ 6460 return -1; 6461 #else 6462 goto consume; 6463 #endif 6464 } 6465 /* "fifth, if neither of the SYN or RST bits is set then 6466 * drop the segment and return." 6467 */ 6468 6469 discard_and_undo: 6470 tcp_clear_options(&tp->rx_opt); 6471 tp->rx_opt.mss_clamp = saved_clamp; 6472 tcp_drop_reason(sk, skb, reason); 6473 return 0; 6474 6475 reset_and_undo: 6476 tcp_clear_options(&tp->rx_opt); 6477 tp->rx_opt.mss_clamp = saved_clamp; 6478 return 1; 6479 } 6480 6481 static void tcp_rcv_synrecv_state_fastopen(struct sock *sk) 6482 { 6483 struct tcp_sock *tp = tcp_sk(sk); 6484 struct request_sock *req; 6485 6486 /* If we are still handling the SYNACK RTO, see if timestamp ECR allows 6487 * undo. If peer SACKs triggered fast recovery, we can't undo here. 6488 */ 6489 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss && !tp->packets_out) 6490 tcp_try_undo_recovery(sk); 6491 6492 /* Reset rtx states to prevent spurious retransmits_timed_out() */ 6493 tp->retrans_stamp = 0; 6494 inet_csk(sk)->icsk_retransmits = 0; 6495 6496 /* Once we leave TCP_SYN_RECV or TCP_FIN_WAIT_1, 6497 * we no longer need req so release it. 6498 */ 6499 req = rcu_dereference_protected(tp->fastopen_rsk, 6500 lockdep_sock_is_held(sk)); 6501 reqsk_fastopen_remove(sk, req, false); 6502 6503 /* Re-arm the timer because data may have been sent out. 6504 * This is similar to the regular data transmission case 6505 * when new data has just been ack'ed. 6506 * 6507 * (TFO) - we could try to be more aggressive and 6508 * retransmitting any data sooner based on when they 6509 * are sent out. 6510 */ 6511 tcp_rearm_rto(sk); 6512 } 6513 6514 /* 6515 * This function implements the receiving procedure of RFC 793 for 6516 * all states except ESTABLISHED and TIME_WAIT. 6517 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be 6518 * address independent. 6519 */ 6520 6521 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb) 6522 { 6523 struct tcp_sock *tp = tcp_sk(sk); 6524 struct inet_connection_sock *icsk = inet_csk(sk); 6525 const struct tcphdr *th = tcp_hdr(skb); 6526 struct request_sock *req; 6527 int queued = 0; 6528 bool acceptable; 6529 SKB_DR(reason); 6530 6531 switch (sk->sk_state) { 6532 case TCP_CLOSE: 6533 SKB_DR_SET(reason, TCP_CLOSE); 6534 goto discard; 6535 6536 case TCP_LISTEN: 6537 if (th->ack) 6538 return 1; 6539 6540 if (th->rst) { 6541 SKB_DR_SET(reason, TCP_RESET); 6542 goto discard; 6543 } 6544 if (th->syn) { 6545 if (th->fin) { 6546 SKB_DR_SET(reason, TCP_FLAGS); 6547 goto discard; 6548 } 6549 /* It is possible that we process SYN packets from backlog, 6550 * so we need to make sure to disable BH and RCU right there. 6551 */ 6552 rcu_read_lock(); 6553 local_bh_disable(); 6554 acceptable = icsk->icsk_af_ops->conn_request(sk, skb) >= 0; 6555 local_bh_enable(); 6556 rcu_read_unlock(); 6557 6558 if (!acceptable) 6559 return 1; 6560 consume_skb(skb); 6561 return 0; 6562 } 6563 SKB_DR_SET(reason, TCP_FLAGS); 6564 goto discard; 6565 6566 case TCP_SYN_SENT: 6567 tp->rx_opt.saw_tstamp = 0; 6568 tcp_mstamp_refresh(tp); 6569 queued = tcp_rcv_synsent_state_process(sk, skb, th); 6570 if (queued >= 0) 6571 return queued; 6572 6573 /* Do step6 onward by hand. */ 6574 tcp_urg(sk, skb, th); 6575 __kfree_skb(skb); 6576 tcp_data_snd_check(sk); 6577 return 0; 6578 } 6579 6580 tcp_mstamp_refresh(tp); 6581 tp->rx_opt.saw_tstamp = 0; 6582 req = rcu_dereference_protected(tp->fastopen_rsk, 6583 