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/proto_memory.h> 76 #include <net/inet_common.h> 77 #include <linux/ipsec.h> 78 #include <asm/unaligned.h> 79 #include <linux/errqueue.h> 80 #include <trace/events/tcp.h> 81 #include <linux/jump_label_ratelimit.h> 82 #include <net/busy_poll.h> 83 #include <net/mptcp.h> 84 85 int sysctl_tcp_max_orphans __read_mostly = NR_FILE; 86 87 #define FLAG_DATA 0x01 /* Incoming frame contained data. */ 88 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */ 89 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */ 90 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */ 91 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */ 92 #define FLAG_DATA_SACKED 0x20 /* New SACK. */ 93 #define FLAG_ECE 0x40 /* ECE in this ACK */ 94 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */ 95 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/ 96 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */ 97 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */ 98 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */ 99 #define FLAG_SET_XMIT_TIMER 0x1000 /* Set TLP or RTO timer */ 100 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */ 101 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */ 102 #define FLAG_NO_CHALLENGE_ACK 0x8000 /* do not call tcp_send_challenge_ack() */ 103 #define FLAG_ACK_MAYBE_DELAYED 0x10000 /* Likely a delayed ACK */ 104 #define FLAG_DSACK_TLP 0x20000 /* DSACK for tail loss probe */ 105 106 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) 107 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) 108 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK) 109 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) 110 111 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) 112 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH)) 113 114 #define REXMIT_NONE 0 /* no loss recovery to do */ 115 #define REXMIT_LOST 1 /* retransmit packets marked lost */ 116 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */ 117 118 #if IS_ENABLED(CONFIG_TLS_DEVICE) 119 static DEFINE_STATIC_KEY_DEFERRED_FALSE(clean_acked_data_enabled, HZ); 120 121 void clean_acked_data_enable(struct inet_connection_sock *icsk, 122 void (*cad)(struct sock *sk, u32 ack_seq)) 123 { 124 icsk->icsk_clean_acked = cad; 125 static_branch_deferred_inc(&clean_acked_data_enabled); 126 } 127 EXPORT_SYMBOL_GPL(clean_acked_data_enable); 128 129 void clean_acked_data_disable(struct inet_connection_sock *icsk) 130 { 131 static_branch_slow_dec_deferred(&clean_acked_data_enabled); 132 icsk->icsk_clean_acked = NULL; 133 } 134 EXPORT_SYMBOL_GPL(clean_acked_data_disable); 135 136 void clean_acked_data_flush(void) 137 { 138 static_key_deferred_flush(&clean_acked_data_enabled); 139 } 140 EXPORT_SYMBOL_GPL(clean_acked_data_flush); 141 #endif 142 143 #ifdef CONFIG_CGROUP_BPF 144 static void bpf_skops_parse_hdr(struct sock *sk, struct sk_buff *skb) 145 { 146 bool unknown_opt = tcp_sk(sk)->rx_opt.saw_unknown && 147 BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), 148 BPF_SOCK_OPS_PARSE_UNKNOWN_HDR_OPT_CB_FLAG); 149 bool parse_all_opt = BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), 150 BPF_SOCK_OPS_PARSE_ALL_HDR_OPT_CB_FLAG); 151 struct bpf_sock_ops_kern sock_ops; 152 153 if (likely(!unknown_opt && !parse_all_opt)) 154 return; 155 156 /* The skb will be handled in the 157 * bpf_skops_established() or 158 * bpf_skops_write_hdr_opt(). 159 */ 160 switch (sk->sk_state) { 161 case TCP_SYN_RECV: 162 case TCP_SYN_SENT: 163 case TCP_LISTEN: 164 return; 165 } 166 167 sock_owned_by_me(sk); 168 169 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp)); 170 sock_ops.op = BPF_SOCK_OPS_PARSE_HDR_OPT_CB; 171 sock_ops.is_fullsock = 1; 172 sock_ops.sk = sk; 173 bpf_skops_init_skb(&sock_ops, skb, tcp_hdrlen(skb)); 174 175 BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops); 176 } 177 178 static void bpf_skops_established(struct sock *sk, int bpf_op, 179 struct sk_buff *skb) 180 { 181 struct bpf_sock_ops_kern sock_ops; 182 183 sock_owned_by_me(sk); 184 185 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp)); 186 sock_ops.op = bpf_op; 187 sock_ops.is_fullsock = 1; 188 sock_ops.sk = sk; 189 /* sk with TCP_REPAIR_ON does not have skb in tcp_finish_connect */ 190 if (skb) 191 bpf_skops_init_skb(&sock_ops, skb, tcp_hdrlen(skb)); 192 193 BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops); 194 } 195 #else 196 static void bpf_skops_parse_hdr(struct sock *sk, struct sk_buff *skb) 197 { 198 } 199 200 static void bpf_skops_established(struct sock *sk, int bpf_op, 201 struct sk_buff *skb) 202 { 203 } 204 #endif 205 206 static __cold void tcp_gro_dev_warn(const struct sock *sk, const struct sk_buff *skb, 207 unsigned int len) 208 { 209 struct net_device *dev; 210 211 rcu_read_lock(); 212 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif); 213 if (!dev || len >= READ_ONCE(dev->mtu)) 214 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n", 215 dev ? dev->name : "Unknown driver"); 216 rcu_read_unlock(); 217 } 218 219 /* Adapt the MSS value used to make delayed ack decision to the 220 * real world. 221 */ 222 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb) 223 { 224 struct inet_connection_sock *icsk = inet_csk(sk); 225 const unsigned int lss = icsk->icsk_ack.last_seg_size; 226 unsigned int len; 227 228 icsk->icsk_ack.last_seg_size = 0; 229 230 /* skb->len may jitter because of SACKs, even if peer 231 * sends good full-sized frames. 232 */ 233 len = skb_shinfo(skb)->gso_size ? : skb->len; 234 if (len >= icsk->icsk_ack.rcv_mss) { 235 /* Note: divides are still a bit expensive. 236 * For the moment, only adjust scaling_ratio 237 * when we update icsk_ack.rcv_mss. 238 */ 239 if (unlikely(len != icsk->icsk_ack.rcv_mss)) { 240 u64 val = (u64)skb->len << TCP_RMEM_TO_WIN_SCALE; 241 242 do_div(val, skb->truesize); 243 tcp_sk(sk)->scaling_ratio = val ? val : 1; 244 } 245 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len, 246 tcp_sk(sk)->advmss); 247 /* Account for possibly-removed options */ 248 DO_ONCE_LITE_IF(len > icsk->icsk_ack.rcv_mss + MAX_TCP_OPTION_SPACE, 249 tcp_gro_dev_warn, sk, skb, len); 250 /* If the skb has a len of exactly 1*MSS and has the PSH bit 251 * set then it is likely the end of an application write. So 252 * more data may not be arriving soon, and yet the data sender 253 * may be waiting for an ACK if cwnd-bound or using TX zero 254 * copy. So we set ICSK_ACK_PUSHED here so that 255 * tcp_cleanup_rbuf() will send an ACK immediately if the app 256 * reads all of the data and is not ping-pong. If len > MSS 257 * then this logic does not matter (and does not hurt) because 258 * tcp_cleanup_rbuf() will always ACK immediately if the app 259 * reads data and there is more than an MSS of unACKed data. 260 */ 261 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_PSH) 262 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 263 } else { 264 /* Otherwise, we make more careful check taking into account, 265 * that SACKs block is variable. 266 * 267 * "len" is invariant segment length, including TCP header. 268 */ 269 len += skb->data - skb_transport_header(skb); 270 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) || 271 /* If PSH is not set, packet should be 272 * full sized, provided peer TCP is not badly broken. 273 * This observation (if it is correct 8)) allows 274 * to handle super-low mtu links fairly. 275 */ 276 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && 277 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) { 278 /* Subtract also invariant (if peer is RFC compliant), 279 * tcp header plus fixed timestamp option length. 280 * Resulting "len" is MSS free of SACK jitter. 281 */ 282 len -= tcp_sk(sk)->tcp_header_len; 283 icsk->icsk_ack.last_seg_size = len; 284 if (len == lss) { 285 icsk->icsk_ack.rcv_mss = len; 286 return; 287 } 288 } 289 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) 290 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; 291 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 292 } 293 } 294 295 static void tcp_incr_quickack(struct sock *sk, unsigned int max_quickacks) 296 { 297 struct inet_connection_sock *icsk = inet_csk(sk); 298 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); 299 300 if (quickacks == 0) 301 quickacks = 2; 302 quickacks = min(quickacks, max_quickacks); 303 if (quickacks > icsk->icsk_ack.quick) 304 icsk->icsk_ack.quick = quickacks; 305 } 306 307 static void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks) 308 { 309 struct inet_connection_sock *icsk = inet_csk(sk); 310 311 tcp_incr_quickack(sk, max_quickacks); 312 inet_csk_exit_pingpong_mode(sk); 313 icsk->icsk_ack.ato = TCP_ATO_MIN; 314 } 315 316 /* Send ACKs quickly, if "quick" count is not exhausted 317 * and the session is not interactive. 318 */ 319 320 static bool tcp_in_quickack_mode(struct sock *sk) 321 { 322 const struct inet_connection_sock *icsk = inet_csk(sk); 323 const struct dst_entry *dst = __sk_dst_get(sk); 324 325 return (dst && dst_metric(dst, RTAX_QUICKACK)) || 326 (icsk->icsk_ack.quick && !inet_csk_in_pingpong_mode(sk)); 327 } 328 329 static void tcp_ecn_queue_cwr(struct tcp_sock *tp) 330 { 331 if (tp->ecn_flags & TCP_ECN_OK) 332 tp->ecn_flags |= TCP_ECN_QUEUE_CWR; 333 } 334 335 static void tcp_ecn_accept_cwr(struct sock *sk, const struct sk_buff *skb) 336 { 337 if (tcp_hdr(skb)->cwr) { 338 tcp_sk(sk)->ecn_flags &= ~TCP_ECN_DEMAND_CWR; 339 340 /* If the sender is telling us it has entered CWR, then its 341 * cwnd may be very low (even just 1 packet), so we should ACK 342 * immediately. 343 */ 344 if (TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) 345 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW; 346 } 347 } 348 349 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp) 350 { 351 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 352 } 353 354 static void __tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb) 355 { 356 struct tcp_sock *tp = tcp_sk(sk); 357 358 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) { 359 case INET_ECN_NOT_ECT: 360 /* Funny extension: if ECT is not set on a segment, 361 * and we already seen ECT on a previous segment, 362 * it is probably a retransmit. 363 */ 364 if (tp->ecn_flags & TCP_ECN_SEEN) 365 tcp_enter_quickack_mode(sk, 2); 366 break; 367 case INET_ECN_CE: 368 if (tcp_ca_needs_ecn(sk)) 369 tcp_ca_event(sk, CA_EVENT_ECN_IS_CE); 370 371 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) { 372 /* Better not delay acks, sender can have a very low cwnd */ 373 tcp_enter_quickack_mode(sk, 2); 374 tp->ecn_flags |= TCP_ECN_DEMAND_CWR; 375 } 376 tp->ecn_flags |= TCP_ECN_SEEN; 377 break; 378 default: 379 if (tcp_ca_needs_ecn(sk)) 380 tcp_ca_event(sk, CA_EVENT_ECN_NO_CE); 381 tp->ecn_flags |= TCP_ECN_SEEN; 382 break; 383 } 384 } 385 386 static void tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb) 387 { 388 if (tcp_sk(sk)->ecn_flags & TCP_ECN_OK) 389 __tcp_ecn_check_ce(sk, skb); 390 } 391 392 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th) 393 { 394 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr)) 395 tp->ecn_flags &= ~TCP_ECN_OK; 396 } 397 398 static void tcp_ecn_rcv_syn(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 bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th) 405 { 406 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK)) 407 return true; 408 return false; 409 } 410 411 /* Buffer size and advertised window tuning. 412 * 413 * 1. Tuning sk->sk_sndbuf, when connection enters established state. 414 */ 415 416 static void tcp_sndbuf_expand(struct sock *sk) 417 { 418 const struct tcp_sock *tp = tcp_sk(sk); 419 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 420 int sndmem, per_mss; 421 u32 nr_segs; 422 423 /* Worst case is non GSO/TSO : each frame consumes one skb 424 * and skb->head is kmalloced using power of two area of memory 425 */ 426 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) + 427 MAX_TCP_HEADER + 428 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 429 430 per_mss = roundup_pow_of_two(per_mss) + 431 SKB_DATA_ALIGN(sizeof(struct sk_buff)); 432 433 nr_segs = max_t(u32, TCP_INIT_CWND, tcp_snd_cwnd(tp)); 434 nr_segs = max_t(u32, nr_segs, tp->reordering + 1); 435 436 /* Fast Recovery (RFC 5681 3.2) : 437 * Cubic needs 1.7 factor, rounded to 2 to include 438 * extra cushion (application might react slowly to EPOLLOUT) 439 */ 440 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2; 441 sndmem *= nr_segs * per_mss; 442 443 if (sk->sk_sndbuf < sndmem) 444 WRITE_ONCE(sk->sk_sndbuf, 445 min(sndmem, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[2]))); 446 } 447 448 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh) 449 * 450 * All tcp_full_space() is split to two parts: "network" buffer, allocated 451 * forward and advertised in receiver window (tp->rcv_wnd) and 452 * "application buffer", required to isolate scheduling/application 453 * latencies from network. 454 * window_clamp is maximal advertised window. It can be less than 455 * tcp_full_space(), in this case tcp_full_space() - window_clamp 456 * is reserved for "application" buffer. The less window_clamp is 457 * the smoother our behaviour from viewpoint of network, but the lower 458 * throughput and the higher sensitivity of the connection to losses. 8) 459 * 460 * rcv_ssthresh is more strict window_clamp used at "slow start" 461 * phase to predict further behaviour of this connection. 462 * It is used for two goals: 463 * - to enforce header prediction at sender, even when application 464 * requires some significant "application buffer". It is check #1. 465 * - to prevent pruning of receive queue because of misprediction 466 * of receiver window. Check #2. 467 * 468 * The scheme does not work when sender sends good segments opening 469 * window and then starts to feed us spaghetti. But it should work 470 * in common situations. Otherwise, we have to rely on queue collapsing. 471 */ 472 473 /* Slow part of check#2. */ 474 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb, 475 unsigned int skbtruesize) 476 { 477 const struct tcp_sock *tp = tcp_sk(sk); 478 /* Optimize this! */ 479 int truesize = tcp_win_from_space(sk, skbtruesize) >> 1; 480 int window = tcp_win_from_space(sk, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2])) >> 1; 481 482 while (tp->rcv_ssthresh <= window) { 483 if (truesize <= skb->len) 484 return 2 * inet_csk(sk)->icsk_ack.rcv_mss; 485 486 truesize >>= 1; 487 window >>= 1; 488 } 489 return 0; 490 } 491 492 /* Even if skb appears to have a bad len/truesize ratio, TCP coalescing 493 * can play nice with us, as sk_buff and skb->head might be either 494 * freed or shared with up to MAX_SKB_FRAGS segments. 495 * Only give a boost to drivers using page frag(s) to hold the frame(s), 496 * and if no payload was pulled in skb->head before reaching us. 497 */ 498 static u32 truesize_adjust(bool adjust, const struct sk_buff *skb) 499 { 500 u32 truesize = skb->truesize; 501 502 if (adjust && !skb_headlen(skb)) { 503 truesize -= SKB_TRUESIZE(skb_end_offset(skb)); 504 /* paranoid check, some drivers might be buggy */ 505 if (unlikely((int)truesize < (int)skb->len)) 506 truesize = skb->truesize; 507 } 508 return truesize; 509 } 510 511 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb, 512 bool adjust) 513 { 514 struct tcp_sock *tp = tcp_sk(sk); 515 int room; 516 517 room = min_t(int, tp->window_clamp, tcp_space(sk)) - tp->rcv_ssthresh; 518 519 if (room <= 0) 520 return; 521 522 /* Check #1 */ 523 if (!tcp_under_memory_pressure(sk)) { 524 unsigned int truesize = truesize_adjust(adjust, skb); 525 int incr; 526 527 /* Check #2. Increase window, if skb with such overhead 528 * will fit to rcvbuf in future. 529 */ 530 if (tcp_win_from_space(sk, truesize) <= skb->len) 531 incr = 2 * tp->advmss; 532 else 533 incr = __tcp_grow_window(sk, skb, truesize); 534 535 if (incr) { 536 incr = max_t(int, incr, 2 * skb->len); 537 tp->rcv_ssthresh += min(room, incr); 538 inet_csk(sk)->icsk_ack.quick |= 1; 539 } 540 } else { 541 /* Under pressure: 542 * Adjust rcv_ssthresh according to reserved mem 543 */ 544 tcp_adjust_rcv_ssthresh(sk); 545 } 546 } 547 548 /* 3. Try to fixup all. It is made immediately after connection enters 549 * established state. 550 */ 551 static void tcp_init_buffer_space(struct sock *sk) 552 { 553 int tcp_app_win = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_app_win); 554 struct tcp_sock *tp = tcp_sk(sk); 555 int maxwin; 556 557 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) 558 tcp_sndbuf_expand(sk); 559 560 tcp_mstamp_refresh(tp); 561 tp->rcvq_space.time = tp->tcp_mstamp; 562 tp->rcvq_space.seq = tp->copied_seq; 563 564 maxwin = tcp_full_space(sk); 565 566 if (tp->window_clamp >= maxwin) { 567 WRITE_ONCE(tp->window_clamp, maxwin); 568 569 if (tcp_app_win && maxwin > 4 * tp->advmss) 570 WRITE_ONCE(tp->window_clamp, 571 max(maxwin - (maxwin >> tcp_app_win), 572 4 * tp->advmss)); 573 } 574 575 /* Force reservation of one segment. */ 576 if (tcp_app_win && 577 tp->window_clamp > 2 * tp->advmss && 578 tp->window_clamp + tp->advmss > maxwin) 579 WRITE_ONCE(tp->window_clamp, 580 max(2 * tp->advmss, maxwin - tp->advmss)); 581 582 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); 583 tp->snd_cwnd_stamp = tcp_jiffies32; 584 tp->rcvq_space.space = min3(tp->rcv_ssthresh, tp->rcv_wnd, 585 (u32)TCP_INIT_CWND * tp->advmss); 586 } 587 588 /* 4. Recalculate window clamp after socket hit its memory bounds. */ 589 static void tcp_clamp_window(struct sock *sk) 590 { 591 struct tcp_sock *tp = tcp_sk(sk); 592 struct inet_connection_sock *icsk = inet_csk(sk); 593 struct net *net = sock_net(sk); 594 int rmem2; 595 596 icsk->icsk_ack.quick = 0; 597 rmem2 = READ_ONCE(net->ipv4.sysctl_tcp_rmem[2]); 598 599 if (sk->sk_rcvbuf < rmem2 && 600 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && 601 !tcp_under_memory_pressure(sk) && 602 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) { 603 WRITE_ONCE(sk->sk_rcvbuf, 604 min(atomic_read(&sk->sk_rmem_alloc), rmem2)); 605 } 606 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) 607 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss); 608 } 609 610 /* Initialize RCV_MSS value. 611 * RCV_MSS is an our guess about MSS used by the peer. 612 * We haven't any direct information about the MSS. 613 * It's better to underestimate the RCV_MSS rather than overestimate. 614 * Overestimations make us ACKing less frequently than needed. 615 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). 616 */ 617 void tcp_initialize_rcv_mss(struct sock *sk) 618 { 619 const struct tcp_sock *tp = tcp_sk(sk); 620 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); 621 622 hint = min(hint, tp->rcv_wnd / 2); 623 hint = min(hint, TCP_MSS_DEFAULT); 624 hint = max(hint, TCP_MIN_MSS); 625 626 inet_csk(sk)->icsk_ack.rcv_mss = hint; 627 } 628 EXPORT_SYMBOL(tcp_initialize_rcv_mss); 629 630 /* Receiver "autotuning" code. 631 * 632 * The algorithm for RTT estimation w/o timestamps is based on 633 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. 634 * <https://public.lanl.gov/radiant/pubs.html#DRS> 635 * 636 * More detail on this code can be found at 637 * <http://staff.psc.edu/jheffner/>, 638 * though this reference is out of date. A new paper 639 * is pending. 640 */ 641 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) 642 { 643 u32 new_sample = tp->rcv_rtt_est.rtt_us; 644 long m = sample; 645 646 if (new_sample != 0) { 647 /* If we sample in larger samples in the non-timestamp 648 * case, we could grossly overestimate the RTT especially 649 * with chatty applications or bulk transfer apps which 650 * are stalled on filesystem I/O. 651 * 652 * Also, since we are only going for a minimum in the 653 * non-timestamp case, we do not smooth things out 654 * else with timestamps disabled convergence takes too 655 * long. 656 */ 657 if (!win_dep) { 658 m -= (new_sample >> 3); 659 new_sample += m; 660 } else { 661 m <<= 3; 662 if (m < new_sample) 663 new_sample = m; 664 } 665 } else { 666 /* No previous measure. */ 667 new_sample = m << 3; 668 } 669 670 tp->rcv_rtt_est.rtt_us = new_sample; 671 } 672 673 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) 674 { 675 u32 delta_us; 676 677 if (tp->rcv_rtt_est.time == 0) 678 goto new_measure; 679 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) 680 return; 681 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time); 682 if (!delta_us) 683 delta_us = 1; 684 tcp_rcv_rtt_update(tp, delta_us, 1); 685 686 new_measure: 687 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; 688 tp->rcv_rtt_est.time = tp->tcp_mstamp; 689 } 690 691 static s32 tcp_rtt_tsopt_us(const struct tcp_sock *tp) 692 { 693 u32 delta, delta_us; 694 695 delta = tcp_time_stamp_ts(tp) - tp->rx_opt.rcv_tsecr; 696 if (tp->tcp_usec_ts) 697 return delta; 698 699 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) { 700 if (!delta) 701 delta = 1; 702 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ); 703 return delta_us; 704 } 705 return -1; 706 } 707 708 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, 709 const struct sk_buff *skb) 710 { 711 struct tcp_sock *tp = tcp_sk(sk); 712 713 if (tp->rx_opt.rcv_tsecr == tp->rcv_rtt_last_tsecr) 714 return; 715 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr; 716 717 if (TCP_SKB_CB(skb)->end_seq - 718 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) { 719 s32 delta = tcp_rtt_tsopt_us(tp); 720 721 if (delta >= 0) 722 tcp_rcv_rtt_update(tp, delta, 0); 723 } 724 } 725 726 /* 727 * This function should be called every time data is copied to user space. 728 * It calculates the appropriate TCP receive buffer space. 729 */ 730 void tcp_rcv_space_adjust(struct sock *sk) 731 { 732 struct tcp_sock *tp = tcp_sk(sk); 733 u32 copied; 734 int time; 735 736 trace_tcp_rcv_space_adjust(sk); 737 738 tcp_mstamp_refresh(tp); 739 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time); 740 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0) 741 return; 742 743 /* Number of bytes copied to user in last RTT */ 744 copied = tp->copied_seq - tp->rcvq_space.seq; 745 if (copied <= tp->rcvq_space.space) 746 goto new_measure; 747 748 /* A bit of theory : 749 * copied = bytes received in previous RTT, our base window 750 * To cope with packet losses, we need a 2x factor 751 * To cope with slow start, and sender growing its cwin by 100 % 752 * every RTT, we need a 4x factor, because the ACK we are sending 753 * now is for the next RTT, not the current one : 754 * <prev RTT . ><current RTT .. ><next RTT .... > 755 */ 756 757 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) && 758 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { 759 u64 rcvwin, grow; 760 int rcvbuf; 761 762 /* minimal window to cope with packet losses, assuming 763 * steady state. Add some cushion because of small variations. 764 */ 765 rcvwin = ((u64)copied << 1) + 16 * tp->advmss; 766 767 /* Accommodate for sender rate increase (eg. slow start) */ 768 grow = rcvwin * (copied - tp->rcvq_space.space); 769 do_div(grow, tp->rcvq_space.space); 770 rcvwin += (grow << 1); 771 772 rcvbuf = min_t(u64, tcp_space_from_win(sk, rcvwin), 773 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2])); 774 if (rcvbuf > sk->sk_rcvbuf) { 775 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf); 776 777 /* Make the window clamp follow along. */ 778 WRITE_ONCE(tp->window_clamp, 779 tcp_win_from_space(sk, rcvbuf)); 780 } 781 } 782 tp->rcvq_space.space = copied; 783 784 new_measure: 785 tp->rcvq_space.seq = tp->copied_seq; 786 tp->rcvq_space.time = tp->tcp_mstamp; 787 } 788 789 static void tcp_save_lrcv_flowlabel(struct sock *sk, const struct sk_buff *skb) 790 { 791 #if IS_ENABLED(CONFIG_IPV6) 792 struct inet_connection_sock *icsk = inet_csk(sk); 793 794 if (skb->protocol == htons(ETH_P_IPV6)) 795 icsk->icsk_ack.lrcv_flowlabel = ntohl(ip6_flowlabel(ipv6_hdr(skb))); 796 #endif 797 } 798 799 /* There is something which you must keep in mind when you analyze the 800 * behavior of the tp->ato delayed ack timeout interval. When a 801 * connection starts up, we want to ack as quickly as possible. The 802 * problem is that "good" TCP's do slow start at the beginning of data 803 * transmission. The means that until we send the first few ACK's the 804 * sender will sit on his end and only queue most of his data, because 805 * he can only send snd_cwnd unacked packets at any given time. For 806 * each ACK we send, he increments snd_cwnd and transmits more of his 807 * queue. -DaveM 808 */ 809 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb) 810 { 811 struct tcp_sock *tp = tcp_sk(sk); 812 struct inet_connection_sock *icsk = inet_csk(sk); 813 u32 now; 814 815 inet_csk_schedule_ack(sk); 816 817 tcp_measure_rcv_mss(sk, skb); 818 819 tcp_rcv_rtt_measure(tp); 820 821 now = tcp_jiffies32; 822 823 if (!icsk->icsk_ack.ato) { 824 /* The _first_ data packet received, initialize 825 * delayed ACK engine. 826 */ 827 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS); 828 icsk->icsk_ack.ato = TCP_ATO_MIN; 829 } else { 830 int m = now - icsk->icsk_ack.lrcvtime; 831 832 if (m <= TCP_ATO_MIN / 2) { 833 /* The fastest case is the first. */ 834 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; 835 } else if (m < icsk->icsk_ack.ato) { 836 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; 837 if (icsk->icsk_ack.ato > icsk->icsk_rto) 838 icsk->icsk_ack.ato = icsk->icsk_rto; 839 } else if (m > icsk->icsk_rto) { 840 /* Too long gap. Apparently sender failed to 841 * restart window, so that we send ACKs quickly. 842 */ 843 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS); 844 } 845 } 846 icsk->icsk_ack.lrcvtime = now; 847 tcp_save_lrcv_flowlabel(sk, skb); 848 849 tcp_ecn_check_ce(sk, skb); 850 851 if (skb->len >= 128) 852 tcp_grow_window(sk, skb, true); 853 } 854 855 /* Called to compute a smoothed rtt estimate. The data fed to this 856 * routine either comes from timestamps, or from segments that were 857 * known _not_ to have been retransmitted [see Karn/Partridge 858 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 859 * piece by Van Jacobson. 860 * NOTE: the next three routines used to be one big routine. 861 * To save cycles in the RFC 1323 implementation it was better to break 862 * it up into three procedures. -- erics 863 */ 864 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us) 865 { 866 struct tcp_sock *tp = tcp_sk(sk); 867 long m = mrtt_us; /* RTT */ 868 u32 srtt = tp->srtt_us; 869 870 /* The following amusing code comes from Jacobson's 871 * article in SIGCOMM '88. Note that rtt and mdev 872 * are scaled versions of rtt and mean deviation. 873 * This is designed to be as fast as possible 874 * m stands for "measurement". 875 * 876 * On a 1990 paper the rto value is changed to: 877 * RTO = rtt + 4 * mdev 878 * 879 * Funny. This algorithm seems to be very broken. 880 * These formulae increase RTO, when it should be decreased, increase 881 * too slowly, when it should be increased quickly, decrease too quickly 882 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely 883 * does not matter how to _calculate_ it. Seems, it was trap 884 * that VJ failed to avoid. 8) 885 */ 886 if (srtt != 0) { 887 m -= (srtt >> 3); /* m is now error in rtt est */ 888 srtt += m; /* rtt = 7/8 rtt + 1/8 new */ 889 if (m < 0) { 890 m = -m; /* m is now abs(error) */ 891 m -= (tp->mdev_us >> 2); /* similar update on mdev */ 892 /* This is similar to one of Eifel findings. 893 * Eifel blocks mdev updates when rtt decreases. 894 * This solution is a bit different: we use finer gain 895 * for mdev in this case (alpha*beta). 896 * Like Eifel it also prevents growth of rto, 897 * but also it limits too fast rto decreases, 898 * happening in pure Eifel. 899 */ 900 if (m > 0) 901 m >>= 3; 902 } else { 903 m -= (tp->mdev_us >> 2); /* similar update on mdev */ 904 } 905 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */ 906 if (tp->mdev_us > tp->mdev_max_us) { 907 tp->mdev_max_us = tp->mdev_us; 908 if (tp->mdev_max_us > tp->rttvar_us) 909 tp->rttvar_us = tp->mdev_max_us; 910 } 911 if (after(tp->snd_una, tp->rtt_seq)) { 912 if (tp->mdev_max_us < tp->rttvar_us) 913 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2; 914 tp->rtt_seq = tp->snd_nxt; 915 tp->mdev_max_us = tcp_rto_min_us(sk); 916 917 tcp_bpf_rtt(sk, mrtt_us, srtt); 918 } 919 } else { 920 /* no previous measure. */ 921 srtt = m << 3; /* take the measured time to be rtt */ 922 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */ 923 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk)); 924 tp->mdev_max_us = tp->rttvar_us; 925 tp->rtt_seq = tp->snd_nxt; 926 927 tcp_bpf_rtt(sk, mrtt_us, srtt); 928 } 929 tp->srtt_us = max(1U, srtt); 930 } 931 932 static void tcp_update_pacing_rate(struct sock *sk) 933 { 934 const struct tcp_sock *tp = tcp_sk(sk); 935 u64 rate; 936 937 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */ 938 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3); 939 940 /* current rate is (cwnd * mss) / srtt 941 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate. 942 * In Congestion Avoidance phase, set it to 120 % the current rate. 943 * 944 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh) 945 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching 946 * end of slow start and should slow down. 947 */ 948 if (tcp_snd_cwnd(tp) < tp->snd_ssthresh / 2) 949 rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio); 950 else 951 rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio); 952 953 rate *= max(tcp_snd_cwnd(tp), tp->packets_out); 954 955 if (likely(tp->srtt_us)) 956 do_div(rate, tp->srtt_us); 957 958 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate 959 * without any lock. We want to make sure compiler wont store 960 * intermediate values in this location. 961 */ 962 WRITE_ONCE(sk->sk_pacing_rate, 963 min_t(u64, rate, READ_ONCE(sk->sk_max_pacing_rate))); 964 } 965 966 /* Calculate rto without backoff. This is the second half of Van Jacobson's 967 * routine referred to above. 968 */ 969 static void tcp_set_rto(struct sock *sk) 970 { 971 const struct tcp_sock *tp = tcp_sk(sk); 972 /* Old crap is replaced with new one. 8) 973 * 974 * More seriously: 975 * 1. If rtt variance happened to be less 50msec, it is hallucination. 976 * It cannot be less due to utterly erratic ACK generation made 977 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ 978 * to do with delayed acks, because at cwnd>2 true delack timeout 979 * is invisible. Actually, Linux-2.4 also generates erratic 980 * ACKs in some circumstances. 981 */ 982 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp); 983 984 /* 2. Fixups made earlier cannot be right. 985 * If we do not estimate RTO correctly without them, 986 * all the algo is pure shit and should be replaced 987 * with correct one. It is exactly, which we pretend to do. 988 */ 989 990 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo 991 * guarantees that rto is higher. 992 */ 993 tcp_bound_rto(sk); 994 } 995 996 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst) 997 { 998 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); 999 1000 if (!cwnd) 1001 cwnd = TCP_INIT_CWND; 1002 return min_t(__u32, cwnd, tp->snd_cwnd_clamp); 1003 } 1004 1005 struct tcp_sacktag_state { 1006 /* Timestamps for earliest and latest never-retransmitted segment 1007 * that was SACKed. RTO needs the earliest RTT to stay conservative, 1008 * but congestion control should still get an accurate delay signal. 1009 */ 1010 u64 first_sackt; 1011 u64 last_sackt; 1012 u32 reord; 1013 u32 sack_delivered; 1014 int flag; 1015 unsigned int mss_now; 1016 struct rate_sample *rate; 1017 }; 1018 1019 /* Take a notice that peer is sending D-SACKs. Skip update of data delivery 1020 * and spurious retransmission information if this DSACK is unlikely caused by 1021 * sender's action: 1022 * - DSACKed sequence range is larger than maximum receiver's window. 1023 * - Total no. of DSACKed segments exceed the total no. of retransmitted segs. 1024 */ 1025 static u32 tcp_dsack_seen(struct tcp_sock *tp, u32 start_seq, 1026 u32 end_seq, struct tcp_sacktag_state *state) 1027 { 1028 u32 seq_len, dup_segs = 1; 1029 1030 if (!before(start_seq, end_seq)) 1031 return 0; 1032 1033 seq_len = end_seq - start_seq; 1034 /* Dubious DSACK: DSACKed range greater than maximum advertised rwnd */ 1035 if (seq_len > tp->max_window) 1036 return 0; 1037 if (seq_len > tp->mss_cache) 1038 dup_segs = DIV_ROUND_UP(seq_len, tp->mss_cache); 1039 else if (tp->tlp_high_seq && tp->tlp_high_seq == end_seq) 1040 state->flag |= FLAG_DSACK_TLP; 1041 1042 tp->dsack_dups += dup_segs; 1043 /* Skip the DSACK if dup segs weren't retransmitted by sender */ 1044 if (tp->dsack_dups > tp->total_retrans) 1045 return 0; 1046 1047 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN; 1048 /* We increase the RACK ordering window in rounds where we receive 1049 * DSACKs that may have been due to reordering causing RACK to trigger 1050 * a spurious fast recovery. Thus RACK ignores DSACKs that happen 1051 * without having seen reordering, or that match TLP probes (TLP 1052 * is timer-driven, not triggered by RACK). 1053 */ 1054 if (tp->reord_seen && !(state->flag & FLAG_DSACK_TLP)) 1055 tp->rack.dsack_seen = 1; 1056 1057 state->flag |= FLAG_DSACKING_ACK; 1058 /* A spurious retransmission is delivered */ 1059 state->sack_delivered += dup_segs; 1060 1061 return dup_segs; 1062 } 1063 1064 /* It's reordering when higher sequence was delivered (i.e. sacked) before 1065 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering 1066 * distance is approximated in full-mss packet distance ("reordering"). 1067 */ 1068 static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq, 1069 const int ts) 1070 { 1071 struct tcp_sock *tp = tcp_sk(sk); 1072 const u32 mss = tp->mss_cache; 1073 u32 fack, metric; 1074 1075 fack = tcp_highest_sack_seq(tp); 1076 if (!before(low_seq, fack)) 1077 return; 1078 1079 metric = fack - low_seq; 1080 if ((metric > tp->reordering * mss) && mss) { 1081 #if FASTRETRANS_DEBUG > 1 1082 pr_debug("Disorder%d %d %u f%u s%u rr%d\n", 1083 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, 1084 tp->reordering, 1085 0, 1086 tp->sacked_out, 1087 tp->undo_marker ? tp->undo_retrans : 0); 1088 #endif 1089 tp->reordering = min_t(u32, (metric + mss - 1) / mss, 1090 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering)); 1091 } 1092 1093 /* This exciting event is worth to be remembered. 8) */ 1094 tp->reord_seen++; 1095 NET_INC_STATS(sock_net(sk), 1096 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER); 1097 } 1098 1099 /* This must be called before lost_out or retrans_out are updated 1100 * on a new loss, because we want to know if all skbs previously 1101 * known to be lost have already been retransmitted, indicating 1102 * that this newly lost skb is our next skb to retransmit. 1103 */ 1104 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb) 1105 { 1106 if ((!tp->retransmit_skb_hint && tp->retrans_out >= tp->lost_out) || 1107 (tp->retransmit_skb_hint && 1108 before(TCP_SKB_CB(skb)->seq, 1109 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))) 1110 tp->retransmit_skb_hint = skb; 1111 } 1112 1113 /* Sum the number of packets on the wire we have marked as lost, and 1114 * notify the congestion control module that the given skb was marked lost. 1115 */ 1116 static void tcp_notify_skb_loss_event(struct tcp_sock *tp, const struct sk_buff *skb) 1117 { 1118 tp->lost += tcp_skb_pcount(skb); 1119 } 1120 1121 void tcp_mark_skb_lost(struct sock *sk, struct sk_buff *skb) 1122 { 1123 __u8 sacked = TCP_SKB_CB(skb)->sacked; 1124 struct tcp_sock *tp = tcp_sk(sk); 1125 1126 if (sacked & TCPCB_SACKED_ACKED) 1127 return; 1128 1129 tcp_verify_retransmit_hint(tp, skb); 1130 if (sacked & TCPCB_LOST) { 1131 if (sacked & TCPCB_SACKED_RETRANS) { 1132 /* Account for retransmits that are lost again */ 1133 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1134 tp->retrans_out -= tcp_skb_pcount(skb); 1135 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT, 1136 tcp_skb_pcount(skb)); 1137 tcp_notify_skb_loss_event(tp, skb); 1138 } 1139 } else { 1140 tp->lost_out += tcp_skb_pcount(skb); 1141 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1142 tcp_notify_skb_loss_event(tp, skb); 1143 } 1144 } 1145 1146 /* Updates the delivered and delivered_ce counts */ 1147 static void tcp_count_delivered(struct tcp_sock *tp, u32 delivered, 1148 bool ece_ack) 1149 { 1150 tp->delivered += delivered; 1151 if (ece_ack) 1152 tp->delivered_ce += delivered; 1153 } 1154 1155 /* This procedure tags the retransmission queue when SACKs arrive. 1156 * 1157 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). 1158 * Packets in queue with these bits set are counted in variables 1159 * sacked_out, retrans_out and lost_out, correspondingly. 1160 * 1161 * Valid combinations are: 1162 * Tag InFlight Description 1163 * 0 1 - orig segment is in flight. 1164 * S 0 - nothing flies, orig reached receiver. 1165 * L 0 - nothing flies, orig lost by net. 1166 * R 2 - both orig and retransmit are in flight. 1167 * L|R 1 - orig is lost, retransmit is in flight. 1168 * S|R 1 - orig reached receiver, retrans is still in flight. 1169 * (L|S|R is logically valid, it could occur when L|R is sacked, 1170 * but it is equivalent to plain S and code short-circuits it to S. 1171 * L|S is logically invalid, it would mean -1 packet in flight 8)) 1172 * 1173 * These 6 states form finite state machine, controlled by the following events: 1174 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) 1175 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) 1176 * 3. Loss detection event of two flavors: 1177 * A. Scoreboard estimator decided the packet is lost. 1178 * A'. Reno "three dupacks" marks head of queue lost. 1179 * B. SACK arrives sacking SND.NXT at the moment, when the 1180 * segment was retransmitted. 1181 * 4. D-SACK added new rule: D-SACK changes any tag to S. 1182 * 1183 * It is pleasant to note, that state diagram turns out to be commutative, 1184 * so that we are allowed not to be bothered by order of our actions, 1185 * when multiple events arrive simultaneously. (see the function below). 1186 * 1187 * Reordering detection. 1188 * -------------------- 1189 * Reordering metric is maximal distance, which a packet can be displaced 1190 * in packet stream. With SACKs we can estimate it: 1191 * 1192 * 1. SACK fills old hole and the corresponding segment was not 1193 * ever retransmitted -> reordering. Alas, we cannot use it 1194 * when segment was retransmitted. 1195 * 2. The last flaw is solved with D-SACK. D-SACK arrives 1196 * for retransmitted and already SACKed segment -> reordering.. 1197 * Both of these heuristics are not used in Loss state, when we cannot 1198 * account for retransmits accurately. 1199 * 1200 * SACK block validation. 1201 * ---------------------- 1202 * 1203 * SACK block range validation checks that the received SACK block fits to 1204 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT. 1205 * Note that SND.UNA is not included to the range though being valid because 1206 * it means that the receiver is rather inconsistent with itself reporting 1207 * SACK reneging when it should advance SND.UNA. Such SACK block this is 1208 * perfectly valid, however, in light of RFC2018 which explicitly states 1209 * that "SACK block MUST reflect the newest segment. Even if the newest 1210 * segment is going to be discarded ...", not that it looks very clever 1211 * in case of head skb. Due to potentional receiver driven attacks, we 1212 * choose to avoid immediate execution of a walk in write queue due to 1213 * reneging and defer head skb's loss recovery to standard loss recovery 1214 * procedure that will eventually trigger (nothing forbids us doing this). 1215 * 1216 * Implements also blockage to start_seq wrap-around. Problem lies in the 1217 * fact that though start_seq (s) is before end_seq (i.e., not reversed), 1218 * there's no guarantee that it will be before snd_nxt (n). The problem 1219 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt 1220 * wrap (s_w): 1221 * 1222 * <- outs wnd -> <- wrapzone -> 1223 * u e n u_w e_w s n_w 1224 * | | | | | | | 1225 * |<------------+------+----- TCP seqno space --------------+---------->| 1226 * ...-- <2^31 ->| |<--------... 1227 * ...---- >2^31 ------>| |<--------... 1228 * 1229 * Current code wouldn't be vulnerable but it's better still to discard such 1230 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat 1231 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in 1232 * snd_nxt wrap -> snd_una region will then become "well defined", i.e., 1233 * equal to the ideal case (infinite seqno space without wrap caused issues). 