lockdep_sock_is_held(sk)); 6584 if (req) { 6585 bool req_stolen; 6586 6587 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV && 6588 sk->sk_state != TCP_FIN_WAIT1); 6589 6590 if (!tcp_check_req(sk, skb, req, true, &req_stolen)) { 6591 SKB_DR_SET(reason, TCP_FASTOPEN); 6592 goto discard; 6593 } 6594 } 6595 6596 if (!th->ack && !th->rst && !th->syn) { 6597 SKB_DR_SET(reason, TCP_FLAGS); 6598 goto discard; 6599 } 6600 if (!tcp_validate_incoming(sk, skb, th, 0)) 6601 return 0; 6602 6603 /* step 5: check the ACK field */ 6604 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH | 6605 FLAG_UPDATE_TS_RECENT | 6606 FLAG_NO_CHALLENGE_ACK) > 0; 6607 6608 if (!acceptable) { 6609 if (sk->sk_state == TCP_SYN_RECV) 6610 return 1; /* send one RST */ 6611 tcp_send_challenge_ack(sk); 6612 SKB_DR_SET(reason, TCP_OLD_ACK); 6613 goto discard; 6614 } 6615 switch (sk->sk_state) { 6616 case TCP_SYN_RECV: 6617 tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */ 6618 if (!tp->srtt_us) 6619 tcp_synack_rtt_meas(sk, req); 6620 6621 if (req) { 6622 tcp_rcv_synrecv_state_fastopen(sk); 6623 } else { 6624 tcp_try_undo_spurious_syn(sk); 6625 tp->retrans_stamp = 0; 6626 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB, 6627 skb); 6628 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 6629 } 6630 smp_mb(); 6631 tcp_set_state(sk, TCP_ESTABLISHED); 6632 sk->sk_state_change(sk); 6633 6634 /* Note, that this wakeup is only for marginal crossed SYN case. 6635 * Passively open sockets are not waked up, because 6636 * sk->sk_sleep == NULL and sk->sk_socket == NULL. 6637 */ 6638 if (sk->sk_socket) 6639 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT); 6640 6641 tp->snd_una = TCP_SKB_CB(skb)->ack_seq; 6642 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale; 6643 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); 6644 6645 if (tp->rx_opt.tstamp_ok) 6646 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 6647 6648 if (!inet_csk(sk)->icsk_ca_ops->cong_control) 6649 tcp_update_pacing_rate(sk); 6650 6651 /* Prevent spurious tcp_cwnd_restart() on first data packet */ 6652 tp->lsndtime = tcp_jiffies32; 6653 6654 tcp_initialize_rcv_mss(sk); 6655 tcp_fast_path_on(tp); 6656 if (sk->sk_shutdown & SEND_SHUTDOWN) 6657 tcp_shutdown(sk, SEND_SHUTDOWN); 6658 break; 6659 6660 case TCP_FIN_WAIT1: { 6661 int tmo; 6662 6663 if (req) 6664 tcp_rcv_synrecv_state_fastopen(sk); 6665 6666 if (tp->snd_una != tp->write_seq) 6667 break; 6668 6669 tcp_set_state(sk, TCP_FIN_WAIT2); 6670 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | SEND_SHUTDOWN); 6671 6672 sk_dst_confirm(sk); 6673 6674 if (!sock_flag(sk, SOCK_DEAD)) { 6675 /* Wake up lingering close() */ 6676 sk->sk_state_change(sk); 6677 break; 6678 } 6679 6680 if (READ_ONCE(tp->linger2) < 0) { 6681 tcp_done(sk); 6682 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 6683 return 1; 6684 } 6685 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 6686 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { 6687 /* Receive out of order FIN after close() */ 6688 if (tp->syn_fastopen && th->fin) 6689 tcp_fastopen_active_disable(sk); 6690 tcp_done(sk); 6691 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 6692 return 1; 6693 } 6694 6695 tmo = tcp_fin_time(sk); 6696 if (tmo > TCP_TIMEWAIT_LEN) { 6697 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN); 6698 } else if (th->fin || sock_owned_by_user(sk)) { 6699 /* Bad case. We could lose such FIN otherwise. 