1234 * 1235 * With D-SACK the lower bound is extended to cover sequence space below 1236 * SND.UNA down to undo_marker, which is the last point of interest. Yet 1237 * again, D-SACK block must not to go across snd_una (for the same reason as 1238 * for the normal SACK blocks, explained above). But there all simplicity 1239 * ends, TCP might receive valid D-SACKs below that. As long as they reside 1240 * fully below undo_marker they do not affect behavior in anyway and can 1241 * therefore be safely ignored. In rare cases (which are more or less 1242 * theoretical ones), the D-SACK will nicely cross that boundary due to skb 1243 * fragmentation and packet reordering past skb's retransmission. To consider 1244 * them correctly, the acceptable range must be extended even more though 1245 * the exact amount is rather hard to quantify. However, tp->max_window can 1246 * be used as an exaggerated estimate. 1247 */ 1248 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack, 1249 u32 start_seq, u32 end_seq) 1250 { 1251 /* Too far in future, or reversed (interpretation is ambiguous) */ 1252 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq)) 1253 return false; 1254 1255 /* Nasty start_seq wrap-around check (see comments above) */ 1256 if (!before(start_seq, tp->snd_nxt)) 1257 return false; 1258 1259 /* In outstanding window? ...This is valid exit for D-SACKs too. 1260 * start_seq == snd_una is non-sensical (see comments above) 1261 */ 1262 if (after(start_seq, tp->snd_una)) 1263 return true; 1264 1265 if (!is_dsack || !tp->undo_marker) 1266 return false; 1267 1268 /* ...Then it's D-SACK, and must reside below snd_una completely */ 1269 if (after(end_seq, tp->snd_una)) 1270 return false; 1271 1272 if (!before(start_seq, tp->undo_marker)) 1273 return true; 1274 1275 /* Too old */ 1276 if (!after(end_seq, tp->undo_marker)) 1277 return false; 1278 1279 /* Undo_marker boundary crossing (overestimates a lot). Known already: 1280 * start_seq < undo_marker and end_seq >= undo_marker. 1281 */ 1282 return !before(start_seq, end_seq - tp->max_window); 1283 } 1284 1285 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb, 1286 struct tcp_sack_block_wire *sp, int num_sacks, 1287 u32 prior_snd_una, struct tcp_sacktag_state *state) 1288 { 1289 struct tcp_sock *tp = tcp_sk(sk); 1290 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq); 1291 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq); 1292 u32 dup_segs; 1293 1294 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) { 1295 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV); 1296 } else if (num_sacks > 1) { 1297 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq); 1298 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq); 1299 1300 if (after(end_seq_0, end_seq_1) || before(start_seq_0, start_seq_1)) 1301 return false; 1302 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKOFORECV); 1303 } else { 1304 return false; 1305 } 1306 1307 dup_segs = tcp_dsack_seen(tp, start_seq_0, end_seq_0, state); 1308 if (!dup_segs) { /* Skip dubious DSACK */ 1309 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKIGNOREDDUBIOUS); 1310 return false; 1311 } 1312 1313 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECVSEGS, dup_segs); 1314 1315 /* D-SACK for already forgotten data... Do dumb counting. */ 1316 if (tp->undo_marker && tp->undo_retrans > 0 && 1317 !after(end_seq_0, prior_snd_una) && 1318 after(end_seq_0, tp->undo_marker)) 1319 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - dup_segs); 1320 1321 return true; 1322 } 1323 1324 /* Check if skb is fully within the SACK block. In presence of GSO skbs, 1325 * the incoming SACK may not exactly match but we can find smaller MSS 1326 * aligned portion of it that matches. Therefore we might need to fragment 1327 * which may fail and creates some hassle (caller must handle error case 1328 * returns). 1329 * 1330 * FIXME: this could be merged to shift decision code 1331 */ 1332 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb, 1333 u32 start_seq, u32 end_seq) 1334 { 1335 int err; 1336 bool in_sack; 1337 unsigned int pkt_len; 1338 unsigned int mss; 1339 1340 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1341 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1342 1343 if (tcp_skb_pcount(skb) > 1 && !in_sack && 1344 after(TCP_SKB_CB(skb)->end_seq, start_seq)) { 1345 mss = tcp_skb_mss(skb); 1346 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); 1347 1348 if (!in_sack) { 1349 pkt_len = start_seq - TCP_SKB_CB(skb)->seq; 1350 if (pkt_len < mss) 1351 pkt_len = mss; 1352 } else { 1353 pkt_len = end_seq - TCP_SKB_CB(skb)->seq; 1354 if (pkt_len < mss) 1355 return -EINVAL; 1356 } 1357 1358 /* Round if necessary so that SACKs cover only full MSSes 1359 * and/or the remaining small portion (if present) 1360 */ 1361 if (pkt_len > mss) { 1362 unsigned int new_len = (pkt_len / mss) * mss; 1363 if (!in_sack && new_len < pkt_len) 1364 new_len += mss; 1365 pkt_len = new_len; 1366 } 1367 1368 if (pkt_len >= skb->len && !in_sack) 1369 return 0; 1370 1371 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, 1372 pkt_len, mss, GFP_ATOMIC); 1373 if (err < 0) 1374 return err; 1375 } 1376 1377 return in_sack; 1378 } 1379 1380 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */ 1381 static u8 tcp_sacktag_one(struct sock *sk, 1382 struct tcp_sacktag_state *state, u8 sacked, 1383 u32 start_seq, u32 end_seq, 1384 int dup_sack, int pcount, 1385 u64 xmit_time) 1386 { 1387 struct tcp_sock *tp = tcp_sk(sk); 1388 1389 /* Account D-SACK for retransmitted packet. */ 1390 if (dup_sack && (sacked & TCPCB_RETRANS)) { 1391 if (tp->undo_marker && tp->undo_retrans > 0 && 1392 after(end_seq, tp->undo_marker)) 1393 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - pcount); 1394 if ((sacked & TCPCB_SACKED_ACKED) && 1395 before(start_seq, state->reord)) 1396 state->reord = start_seq; 1397 } 1398 1399 /* Nothing to do; acked frame is about to be dropped (was ACKed). */ 1400 if (!after(end_seq, tp->snd_una)) 1401 return sacked; 1402 1403 if (!(sacked & TCPCB_SACKED_ACKED)) { 1404 tcp_rack_advance(tp, sacked, end_seq, xmit_time); 1405 1406 if (sacked & TCPCB_SACKED_RETRANS) { 1407 /* If the segment is not tagged as lost, 1408 * we do not clear RETRANS, believing 1409 * that retransmission is still in flight. 1410 */ 1411 if (sacked & TCPCB_LOST) { 1412 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); 1413 tp->lost_out -= pcount; 1414 tp->retrans_out -= pcount; 1415 } 1416 } else { 1417 if (!(sacked & TCPCB_RETRANS)) { 1418 /* New sack for not retransmitted frame, 1419 * which was in hole. It is reordering. 1420 */ 1421 if (before(start_seq, 1422 tcp_highest_sack_seq(tp)) && 1423 before(start_seq, state->reord)) 1424 state->reord = start_seq; 1425 1426 if (!after(end_seq, tp->high_seq)) 1427 state->flag |= FLAG_ORIG_SACK_ACKED; 1428 if (state->first_sackt == 0) 1429 state->first_sackt = xmit_time; 1430 state->last_sackt = xmit_time; 1431 } 1432 1433 if (sacked & TCPCB_LOST) { 1434 sacked &= ~TCPCB_LOST; 1435 tp->lost_out -= pcount; 1436 } 1437 } 1438 1439 sacked |= TCPCB_SACKED_ACKED; 1440 state->flag |= FLAG_DATA_SACKED; 1441 tp->sacked_out += pcount; 1442 /* Out-of-order packets delivered */ 1443 state->sack_delivered += pcount; 1444 1445 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */ 1446 if (tp->lost_skb_hint && 1447 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq)) 1448 tp->lost_cnt_hint += pcount; 1449 } 1450 1451 /* D-SACK. We can detect redundant retransmission in S|R and plain R 1452 * frames and clear it. undo_retrans is decreased above, L|R frames 1453 * are accounted above as well. 1454 */ 1455 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) { 1456 sacked &= ~TCPCB_SACKED_RETRANS; 1457 tp->retrans_out -= pcount; 1458 } 1459 1460 return sacked; 1461 } 1462 1463 /* Shift newly-SACKed bytes from this skb to the immediately previous 1464 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such. 1465 */ 1466 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev, 1467 struct sk_buff *skb, 1468 struct tcp_sacktag_state *state, 1469 unsigned int pcount, int shifted, int mss, 1470 bool dup_sack) 1471 { 1472 struct tcp_sock *tp = tcp_sk(sk); 1473 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */ 1474 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */ 1475 1476 BUG_ON(!pcount); 1477 1478 /* Adjust counters and hints for the newly sacked sequence 1479 * range but discard the return value since prev is already 1480 * marked. We must tag the range first because the seq 1481 * advancement below implicitly advances 1482 * tcp_highest_sack_seq() when skb is highest_sack. 1483 */ 1484 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked, 1485 start_seq, end_seq, dup_sack, pcount, 1486 tcp_skb_timestamp_us(skb)); 1487 tcp_rate_skb_delivered(sk, skb, state->rate); 1488 1489 if (skb == tp->lost_skb_hint) 1490 tp->lost_cnt_hint += pcount; 1491 1492 TCP_SKB_CB(prev)->end_seq += shifted; 1493 TCP_SKB_CB(skb)->seq += shifted; 1494 1495 tcp_skb_pcount_add(prev, pcount); 1496 WARN_ON_ONCE(tcp_skb_pcount(skb) < pcount); 1497 tcp_skb_pcount_add(skb, -pcount); 1498 1499 /* When we're adding to gso_segs == 1, gso_size will be zero, 1500 * in theory this shouldn't be necessary but as long as DSACK 1501 * code can come after this skb later on it's better to keep 1502 * setting gso_size to something. 1503 */ 1504 if (!TCP_SKB_CB(prev)->tcp_gso_size) 1505 TCP_SKB_CB(prev)->tcp_gso_size = mss; 1506 1507 /* CHECKME: To clear or not to clear? Mimics normal skb currently */ 1508 if (tcp_skb_pcount(skb) <= 1) 1509 TCP_SKB_CB(skb)->tcp_gso_size = 0; 1510 1511 /* Difference in this won't matter, both ACKed by the same cumul. ACK */ 1512 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS); 1513 1514 if (skb->len > 0) { 1515 BUG_ON(!tcp_skb_pcount(skb)); 1516 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED); 1517 return false; 1518 } 1519 1520 /* Whole SKB was eaten :-) */ 1521 1522 if (skb == tp->retransmit_skb_hint) 1523 tp->retransmit_skb_hint = prev; 1524 if (skb == tp->lost_skb_hint) { 1525 tp->lost_skb_hint = prev; 1526 tp->lost_cnt_hint -= tcp_skb_pcount(prev); 1527 } 1528 1529 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 1530 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor; 1531 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1532 TCP_SKB_CB(prev)->end_seq++; 1533 1534 if (skb == tcp_highest_sack(sk)) 1535 tcp_advance_highest_sack(sk, skb); 1536 1537 tcp_skb_collapse_tstamp(prev, skb); 1538 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp)) 1539 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0; 1540 1541 tcp_rtx_queue_unlink_and_free(skb, sk); 1542 1543 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED); 1544 1545 return true; 1546 } 1547 1548 /* I wish gso_size would have a bit more sane initialization than 1549 * something-or-zero which complicates things 1550 */ 1551 static int tcp_skb_seglen(const struct sk_buff *skb) 1552 { 1553 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb); 1554 } 1555 1556 /* Shifting pages past head area doesn't work */ 1557 static int skb_can_shift(const struct sk_buff *skb) 1558 { 1559 return !skb_headlen(skb) && skb_is_nonlinear(skb); 1560 } 1561 1562 int tcp_skb_shift(struct sk_buff *to, struct sk_buff *from, 1563 int pcount, int shiftlen) 1564 { 1565 /* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE) 1566 * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need 1567 * to make sure not storing more than 65535 * 8 bytes per skb, 1568 * even if current MSS is bigger. 1569 */ 1570 if (unlikely(to->len + shiftlen >= 65535 * TCP_MIN_GSO_SIZE)) 1571 return 0; 1572 if (unlikely(tcp_skb_pcount(to) + pcount > 65535)) 1573 return 0; 1574 return skb_shift(to, from, shiftlen); 1575 } 1576 1577 /* Try collapsing SACK blocks spanning across multiple skbs to a single 1578 * skb. 1579 */ 1580 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb, 1581 struct tcp_sacktag_state *state, 1582 u32 start_seq, u32 end_seq, 1583 bool dup_sack) 1584 { 1585 struct tcp_sock *tp = tcp_sk(sk); 1586 struct sk_buff *prev; 1587 int mss; 1588 int pcount = 0; 1589 int len; 1590 int in_sack; 1591 1592 /* Normally R but no L won't result in plain S */ 1593 if (!dup_sack && 1594 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS) 1595 goto fallback; 1596 if (!skb_can_shift(skb)) 1597 goto fallback; 1598 /* This frame is about to be dropped (was ACKed). */ 1599 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1600 goto fallback; 1601 1602 /* Can only happen with delayed DSACK + discard craziness */ 1603 prev = skb_rb_prev(skb); 1604 if (!prev) 1605 goto fallback; 1606 1607 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) 1608 goto fallback; 1609 1610 if (!tcp_skb_can_collapse(prev, skb)) 1611 goto fallback; 1612 1613 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1614 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1615 1616 if (in_sack) { 1617 len = skb->len; 1618 pcount = tcp_skb_pcount(skb); 1619 mss = tcp_skb_seglen(skb); 1620 1621 /* TODO: Fix DSACKs to not fragment already SACKed and we can 1622 * drop this restriction as unnecessary 1623 */ 1624 if (mss != tcp_skb_seglen(prev)) 1625 goto fallback; 1626 } else { 1627 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq)) 1628 goto noop; 1629 /* CHECKME: This is non-MSS split case only?, this will 1630 * cause skipped skbs due to advancing loop btw, original 1631 * has that feature too 1632 */ 1633 if (tcp_skb_pcount(skb) <= 1) 1634 goto noop; 1635 1636 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); 1637 if (!in_sack) { 1638 /* TODO: head merge to next could be attempted here 1639 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)), 1640 * though it might not be worth of the additional hassle 1641 * 1642 * ...we can probably just fallback to what was done 1643 * previously. We could try merging non-SACKed ones 1644 * as well but it probably isn't going to buy off 1645 * because later SACKs might again split them, and 1646 * it would make skb timestamp tracking considerably 1647 * harder problem. 1648 */ 1649 goto fallback; 1650 } 1651 1652 len = end_seq - TCP_SKB_CB(skb)->seq; 1653 BUG_ON(len < 0); 1654 BUG_ON(len > skb->len); 1655 1656 /* MSS boundaries should be honoured or else pcount will 1657 * severely break even though it makes things bit trickier. 1658 * Optimize common case to avoid most of the divides 1659 */ 1660 mss = tcp_skb_mss(skb); 1661 1662 /* TODO: Fix DSACKs to not fragment already SACKed and we can 1663 * drop this restriction as unnecessary 1664 */ 1665 if (mss != tcp_skb_seglen(prev)) 1666 goto fallback; 1667 1668 if (len == mss) { 1669 pcount = 1; 1670 } else if (len < mss) { 1671 goto noop; 1672 } else { 1673 pcount = len / mss; 1674 len = pcount * mss; 1675 } 1676 } 1677 1678 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */ 1679 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una)) 1680 goto fallback; 1681 1682 if (!tcp_skb_shift(prev, skb, pcount, len)) 1683 goto fallback; 1684 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack)) 1685 goto out; 1686 1687 /* Hole filled allows collapsing with the next as well, this is very 1688 * useful when hole on every nth skb pattern happens 1689 */ 1690 skb = skb_rb_next(prev); 1691 if (!skb) 1692 goto out; 1693 1694 if (!skb_can_shift(skb) || 1695 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) || 1696 (mss != tcp_skb_seglen(skb))) 1697 goto out; 1698 1699 if (!tcp_skb_can_collapse(prev, skb)) 1700 goto out; 1701 len = skb->len; 1702 pcount = tcp_skb_pcount(skb); 1703 if (tcp_skb_shift(prev, skb, pcount, len)) 1704 tcp_shifted_skb(sk, prev, skb, state, pcount, 1705 len, mss, 0); 1706 1707 out: 1708 return prev; 1709 1710 noop: 1711 return skb; 1712 1713 fallback: 1714 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK); 1715 return NULL; 1716 } 1717 1718 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk, 1719 struct tcp_sack_block *next_dup, 1720 struct tcp_sacktag_state *state, 1721 u32 start_seq, u32 end_seq, 1722 bool dup_sack_in) 1723 { 1724 struct tcp_sock *tp = tcp_sk(sk); 1725 struct sk_buff *tmp; 1726 1727 skb_rbtree_walk_from(skb) { 1728 int in_sack = 0; 1729 bool dup_sack = dup_sack_in; 1730 1731 /* queue is in-order => we can short-circuit the walk early */ 1732 if (!before(TCP_SKB_CB(skb)->seq, end_seq)) 1733 break; 1734 1735 if (next_dup && 1736 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) { 1737 in_sack = tcp_match_skb_to_sack(sk, skb, 1738 next_dup->start_seq, 1739 next_dup->end_seq); 1740 if (in_sack > 0) 1741 dup_sack = true; 1742 } 1743 1744 /* skb reference here is a bit tricky to get right, since 1745 * shifting can eat and free both this skb and the next, 1746 * so not even _safe variant of the loop is enough. 1747 */ 1748 if (in_sack <= 0) { 1749 tmp = tcp_shift_skb_data(sk, skb, state, 1750 start_seq, end_seq, dup_sack); 1751 if (tmp) { 1752 if (tmp != skb) { 1753 skb = tmp; 1754 continue; 1755 } 1756 1757 in_sack = 0; 1758 } else { 1759 in_sack = tcp_match_skb_to_sack(sk, skb, 1760 start_seq, 1761 end_seq); 1762 } 1763 } 1764 1765 if (unlikely(in_sack < 0)) 1766 break; 1767 1768 if (in_sack) { 1769 TCP_SKB_CB(skb)->sacked = 1770 tcp_sacktag_one(sk, 1771 state, 1772 TCP_SKB_CB(skb)->sacked, 1773 TCP_SKB_CB(skb)->seq, 1774 TCP_SKB_CB(skb)->end_seq, 1775 dup_sack, 1776 tcp_skb_pcount(skb), 1777 tcp_skb_timestamp_us(skb)); 1778 tcp_rate_skb_delivered(sk, skb, state->rate); 1779 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1780 list_del_init(&skb->tcp_tsorted_anchor); 1781 1782 if (!before(TCP_SKB_CB(skb)->seq, 1783 tcp_highest_sack_seq(tp))) 1784 tcp_advance_highest_sack(sk, skb); 1785 } 1786 } 1787 return skb; 1788 } 1789 1790 static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk, u32 seq) 1791 { 1792 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node; 1793 struct sk_buff *skb; 1794 1795 while (*p) { 1796 parent = *p; 1797 skb = rb_to_skb(parent); 1798 if (before(seq, TCP_SKB_CB(skb)->seq)) { 1799 p = &parent->rb_left; 1800 continue; 1801 } 1802 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) { 1803 p = &parent->rb_right; 1804 continue; 1805 } 1806 return skb; 1807 } 1808 return NULL; 1809 } 1810 1811 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk, 1812 u32 skip_to_seq) 1813 { 1814 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq)) 1815 return skb; 1816 1817 return tcp_sacktag_bsearch(sk, skip_to_seq); 1818 } 1819 1820 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb, 1821 struct sock *sk, 1822 struct tcp_sack_block *next_dup, 1823 struct tcp_sacktag_state *state, 1824 u32 skip_to_seq) 1825 { 1826 if (!next_dup) 1827 return skb; 1828 1829 if (before(next_dup->start_seq, skip_to_seq)) { 1830 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq); 1831 skb = tcp_sacktag_walk(skb, sk, NULL, state, 1832 next_dup->start_seq, next_dup->end_seq, 1833 1); 1834 } 1835 1836 return skb; 1837 } 1838 1839 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache) 1840 { 1841 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); 1842 } 1843 1844 static int 1845 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb, 1846 u32 prior_snd_una, struct tcp_sacktag_state *state) 1847 { 1848 struct tcp_sock *tp = tcp_sk(sk); 1849 const unsigned char *ptr = (skb_transport_header(ack_skb) + 1850 TCP_SKB_CB(ack_skb)->sacked); 1851 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2); 1852 struct tcp_sack_block sp[TCP_NUM_SACKS]; 1853 struct tcp_sack_block *cache; 1854 struct sk_buff *skb; 1855 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3); 1856 int used_sacks; 1857 bool found_dup_sack = false; 1858 int i, j; 1859 int first_sack_index; 1860 1861 state->flag = 0; 1862 state->reord = tp->snd_nxt; 1863 1864 if (!tp->sacked_out) 1865 tcp_highest_sack_reset(sk); 1866 1867 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire, 1868 num_sacks, prior_snd_una, state); 1869 1870 /* Eliminate too old ACKs, but take into 1871 * account more or less fresh ones, they can 1872 * contain valid SACK info. 1873 */ 1874 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) 1875 return 0; 1876 1877 if (!tp->packets_out) 1878 goto out; 1879 1880 used_sacks = 0; 1881 first_sack_index = 0; 1882 for (i = 0; i < num_sacks; i++) { 1883 bool dup_sack = !i && found_dup_sack; 1884 1885 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq); 1886 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq); 1887 1888 if (!tcp_is_sackblock_valid(tp, dup_sack, 1889 sp[used_sacks].start_seq, 1890 sp[used_sacks].end_seq)) { 1891 int mib_idx; 1892 1893 if (dup_sack) { 1894 if (!tp->undo_marker) 1895 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO; 1896 else 1897 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD; 