6700 * It is not a big problem, but it looks confusing 6701 * and not so rare event. We still can lose it now, 6702 * if it spins in bh_lock_sock(), but it is really 6703 * marginal case. 6704 */ 6705 inet_csk_reset_keepalive_timer(sk, tmo); 6706 } else { 6707 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 6708 goto consume; 6709 } 6710 break; 6711 } 6712 6713 case TCP_CLOSING: 6714 if (tp->snd_una == tp->write_seq) { 6715 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 6716 goto consume; 6717 } 6718 break; 6719 6720 case TCP_LAST_ACK: 6721 if (tp->snd_una == tp->write_seq) { 6722 tcp_update_metrics(sk); 6723 tcp_done(sk); 6724 goto consume; 6725 } 6726 break; 6727 } 6728 6729 /* step 6: check the URG bit */ 6730 tcp_urg(sk, skb, th); 6731 6732 /* step 7: process the segment text */ 6733 switch (sk->sk_state) { 6734 case TCP_CLOSE_WAIT: 6735 case TCP_CLOSING: 6736 case TCP_LAST_ACK: 6737 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 6738 /* If a subflow has been reset, the packet should not 6739 * continue to be processed, drop the packet. 6740 */ 6741 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb)) 6742 goto discard; 6743 break; 6744 } 6745 fallthrough; 6746 case TCP_FIN_WAIT1: 6747 case TCP_FIN_WAIT2: 6748 /* RFC 793 says to queue data in these states, 6749 * RFC 1122 says we MUST send a reset. 6750 * BSD 4.4 also does reset. 6751 */ 6752 if (sk->sk_shutdown & RCV_SHUTDOWN) { 6753 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 6754 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { 6755 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 6756 tcp_reset(sk, skb); 6757 return 1; 6758 } 6759 } 6760 fallthrough; 6761 case TCP_ESTABLISHED: 6762 tcp_data_queue(sk, skb); 6763 queued = 1; 6764 break; 6765 } 6766 6767 /* tcp_data could move socket to TIME-WAIT */ 6768 if (sk->sk_state != TCP_CLOSE) { 6769 tcp_data_snd_check(sk); 6770 tcp_ack_snd_check(sk); 6771 } 6772 6773 if (!queued) { 6774 discard: 6775 tcp_drop_reason(sk, skb, reason); 6776 } 6777 return 0; 6778 6779 consume: 6780 __kfree_skb(skb); 6781 return 0; 6782 } 6783 EXPORT_SYMBOL(tcp_rcv_state_process); 6784 6785 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family) 6786 { 6787 struct inet_request_sock *ireq = inet_rsk(req); 6788 6789 if (family == AF_INET) 6790 net_dbg_ratelimited("drop open request from %pI4/%u\n", 6791 &ireq->ir_rmt_addr, port); 6792 #if IS_ENABLED(CONFIG_IPV6) 6793 else if (family == AF_INET6) 6794 net_dbg_ratelimited("drop open request from %pI6/%u\n", 6795 &ireq->ir_v6_rmt_addr, port); 6796 #endif 6797 } 6798 6799 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set 6800 * 6801 * If we receive a SYN packet with these bits set, it means a 6802 * network is playing bad games with TOS bits. In order to 6803 * avoid possible false congestion notifications, we disable 6804 * TCP ECN negotiation. 6805 * 6806 * Exception: tcp_ca wants ECN. This is required for DCTCP 6807 * congestion control: Linux DCTCP asserts ECT on all packets, 6808 * including SYN, which is most optimal solution; however, 6809 * others, such as FreeBSD do not. 6810 * 6811 * Exception: At least one of the reserved bits of the TCP header (th->res1) is 6812 * set, indicating the use of a future TCP extension (such as AccECN). See 6813 * RFC8311 §4.3 which updates RFC3168 to allow the development of such 6814 * extensions. 6815 */ 6816 static void tcp_ecn_create_request(struct request_sock *req, 6817 const struct sk_buff *skb, 6818 const struct sock *listen_sk, 6819 const struct dst_entry *dst) 6820 { 6821 const struct tcphdr *th = tcp_hdr(skb); 6822 const struct net *net = sock_net(listen_sk); 6823 bool th_ecn = th->ece && th->cwr; 6824 bool ect, ecn_ok; 6825 u32 ecn_ok_dst; 6826 6827 if (!th_ecn) 6828 return; 6829 6830 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield); 6831 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK); 6832 ecn_ok = READ_ONCE(net->ipv4.sysctl_tcp_ecn) || ecn_ok_dst; 6833 6834 if (((!ect || th->res1) && ecn_ok) || tcp_ca_needs_ecn(listen_sk) || 6835 (ecn_ok_dst & DST_FEATURE_ECN_CA) || 6836 tcp_bpf_ca_needs_ecn((struct sock *)req)) 6837 inet_rsk(req)->ecn_ok = 1; 6838 } 6839 6840 static void tcp_openreq_init(struct request_sock *req, 6841 const struct tcp_options_received *rx_opt, 6842 struct sk_buff *skb, const struct sock *sk) 6843 { 6844 struct inet_request_sock *ireq = inet_rsk(req); 6845 6846 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */ 6847 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq; 6848 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 6849 tcp_rsk(req)->snt_synack = 0; 6850 tcp_rsk(req)->last_oow_ack_time = 0; 6851 req->mss = rx_opt->mss_clamp; 6852 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0; 6853 ireq->tstamp_ok = rx_opt->tstamp_ok; 6854 ireq->sack_ok = rx_opt->sack_ok; 6855 ireq->snd_wscale = rx_opt->snd_wscale; 6856 ireq->wscale_ok = rx_opt->wscale_ok; 6857 ireq->acked = 0; 6858 ireq->ecn_ok = 0; 6859 ireq->ir_rmt_port = tcp_hdr(skb)->source; 6860 ireq->ir_num = ntohs(tcp_hdr(skb)->dest); 6861 ireq->ir_mark = inet_request_mark(sk, skb); 6862 #if IS_ENABLED(CONFIG_SMC) 6863 ireq->smc_ok = rx_opt->smc_ok && !(tcp_sk(sk)->smc_hs_congested && 6864 tcp_sk(sk)->smc_hs_congested(sk)); 6865 #endif 6866 } 6867 6868 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops, 6869 struct sock *sk_listener, 6870 bool attach_listener) 6871 { 6872 struct request_sock *req = reqsk_alloc(ops, sk_listener, 6873 attach_listener); 6874 6875 if (req) { 6876 struct inet_request_sock *ireq = inet_rsk(req); 6877 6878 ireq->ireq_opt = NULL; 6879 #if IS_ENABLED(CONFIG_IPV6) 6880 ireq->pktopts = NULL; 6881 #endif 6882 atomic64_set(&ireq->ir_cookie, 0); 6883 ireq->ireq_state = TCP_NEW_SYN_RECV; 6884 write_pnet(&ireq->ireq_net, sock_net(sk_listener)); 6885 ireq->ireq_family = sk_listener->sk_family; 6886 req->timeout = TCP_TIMEOUT_INIT; 6887 } 6888 6889 return req; 6890 } 6891 EXPORT_SYMBOL(inet_reqsk_alloc); 6892 6893 /* 6894 * Return true if a syncookie should be sent 6895 */ 6896 static bool tcp_syn_flood_action(const struct sock *sk, const char *proto) 6897 { 6898 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue; 6899 const char *msg = "Dropping request"; 6900 struct net *net = sock_net(sk); 6901 bool want_cookie = false; 6902 u8 syncookies; 6903 6904 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies); 6905 6906 #ifdef CONFIG_SYN_COOKIES 6907 if (syncookies) { 6908 msg = "Sending cookies"; 6909 want_cookie = true; 6910 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES); 6911 } else 6912 #endif 6913 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP); 6914 6915 if (!READ_ONCE(queue->synflood_warned) && syncookies != 2 && 6916 xchg(&queue->synflood_warned, 1) == 0) { 6917 if (IS_ENABLED(CONFIG_IPV6) && sk->sk_family == AF_INET6) { 6918 net_info_ratelimited("%s: Possible SYN flooding on port [%pI6c]:%u. %s.