1898 } else { 1899 /* Don't count olds caused by ACK reordering */ 1900 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) && 1901 !after(sp[used_sacks].end_seq, tp->snd_una)) 1902 continue; 1903 mib_idx = LINUX_MIB_TCPSACKDISCARD; 1904 } 1905 1906 NET_INC_STATS(sock_net(sk), mib_idx); 1907 if (i == 0) 1908 first_sack_index = -1; 1909 continue; 1910 } 1911 1912 /* Ignore very old stuff early */ 1913 if (!after(sp[used_sacks].end_seq, prior_snd_una)) { 1914 if (i == 0) 1915 first_sack_index = -1; 1916 continue; 1917 } 1918 1919 used_sacks++; 1920 } 1921 1922 /* order SACK blocks to allow in order walk of the retrans queue */ 1923 for (i = used_sacks - 1; i > 0; i--) { 1924 for (j = 0; j < i; j++) { 1925 if (after(sp[j].start_seq, sp[j + 1].start_seq)) { 1926 swap(sp[j], sp[j + 1]); 1927 1928 /* Track where the first SACK block goes to */ 1929 if (j == first_sack_index) 1930 first_sack_index = j + 1; 1931 } 1932 } 1933 } 1934 1935 state->mss_now = tcp_current_mss(sk); 1936 skb = NULL; 1937 i = 0; 1938 1939 if (!tp->sacked_out) { 1940 /* It's already past, so skip checking against it */ 1941 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); 1942 } else { 1943 cache = tp->recv_sack_cache; 1944 /* Skip empty blocks in at head of the cache */ 1945 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq && 1946 !cache->end_seq) 1947 cache++; 1948 } 1949 1950 while (i < used_sacks) { 1951 u32 start_seq = sp[i].start_seq; 1952 u32 end_seq = sp[i].end_seq; 1953 bool dup_sack = (found_dup_sack && (i == first_sack_index)); 1954 struct tcp_sack_block *next_dup = NULL; 1955 1956 if (found_dup_sack && ((i + 1) == first_sack_index)) 1957 next_dup = &sp[i + 1]; 1958 1959 /* Skip too early cached blocks */ 1960 while (tcp_sack_cache_ok(tp, cache) && 1961 !before(start_seq, cache->end_seq)) 1962 cache++; 1963 1964 /* Can skip some work by looking recv_sack_cache? */ 1965 if (tcp_sack_cache_ok(tp, cache) && !dup_sack && 1966 after(end_seq, cache->start_seq)) { 1967 1968 /* Head todo? */ 1969 if (before(start_seq, cache->start_seq)) { 1970 skb = tcp_sacktag_skip(skb, sk, start_seq); 1971 skb = tcp_sacktag_walk(skb, sk, next_dup, 1972 state, 1973 start_seq, 1974 cache->start_seq, 1975 dup_sack); 1976 } 1977 1978 /* Rest of the block already fully processed? */ 1979 if (!after(end_seq, cache->end_seq)) 1980 goto advance_sp; 1981 1982 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup, 1983 state, 1984 cache->end_seq); 1985 1986 /* ...tail remains todo... */ 1987 if (tcp_highest_sack_seq(tp) == cache->end_seq) { 1988 /* ...but better entrypoint exists! */ 1989 skb = tcp_highest_sack(sk); 1990 if (!skb) 1991 break; 1992 cache++; 1993 goto walk; 1994 } 1995 1996 skb = tcp_sacktag_skip(skb, sk, cache->end_seq); 1997 /* Check overlap against next cached too (past this one already) */ 1998 cache++; 1999 continue; 2000 } 2001 2002 if (!before(start_seq, tcp_highest_sack_seq(tp))) { 2003 skb = tcp_highest_sack(sk); 2004 if (!skb) 2005 break; 2006 } 2007 skb = tcp_sacktag_skip(skb, sk, start_seq); 2008 2009 walk: 2010 skb = tcp_sacktag_walk(skb, sk, next_dup, state, 2011 start_seq, end_seq, dup_sack); 2012 2013 advance_sp: 2014 i++; 2015 } 2016 2017 /* Clear the head of the cache sack blocks so we can skip it next time */ 2018 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) { 2019 tp->recv_sack_cache[i].start_seq = 0; 2020 tp->recv_sack_cache[i].end_seq = 0; 2021 } 2022 for (j = 0; j < used_sacks; j++) 2023 tp->recv_sack_cache[i++] = sp[j]; 2024 2025 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker) 2026 tcp_check_sack_reordering(sk, state->reord, 0); 2027 2028 tcp_verify_left_out(tp); 2029 out: 2030 2031 #if FASTRETRANS_DEBUG > 0 2032 WARN_ON((int)tp->sacked_out < 0); 2033 WARN_ON((int)tp->lost_out < 0); 2034 WARN_ON((int)tp->retrans_out < 0); 2035 WARN_ON((int)tcp_packets_in_flight(tp) < 0); 2036 #endif 2037 return state->flag; 2038 } 2039 2040 /* Limits sacked_out so that sum with lost_out isn't ever larger than 2041 * packets_out. Returns false if sacked_out adjustement wasn't necessary. 2042 */ 2043 static bool tcp_limit_reno_sacked(struct tcp_sock *tp) 2044 { 2045 u32 holes; 2046 2047 holes = max(tp->lost_out, 1U); 2048 holes = min(holes, tp->packets_out); 2049 2050 if ((tp->sacked_out + holes) > tp->packets_out) { 2051 tp->sacked_out = tp->packets_out - holes; 2052 return true; 2053 } 2054 return false; 2055 } 2056 2057 /* If we receive more dupacks than we expected counting segments 2058 * in assumption of absent reordering, interpret this as reordering. 2059 * The only another reason could be bug in receiver TCP. 2060 */ 2061 static void tcp_check_reno_reordering(struct sock *sk, const int addend) 2062 { 2063 struct tcp_sock *tp = tcp_sk(sk); 2064 2065 if (!tcp_limit_reno_sacked(tp)) 2066 return; 2067 2068 tp->reordering = min_t(u32, tp->packets_out + addend, 2069 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering)); 2070 tp->reord_seen++; 2071 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER); 2072 } 2073 2074 /* Emulate SACKs for SACKless connection: account for a new dupack. */ 2075 2076 static void tcp_add_reno_sack(struct sock *sk, int num_dupack, bool ece_ack) 2077 { 2078 if (num_dupack) { 2079 struct tcp_sock *tp = tcp_sk(sk); 2080 u32 prior_sacked = tp->sacked_out; 2081 s32 delivered; 2082 2083 tp->sacked_out += num_dupack; 2084 tcp_check_reno_reordering(sk, 0); 2085 delivered = tp->sacked_out - prior_sacked; 2086 if (delivered > 0) 2087 tcp_count_delivered(tp, delivered, ece_ack); 2088 tcp_verify_left_out(tp); 2089 } 2090 } 2091 2092 /* Account for ACK, ACKing some data in Reno Recovery phase. */ 2093 2094 static void tcp_remove_reno_sacks(struct sock *sk, int acked, bool ece_ack) 2095 { 2096 struct tcp_sock *tp = tcp_sk(sk); 2097 2098 if (acked > 0) { 2099 /* One ACK acked hole. The rest eat duplicate ACKs. */ 2100 tcp_count_delivered(tp, max_t(int, acked - tp->sacked_out, 1), 2101 ece_ack); 2102 if (acked - 1 >= tp->sacked_out) 2103 tp->sacked_out = 0; 2104 else 2105 tp->sacked_out -= acked - 1; 2106 } 2107 tcp_check_reno_reordering(sk, acked); 2108 tcp_verify_left_out(tp); 2109 } 2110 2111 static inline void tcp_reset_reno_sack(struct tcp_sock *tp) 2112 { 2113 tp->sacked_out = 0; 2114 } 2115 2116 void tcp_clear_retrans(struct tcp_sock *tp) 2117 { 2118 tp->retrans_out = 0; 2119 tp->lost_out = 0; 2120 tp->undo_marker = 0; 2121 tp->undo_retrans = -1; 2122 tp->sacked_out = 0; 2123 tp->rto_stamp = 0; 2124 tp->total_rto = 0; 2125 tp->total_rto_recoveries = 0; 2126 tp->total_rto_time = 0; 2127 } 2128 2129 static inline void tcp_init_undo(struct tcp_sock *tp) 2130 { 2131 tp->undo_marker = tp->snd_una; 2132 /* Retransmission still in flight may cause DSACKs later. */ 2133 tp->undo_retrans = tp->retrans_out ? : -1; 2134 } 2135 2136 static bool tcp_is_rack(const struct sock *sk) 2137 { 2138 return READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_recovery) & 2139 TCP_RACK_LOSS_DETECTION; 2140 } 2141 2142 /* If we detect SACK reneging, forget all SACK information 2143 * and reset tags completely, otherwise preserve SACKs. If receiver 2144 * dropped its ofo queue, we will know this due to reneging detection. 2145 */ 2146 static void tcp_timeout_mark_lost(struct sock *sk) 2147 { 2148 struct tcp_sock *tp = tcp_sk(sk); 2149 struct sk_buff *skb, *head; 2150 bool is_reneg; /* is receiver reneging on SACKs? */ 2151 2152 head = tcp_rtx_queue_head(sk); 2153 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED); 2154 if (is_reneg) { 2155 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING); 2156 tp->sacked_out = 0; 2157 /* Mark SACK reneging until we recover from this loss event. */ 2158 tp->is_sack_reneg = 1; 2159 } else if (tcp_is_reno(tp)) { 2160 tcp_reset_reno_sack(tp); 2161 } 2162 2163 skb = head; 2164 skb_rbtree_walk_from(skb) { 2165 if (is_reneg) 2166 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED; 2167 else if (tcp_is_rack(sk) && skb != head && 2168 tcp_rack_skb_timeout(tp, skb, 0) > 0) 2169 continue; /* Don't mark recently sent ones lost yet */ 2170 tcp_mark_skb_lost(sk, skb); 2171 } 2172 tcp_verify_left_out(tp); 2173 tcp_clear_all_retrans_hints(tp); 2174 } 2175 2176 /* Enter Loss state. */ 2177 void tcp_enter_loss(struct sock *sk) 2178 { 2179 const struct inet_connection_sock *icsk = inet_csk(sk); 2180 struct tcp_sock *tp = tcp_sk(sk); 2181 struct net *net = sock_net(sk); 2182 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery; 2183 u8 reordering; 2184 2185 tcp_timeout_mark_lost(sk); 2186 2187 /* Reduce ssthresh if it has not yet been made inside this window. */ 2188 if (icsk->icsk_ca_state <= TCP_CA_Disorder || 2189 !after(tp->high_seq, tp->snd_una) || 2190 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) { 2191 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2192 tp->prior_cwnd = tcp_snd_cwnd(tp); 2193 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 2194 tcp_ca_event(sk, CA_EVENT_LOSS); 2195 tcp_init_undo(tp); 2196 } 2197 tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + 1); 2198 tp->snd_cwnd_cnt = 0; 2199 tp->snd_cwnd_stamp = tcp_jiffies32; 2200 2201 /* Timeout in disordered state after receiving substantial DUPACKs 2202 * suggests that the degree of reordering is over-estimated. 2203 */ 2204 reordering = READ_ONCE(net->ipv4.sysctl_tcp_reordering); 2205 if (icsk->icsk_ca_state <= TCP_CA_Disorder && 2206 tp->sacked_out >= reordering) 2207 tp->reordering = min_t(unsigned int, tp->reordering, 2208 reordering); 2209 2210 tcp_set_ca_state(sk, TCP_CA_Loss); 2211 tp->high_seq = tp->snd_nxt; 2212 tcp_ecn_queue_cwr(tp); 2213 2214 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous 2215 * loss recovery is underway except recurring timeout(s) on 2216 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing 2217 */ 2218 tp->frto = READ_ONCE(net->ipv4.sysctl_tcp_frto) && 2219 (new_recovery || icsk->icsk_retransmits) && 2220 !inet_csk(sk)->icsk_mtup.probe_size; 2221 } 2222 2223 /* If ACK arrived pointing to a remembered SACK, it means that our 2224 * remembered SACKs do not reflect real state of receiver i.e. 2225 * receiver _host_ is heavily congested (or buggy). 2226 * 2227 * To avoid big spurious retransmission bursts due to transient SACK 2228 * scoreboard oddities that look like reneging, we give the receiver a 2229 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will 2230 * restore sanity to the SACK scoreboard. If the apparent reneging 2231 * persists until this RTO then we'll clear the SACK scoreboard. 2232 */ 2233 static bool tcp_check_sack_reneging(struct sock *sk, int *ack_flag) 2234 { 2235 if (*ack_flag & FLAG_SACK_RENEGING && 2236 *ack_flag & FLAG_SND_UNA_ADVANCED) { 2237 struct tcp_sock *tp = tcp_sk(sk); 2238 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4), 2239 msecs_to_jiffies(10)); 2240 2241 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 2242 delay, TCP_RTO_MAX); 2243 *ack_flag &= ~FLAG_SET_XMIT_TIMER; 2244 return true; 2245 } 2246 return false; 2247 } 2248 2249 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs 2250 * counter when SACK is enabled (without SACK, sacked_out is used for 2251 * that purpose). 2252 * 2253 * With reordering, holes may still be in flight, so RFC3517 recovery 2254 * uses pure sacked_out (total number of SACKed segments) even though 2255 * it violates the RFC that uses duplicate ACKs, often these are equal 2256 * but when e.g. out-of-window ACKs or packet duplication occurs, 2257 * they differ. Since neither occurs due to loss, TCP should really 2258 * ignore them. 2259 */ 2260 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp) 2261 { 2262 return tp->sacked_out + 1; 2263 } 2264 2265 /* Linux NewReno/SACK/ECN state machine. 2266 * -------------------------------------- 2267 * 2268 * "Open" Normal state, no dubious events, fast path. 2269 * "Disorder" In all the respects it is "Open", 2270 * but requires a bit more attention. It is entered when 2271 * we see some SACKs or dupacks. It is split of "Open" 2272 * mainly to move some processing from fast path to slow one. 2273 * "CWR" CWND was reduced due to some Congestion Notification event. 2274 * It can be ECN, ICMP source quench, local device congestion. 2275 * "Recovery" CWND was reduced, we are fast-retransmitting. 2276 * "Loss" CWND was reduced due to RTO timeout or SACK reneging. 2277 * 2278 * tcp_fastretrans_alert() is entered: 2279 * - each incoming ACK, if state is not "Open" 2280 * - when arrived ACK is unusual, namely: 2281 * * SACK 2282 * * Duplicate ACK. 2283 * * ECN ECE. 2284 * 2285 * Counting packets in flight is pretty simple. 2286 * 2287 * in_flight = packets_out - left_out + retrans_out 2288 * 2289 * packets_out is SND.NXT-SND.UNA counted in packets. 2290 * 2291 * retrans_out is number of retransmitted segments. 2292 * 2293 * left_out is number of segments left network, but not ACKed yet. 2294 * 2295 * left_out = sacked_out + lost_out 2296 * 2297 * sacked_out: Packets, which arrived to receiver out of order 2298 * and hence not ACKed. With SACKs this number is simply 2299 * amount of SACKed data. Even without SACKs 2300 * it is easy to give pretty reliable estimate of this number, 2301 * counting duplicate ACKs. 2302 * 2303 * lost_out: Packets lost by network. TCP has no explicit 2304 * "loss notification" feedback from network (for now). 2305 * It means that this number can be only _guessed_. 2306 * Actually, it is the heuristics to predict lossage that 2307 * distinguishes different algorithms. 2308 * 2309 * F.e. after RTO, when all the queue is considered as lost, 2310 * lost_out = packets_out and in_flight = retrans_out. 2311 * 2312 * Essentially, we have now a few algorithms detecting 2313 * lost packets. 2314 * 2315 * If the receiver supports SACK: 2316 * 2317 * RFC6675/3517: It is the conventional algorithm. A packet is 2318 * considered lost if the number of higher sequence packets 2319 * SACKed is greater than or equal the DUPACK thoreshold 2320 * (reordering). This is implemented in tcp_mark_head_lost and 2321 * tcp_update_scoreboard. 2322 * 2323 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm 2324 * (2017-) that checks timing instead of counting DUPACKs. 2325 * Essentially a packet is considered lost if it's not S/ACKed 2326 * after RTT + reordering_window, where both metrics are 2327 * dynamically measured and adjusted. This is implemented in 2328 * tcp_rack_mark_lost. 2329 * 2330 * If the receiver does not support SACK: 2331 * 2332 * NewReno (RFC6582): in Recovery we assume that one segment 2333 * is lost (classic Reno). While we are in Recovery and 2334 * a partial ACK arrives, we assume that one more packet 2335 * is lost (NewReno). This heuristics are the same in NewReno 2336 * and SACK. 2337 * 2338 * Really tricky (and requiring careful tuning) part of algorithm 2339 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue(). 2340 * The first determines the moment _when_ we should reduce CWND and, 2341 * hence, slow down forward transmission. In fact, it determines the moment 2342 * when we decide that hole is caused by loss, rather than by a reorder. 2343 * 2344 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill 2345 * holes, caused by lost packets. 2346 * 2347 * And the most logically complicated part of algorithm is undo 2348 * heuristics. We detect false retransmits due to both too early 2349 * fast retransmit (reordering) and underestimated RTO, analyzing 2350 * timestamps and D-SACKs. When we detect that some segments were 2351 * retransmitted by mistake and CWND reduction was wrong, we undo 2352 * window reduction and abort recovery phase. This logic is hidden 2353 * inside several functions named tcp_try_undo_<something>. 2354 */ 2355 2356 /* This function decides, when we should leave Disordered state 2357 * and enter Recovery phase, reducing congestion window. 2358 * 2359 * Main question: may we further continue forward transmission 2360 * with the same cwnd? 2361 */ 2362 static bool tcp_time_to_recover(struct sock *sk, int flag) 2363 { 2364 struct tcp_sock *tp = tcp_sk(sk); 2365 2366 /* Trick#1: The loss is proven. */ 2367 if (tp->lost_out) 2368 return true; 2369 2370 /* Not-A-Trick#2 : Classic rule... */ 2371 if (!tcp_is_rack(sk) && tcp_dupack_heuristics(tp) > tp->reordering) 2372 return true; 2373 2374 return false; 2375 } 2376 2377 /* Detect loss in event "A" above by marking head of queue up as lost. 2378 * For RFC3517 SACK, a segment is considered lost if it 2379 * has at least tp->reordering SACKed seqments above it; "packets" refers to 2380 * the maximum SACKed segments to pass before reaching this limit. 2381 */ 2382 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head) 2383 { 2384 struct tcp_sock *tp = tcp_sk(sk); 2385 struct sk_buff *skb; 2386 int cnt; 2387 /* Use SACK to deduce losses of new sequences sent during recovery */ 2388 const u32 loss_high = tp->snd_nxt; 2389 2390 WARN_ON(packets > tp->packets_out); 2391 skb = tp->lost_skb_hint; 2392 if (skb) { 2393 /* Head already handled? */ 2394 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una)) 2395 return; 2396 cnt = tp->lost_cnt_hint; 2397 } else { 2398 skb = tcp_rtx_queue_head(sk); 2399 cnt = 0; 2400 } 2401 2402 skb_rbtree_walk_from(skb) { 2403 /* TODO: do this better */ 2404 /* this is not the most efficient way to do this... */ 2405 tp->lost_skb_hint = skb; 2406 tp->lost_cnt_hint = cnt; 2407 2408 if (after(TCP_SKB_CB(skb)->end_seq, loss_high)) 2409 break; 2410 2411 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2412 cnt += tcp_skb_pcount(skb); 2413 2414 if (cnt > packets) 2415 break; 2416 2417 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_LOST)) 2418 tcp_mark_skb_lost(sk, skb); 2419 2420 if (mark_head) 2421 break; 2422 } 2423 tcp_verify_left_out(tp); 2424 } 2425 2426 /* Account newly detected lost packet(s) */ 2427 2428 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit) 2429 { 2430 struct tcp_sock *tp = tcp_sk(sk); 2431 2432 if (tcp_is_sack(tp)) { 2433 int sacked_upto = tp->sacked_out - tp->reordering; 2434 if (sacked_upto >= 0) 2435 tcp_mark_head_lost(sk, sacked_upto, 0); 2436 else if (fast_rexmit) 2437 tcp_mark_head_lost(sk, 1, 1); 2438 } 2439 } 2440 2441 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when) 2442 { 2443 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 2444 before(tp->rx_opt.rcv_tsecr, when); 2445 } 2446 2447 /* skb is spurious retransmitted if the returned timestamp echo 2448 * reply is prior to the skb transmission time 2449 */ 2450 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp, 2451 const struct sk_buff *skb) 2452 { 2453 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) && 2454 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb)); 2455 } 2456 2457 /* Nothing was retransmitted or returned timestamp is less 2458 * than timestamp of the first retransmission. 2459 */ 2460 static inline bool tcp_packet_delayed(const struct tcp_sock *tp) 2461 { 2462 return tp->retrans_stamp && 2463 tcp_tsopt_ecr_before(tp, tp->retrans_stamp); 2464 } 2465 2466 /* Undo procedures. */ 2467 2468 /* We can clear retrans_stamp when there are no retransmissions in the 2469 * window. It would seem that it is trivially available for us in 2470 * tp->retrans_out, however, that kind of assumptions doesn't consider 2471 * what will happen if errors occur when sending retransmission for the 2472 * second time. ...It could the that such segment has only 2473 * TCPCB_EVER_RETRANS set at the present time. It seems that checking 2474 * the head skb is enough except for some reneging corner cases that 2475 * are not worth the effort. 2476 * 2477 * Main reason for all this complexity is the fact that connection dying 2478 * time now depends on the validity of the retrans_stamp, in particular, 2479 * that successive retransmissions of a segment must not advance 2480 * retrans_stamp under any conditions. 2481 */ 2482 static bool tcp_any_retrans_done(const struct sock *sk) 2483 { 2484 const struct tcp_sock *tp = tcp_sk(sk); 2485 struct sk_buff *skb; 2486 2487 if (tp->retrans_out) 2488 return true; 2489 2490 skb = tcp_rtx_queue_head(sk); 2491 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS)) 2492 return true; 2493 2494 return false; 2495 } 2496 2497 static void DBGUNDO(struct sock *sk, const char *msg) 2498 { 2499 #if FASTRETRANS_DEBUG > 1 2500 struct tcp_sock *tp = tcp_sk(sk); 2501 struct inet_sock *inet = inet_sk(sk); 2502 2503 if (sk->sk_family == AF_INET) { 2504 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n", 2505 msg, 2506 &inet->inet_daddr, ntohs(inet->inet_dport), 2507 tcp_snd_cwnd(tp), tcp_left_out(tp), 2508 tp->snd_ssthresh, tp->prior_ssthresh, 2509 tp->packets_out); 2510 } 2511 #if IS_ENABLED(CONFIG_IPV6) 2512 else if (sk->sk_family == AF_INET6) { 2513 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n", 2514 msg, 2515 &sk->sk_v6_daddr, ntohs(inet->inet_dport), 2516 tcp_snd_cwnd(tp), tcp_left_out(tp), 2517 tp->snd_ssthresh, tp->prior_ssthresh, 2518 tp->packets_out); 2519 } 2520 #endif 2521 #endif 2522 } 2523 2524 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss) 2525 { 2526 struct tcp_sock *tp = tcp_sk(sk); 2527 2528 if (unmark_loss) { 2529 struct sk_buff *skb; 2530 2531 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) { 2532 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 2533 } 2534 tp->lost_out = 0; 2535 tcp_clear_all_retrans_hints(tp); 2536 } 2537 2538 if (tp->prior_ssthresh) { 2539 const struct inet_connection_sock *icsk = inet_csk(sk); 2540 2541 tcp_snd_cwnd_set(tp, icsk->icsk_ca_ops->undo_cwnd(sk)); 2542 2543 if (tp->prior_ssthresh > tp->snd_ssthresh) { 2544 tp->snd_ssthresh = tp->prior_ssthresh; 2545 tcp_ecn_withdraw_cwr(tp); 2546 } 2547 } 2548 tp->snd_cwnd_stamp = tcp_jiffies32; 2549 tp->undo_marker = 0; 2550 tp->rack.advanced = 1; /* Force RACK to re-exam losses */ 2551 } 2552 2553 static inline bool tcp_may_undo(const struct tcp_sock *tp) 2554 { 2555 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp)); 2556 } 2557 2558 static bool tcp_is_non_sack_preventing_reopen(struct sock *sk) 2559 { 2560 struct tcp_sock *tp = tcp_sk(sk); 2561 2562 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) { 2563 /* Hold old state until something *above* high_seq 2564 * is ACKed. For Reno it is MUST to prevent false 2565 * fast retransmits (RFC2582). SACK TCP is safe. */ 2566 if (!tcp_any_retrans_done(sk)) 2567 tp->retrans_stamp = 0; 2568 return true; 2569 } 2570 return false; 2571 } 2572 2573 /* People celebrate: "We love our President!" */ 2574 static bool tcp_try_undo_recovery(struct sock *sk) 2575 { 2576 struct tcp_sock *tp = tcp_sk(sk); 2577 2578 if (tcp_may_undo(tp)) { 2579 int mib_idx; 2580 2581 /* Happy end! We did not retransmit anything 2582 * or our original transmission succeeded. 2583 */ 2584 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans"); 2585 tcp_undo_cwnd_reduction(sk, false); 2586 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss) 2587 mib_idx = LINUX_MIB_TCPLOSSUNDO; 2588 else 2589 mib_idx = LINUX_MIB_TCPFULLUNDO; 2590 2591 NET_INC_STATS(sock_net(sk), mib_idx); 2592 } else if (tp->rack.reo_wnd_persist) { 2593 tp->rack.reo_wnd_persist--; 2594 } 2595 if (tcp_is_non_sack_preventing_reopen(sk)) 2596 return true; 2597 tcp_set_ca_state(sk, TCP_CA_Open); 2598 tp->is_sack_reneg = 0; 2599 return false; 2600 } 2601 2602 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */ 2603 static bool tcp_try_undo_dsack(struct sock *sk) 2604 { 2605 struct tcp_sock *tp = tcp_sk(sk); 2606 2607 if (tp->undo_marker && !tp->undo_retrans) { 2608 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH, 2609 tp->rack.reo_wnd_persist + 1); 2610 DBGUNDO(sk, "D-SACK"); 2611 tcp_undo_cwnd_reduction(sk, false); 2612 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO); 2613 return true; 2614 } 2615 return false; 2616 } 2617 2618 /* Undo during loss recovery after partial ACK or using F-RTO. */ 2619 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo) 2620 { 2621 struct tcp_sock *tp = tcp_sk(sk); 2622 2623 if (frto_undo || tcp_may_undo(tp)) { 2624 tcp_undo_cwnd_reduction(sk, true); 2625 2626 DBGUNDO(sk, "partial loss"); 2627 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO); 2628 if (frto_undo) 2629 NET_INC_STATS(sock_net(sk), 2630 LINUX_MIB_TCPSPURIOUSRTOS); 2631 inet_csk(sk)->icsk_retransmits = 0; 2632 if (tcp_is_non_sack_preventing_reopen(sk)) 2633 return true; 2634 if (frto_undo || tcp_is_sack(tp)) { 2635 tcp_set_ca_state(sk, TCP_CA_Open); 2636 tp->is_sack_reneg = 0; 2637 } 2638 return true; 2639 } 2640 return false; 2641 } 2642 2643 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937. 2644 * It computes the number of packets to send (sndcnt) based on packets newly 2645 * delivered: 2646 * 1) If the packets in flight is larger than ssthresh, PRR spreads the 2647 * cwnd reductions across a full RTT. 2648 * 2) Otherwise PRR uses packet conservation to send as much as delivered. 2649 * But when SND_UNA is acked without further losses, 2650 * slow starts cwnd up to ssthresh to speed up the recovery. 2651 */ 2652 static void tcp_init_cwnd_reduction(struct sock *sk) 2653 { 2654 struct tcp_sock *tp = tcp_sk(sk); 2655 2656 tp->high_seq = tp->snd_nxt; 2657 tp->tlp_high_seq = 0; 2658 tp->snd_cwnd_cnt = 0; 2659 tp->prior_cwnd = tcp_snd_cwnd(tp); 2660 tp->prr_delivered = 0; 2661 tp->prr_out = 0; 2662 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk); 2663 tcp_ecn_queue_cwr(tp); 2664 } 2665 2666 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int newly_lost, int flag) 2667 { 2668 struct tcp_sock *tp = tcp_sk(sk); 2669 int sndcnt = 0; 2670 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp); 2671 2672 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd)) 2673 return; 2674 2675 tp->prr_delivered += newly_acked_sacked; 2676 if (delta < 0) { 2677 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered + 2678 tp->prior_cwnd - 1; 2679 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out; 2680 } else { 2681 sndcnt = max_t(int, tp->prr_delivered - tp->prr_out, 2682 newly_acked_sacked); 2683 if (flag & FLAG_SND_UNA_ADVANCED && !newly_lost) 2684 sndcnt++; 2685 sndcnt = min(delta, sndcnt); 2686 } 2687 /* Force a fast retransmit upon entering fast recovery */ 2688 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1)); 2689 tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + sndcnt); 2690 } 2691 2692 static inline void tcp_end_cwnd_reduction(struct sock *sk) 2693 { 2694 struct tcp_sock *tp = tcp_sk(sk); 2695 2696 if (inet_csk(sk)->icsk_ca_ops->cong_control) 2697 return; 2698 2699 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */ 2700 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH && 2701 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) { 2702 tcp_snd_cwnd_set(tp, tp->snd_ssthresh); 2703 tp->snd_cwnd_stamp = tcp_jiffies32; 2704 } 2705 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR); 2706 } 2707 2708 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */ 2709 void tcp_enter_cwr(struct sock *sk) 2710 { 2711 struct tcp_sock *tp = tcp_sk(sk); 2712 2713 tp->prior_ssthresh = 0; 2714 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) { 2715 tp->undo_marker = 0; 2716 tcp_init_cwnd_reduction(sk); 2717 tcp_set_ca_state(sk, TCP_CA_CWR); 2718 } 2719 } 2720 EXPORT_SYMBOL(tcp_enter_cwr); 2721 2722 static void tcp_try_keep_open(struct sock *sk) 2723 { 2724 struct tcp_sock *tp = tcp_sk(sk); 2725 int state = TCP_CA_Open; 2726 2727 if (tcp_left_out(tp) || tcp_any_retrans_done(sk)) 2728 state = TCP_CA_Disorder; 2729 2730 if (inet_csk(sk)->icsk_ca_state != state) { 2731 tcp_set_ca_state(sk, state); 2732 tp->high_seq = tp->snd_nxt; 2733 } 2734 } 2735 2736 static void tcp_try_to_open(struct sock *sk, int flag) 2737 { 2738 struct tcp_sock *tp = tcp_sk(sk); 2739 2740 tcp_verify_left_out(tp); 2741 2742 if (!tcp_any_retrans_done(sk)) 2743 tp->retrans_stamp = 0; 2744 2745 if (flag & FLAG_ECE) 2746 tcp_enter_cwr(sk); 2747 2748 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) { 2749 tcp_try_keep_open(sk); 2750 } 2751 } 2752 2753 static void tcp_mtup_probe_failed(struct sock *sk) 2754 { 2755 struct inet_connection_sock *icsk = inet_csk(sk); 2756 2757 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1; 2758 icsk->icsk_mtup.probe_size = 0; 2759 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL); 2760 } 2761 2762 static void tcp_mtup_probe_success(struct sock *sk) 2763 { 2764 struct tcp_sock *tp = tcp_sk(sk); 2765 struct inet_connection_sock *icsk = inet_csk(sk); 2766 u64 val; 2767 2768 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2769 2770 val = (u64)tcp_snd_cwnd(tp) * tcp_mss_to_mtu(sk, tp->mss_cache); 2771 do_div(val, icsk->icsk_mtup.probe_size); 2772 DEBUG_NET_WARN_ON_ONCE((u32)val != val); 2773 tcp_snd_cwnd_set(tp, max_t(u32, 1U, val)); 2774 2775 tp->snd_cwnd_cnt = 0; 2776 tp->snd_cwnd_stamp = tcp_jiffies32; 2777 tp->snd_ssthresh = tcp_current_ssthresh(sk); 2778 2779 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size; 2780 icsk->icsk_mtup.probe_size = 0; 2781 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 2782 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS); 2783 } 2784 2785 /* Sometimes we deduce that packets have been dropped due to reasons other than 2786 * congestion, like path MTU reductions or failed client TFO attempts. In these 2787 * cases we call this function to retransmit as many packets as cwnd allows, 2788 * without reducing cwnd. Given that retransmits will set retrans_stamp to a 2789 * non-zero value (and may do so in a later calling context due to TSQ), we 2790 * also enter CA_Loss so that we track when all retransmitted packets are ACKed 2791 * and clear retrans_stamp when that happens (to ensure later recurring RTOs 2792 * are using the correct retrans_stamp and don't declare ETIMEDOUT 2793 * prematurely). 2794 */ 2795 static void tcp_non_congestion_loss_retransmit(struct sock *sk) 2796 { 2797 const struct inet_connection_sock *icsk = inet_csk(sk); 2798 struct tcp_sock *tp = tcp_sk(sk); 2799 2800 if (icsk->icsk_ca_state != TCP_CA_Loss) { 2801 tp->high_seq = tp->snd_nxt; 2802 tp->snd_ssthresh = tcp_current_ssthresh(sk); 2803 tp->prior_ssthresh = 0; 2804 tp->undo_marker = 0; 2805 tcp_set_ca_state(sk, TCP_CA_Loss); 2806 } 2807 tcp_xmit_retransmit_queue(sk); 2808 } 2809 2810 /* Do a simple retransmit without using the backoff mechanisms in 2811 * tcp_timer. This is used for path mtu discovery. 2812 * The socket is already locked here. 2813 */ 2814 void tcp_simple_retransmit(struct sock *sk) 2815 { 2816 struct tcp_sock *tp = tcp_sk(sk); 2817 struct sk_buff *skb; 2818 int mss; 2819 2820 /* A fastopen SYN request is stored as two separate packets within 2821 * the retransmit queue, this is done by tcp_send_syn_data(). 2822 * As a result simply checking the MSS of the frames in the queue 2823 * will not work for the SYN packet. 2824 * 2825 * Us being here is an indication of a path MTU issue so we can 2826 * assume that the fastopen SYN was lost and just mark all the 2827 * frames in the retransmit queue as lost. We will use an MSS of 2828 * -1 to mark all frames as lost, otherwise compute the current MSS. 2829 */ 2830 if (tp->syn_data && sk->sk_state == TCP_SYN_SENT) 2831 mss = -1; 2832 else 2833 mss = tcp_current_mss(sk); 2834 2835 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) { 2836 if (tcp_skb_seglen(skb) > mss) 2837 tcp_mark_skb_lost(sk, skb); 2838 } 2839 2840 tcp_clear_retrans_hints_partial(tp); 2841 2842 if (!tp->lost_out) 2843 return; 2844 2845 if (tcp_is_reno(tp)) 2846 tcp_limit_reno_sacked(tp); 2847 2848 tcp_verify_left_out(tp); 2849 2850 /* Don't muck with the congestion window here. 2851 * Reason is that we do not increase amount of _data_ 2852 * in network, but units changed and effective 2853 * cwnd/ssthresh really reduced now. 2854 */ 2855 tcp_non_congestion_loss_retransmit(sk); 2856 } 2857 EXPORT_SYMBOL(tcp_simple_retransmit); 2858 2859 void tcp_enter_recovery(struct sock *sk, bool ece_ack) 2860 { 2861 struct tcp_sock *tp = tcp_sk(sk); 2862 int mib_idx; 2863 2864 if (tcp_is_reno(tp)) 2865 mib_idx = LINUX_MIB_TCPRENORECOVERY; 2866 else 2867 mib_idx = LINUX_MIB_TCPSACKRECOVERY; 2868 2869 NET_INC_STATS(sock_net(sk), mib_idx); 2870 2871 tp->prior_ssthresh = 0; 2872 tcp_init_undo(tp); 2873 2874 if (!tcp_in_cwnd_reduction(sk)) { 2875 if (!ece_ack) 2876 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2877 tcp_init_cwnd_reduction(sk); 2878 } 2879 tcp_set_ca_state(sk, TCP_CA_Recovery); 2880 } 2881 2882 static void tcp_update_rto_time(struct tcp_sock *tp) 2883 { 2884 if (tp->rto_stamp) { 2885 tp->total_rto_time += tcp_time_stamp_ms(tp) - tp->rto_stamp; 2886 tp->rto_stamp = 0; 2887 } 2888 } 2889 2890 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are 2891 * recovered or spurious. Otherwise retransmits more on partial ACKs. 2892 */ 2893 static void tcp_process_loss(struct sock *sk, int flag, int num_dupack, 2894 int *rexmit) 2895 { 2896 struct tcp_sock *tp = tcp_sk(sk); 2897 bool recovered = !before(tp->snd_una, tp->high_seq); 2898 2899 if ((flag & FLAG_SND_UNA_ADVANCED || rcu_access_pointer(tp->fastopen_rsk)) && 2900 tcp_try_undo_loss(sk, false)) 2901 return; 2902 2903 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */ 2904 /* Step 3.b. A timeout is spurious if not all data are 2905 * lost, i.e., never-retransmitted data are (s)acked. 2906 */ 2907 if ((flag & FLAG_ORIG_SACK_ACKED) && 2908 tcp_try_undo_loss(sk, true)) 2909 return; 2910 2911 if (after(tp->snd_nxt, tp->high_seq)) { 2912 if (flag & FLAG_DATA_SACKED || num_dupack) 2913 tp->frto = 0; /* Step 3.a. loss was real */ 2914 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) { 2915 tp->high_seq = tp->snd_nxt; 2916 /* Step 2.b. Try send new data (but deferred until cwnd 2917 * is updated in tcp_ack()). Otherwise fall back to 2918 * the conventional recovery. 2919 */ 2920 if (!tcp_write_queue_empty(sk) && 2921 after(tcp_wnd_end(tp), tp->snd_nxt)) { 2922 *rexmit = REXMIT_NEW; 2923 return; 2924 } 2925 tp->frto = 0; 2926 } 2927 } 2928 2929 if (recovered) { 2930 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */ 2931 tcp_try_undo_recovery(sk); 2932 return; 2933 } 2934 if (tcp_is_reno(tp)) { 2935 /* A Reno DUPACK means new data in F-RTO step 2.b above are 2936 * delivered. Lower inflight to clock out (re)transmissions. 2937 */ 2938 if (after(tp->snd_nxt, tp->high_seq) && num_dupack) 2939 tcp_add_reno_sack(sk, num_dupack, flag & FLAG_ECE); 2940 else if (flag & FLAG_SND_UNA_ADVANCED) 2941 tcp_reset_reno_sack(tp); 2942 } 2943 *rexmit = REXMIT_LOST; 2944 } 2945 2946 static bool tcp_force_fast_retransmit(struct sock *sk) 2947 { 2948 struct tcp_sock *tp = tcp_sk(sk); 2949 2950 return after(tcp_highest_sack_seq(tp), 2951 tp->snd_una + tp->reordering * tp->mss_cache); 2952 } 2953 2954 /* Undo during fast recovery after partial ACK. */ 2955 static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una, 2956 bool *do_lost) 2957 { 2958 struct tcp_sock *tp = tcp_sk(sk); 2959 2960 if (tp->undo_marker && tcp_packet_delayed(tp)) { 2961 /* Plain luck! Hole if filled with delayed 2962 * packet, rather than with a retransmit. Check reordering. 2963 */ 2964 tcp_check_sack_reordering(sk, prior_snd_una, 1); 2965 2966 /* We are getting evidence that the reordering degree is higher 2967 * than we realized. If there are no retransmits out then we 2968 * can undo. Otherwise we clock out new packets but do not 2969 * mark more packets lost or retransmit more. 2970 */ 2971 if (tp->retrans_out) 2972 return true; 2973 2974 if (!tcp_any_retrans_done(sk)) 2975 tp->retrans_stamp = 0; 2976 2977 DBGUNDO(sk, "partial recovery"); 2978 tcp_undo_cwnd_reduction(sk, true); 2979 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO); 2980 tcp_try_keep_open(sk); 2981 } else { 2982 /* Partial ACK arrived. Force fast retransmit. */ 2983 *do_lost = tcp_force_fast_retransmit(sk); 2984 } 2985 return false; 2986 } 2987 2988 static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag) 2989 { 2990 struct tcp_sock *tp = tcp_sk(sk); 2991 2992 if (tcp_rtx_queue_empty(sk)) 2993 return; 2994 2995 if (unlikely(tcp_is_reno(tp))) { 2996 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED); 2997 } else if (tcp_is_rack(sk)) { 2998 u32 prior_retrans = tp->retrans_out; 2999 3000 if (tcp_rack_mark_lost(sk)) 3001 *ack_flag &= ~FLAG_SET_XMIT_TIMER; 3002 if (prior_retrans > tp->retrans_out) 3003 *ack_flag |= FLAG_LOST_RETRANS; 3004 } 3005 } 3006 3007 /* Process an event, which can update packets-in-flight not trivially. 3008 * Main goal of this function is to calculate new estimate for left_out, 3009 * taking into account both packets sitting in receiver's buffer and 3010 * packets lost by network. 3011 * 3012 * Besides that it updates the congestion state when packet loss or ECN 3013 * is detected. But it does not reduce the cwnd, it is done by the 3014 * congestion control later. 3015 * 3016 * It does _not_ decide what to send, it is made in function 3017 * tcp_xmit_retransmit_queue(). 3018 */ 3019 static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una, 3020 int num_dupack, int *ack_flag, int *rexmit) 3021 { 3022 struct inet_connection_sock *icsk = inet_csk(sk); 3023 struct tcp_sock *tp = tcp_sk(sk); 3024 int fast_rexmit = 0, flag = *ack_flag; 3025 bool ece_ack = flag & FLAG_ECE; 3026 bool do_lost = num_dupack || ((flag & FLAG_DATA_SACKED) && 3027 tcp_force_fast_retransmit(sk)); 3028 3029 if (!tp->packets_out && tp->sacked_out) 3030 tp->sacked_out = 0; 3031 3032 /* Now state machine starts. 3033 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */ 3034 if (ece_ack) 3035 tp->prior_ssthresh = 0; 3036 3037 /* B. In all the states check for reneging SACKs. */ 3038 if (tcp_check_sack_reneging(sk, ack_flag)) 3039 return; 3040 3041 /* C. Check consistency of the current state. */ 3042 tcp_verify_left_out(tp); 3043 3044 /* D. Check state exit conditions. State can be terminated 3045 * when high_seq is ACKed. */ 3046 if (icsk->icsk_ca_state == TCP_CA_Open) { 3047 WARN_ON(tp->retrans_out != 0 && !tp->syn_data); 3048 tp->retrans_stamp = 0; 3049 } else if (!before(tp->snd_una, tp->high_seq)) { 3050 switch (icsk->icsk_ca_state) { 3051 case TCP_CA_CWR: 3052 /* CWR is to be held something *above* high_seq 3053 * is ACKed for CWR bit to reach receiver. */ 3054 if (tp->snd_una != tp->high_seq) { 3055 tcp_end_cwnd_reduction(sk); 3056 tcp_set_ca_state(sk, TCP_CA_Open); 3057 } 3058 break; 3059 3060 case TCP_CA_Recovery: 3061 if (tcp_is_reno(tp)) 3062 tcp_reset_reno_sack(tp); 3063 if (tcp_try_undo_recovery(sk)) 3064 return; 3065 tcp_end_cwnd_reduction(sk); 3066 break; 3067 } 3068 } 3069 3070 /* E. Process state. */ 3071 switch (icsk->icsk_ca_state) { 3072 case TCP_CA_Recovery: 3073 if (!(flag & FLAG_SND_UNA_ADVANCED)) { 3074 if (tcp_is_reno(tp)) 3075 tcp_add_reno_sack(sk, num_dupack, ece_ack); 3076 } else if (tcp_try_undo_partial(sk, prior_snd_una, &do_lost)) 3077 return; 3078 3079 if (tcp_try_undo_dsack(sk)) 3080 tcp_try_keep_open(sk); 3081 3082 tcp_identify_packet_loss(sk, ack_flag); 3083 if (icsk->icsk_ca_state != TCP_CA_Recovery) { 3084 if (!tcp_time_to_recover(sk, flag)) 3085 return; 3086 /* Undo reverts the recovery state. If loss is evident, 3087 * starts a new recovery (e.g. reordering then loss); 3088 */ 3089 tcp_enter_recovery(sk, ece_ack); 3090 } 3091 break; 3092 case TCP_CA_Loss: 3093 tcp_process_loss(sk, flag, num_dupack, rexmit); 3094 if (icsk->icsk_ca_state != TCP_CA_Loss) 3095 tcp_update_rto_time(tp); 3096 tcp_identify_packet_loss(sk, ack_flag); 3097 if (!(icsk->icsk_ca_state == TCP_CA_Open || 3098 (*ack_flag & FLAG_LOST_RETRANS))) 3099 return; 3100 /* Change state if cwnd is undone or retransmits are lost */ 3101 fallthrough; 3102 default: 3103 if (tcp_is_reno(tp)) { 3104 if (flag & FLAG_SND_UNA_ADVANCED) 3105 tcp_reset_reno_sack(tp); 3106 tcp_add_reno_sack(sk, num_dupack, ece_ack); 3107 } 3108 3109 if (icsk->icsk_ca_state <= TCP_CA_Disorder) 3110 tcp_try_undo_dsack(sk); 3111 3112 tcp_identify_packet_loss(sk, ack_flag); 3113 if (!tcp_time_to_recover(sk, flag)) { 3114 tcp_try_to_open(sk, flag); 3115 return; 3116 } 3117 3118 /* MTU probe failure: don't reduce cwnd */ 3119 if (icsk->icsk_ca_state < TCP_CA_CWR && 3120 icsk->icsk_mtup.probe_size && 3121 tp->snd_una == tp->mtu_probe.probe_seq_start) { 3122 tcp_mtup_probe_failed(sk); 3123 /* Restores the reduction we did in tcp_mtup_probe() */ 3124 tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) + 1); 3125 tcp_simple_retransmit(sk); 3126 return; 3127 } 3128 3129 /* Otherwise enter Recovery state */ 3130 tcp_enter_recovery(sk, ece_ack); 3131 fast_rexmit = 1; 3132 } 3133 3134 if (!tcp_is_rack(sk) && do_lost) 3135 tcp_update_scoreboard(sk, fast_rexmit); 3136 *rexmit = REXMIT_LOST; 3137 } 3138 3139 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag) 3140 { 3141 u32 wlen = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen) * HZ; 3142 struct tcp_sock *tp = tcp_sk(sk); 3143 3144 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) { 3145 /* If the remote keeps returning delayed ACKs, eventually 3146 * the min filter would pick it up and overestimate the 3147 * prop. delay when it expires. Skip suspected delayed ACKs. 3148 */ 3149 return; 3150 } 3151 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32, 3152 rtt_us ? : jiffies_to_usecs(1)); 3153 } 3154 3155 static bool tcp_ack_update_rtt(struct sock *sk, const int flag, 3156 long seq_rtt_us, long sack_rtt_us, 3157 long ca_rtt_us, struct rate_sample *rs) 3158 { 3159 const struct tcp_sock *tp = tcp_sk(sk); 3160 3161 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because 3162 * broken middle-boxes or peers may corrupt TS-ECR fields. But 3163 * Karn's algorithm forbids taking RTT if some retransmitted data 3164 * is acked (RFC6298). 