\n", 6919 proto, inet6_rcv_saddr(sk), 6920 sk->sk_num, msg); 6921 } else { 6922 net_info_ratelimited("%s: Possible SYN flooding on port %pI4:%u. %s.\n", 6923 proto, &sk->sk_rcv_saddr, 6924 sk->sk_num, msg); 6925 } 6926 } 6927 6928 return want_cookie; 6929 } 6930 6931 static void tcp_reqsk_record_syn(const struct sock *sk, 6932 struct request_sock *req, 6933 const struct sk_buff *skb) 6934 { 6935 if (tcp_sk(sk)->save_syn) { 6936 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb); 6937 struct saved_syn *saved_syn; 6938 u32 mac_hdrlen; 6939 void *base; 6940 6941 if (tcp_sk(sk)->save_syn == 2) { /* Save full header. */ 6942 base = skb_mac_header(skb); 6943 mac_hdrlen = skb_mac_header_len(skb); 6944 len += mac_hdrlen; 6945 } else { 6946 base = skb_network_header(skb); 6947 mac_hdrlen = 0; 6948 } 6949 6950 saved_syn = kmalloc(struct_size(saved_syn, data, len), 6951 GFP_ATOMIC); 6952 if (saved_syn) { 6953 saved_syn->mac_hdrlen = mac_hdrlen; 6954 saved_syn->network_hdrlen = skb_network_header_len(skb); 6955 saved_syn->tcp_hdrlen = tcp_hdrlen(skb); 6956 memcpy(saved_syn->data, base, len); 6957 req->saved_syn = saved_syn; 6958 } 6959 } 6960 } 6961 6962 /* If a SYN cookie is required and supported, returns a clamped MSS value to be 6963 * used for SYN cookie generation. 6964 */ 6965 u16 tcp_get_syncookie_mss(struct request_sock_ops *rsk_ops, 6966 const struct tcp_request_sock_ops *af_ops, 6967 struct sock *sk, struct tcphdr *th) 6968 { 6969 struct tcp_sock *tp = tcp_sk(sk); 6970 u16 mss; 6971 6972 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_syncookies) != 2 && 6973 !inet_csk_reqsk_queue_is_full(sk)) 6974 return 0; 6975 6976 if (!tcp_syn_flood_action(sk, rsk_ops->slab_name)) 6977 return 0; 6978 6979 if (sk_acceptq_is_full(sk)) { 6980 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS); 6981 return 0; 6982 } 6983 6984 mss = tcp_parse_mss_option(th, tp->rx_opt.user_mss); 6985 if (!mss) 6986 mss = af_ops->mss_clamp; 6987 6988 return mss; 6989 } 6990 EXPORT_SYMBOL_GPL(tcp_get_syncookie_mss); 6991 6992 int tcp_conn_request(struct request_sock_ops *rsk_ops, 6993 const struct tcp_request_sock_ops *af_ops, 6994 struct sock *sk, struct sk_buff *skb) 6995 { 6996 struct tcp_fastopen_cookie foc = { .len = -1 }; 6997 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn; 6998 struct tcp_options_received tmp_opt; 6999 struct tcp_sock *tp = tcp_sk(sk); 7000 struct net *net = sock_net(sk); 7001 struct sock *fastopen_sk = NULL; 7002 struct request_sock *req; 7003 bool want_cookie = false; 7004 struct dst_entry *dst; 7005 struct flowi fl; 7006 u8 syncookies; 7007 7008 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies); 7009 7010 /* TW buckets are converted to open requests without 7011 * limitations, they conserve resources and peer is 7012 * evidently real one. 7013 */ 7014 if ((syncookies == 2 || inet_csk_reqsk_queue_is_full(sk)) && !isn) { 7015 want_cookie = tcp_syn_flood_action(sk, rsk_ops->slab_name); 7016 if (!want_cookie) 7017 goto drop; 7018 } 7019 7020 if (sk_acceptq_is_full(sk)) { 7021 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS); 7022 goto drop; 7023 } 7024 7025 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie); 7026 if (!req) 7027 goto drop; 7028 7029 req->syncookie = want_cookie; 7030 tcp_rsk(req)->af_specific = af_ops; 7031 tcp_rsk(req)->ts_off = 0; 7032 #if IS_ENABLED(CONFIG_MPTCP) 7033 tcp_rsk(req)->is_mptcp = 0; 7034 #endif 7035 7036 tcp_clear_options(&tmp_opt); 7037 tmp_opt.mss_clamp = af_ops->mss_clamp; 7038 tmp_opt.user_mss = tp->rx_opt.user_mss; 7039 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0, 7040 want_cookie ? NULL : &foc); 7041 7042 if (want_cookie && !tmp_opt.saw_tstamp) 7043 tcp_clear_options(&tmp_opt); 7044 7045 if (IS_ENABLED(CONFIG_SMC) && want_cookie) 7046 tmp_opt.smc_ok = 0; 7047 7048 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp; 7049 tcp_openreq_init(req, &tmp_opt, skb, sk); 7050 inet_rsk(req)->no_srccheck = inet_test_bit(TRANSPARENT, sk); 7051 7052 /* Note: tcp_v6_init_req() might override ir_iif for link locals */ 7053 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb); 7054 7055 dst = af_ops->route_req(sk, skb, &fl, req); 7056 if (!dst) 7057 goto drop_and_free; 7058 7059 if (tmp_opt.tstamp_ok) 7060 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb); 7061 7062 if (!want_cookie && !isn) { 7063 int max_syn_backlog = READ_ONCE(net->ipv4.sysctl_max_syn_backlog); 7064 7065 /* Kill the following clause, if you dislike this way. */ 7066 if (!syncookies && 7067 (max_syn_backlog - inet_csk_reqsk_queue_len(sk) < 7068 (max_syn_backlog >> 2)) && 7069 !tcp_peer_is_proven(req, dst)) { 7070 /* Without syncookies last quarter of 7071 * backlog is filled with destinations, 7072 * proven to be alive. 7073 * It means that we continue to communicate 7074 * to destinations, already remembered 7075 * to the moment of synflood. 7076 */ 7077 pr_drop_req(req, ntohs(tcp_hdr(skb)->source), 7078 rsk_ops->family); 7079 goto drop_and_release; 7080 } 7081 7082 isn = af_ops->init_seq(skb); 7083 } 7084 7085 tcp_ecn_create_request(req, skb, sk, dst); 7086 7087 if (want_cookie) { 7088 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss); 7089 if (!tmp_opt.tstamp_ok) 7090 inet_rsk(req)->ecn_ok = 0; 7091 } 7092 7093 tcp_rsk(req)->snt_isn = isn; 7094 tcp_rsk(req)->txhash = net_tx_rndhash(); 7095 tcp_rsk(req)->syn_tos = TCP_SKB_CB(skb)->ip_dsfield; 7096 tcp_openreq_init_rwin(req, sk, dst); 7097 sk_rx_queue_set(req_to_sk(req), skb); 7098 if (!want_cookie) { 7099 tcp_reqsk_record_syn(sk, req, skb); 7100 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst); 7101 } 7102 if (fastopen_sk) { 7103 af_ops->send_synack(fastopen_sk, dst, &fl, req, 7104 &foc, TCP_SYNACK_FASTOPEN, skb); 7105 /* Add the child socket directly into the accept queue */ 7106 if (!inet_csk_reqsk_queue_add(sk, req, fastopen_sk)) { 7107 reqsk_fastopen_remove(fastopen_sk, req, false); 7108 bh_unlock_sock(fastopen_sk); 7109 sock_put(fastopen_sk); 7110 goto drop_and_free; 7111 } 7112 sk->sk_data_ready(sk); 7113 bh_unlock_sock(fastopen_sk); 7114 sock_put(fastopen_sk); 7115 } else { 7116 tcp_rsk(req)->tfo_listener = false; 7117 if (!want_cookie) { 7118 req->timeout = tcp_timeout_init((struct sock *)req); 7119 if (unlikely(!inet_csk_reqsk_queue_hash_add(sk, req, 7120 req->timeout))) { 7121 reqsk_free(req); 7122 return 0; 7123 } 7124 7125 } 7126 af_ops->send_synack(sk, dst, &fl, req, &foc, 7127 !want_cookie ? TCP_SYNACK_NORMAL : 7128 TCP_SYNACK_COOKIE, 7129 skb); 7130 if (want_cookie) { 7131 reqsk_free(req); 7132 return 0; 7133 } 7134 } 7135 reqsk_put(req); 7136 return 0; 7137 7138 drop_and_release: 7139 dst_release(dst); 7140 drop_and_free: 7141 __reqsk_free(req); 7142 drop: 7143 tcp_listendrop(sk); 7144 return 0; 7145 } 7146 EXPORT_SYMBOL(tcp_conn_request); 7147