3165 */ 3166 if (seq_rtt_us < 0) 3167 seq_rtt_us = sack_rtt_us; 3168 3169 /* RTTM Rule: A TSecr value received in a segment is used to 3170 * update the averaged RTT measurement only if the segment 3171 * acknowledges some new data, i.e., only if it advances the 3172 * left edge of the send window. 3173 * See draft-ietf-tcplw-high-performance-00, section 3.3. 3174 */ 3175 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && 3176 tp->rx_opt.rcv_tsecr && flag & FLAG_ACKED) 3177 seq_rtt_us = ca_rtt_us = tcp_rtt_tsopt_us(tp); 3178 3179 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */ 3180 if (seq_rtt_us < 0) 3181 return false; 3182 3183 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is 3184 * always taken together with ACK, SACK, or TS-opts. Any negative 3185 * values will be skipped with the seq_rtt_us < 0 check above. 3186 */ 3187 tcp_update_rtt_min(sk, ca_rtt_us, flag); 3188 tcp_rtt_estimator(sk, seq_rtt_us); 3189 tcp_set_rto(sk); 3190 3191 /* RFC6298: only reset backoff on valid RTT measurement. */ 3192 inet_csk(sk)->icsk_backoff = 0; 3193 return true; 3194 } 3195 3196 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */ 3197 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req) 3198 { 3199 struct rate_sample rs; 3200 long rtt_us = -1L; 3201 3202 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack) 3203 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack); 3204 3205 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs); 3206 } 3207 3208 3209 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked) 3210 { 3211 const struct inet_connection_sock *icsk = inet_csk(sk); 3212 3213 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked); 3214 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32; 3215 } 3216 3217 /* Restart timer after forward progress on connection. 3218 * RFC2988 recommends to restart timer to now+rto. 3219 */ 3220 void tcp_rearm_rto(struct sock *sk) 3221 { 3222 const struct inet_connection_sock *icsk = inet_csk(sk); 3223 struct tcp_sock *tp = tcp_sk(sk); 3224 3225 /* If the retrans timer is currently being used by Fast Open 3226 * for SYN-ACK retrans purpose, stay put. 3227 */ 3228 if (rcu_access_pointer(tp->fastopen_rsk)) 3229 return; 3230 3231 if (!tp->packets_out) { 3232 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS); 3233 } else { 3234 u32 rto = inet_csk(sk)->icsk_rto; 3235 /* Offset the time elapsed after installing regular RTO */ 3236 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT || 3237 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) { 3238 s64 delta_us = tcp_rto_delta_us(sk); 3239 /* delta_us may not be positive if the socket is locked 3240 * when the retrans timer fires and is rescheduled. 3241 */ 3242 rto = usecs_to_jiffies(max_t(int, delta_us, 1)); 3243 } 3244 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto, 3245 TCP_RTO_MAX); 3246 } 3247 } 3248 3249 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */ 3250 static void tcp_set_xmit_timer(struct sock *sk) 3251 { 3252 if (!tcp_schedule_loss_probe(sk, true)) 3253 tcp_rearm_rto(sk); 3254 } 3255 3256 /* If we get here, the whole TSO packet has not been acked. */ 3257 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb) 3258 { 3259 struct tcp_sock *tp = tcp_sk(sk); 3260 u32 packets_acked; 3261 3262 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)); 3263 3264 packets_acked = tcp_skb_pcount(skb); 3265 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 3266 return 0; 3267 packets_acked -= tcp_skb_pcount(skb); 3268 3269 if (packets_acked) { 3270 BUG_ON(tcp_skb_pcount(skb) == 0); 3271 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)); 3272 } 3273 3274 return packets_acked; 3275 } 3276 3277 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb, 3278 const struct sk_buff *ack_skb, u32 prior_snd_una) 3279 { 3280 const struct skb_shared_info *shinfo; 3281 3282 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */ 3283 if (likely(!TCP_SKB_CB(skb)->txstamp_ack)) 3284 return; 3285 3286 shinfo = skb_shinfo(skb); 3287 if (!before(shinfo->tskey, prior_snd_una) && 3288 before(shinfo->tskey, tcp_sk(sk)->snd_una)) { 3289 tcp_skb_tsorted_save(skb) { 3290 __skb_tstamp_tx(skb, ack_skb, NULL, sk, SCM_TSTAMP_ACK); 3291 } tcp_skb_tsorted_restore(skb); 3292 } 3293 } 3294 3295 /* Remove acknowledged frames from the retransmission queue. If our packet 3296 * is before the ack sequence we can discard it as it's confirmed to have 3297 * arrived at the other end. 3298 */ 3299 static int tcp_clean_rtx_queue(struct sock *sk, const struct sk_buff *ack_skb, 3300 u32 prior_fack, u32 prior_snd_una, 3301 struct tcp_sacktag_state *sack, bool ece_ack) 3302 { 3303 const struct inet_connection_sock *icsk = inet_csk(sk); 3304 u64 first_ackt, last_ackt; 3305 struct tcp_sock *tp = tcp_sk(sk); 3306 u32 prior_sacked = tp->sacked_out; 3307 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */ 3308 struct sk_buff *skb, *next; 3309 bool fully_acked = true; 3310 long sack_rtt_us = -1L; 3311 long seq_rtt_us = -1L; 3312 long ca_rtt_us = -1L; 3313 u32 pkts_acked = 0; 3314 bool rtt_update; 3315 int flag = 0; 3316 3317 first_ackt = 0; 3318 3319 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) { 3320 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 3321 const u32 start_seq = scb->seq; 3322 u8 sacked = scb->sacked; 3323 u32 acked_pcount; 3324 3325 /* Determine how many packets and what bytes were acked, tso and else */ 3326 if (after(scb->end_seq, tp->snd_una)) { 3327 if (tcp_skb_pcount(skb) == 1 || 3328 !after(tp->snd_una, scb->seq)) 3329 break; 3330 3331 acked_pcount = tcp_tso_acked(sk, skb); 3332 if (!acked_pcount) 3333 break; 3334 fully_acked = false; 3335 } else { 3336 acked_pcount = tcp_skb_pcount(skb); 3337 } 3338 3339 if (unlikely(sacked & TCPCB_RETRANS)) { 3340 if (sacked & TCPCB_SACKED_RETRANS) 3341 tp->retrans_out -= acked_pcount; 3342 flag |= FLAG_RETRANS_DATA_ACKED; 3343 } else if (!(sacked & TCPCB_SACKED_ACKED)) { 3344 last_ackt = tcp_skb_timestamp_us(skb); 3345 WARN_ON_ONCE(last_ackt == 0); 3346 if (!first_ackt) 3347 first_ackt = last_ackt; 3348 3349 if (before(start_seq, reord)) 3350 reord = start_seq; 3351 if (!after(scb->end_seq, tp->high_seq)) 3352 flag |= FLAG_ORIG_SACK_ACKED; 3353 } 3354 3355 if (sacked & TCPCB_SACKED_ACKED) { 3356 tp->sacked_out -= acked_pcount; 3357 } else if (tcp_is_sack(tp)) { 3358 tcp_count_delivered(tp, acked_pcount, ece_ack); 3359 if (!tcp_skb_spurious_retrans(tp, skb)) 3360 tcp_rack_advance(tp, sacked, scb->end_seq, 3361 tcp_skb_timestamp_us(skb)); 3362 } 3363 if (sacked & TCPCB_LOST) 3364 tp->lost_out -= acked_pcount; 3365 3366 tp->packets_out -= acked_pcount; 3367 pkts_acked += acked_pcount; 3368 tcp_rate_skb_delivered(sk, skb, sack->rate); 3369 3370 /* Initial outgoing SYN's get put onto the write_queue 3371 * just like anything else we transmit. It is not 3372 * true data, and if we misinform our callers that 3373 * this ACK acks real data, we will erroneously exit 3374 * connection startup slow start one packet too 3375 * quickly. This is severely frowned upon behavior. 3376 */ 3377 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) { 3378 flag |= FLAG_DATA_ACKED; 3379 } else { 3380 flag |= FLAG_SYN_ACKED; 3381 tp->retrans_stamp = 0; 3382 } 3383 3384 if (!fully_acked) 3385 break; 3386 3387 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una); 3388 3389 next = skb_rb_next(skb); 3390 if (unlikely(skb == tp->retransmit_skb_hint)) 3391 tp->retransmit_skb_hint = NULL; 3392 if (unlikely(skb == tp->lost_skb_hint)) 3393 tp->lost_skb_hint = NULL; 3394 tcp_highest_sack_replace(sk, skb, next); 3395 tcp_rtx_queue_unlink_and_free(skb, sk); 3396 } 3397 3398 if (!skb) 3399 tcp_chrono_stop(sk, TCP_CHRONO_BUSY); 3400 3401 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una))) 3402 tp->snd_up = tp->snd_una; 3403 3404 if (skb) { 3405 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una); 3406 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 3407 flag |= FLAG_SACK_RENEGING; 3408 } 3409 3410 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) { 3411 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt); 3412 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt); 3413 3414 if (pkts_acked == 1 && fully_acked && !prior_sacked && 3415 (tp->snd_una - prior_snd_una) < tp->mss_cache && 3416 sack->rate->prior_delivered + 1 == tp->delivered && 3417 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) { 3418 /* Conservatively mark a delayed ACK. It's typically 3419 * from a lone runt packet over the round trip to 3420 * a receiver w/o out-of-order or CE events. 3421 */ 3422 flag |= FLAG_ACK_MAYBE_DELAYED; 3423 } 3424 } 3425 if (sack->first_sackt) { 3426 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt); 3427 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt); 3428 } 3429 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us, 3430 ca_rtt_us, sack->rate); 3431 3432 if (flag & FLAG_ACKED) { 3433 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */ 3434 if (unlikely(icsk->icsk_mtup.probe_size && 3435 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) { 3436 tcp_mtup_probe_success(sk); 3437 } 3438 3439 if (tcp_is_reno(tp)) { 3440 tcp_remove_reno_sacks(sk, pkts_acked, ece_ack); 3441 3442 /* If any of the cumulatively ACKed segments was 3443 * retransmitted, non-SACK case cannot confirm that 3444 * progress was due to original transmission due to 3445 * lack of TCPCB_SACKED_ACKED bits even if some of 3446 * the packets may have been never retransmitted. 3447 */ 3448 if (flag & FLAG_RETRANS_DATA_ACKED) 3449 flag &= ~FLAG_ORIG_SACK_ACKED; 3450 } else { 3451 int delta; 3452 3453 /* Non-retransmitted hole got filled? That's reordering */ 3454 if (before(reord, prior_fack)) 3455 tcp_check_sack_reordering(sk, reord, 0); 3456 3457 delta = prior_sacked - tp->sacked_out; 3458 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta); 3459 } 3460 } else if (skb && rtt_update && sack_rtt_us >= 0 && 3461 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp, 3462 tcp_skb_timestamp_us(skb))) { 3463 /* Do not re-arm RTO if the sack RTT is measured from data sent 3464 * after when the head was last (re)transmitted. Otherwise the 3465 * timeout may continue to extend in loss recovery. 3466 */ 3467 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */ 3468 } 3469 3470 if (icsk->icsk_ca_ops->pkts_acked) { 3471 struct ack_sample sample = { .pkts_acked = pkts_acked, 3472 .rtt_us = sack->rate->rtt_us }; 3473 3474 sample.in_flight = tp->mss_cache * 3475 (tp->delivered - sack->rate->prior_delivered); 3476 icsk->icsk_ca_ops->pkts_acked(sk, &sample); 3477 } 3478 3479 #if FASTRETRANS_DEBUG > 0 3480 WARN_ON((int)tp->sacked_out < 0); 3481 WARN_ON((int)tp->lost_out < 0); 3482 WARN_ON((int)tp->retrans_out < 0); 3483 if (!tp->packets_out && tcp_is_sack(tp)) { 3484 icsk = inet_csk(sk); 3485 if (tp->lost_out) { 3486 pr_debug("Leak l=%u %d\n", 3487 tp->lost_out, icsk->icsk_ca_state); 3488 tp->lost_out = 0; 3489 } 3490 if (tp->sacked_out) { 3491 pr_debug("Leak s=%u %d\n", 3492 tp->sacked_out, icsk->icsk_ca_state); 3493 tp->sacked_out = 0; 3494 } 3495 if (tp->retrans_out) { 3496 pr_debug("Leak r=%u %d\n", 3497 tp->retrans_out, icsk->icsk_ca_state); 3498 tp->retrans_out = 0; 3499 } 3500 } 3501 #endif 3502 return flag; 3503 } 3504 3505 static void tcp_ack_probe(struct sock *sk) 3506 { 3507 struct inet_connection_sock *icsk = inet_csk(sk); 3508 struct sk_buff *head = tcp_send_head(sk); 3509 const struct tcp_sock *tp = tcp_sk(sk); 3510 3511 /* Was it a usable window open? */ 3512 if (!head) 3513 return; 3514 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) { 3515 icsk->icsk_backoff = 0; 3516 icsk->icsk_probes_tstamp = 0; 3517 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0); 3518 /* Socket must be waked up by subsequent tcp_data_snd_check(). 3519 * This function is not for random using! 3520 */ 3521 } else { 3522 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX); 3523 3524 when = tcp_clamp_probe0_to_user_timeout(sk, when); 3525 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, when, TCP_RTO_MAX); 3526 } 3527 } 3528 3529 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag) 3530 { 3531 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) || 3532 inet_csk(sk)->icsk_ca_state != TCP_CA_Open; 3533 } 3534 3535 /* Decide wheather to run the increase function of congestion control. */ 3536 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag) 3537 { 3538 /* If reordering is high then always grow cwnd whenever data is 3539 * delivered regardless of its ordering. Otherwise stay conservative 3540 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/ 3541 * new SACK or ECE mark may first advance cwnd here and later reduce 3542 * cwnd in tcp_fastretrans_alert() based on more states. 3543 */ 3544 if (tcp_sk(sk)->reordering > 3545 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reordering)) 3546 return flag & FLAG_FORWARD_PROGRESS; 3547 3548 return flag & FLAG_DATA_ACKED; 3549 } 3550 3551 /* The "ultimate" congestion control function that aims to replace the rigid 3552 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction). 3553 * It's called toward the end of processing an ACK with precise rate 3554 * information. All transmission or retransmission are delayed afterwards. 3555 */ 3556 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked, 3557 int flag, const struct rate_sample *rs) 3558 { 3559 const struct inet_connection_sock *icsk = inet_csk(sk); 3560 3561 if (icsk->icsk_ca_ops->cong_control) { 3562 icsk->icsk_ca_ops->cong_control(sk, ack, flag, rs); 3563 return; 3564 } 3565 3566 if (tcp_in_cwnd_reduction(sk)) { 3567 /* Reduce cwnd if state mandates */ 3568 tcp_cwnd_reduction(sk, acked_sacked, rs->losses, flag); 3569 } else if (tcp_may_raise_cwnd(sk, flag)) { 3570 /* Advance cwnd if state allows */ 3571 tcp_cong_avoid(sk, ack, acked_sacked); 3572 } 3573 tcp_update_pacing_rate(sk); 3574 } 3575 3576 /* Check that window update is acceptable. 3577 * The function assumes that snd_una<=ack<=snd_next. 3578 */ 3579 static inline bool tcp_may_update_window(const struct tcp_sock *tp, 3580 const u32 ack, const u32 ack_seq, 3581 const u32 nwin) 3582 { 3583 return after(ack, tp->snd_una) || 3584 after(ack_seq, tp->snd_wl1) || 3585 (ack_seq == tp->snd_wl1 && (nwin > tp->snd_wnd || !nwin)); 3586 } 3587 3588 static void tcp_snd_sne_update(struct tcp_sock *tp, u32 ack) 3589 { 3590 #ifdef CONFIG_TCP_AO 3591 struct tcp_ao_info *ao; 3592 3593 if (!static_branch_unlikely(&tcp_ao_needed.key)) 3594 return; 3595 3596 ao = rcu_dereference_protected(tp->ao_info, 3597 lockdep_sock_is_held((struct sock *)tp)); 3598 if (ao && ack < tp->snd_una) { 3599 ao->snd_sne++; 3600 trace_tcp_ao_snd_sne_update((struct sock *)tp, ao->snd_sne); 3601 } 3602 #endif 3603 } 3604 3605 /* If we update tp->snd_una, also update tp->bytes_acked */ 3606 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack) 3607 { 3608 u32 delta = ack - tp->snd_una; 3609 3610 sock_owned_by_me((struct sock *)tp); 3611 tp->bytes_acked += delta; 3612 tcp_snd_sne_update(tp, ack); 3613 tp->snd_una = ack; 3614 } 3615 3616 static void tcp_rcv_sne_update(struct tcp_sock *tp, u32 seq) 3617 { 3618 #ifdef CONFIG_TCP_AO 3619 struct tcp_ao_info *ao; 3620 3621 if (!static_branch_unlikely(&tcp_ao_needed.key)) 3622 return; 3623 3624 ao = rcu_dereference_protected(tp->ao_info, 3625 lockdep_sock_is_held((struct sock *)tp)); 3626 if (ao && seq < tp->rcv_nxt) { 3627 ao->rcv_sne++; 3628 trace_tcp_ao_rcv_sne_update((struct sock *)tp, ao->rcv_sne); 3629 } 3630 #endif 3631 } 3632 3633 /* If we update tp->rcv_nxt, also update tp->bytes_received */ 3634 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq) 3635 { 3636 u32 delta = seq - tp->rcv_nxt; 3637 3638 sock_owned_by_me((struct sock *)tp); 3639 tp->bytes_received += delta; 3640 tcp_rcv_sne_update(tp, seq); 3641 WRITE_ONCE(tp->rcv_nxt, seq); 3642 } 3643 3644 /* Update our send window. 3645 * 3646 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2 3647 * and in FreeBSD. NetBSD's one is even worse.) is wrong. 3648 */ 3649 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack, 3650 u32 ack_seq) 3651 { 3652 struct tcp_sock *tp = tcp_sk(sk); 3653 int flag = 0; 3654 u32 nwin = ntohs(tcp_hdr(skb)->window); 3655 3656 if (likely(!tcp_hdr(skb)->syn)) 3657 nwin <<= tp->rx_opt.snd_wscale; 3658 3659 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) { 3660 flag |= FLAG_WIN_UPDATE; 3661 tcp_update_wl(tp, ack_seq); 3662 3663 if (tp->snd_wnd != nwin) { 3664 tp->snd_wnd = nwin; 3665 3666 /* Note, it is the only place, where 3667 * fast path is recovered for sending TCP. 3668 */ 3669 tp->pred_flags = 0; 3670 tcp_fast_path_check(sk); 3671 3672 if (!tcp_write_queue_empty(sk)) 3673 tcp_slow_start_after_idle_check(sk); 3674 3675 if (nwin > tp->max_window) { 3676 tp->max_window = nwin; 3677 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie); 3678 } 3679 } 3680 } 3681 3682 tcp_snd_una_update(tp, ack); 3683 3684 return flag; 3685 } 3686 3687 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx, 3688 u32 *last_oow_ack_time) 3689 { 3690 /* Paired with the WRITE_ONCE() in this function. */ 3691 u32 val = READ_ONCE(*last_oow_ack_time); 3692 3693 if (val) { 3694 s32 elapsed = (s32)(tcp_jiffies32 - val); 3695 3696 if (0 <= elapsed && 3697 elapsed < READ_ONCE(net->ipv4.sysctl_tcp_invalid_ratelimit)) { 3698 NET_INC_STATS(net, mib_idx); 3699 return true; /* rate-limited: don't send yet! */ 3700 } 3701 } 3702 3703 /* Paired with the prior READ_ONCE() and with itself, 3704 * as we might be lockless. 3705 */ 3706 WRITE_ONCE(*last_oow_ack_time, tcp_jiffies32); 3707 3708 return false; /* not rate-limited: go ahead, send dupack now! */ 3709 } 3710 3711 /* Return true if we're currently rate-limiting out-of-window ACKs and 3712 * thus shouldn't send a dupack right now. We rate-limit dupacks in 3713 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS 3714 * attacks that send repeated SYNs or ACKs for the same connection. To 3715 * do this, we do not send a duplicate SYNACK or ACK if the remote 3716 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate. 3717 */ 3718 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb, 3719 int mib_idx, u32 *last_oow_ack_time) 3720 { 3721 /* Data packets without SYNs are not likely part of an ACK loop. */ 3722 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) && 3723 !tcp_hdr(skb)->syn) 3724 return false; 3725 3726 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time); 3727 } 3728 3729 /* RFC 5961 7 [ACK Throttling] */ 3730 static void tcp_send_challenge_ack(struct sock *sk) 3731 { 3732 struct tcp_sock *tp = tcp_sk(sk); 3733 struct net *net = sock_net(sk); 3734 u32 count, now, ack_limit; 3735 3736 /* First check our per-socket dupack rate limit. */ 3737 if (__tcp_oow_rate_limited(net, 3738 LINUX_MIB_TCPACKSKIPPEDCHALLENGE, 3739 &tp->last_oow_ack_time)) 3740 return; 3741 3742 ack_limit = READ_ONCE(net->ipv4.sysctl_tcp_challenge_ack_limit); 3743 if (ack_limit == INT_MAX) 3744 goto send_ack; 3745 3746 /* Then check host-wide RFC 5961 rate limit. */ 3747 now = jiffies / HZ; 3748 if (now != READ_ONCE(net->ipv4.tcp_challenge_timestamp)) { 3749 u32 half = (ack_limit + 1) >> 1; 3750 3751 WRITE_ONCE(net->ipv4.tcp_challenge_timestamp, now); 3752 WRITE_ONCE(net->ipv4.tcp_challenge_count, 3753 get_random_u32_inclusive(half, ack_limit + half - 1)); 3754 } 3755 count = READ_ONCE(net->ipv4.tcp_challenge_count); 3756 if (count > 0) { 3757 WRITE_ONCE(net->ipv4.tcp_challenge_count, count - 1); 3758 send_ack: 3759 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK); 3760 tcp_send_ack(sk); 3761 } 3762 } 3763 3764 static void tcp_store_ts_recent(struct tcp_sock *tp) 3765 { 3766 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval; 3767 tp->rx_opt.ts_recent_stamp = ktime_get_seconds(); 3768 } 3769 3770 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq) 3771 { 3772 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) { 3773 /* PAWS bug workaround wrt. ACK frames, the PAWS discard 3774 * extra check below makes sure this can only happen 3775 * for pure ACK frames. -DaveM 3776 * 3777 * Not only, also it occurs for expired timestamps. 3778 */ 3779 3780 if (tcp_paws_check(&tp->rx_opt, 0)) 3781 tcp_store_ts_recent(tp); 3782 } 3783 } 3784 3785 /* This routine deals with acks during a TLP episode and ends an episode by 3786 * resetting tlp_high_seq. Ref: TLP algorithm in draft-ietf-tcpm-rack 3787 */ 3788 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag) 3789 { 3790 struct tcp_sock *tp = tcp_sk(sk); 3791 3792 if (before(ack, tp->tlp_high_seq)) 3793 return; 3794 3795 if (!tp->tlp_retrans) { 3796 /* TLP of new data has been acknowledged */ 3797 tp->tlp_high_seq = 0; 3798 } else if (flag & FLAG_DSACK_TLP) { 3799 /* This DSACK means original and TLP probe arrived; no loss */ 3800 tp->tlp_high_seq = 0; 3801 } else if (after(ack, tp->tlp_high_seq)) { 3802 /* ACK advances: there was a loss, so reduce cwnd. Reset 3803 * tlp_high_seq in tcp_init_cwnd_reduction() 3804 */ 3805 tcp_init_cwnd_reduction(sk); 3806 tcp_set_ca_state(sk, TCP_CA_CWR); 3807 tcp_end_cwnd_reduction(sk); 3808 tcp_try_keep_open(sk); 3809 NET_INC_STATS(sock_net(sk), 3810 LINUX_MIB_TCPLOSSPROBERECOVERY); 3811 } else if (!(flag & (FLAG_SND_UNA_ADVANCED | 3812 FLAG_NOT_DUP | FLAG_DATA_SACKED))) { 3813 /* Pure dupack: original and TLP probe arrived; no loss */ 3814 tp->tlp_high_seq = 0; 3815 } 3816 } 3817 3818 static inline void tcp_in_ack_event(struct sock *sk, u32 flags) 3819 { 3820 const struct inet_connection_sock *icsk = inet_csk(sk); 3821 3822 if (icsk->icsk_ca_ops->in_ack_event) 3823 icsk->icsk_ca_ops->in_ack_event(sk, flags); 3824 } 3825 3826 /* Congestion control has updated the cwnd already. So if we're in 3827 * loss recovery then now we do any new sends (for FRTO) or 3828 * retransmits (for CA_Loss or CA_recovery) that make sense. 3829 */ 3830 static void tcp_xmit_recovery(struct sock *sk, int rexmit) 3831 { 3832 struct tcp_sock *tp = tcp_sk(sk); 3833 3834 if (rexmit == REXMIT_NONE || sk->sk_state == TCP_SYN_SENT) 3835 return; 3836 3837 if (unlikely(rexmit == REXMIT_NEW)) { 3838 __tcp_push_pending_frames(sk, tcp_current_mss(sk), 3839 TCP_NAGLE_OFF); 3840 if (after(tp->snd_nxt, tp->high_seq)) 3841 return; 3842 tp->frto = 0; 3843 } 3844 tcp_xmit_retransmit_queue(sk); 3845 } 3846 3847 /* Returns the number of packets newly acked or sacked by the current ACK */ 3848 static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, int flag) 3849 { 3850 const struct net *net = sock_net(sk); 3851 struct tcp_sock *tp = tcp_sk(sk); 3852 u32 delivered; 3853 3854 delivered = tp->delivered - prior_delivered; 3855 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered); 3856 if (flag & FLAG_ECE) 3857 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, delivered); 3858 3859 return delivered; 3860 } 3861 3862 /* This routine deals with incoming acks, but not outgoing ones. */ 3863 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag) 3864 { 3865 struct inet_connection_sock *icsk = inet_csk(sk); 3866 struct tcp_sock *tp = tcp_sk(sk); 3867 struct tcp_sacktag_state sack_state; 3868 struct rate_sample rs = { .prior_delivered = 0 }; 3869 u32 prior_snd_una = tp->snd_una; 3870 bool is_sack_reneg = tp->is_sack_reneg; 3871 u32 ack_seq = TCP_SKB_CB(skb)->seq; 3872 u32 ack = TCP_SKB_CB(skb)->ack_seq; 3873 int num_dupack = 0; 3874 int prior_packets = tp->packets_out; 3875 u32 delivered = tp->delivered; 3876 u32 lost = tp->lost; 3877 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */ 3878 u32 prior_fack; 3879 3880 sack_state.first_sackt = 0; 3881 sack_state.rate = &rs; 3882 sack_state.sack_delivered = 0; 3883 3884 /* We very likely will need to access rtx queue. */ 3885 prefetch(sk->tcp_rtx_queue.rb_node); 3886 3887 /* If the ack is older than previous acks 3888 * then we can probably ignore it. 3889 */ 3890 if (before(ack, prior_snd_una)) { 3891 u32 max_window; 3892 3893 /* do not accept ACK for bytes we never sent. */ 3894 max_window = min_t(u64, tp->max_window, tp->bytes_acked); 3895 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */ 3896 if (before(ack, prior_snd_una - max_window)) { 3897 if (!(flag & FLAG_NO_CHALLENGE_ACK)) 3898 tcp_send_challenge_ack(sk); 3899 return -SKB_DROP_REASON_TCP_TOO_OLD_ACK; 3900 } 3901 goto old_ack; 3902 } 3903 3904 /* If the ack includes data we haven't sent yet, discard 3905 * this segment (RFC793 Section 3.9). 3906 */ 3907 if (after(ack, tp->snd_nxt)) 3908 return -SKB_DROP_REASON_TCP_ACK_UNSENT_DATA; 3909 3910 if (after(ack, prior_snd_una)) { 3911 flag |= FLAG_SND_UNA_ADVANCED; 3912 icsk->icsk_retransmits = 0; 3913 3914 #if IS_ENABLED(CONFIG_TLS_DEVICE) 3915 if (static_branch_unlikely(&clean_acked_data_enabled.key)) 3916 if (icsk->icsk_clean_acked) 3917 icsk->icsk_clean_acked(sk, ack); 3918 #endif 3919 } 3920 3921 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una; 3922 rs.prior_in_flight = tcp_packets_in_flight(tp); 3923 3924 /* ts_recent update must be made after we are sure that the packet 3925 * is in window. 3926 */ 3927 if (flag & FLAG_UPDATE_TS_RECENT) 3928 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); 3929 3930 if ((flag & (FLAG_SLOWPATH | FLAG_SND_UNA_ADVANCED)) == 3931 FLAG_SND_UNA_ADVANCED) { 3932 /* Window is constant, pure forward advance. 3933 * No more checks are required. 3934 * Note, we use the fact that SND.UNA>=SND.WL2. 3935 */ 3936 tcp_update_wl(tp, ack_seq); 3937 tcp_snd_una_update(tp, ack); 3938 flag |= FLAG_WIN_UPDATE; 3939 3940 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE); 3941 3942 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS); 3943 } else { 3944 u32 ack_ev_flags = CA_ACK_SLOWPATH; 3945 3946 if (ack_seq != TCP_SKB_CB(skb)->end_seq) 3947 flag |= FLAG_DATA; 3948 else 3949 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS); 3950 3951 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq); 3952 3953 if (TCP_SKB_CB(skb)->sacked) 3954 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una, 3955 &sack_state); 3956 3957 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) { 3958 flag |= FLAG_ECE; 3959 ack_ev_flags |= CA_ACK_ECE; 3960 } 3961 3962 if (sack_state.sack_delivered) 3963 tcp_count_delivered(tp, sack_state.sack_delivered, 3964 flag & FLAG_ECE); 3965 3966 if (flag & FLAG_WIN_UPDATE) 3967 ack_ev_flags |= CA_ACK_WIN_UPDATE; 3968 3969 tcp_in_ack_event(sk, ack_ev_flags); 3970 } 3971 3972 /* This is a deviation from RFC3168 since it states that: 3973 * "When the TCP data sender is ready to set the CWR bit after reducing 3974 * the congestion window, it SHOULD set the CWR bit only on the first 3975 * new data packet that it transmits." 3976 * We accept CWR on pure ACKs to be more robust 3977 * with widely-deployed TCP implementations that do this. 3978 */ 3979 tcp_ecn_accept_cwr(sk, skb); 3980 3981 /* We passed data and got it acked, remove any soft error 3982 * log. Something worked... 3983 */ 3984 WRITE_ONCE(sk->sk_err_soft, 0); 3985 icsk->icsk_probes_out = 0; 3986 tp->rcv_tstamp = tcp_jiffies32; 3987 if (!prior_packets) 3988 goto no_queue; 3989 3990 /* See if we can take anything off of the retransmit queue. */ 3991 flag |= tcp_clean_rtx_queue(sk, skb, prior_fack, prior_snd_una, 3992 &sack_state, flag & FLAG_ECE); 3993 3994 tcp_rack_update_reo_wnd(sk, &rs); 3995 3996 if (tp->tlp_high_seq) 3997 tcp_process_tlp_ack(sk, ack, flag); 3998 3999 if (tcp_ack_is_dubious(sk, flag)) { 4000 if (!(flag & (FLAG_SND_UNA_ADVANCED | 4001 FLAG_NOT_DUP | FLAG_DSACKING_ACK))) { 4002 num_dupack = 1; 4003 /* Consider if pure acks were aggregated in tcp_add_backlog() */ 4004 if (!(flag & FLAG_DATA)) 4005 num_dupack = max_t(u16, 1, skb_shinfo(skb)->gso_segs); 4006 } 4007 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag, 4008 &rexmit); 4009 } 4010 4011 /* If needed, reset TLP/RTO timer when RACK doesn't set. */ 4012 if (flag & FLAG_SET_XMIT_TIMER) 4013 tcp_set_xmit_timer(sk); 4014 4015 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) 4016 sk_dst_confirm(sk); 4017 4018 delivered = tcp_newly_delivered(sk, delivered, flag); 4019 lost = tp->lost - lost; /* freshly marked lost */ 4020 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED); 4021 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate); 4022 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate); 4023 tcp_xmit_recovery(sk, rexmit); 4024 return 1; 4025 4026 no_queue: 4027 /* If data was DSACKed, see if we can undo a cwnd reduction. */ 4028 if (flag & FLAG_DSACKING_ACK) { 4029 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag, 4030 &rexmit); 4031 tcp_newly_delivered(sk, delivered, flag); 4032 } 4033 /* If this ack opens up a zero window, clear backoff. It was 4034 * being used to time the probes, and is probably far higher than 4035 * it needs to be for normal retransmission. 4036 */ 4037 tcp_ack_probe(sk); 4038 4039 if (tp->tlp_high_seq) 4040 tcp_process_tlp_ack(sk, ack, flag); 4041 return 1; 4042 4043 old_ack: 4044 /* If data was SACKed, tag it and see if we should send more data. 4045 * If data was DSACKed, see if we can undo a cwnd reduction. 4046 */ 4047 if (TCP_SKB_CB(skb)->sacked) { 4048 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una, 4049 &sack_state); 4050 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag, 4051 &rexmit); 4052 tcp_newly_delivered(sk, delivered, flag); 4053 tcp_xmit_recovery(sk, rexmit); 4054 } 4055 4056 return 0; 4057 } 4058 4059 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie, 4060 bool syn, struct tcp_fastopen_cookie *foc, 4061 bool exp_opt) 4062 { 4063 /* Valid only in SYN or SYN-ACK with an even length. */ 4064 if (!foc || !syn || len < 0 || (len & 1)) 4065 return; 4066 4067 if (len >= TCP_FASTOPEN_COOKIE_MIN && 4068 len <= TCP_FASTOPEN_COOKIE_MAX) 4069 memcpy(foc->val, cookie, len); 4070 else if (len != 0) 4071 len = -1; 4072 foc->len = len; 4073 foc->exp = exp_opt; 4074 } 4075 4076 static bool smc_parse_options(const struct tcphdr *th, 4077 struct tcp_options_received *opt_rx, 4078 const unsigned char *ptr, 4079 int opsize) 4080 { 4081 #if IS_ENABLED(CONFIG_SMC) 4082 if (static_branch_unlikely(&tcp_have_smc)) { 4083 if (th->syn && !(opsize & 1) && 4084 opsize >= TCPOLEN_EXP_SMC_BASE && 4085 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC) { 4086 opt_rx->smc_ok = 1; 4087 return true; 4088 } 4089 } 4090 #endif 4091 return false; 4092 } 4093 4094 /* Try to parse the MSS option from the TCP header. Return 0 on failure, clamped 4095 * value on success. 4096 */ 4097 u16 tcp_parse_mss_option(const struct tcphdr *th, u16 user_mss) 4098 { 4099 const unsigned char *ptr = (const unsigned char *)(th + 1); 4100 int length = (th->doff * 4) - sizeof(struct tcphdr); 4101 u16 mss = 0; 4102 4103 while (length > 0) { 4104 int opcode = *ptr++; 4105 int opsize; 4106 4107 switch (opcode) { 4108 case TCPOPT_EOL: 4109 return mss; 4110 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ 4111 length--; 4112 continue; 4113 default: 4114 if (length < 2) 4115 return mss; 4116 opsize = *ptr++; 4117 if (opsize < 2) /* "silly options" */ 4118 return mss; 4119 if (opsize > length) 4120 return mss; /* fail on partial options */ 4121 if (opcode == TCPOPT_MSS && opsize == TCPOLEN_MSS) { 4122 u16 in_mss = get_unaligned_be16(ptr); 4123 4124 if (in_mss) { 4125 if (user_mss && user_mss < in_mss) 4126 in_mss = user_mss; 4127 mss = in_mss; 4128 } 4129 } 4130 ptr += opsize - 2; 4131 length -= opsize; 4132 } 4133 } 4134 return mss; 4135 } 4136 EXPORT_SYMBOL_GPL(tcp_parse_mss_option); 4137 4138 /* Look for tcp options. Normally only called on SYN and SYNACK packets. 4139 * But, this can also be called on packets in the established flow when 4140 * the fast version below fails. 4141 */ 4142 void tcp_parse_options(const struct net *net, 4143 const struct sk_buff *skb, 4144 struct tcp_options_received *opt_rx, int estab, 4145 struct tcp_fastopen_cookie *foc) 4146 { 4147 const unsigned char *ptr; 4148 const struct tcphdr *th = tcp_hdr(skb); 4149 int length = (th->doff * 4) - sizeof(struct tcphdr); 4150 4151 ptr = (const unsigned char *)(th + 1); 4152 opt_rx->saw_tstamp = 0; 4153 opt_rx->saw_unknown = 0; 4154 4155 while (length > 0) { 4156 int opcode = *ptr++; 4157 int opsize; 4158 4159 switch (opcode) { 4160 case TCPOPT_EOL: 4161 return; 4162 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ 4163 length--; 4164 continue; 4165 default: 4166 if (length < 2) 4167 return; 4168 opsize = *ptr++; 4169 if (opsize < 2) /* "silly options" */ 4170 return; 4171 if (opsize > length) 4172 return; /* don't parse partial options */ 4173 switch (opcode) { 4174 case TCPOPT_MSS: 4175 if (opsize == TCPOLEN_MSS && th->syn && !estab) { 4176 u16 in_mss = get_unaligned_be16(ptr); 4177 if (in_mss) { 4178 if (opt_rx->user_mss && 4179 opt_rx->user_mss < in_mss) 4180 in_mss = opt_rx->user_mss; 4181 opt_rx->mss_clamp = in_mss; 4182 } 4183 } 4184 break; 4185 case TCPOPT_WINDOW: 4186 if (opsize == TCPOLEN_WINDOW && th->syn && 4187 !estab && READ_ONCE(net->ipv4.sysctl_tcp_window_scaling)) { 4188 __u8 snd_wscale = *(__u8 *)ptr; 4189 opt_rx->wscale_ok = 1; 4190 if (snd_wscale > TCP_MAX_WSCALE) { 4191 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n", 4192 __func__, 4193 snd_wscale, 4194 TCP_MAX_WSCALE); 4195 snd_wscale = TCP_MAX_WSCALE; 4196 } 4197 opt_rx->snd_wscale = snd_wscale; 4198 } 4199 break; 4200 case TCPOPT_TIMESTAMP: 4201 if ((opsize == TCPOLEN_TIMESTAMP) && 4202 ((estab && opt_rx->tstamp_ok) || 4203 (!estab && READ_ONCE(net->ipv4.sysctl_tcp_timestamps)))) { 4204 opt_rx->saw_tstamp = 1; 4205 opt_rx->rcv_tsval = get_unaligned_be32(ptr); 4206 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4); 4207 } 4208 break; 4209 case TCPOPT_SACK_PERM: 4210 if (opsize == TCPOLEN_SACK_PERM && th->syn && 4211 !estab && READ_ONCE(net->ipv4.sysctl_tcp_sack)) { 4212 opt_rx->sack_ok = TCP_SACK_SEEN; 4213 tcp_sack_reset(opt_rx); 4214 } 4215 break; 4216 4217 case TCPOPT_SACK: 4218 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) && 4219 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) && 4220 opt_rx->sack_ok) { 4221 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th; 4222 } 4223 break; 4224 #ifdef CONFIG_TCP_MD5SIG 4225 case TCPOPT_MD5SIG: 4226 /* The MD5 Hash has already been 4227 * checked (see tcp_v{4,6}_rcv()). 4228 */ 4229 break; 4230 #endif 4231 case TCPOPT_FASTOPEN: 4232 tcp_parse_fastopen_option( 4233 opsize - TCPOLEN_FASTOPEN_BASE, 4234 ptr, th->syn, foc, false); 4235 break; 4236 4237 case TCPOPT_EXP: 4238 /* Fast Open option shares code 254 using a 4239 * 16 bits magic number. 4240 */ 4241 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE && 4242 get_unaligned_be16(ptr) == 4243 TCPOPT_FASTOPEN_MAGIC) { 4244 tcp_parse_fastopen_option(opsize - 4245 TCPOLEN_EXP_FASTOPEN_BASE, 4246 ptr + 2, th->syn, foc, true); 4247 break; 4248 } 4249 4250 if (smc_parse_options(th, opt_rx, ptr, opsize)) 4251 break; 4252 4253 opt_rx->saw_unknown = 1; 4254 break; 4255 4256 default: 4257 opt_rx->saw_unknown = 1; 4258 } 4259 ptr += opsize-2; 4260 length -= opsize; 4261 } 4262 } 4263 } 4264 EXPORT_SYMBOL(tcp_parse_options); 4265 4266 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th) 4267 { 4268 const __be32 *ptr = (const __be32 *)(th + 1); 4269 4270 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) 4271 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) { 4272 tp->rx_opt.saw_tstamp = 1; 4273 ++ptr; 4274 tp->rx_opt.rcv_tsval = ntohl(*ptr); 4275 ++ptr; 4276 if (*ptr) 4277 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset; 4278 else 4279 tp->rx_opt.rcv_tsecr = 0; 4280 return true; 4281 } 4282 return false; 4283 } 4284 4285 /* Fast parse options. This hopes to only see timestamps. 4286 * If it is wrong it falls back on tcp_parse_options(). 4287 */ 4288 static bool tcp_fast_parse_options(const struct net *net, 4289 const struct sk_buff *skb, 4290 const struct tcphdr *th, struct tcp_sock *tp) 4291 { 4292 /* In the spirit of fast parsing, compare doff directly to constant 4293 * values. Because equality is used, short doff can be ignored here. 4294 */ 4295 if (th->doff == (sizeof(*th) / 4)) { 4296 tp->rx_opt.saw_tstamp = 0; 4297 return false; 4298 } else if (tp->rx_opt.tstamp_ok && 4299 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) { 4300 if (tcp_parse_aligned_timestamp(tp, th)) 4301 return true; 4302 } 4303 4304 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL); 4305 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) 4306 tp->rx_opt.rcv_tsecr -= tp->tsoffset; 4307 4308 return true; 4309 } 4310 4311 #if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO) 4312 /* 4313 * Parse Signature options 4314 */ 4315 int tcp_do_parse_auth_options(const struct tcphdr *th, 4316 const u8 **md5_hash, const u8 **ao_hash) 4317 { 4318 int length = (th->doff << 2) - sizeof(*th); 4319 const u8 *ptr = (const u8 *)(th + 1); 4320 unsigned int minlen = TCPOLEN_MD5SIG; 4321 4322 if (IS_ENABLED(CONFIG_TCP_AO)) 4323 minlen = sizeof(struct tcp_ao_hdr) + 1; 4324 4325 *md5_hash = NULL; 4326 *ao_hash = NULL; 4327 4328 /* If not enough data remaining, we can short cut */ 4329 while (length >= minlen) { 4330 int opcode = *ptr++; 4331 int opsize; 4332 4333 switch (opcode) { 4334 case TCPOPT_EOL: 4335 return 0; 4336 case TCPOPT_NOP: 4337 length--; 4338 continue; 4339 default: 4340 opsize = *ptr++; 4341 if (opsize < 2 || opsize > length) 4342 return -EINVAL; 4343 if (opcode == TCPOPT_MD5SIG) { 4344 if (opsize != TCPOLEN_MD5SIG) 4345 return -EINVAL; 4346 if (unlikely(*md5_hash || *ao_hash)) 4347 return -EEXIST; 4348 *md5_hash = ptr; 4349 } else if (opcode == TCPOPT_AO) { 4350 if (opsize <= sizeof(struct tcp_ao_hdr)) 4351 return -EINVAL; 4352 if (unlikely(*md5_hash || *ao_hash)) 4353 return -EEXIST; 4354 *ao_hash = ptr; 4355 } 4356 } 4357 ptr += opsize - 2; 4358 length -= opsize; 4359 } 4360 return 0; 4361 } 4362 EXPORT_SYMBOL(tcp_do_parse_auth_options); 4363 #endif 4364 4365 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM 4366 * 4367 * It is not fatal. If this ACK does _not_ change critical state (seqs, window) 4368 * it can pass through stack. So, the following predicate verifies that 4369 * this segment is not used for anything but congestion avoidance or 4370 * fast retransmit. Moreover, we even are able to eliminate most of such 4371 * second order effects, if we apply some small "replay" window (~RTO) 4372 * to timestamp space. 4373 * 4374 * All these measures still do not guarantee that we reject wrapped ACKs 4375 * on networks with high bandwidth, when sequence space is recycled fastly, 4376 * but it guarantees that such events will be very rare and do not affect 4377 * connection seriously. This doesn't look nice, but alas, PAWS is really 4378 * buggy extension. 4379 * 4380 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC 4381 * states that events when retransmit arrives after original data are rare. 4382 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is 4383 * the biggest problem on large power networks even with minor reordering. 4384 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe 4385 * up to bandwidth of 18Gigabit/sec. 8) ] 4386 */ 4387 4388 /* Estimates max number of increments of remote peer TSval in 4389 * a replay window (based on our current RTO estimation). 4390 */ 4391 static u32 tcp_tsval_replay(const struct sock *sk) 4392 { 4393 /* If we use usec TS resolution, 4394 * then expect the remote peer to use the same resolution. 4395 */ 4396 if (tcp_sk(sk)->tcp_usec_ts) 4397 return inet_csk(sk)->icsk_rto * (USEC_PER_SEC / HZ); 4398 4399 /* RFC 7323 recommends a TSval clock between 1ms and 1sec. 4400 * We know that some OS (including old linux) can use 1200 Hz. 4401 */ 4402 return inet_csk(sk)->icsk_rto * 1200 / HZ; 4403 } 4404 4405 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb) 4406 { 4407 const struct tcp_sock *tp = tcp_sk(sk); 4408 const struct tcphdr *th = tcp_hdr(skb); 4409 u32 seq = TCP_SKB_CB(skb)->seq; 4410 u32 ack = TCP_SKB_CB(skb)->ack_seq; 4411 4412 return /* 1. Pure ACK with correct sequence number. */ 4413 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) && 4414 4415 /* 2. ... and duplicate ACK. */ 4416 ack == tp->snd_una && 4417 4418 /* 3. ... and does not update window. */ 4419 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) && 4420 4421 /* 4. ... and sits in replay window. */ 4422 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= 4423 tcp_tsval_replay(sk); 4424 } 4425 4426 static inline bool tcp_paws_discard(const struct sock *sk, 4427 const struct sk_buff *skb) 4428 { 4429 const struct tcp_sock *tp = tcp_sk(sk); 4430 4431 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) && 4432 !tcp_disordered_ack(sk, skb); 4433 } 4434 4435 /* Check segment sequence number for validity. 4436 * 4437 * Segment controls are considered valid, if the segment 4438 * fits to the window after truncation to the window. Acceptability 4439 * of data (and SYN, FIN, of course) is checked separately. 4440 * See tcp_data_queue(), for example. 4441 * 4442 * Also, controls (RST is main one) are accepted using RCV.WUP instead 4443 * of RCV.NXT. Peer still did not advance his SND.UNA when we 4444 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP. 4445 * (borrowed from freebsd) 4446 */ 4447 4448 static enum skb_drop_reason tcp_sequence(const struct tcp_sock *tp, 4449 u32 seq, u32 end_seq) 4450 { 4451 if (before(end_seq, tp->rcv_wup)) 4452 return SKB_DROP_REASON_TCP_OLD_SEQUENCE; 4453 4454 if (after(seq, tp->rcv_nxt + tcp_receive_window(tp))) 4455 return SKB_DROP_REASON_TCP_INVALID_SEQUENCE; 4456 4457 return SKB_NOT_DROPPED_YET; 4458 } 4459 4460 4461 void tcp_done_with_error(struct sock *sk, int err) 4462 { 4463 /* This barrier is coupled with smp_rmb() in tcp_poll() */ 4464 WRITE_ONCE(sk->sk_err, err); 4465 smp_wmb(); 4466 4467 tcp_write_queue_purge(sk); 4468 tcp_done(sk); 4469 4470 if (!sock_flag(sk, SOCK_DEAD)) 4471 sk_error_report(sk); 4472 } 4473 EXPORT_SYMBOL(tcp_done_with_error); 4474 4475 /* When we get a reset we do this. */ 4476 void tcp_reset(struct sock *sk, struct sk_buff *skb) 4477 { 4478 int err; 4479 4480 trace_tcp_receive_reset(sk); 4481 4482 /* mptcp can't tell us to ignore reset pkts, 4483 * so just ignore the return value of mptcp_incoming_options(). 4484 */ 4485 if (sk_is_mptcp(sk)) 4486 mptcp_incoming_options(sk, skb); 4487 4488 /* We want the right error as BSD sees it (and indeed as we do). */ 4489 switch (sk->sk_state) { 4490 case TCP_SYN_SENT: 4491 err = ECONNREFUSED; 4492 break; 4493 case TCP_CLOSE_WAIT: 4494 err = EPIPE; 4495 break; 4496 case TCP_CLOSE: 4497 return; 4498 default: 4499 err = ECONNRESET; 4500 } 4501 tcp_done_with_error(sk, err); 4502 } 4503 4504 /* 4505 * Process the FIN bit. This now behaves as it is supposed to work 4506 * and the FIN takes effect when it is validly part of sequence 4507 * space. Not before when we get holes. 4508 * 4509 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT 4510 * (and thence onto LAST-ACK and finally, CLOSE, we never enter 4511 * TIME-WAIT) 4512 * 4513 * If we are in FINWAIT-1, a received FIN indicates simultaneous 4514 * close and we go into CLOSING (and later onto TIME-WAIT) 4515 * 4516 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT. 4517 */ 4518 void tcp_fin(struct sock *sk) 4519 { 4520 struct tcp_sock *tp = tcp_sk(sk); 4521 4522 inet_csk_schedule_ack(sk); 4523 4524 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN); 4525 sock_set_flag(sk, SOCK_DONE); 4526 4527 switch (sk->sk_state) { 4528 case TCP_SYN_RECV: 4529 case TCP_ESTABLISHED: 4530 /* Move to CLOSE_WAIT */ 4531 tcp_set_state(sk, TCP_CLOSE_WAIT); 4532 inet_csk_enter_pingpong_mode(sk); 4533 break; 4534 4535 case TCP_CLOSE_WAIT: 4536 case TCP_CLOSING: 4537 /* Received a retransmission of the FIN, do 4538 * nothing. 4539 */ 4540 break; 4541 case TCP_LAST_ACK: 4542 /* RFC793: Remain in the LAST-ACK state. */ 4543 break; 4544 4545 case TCP_FIN_WAIT1: 4546 /* This case occurs when a simultaneous close 4547 * happens, we must ack the received FIN and 4548 * enter the CLOSING state. 4549 */ 4550 tcp_send_ack(sk); 4551 tcp_set_state(sk, TCP_CLOSING); 4552 break; 4553 case TCP_FIN_WAIT2: 4554 /* Received a FIN -- send ACK and enter TIME_WAIT. */ 4555 tcp_send_ack(sk); 4556 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 4557 break; 4558 default: 4559 /* Only TCP_LISTEN and TCP_CLOSE are left, in these 4560 * cases we should never reach this piece of code. 4561 */ 4562 pr_err("%s: Impossible, sk->sk_state=%d\n", 4563 __func__, sk->sk_state); 4564 break; 4565 } 4566 4567 /* It _is_ possible, that we have something out-of-order _after_ FIN. 4568 * Probably, we should reset in this case. For now drop them. 4569 */ 4570 skb_rbtree_purge(&tp->out_of_order_queue); 4571 if (tcp_is_sack(tp)) 4572 tcp_sack_reset(&tp->rx_opt); 4573 4574 if (!sock_flag(sk, SOCK_DEAD)) { 4575 sk->sk_state_change(sk); 4576 4577 /* Do not send POLL_HUP for half duplex close. */ 4578 if (sk->sk_shutdown == SHUTDOWN_MASK || 4579 sk->sk_state == TCP_CLOSE) 4580 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP); 4581 else 4582 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN); 4583 } 4584 } 4585 4586 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, 4587 u32 end_seq) 4588 { 4589 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) { 4590 if (before(seq, sp->start_seq)) 4591 sp->start_seq = seq; 4592 if (after(end_seq, sp->end_seq)) 4593 sp->end_seq = end_seq; 4594 return true; 4595 } 4596 return false; 4597 } 4598 4599 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq) 4600 { 4601 struct tcp_sock *tp = tcp_sk(sk); 4602 4603 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) { 4604 int mib_idx; 4605 4606 if (before(seq, tp->rcv_nxt)) 4607 mib_idx = LINUX_MIB_TCPDSACKOLDSENT; 4608 else 4609 mib_idx = LINUX_MIB_TCPDSACKOFOSENT; 4610 4611 NET_INC_STATS(sock_net(sk), mib_idx); 4612 4613 tp->rx_opt.dsack = 1; 4614 tp->duplicate_sack[0].start_seq = seq; 4615 tp->duplicate_sack[0].end_seq = end_seq; 4616 } 4617 } 4618 4619 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq) 4620 { 4621 struct tcp_sock *tp = tcp_sk(sk); 4622 4623 if (!tp->rx_opt.dsack) 4624 tcp_dsack_set(sk, seq, end_seq); 4625 else 4626 tcp_sack_extend(tp->duplicate_sack, seq, end_seq); 4627 } 4628 4629 static void tcp_rcv_spurious_retrans(struct sock *sk, const struct sk_buff *skb) 4630 { 4631 /* When the ACK path fails or drops most ACKs, the sender would 4632 * timeout and spuriously retransmit the same segment repeatedly. 4633 * If it seems our ACKs are not reaching the other side, 4634 * based on receiving a duplicate data segment with new flowlabel 4635 * (suggesting the sender suffered an RTO), and we are not already 4636 * repathing due to our own RTO, then rehash the socket to repath our 4637 * packets. 4638 */ 4639 #if IS_ENABLED(CONFIG_IPV6) 4640 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss && 4641 skb->protocol == htons(ETH_P_IPV6) && 4642 (tcp_sk(sk)->inet_conn.icsk_ack.lrcv_flowlabel != 4643 ntohl(ip6_flowlabel(ipv6_hdr(skb)))) && 4644 sk_rethink_txhash(sk)) 4645 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDUPLICATEDATAREHASH); 4646 4647 /* Save last flowlabel after a spurious retrans. */ 4648 tcp_save_lrcv_flowlabel(sk, skb); 4649 #endif 4650 } 4651 4652 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb) 4653 { 4654 struct tcp_sock *tp = tcp_sk(sk); 4655 4656 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 4657 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4658 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); 4659 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS); 4660 4661 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) { 4662 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 4663 4664 tcp_rcv_spurious_retrans(sk, skb); 4665 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) 4666 end_seq = tp->rcv_nxt; 4667 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq); 4668 } 4669 } 4670 4671 tcp_send_ack(sk); 4672 } 4673 4674 /* These routines update the SACK block as out-of-order packets arrive or 4675 * in-order packets close up the sequence space. 4676 */ 4677 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp) 4678 { 4679 int this_sack; 4680 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4681 struct tcp_sack_block *swalk = sp + 1; 4682 4683 /* See if the recent change to the first SACK eats into 4684 * or hits the sequence space of other SACK blocks, if so coalesce. 4685 */ 4686 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) { 4687 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) { 4688 int i; 4689 4690 /* Zap SWALK, by moving every further SACK up by one slot. 4691 * Decrease num_sacks. 4692 */ 4693 tp->rx_opt.num_sacks--; 4694 for (i = this_sack; i < tp->rx_opt.num_sacks; i++) 4695 sp[i] = sp[i + 1]; 4696 continue; 4697 } 4698 this_sack++; 4699 swalk++; 4700 } 4701 } 4702 4703 void tcp_sack_compress_send_ack(struct sock *sk) 4704 { 4705 struct tcp_sock *tp = tcp_sk(sk); 4706 4707 if (!tp->compressed_ack) 4708 return; 4709 4710 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1) 4711 __sock_put(sk); 4712 4713 /* Since we have to send one ack finally, 4714 * substract one from tp->compressed_ack to keep 4715 * LINUX_MIB_TCPACKCOMPRESSED accurate. 4716 */ 4717 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED, 4718 tp->compressed_ack - 1); 4719 4720 tp->compressed_ack = 0; 4721 tcp_send_ack(sk); 4722 } 4723 4724 /* Reasonable amount of sack blocks included in TCP SACK option 4725 * The max is 4, but this becomes 3 if TCP timestamps are there. 4726 * Given that SACK packets might be lost, be conservative and use 2. 4727 */ 4728 #define TCP_SACK_BLOCKS_EXPECTED 2 4729 4730 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq) 4731 { 4732 struct tcp_sock *tp = tcp_sk(sk); 4733 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4734 int cur_sacks = tp->rx_opt.num_sacks; 4735 int this_sack; 4736 4737 if (!cur_sacks) 4738 goto new_sack; 4739 4740 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) { 4741 if (tcp_sack_extend(sp, seq, end_seq)) { 4742 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED) 4743 tcp_sack_compress_send_ack(sk); 4744 /* Rotate this_sack to the first one. */ 4745 for (; this_sack > 0; this_sack--, sp--) 4746 swap(*sp, *(sp - 1)); 4747 if (cur_sacks > 1) 4748 tcp_sack_maybe_coalesce(tp); 4749 return; 4750 } 4751 } 4752 4753 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED) 4754 tcp_sack_compress_send_ack(sk); 4755 4756 /* Could not find an adjacent existing SACK, build a new one, 4757 * put it at the front, and shift everyone else down. We 4758 * always know there is at least one SACK present already here. 4759 * 4760 * If the sack array is full, forget about the last one. 4761 */ 4762 if (this_sack >= TCP_NUM_SACKS) { 4763 this_sack--; 4764 tp->rx_opt.num_sacks--; 4765 sp--; 4766 } 4767 for (; this_sack > 0; this_sack--, sp--) 4768 *sp = *(sp - 1); 4769 4770 new_sack: 4771 /* Build the new head SACK, and we're done. */ 4772 sp->start_seq = seq; 4773 sp->end_seq = end_seq; 4774 tp->rx_opt.num_sacks++; 4775 } 4776 4777 /* RCV.NXT advances, some SACKs should be eaten. */ 4778 4779 static void tcp_sack_remove(struct tcp_sock *tp) 4780 { 4781 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4782 int num_sacks = tp->rx_opt.num_sacks; 4783 int this_sack; 4784 4785 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */ 4786 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) { 4787 tp->rx_opt.num_sacks = 0; 4788 return; 4789 } 4790 4791 for (this_sack = 0; this_sack < num_sacks;) { 4792 /* Check if the start of the sack is covered by RCV.NXT. */ 4793 if (!before(tp->rcv_nxt, sp->start_seq)) { 4794 int i; 4795 4796 /* RCV.NXT must cover all the block! */ 4797 WARN_ON(before(tp->rcv_nxt, sp->end_seq)); 4798 4799 /* Zap this SACK, by moving forward any other SACKS. */ 4800 for (i = this_sack+1; i < num_sacks; i++) 4801 tp->selective_acks[i-1] = tp->selective_acks[i]; 4802 num_sacks--; 4803 continue; 4804 } 4805 this_sack++; 4806 sp++; 4807 } 4808 tp->rx_opt.num_sacks = num_sacks; 4809 } 4810 4811 /** 4812 * tcp_try_coalesce - try to merge skb to prior one 4813 * @sk: socket 4814 * @to: prior buffer 4815 * @from: buffer to add in queue 4816 * @fragstolen: pointer to boolean 4817 * 4818 * Before queueing skb @from after @to, try to merge them 4819 * to reduce overall memory use and queue lengths, if cost is small. 4820 * Packets in ofo or receive queues can stay a long time. 4821 * Better try to coalesce them right now to avoid future collapses. 4822 * Returns true if caller should free @from instead of queueing it 4823 */ 4824 static bool tcp_try_coalesce(struct sock *sk, 4825 struct sk_buff *to, 4826 struct sk_buff *from, 4827 bool *fragstolen) 4828 { 4829 int delta; 4830 4831 *fragstolen = false; 4832 4833 /* Its possible this segment overlaps with prior segment in queue */ 4834 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq) 4835 return false; 4836 4837 if (!tcp_skb_can_collapse_rx(to, from)) 4838 return false; 4839 4840 if (!skb_try_coalesce(to, from, fragstolen, &delta)) 4841 return false; 4842 4843 atomic_add(delta, &sk->sk_rmem_alloc); 4844 sk_mem_charge(sk, delta); 4845 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE); 4846 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq; 4847 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq; 4848 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags; 4849 4850 if (TCP_SKB_CB(from)->has_rxtstamp) { 4851 TCP_SKB_CB(to)->has_rxtstamp = true; 4852 to->tstamp = from->tstamp; 4853 skb_hwtstamps(to)->hwtstamp = skb_hwtstamps(from)->hwtstamp; 4854 } 4855 4856 return true; 4857 } 4858 4859 static bool tcp_ooo_try_coalesce(struct sock *sk, 4860 struct sk_buff *to, 4861 struct sk_buff *from, 4862 bool *fragstolen) 4863 { 4864 bool res = tcp_try_coalesce(sk, to, from, fragstolen); 4865 4866 /* In case tcp_drop_reason() is called later, update to->gso_segs */ 4867 if (res) { 4868 u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) + 4869 max_t(u16, 1, skb_shinfo(from)->gso_segs); 4870 4871 skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF); 4872 } 4873 return res; 4874 } 4875 4876 static void tcp_drop_reason(struct sock *sk, struct sk_buff *skb, 4877 enum skb_drop_reason reason) 4878 { 4879 sk_drops_add(sk, skb); 4880 sk_skb_reason_drop(sk, skb, reason); 4881 } 4882 4883 /* This one checks to see if we can put data from the 4884 * out_of_order queue into the receive_queue. 4885 */ 4886 static void tcp_ofo_queue(struct sock *sk) 4887 { 4888 struct tcp_sock *tp = tcp_sk(sk); 4889 __u32 dsack_high = tp->rcv_nxt; 4890 bool fin, fragstolen, eaten; 4891 struct sk_buff *skb, *tail; 4892 struct rb_node *p; 4893 4894 p = rb_first(&tp->out_of_order_queue); 4895 while (p) { 4896 skb = rb_to_skb(p); 4897 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 4898 break; 4899 4900 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) { 4901 __u32 dsack = dsack_high; 4902 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high)) 4903 dsack_high = TCP_SKB_CB(skb)->end_seq; 4904 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack); 4905 } 4906 p = rb_next(p); 4907 rb_erase(&skb->rbnode, &tp->out_of_order_queue); 4908 4909 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) { 4910 tcp_drop_reason(sk, skb, SKB_DROP_REASON_TCP_OFO_DROP); 4911 continue; 4912 } 4913 4914 tail = skb_peek_tail(&sk->sk_receive_queue); 4915 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen); 4916 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq); 4917 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN; 4918 if (!eaten) 4919 __skb_queue_tail(&sk->sk_receive_queue, skb); 4920 else 4921 kfree_skb_partial(skb, fragstolen); 4922 4923 if (unlikely(fin)) { 4924 tcp_fin(sk); 4925 /* tcp_fin() purges tp->out_of_order_queue, 4926 * so we must end this loop right now. 4927 */ 4928 break; 4929 } 4930 } 4931 } 4932 4933 static bool tcp_prune_ofo_queue(struct sock *sk, const struct sk_buff *in_skb); 4934 static int tcp_prune_queue(struct sock *sk, const struct sk_buff *in_skb); 4935 4936 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb, 4937 unsigned int size) 4938 { 4939 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || 4940 !sk_rmem_schedule(sk, skb, size)) { 4941 4942 if (tcp_prune_queue(sk, skb) < 0) 4943 return -1; 4944 4945 while (!sk_rmem_schedule(sk, skb, size)) { 4946 if (!tcp_prune_ofo_queue(sk, skb)) 4947 return -1; 4948 } 4949 } 4950 return 0; 4951 } 4952 4953 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb) 4954 { 4955 struct tcp_sock *tp = tcp_sk(sk); 4956 struct rb_node **p, *parent; 4957 struct sk_buff *skb1; 4958 u32 seq, end_seq; 4959 bool fragstolen; 4960 4961 tcp_save_lrcv_flowlabel(sk, skb); 4962 tcp_ecn_check_ce(sk, skb); 4963 4964 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) { 4965 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP); 4966 sk->sk_data_ready(sk); 4967 tcp_drop_reason(sk, skb, SKB_DROP_REASON_PROTO_MEM); 4968 return; 4969 } 4970 4971 /* Disable header prediction. */ 4972 tp->pred_flags = 0; 4973 inet_csk_schedule_ack(sk); 4974 4975 tp->rcv_ooopack += max_t(u16, 1, skb_shinfo(skb)->gso_segs); 4976 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE); 4977 seq = TCP_SKB_CB(skb)->seq; 4978 end_seq = TCP_SKB_CB(skb)->end_seq; 4979 4980 p = &tp->out_of_order_queue.rb_node; 4981 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) { 4982 /* Initial out of order segment, build 1 SACK. */ 4983 if (tcp_is_sack(tp)) { 4984 tp->rx_opt.num_sacks = 1; 4985 tp->selective_acks[0].start_seq = seq; 4986 tp->selective_acks[0].end_seq = end_seq; 4987 } 4988 rb_link_node(&skb->rbnode, NULL, p); 4989 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue); 4990 tp->ooo_last_skb = skb; 4991 goto end; 4992 } 4993 4994 /* In the typical case, we are adding an skb to the end of the list. 4995 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup. 4996 */ 4997 if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb, 4998 skb, &fragstolen)) { 4999 coalesce_done: 5000 /* For non sack flows, do not grow window to force DUPACK 5001 * and trigger fast retransmit. 5002 */ 5003 if (tcp_is_sack(tp)) 5004 tcp_grow_window(sk, skb, true); 5005 kfree_skb_partial(skb, fragstolen); 5006 skb = NULL; 5007 goto add_sack; 5008 } 5009 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */ 5010 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) { 5011 parent = &tp->ooo_last_skb->rbnode; 5012 p = &parent->rb_right; 5013 goto insert; 5014 } 5015 5016 /* Find place to insert this segment. Handle overlaps on the way. */ 5017 parent = NULL; 5018 while (*p) { 5019 parent = *p; 5020 skb1 = rb_to_skb(parent); 5021 if (before(seq, TCP_SKB_CB(skb1)->seq)) { 5022 p = &parent->rb_left; 5023 continue; 5024 } 5025 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) { 5026 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 5027 /* All the bits are present. Drop. */ 5028 NET_INC_STATS(sock_net(sk), 5029 LINUX_MIB_TCPOFOMERGE); 5030 tcp_drop_reason(sk, skb, 5031 SKB_DROP_REASON_TCP_OFOMERGE); 5032 skb = NULL; 5033 tcp_dsack_set(sk, seq, end_seq); 5034 goto add_sack; 5035 } 5036 if (after(seq, TCP_SKB_CB(skb1)->seq)) { 5037 /* Partial overlap. */ 5038 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq); 5039 } else { 5040 /* skb's seq == skb1's seq and skb covers skb1. 5041 * Replace skb1 with skb. 5042 */ 5043 rb_replace_node(&skb1->rbnode, &skb->rbnode, 5044 &tp->out_of_order_queue); 5045 tcp_dsack_extend(sk, 5046 TCP_SKB_CB(skb1)->seq, 5047 TCP_SKB_CB(skb1)->end_seq); 5048 NET_INC_STATS(sock_net(sk), 5049 LINUX_MIB_TCPOFOMERGE); 5050 tcp_drop_reason(sk, skb1, 5051 SKB_DROP_REASON_TCP_OFOMERGE); 5052 goto merge_right; 5053 } 5054 } else if (tcp_ooo_try_coalesce(sk, skb1, 5055 skb, &fragstolen)) { 5056 goto coalesce_done; 5057 } 5058 p = &parent->rb_right; 5059 } 5060 insert: 5061 /* Insert segment into RB tree. */ 5062 rb_link_node(&skb->rbnode, parent, p); 5063 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue); 5064 5065 merge_right: 5066 /* Remove other segments covered by skb. */ 5067 while ((skb1 = skb_rb_next(skb)) != NULL) { 5068 if (!after(end_seq, TCP_SKB_CB(skb1)->seq)) 5069 break; 5070 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 5071 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, 5072 end_seq); 5073 break; 5074 } 5075 rb_erase(&skb1->rbnode, &tp->out_of_order_queue); 5076 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, 5077 TCP_SKB_CB(skb1)->end_seq); 5078 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE); 5079 tcp_drop_reason(sk, skb1, SKB_DROP_REASON_TCP_OFOMERGE); 5080 } 5081 /* If there is no skb after us, we are the last_skb ! */ 5082 if (!skb1) 5083 tp->ooo_last_skb = skb; 5084 5085 add_sack: 5086 if (tcp_is_sack(tp)) 5087 tcp_sack_new_ofo_skb(sk, seq, end_seq); 5088 end: 5089 if (skb) { 5090 /* For non sack flows, do not grow window to force DUPACK 5091 * and trigger fast retransmit. 5092 */ 5093 if (tcp_is_sack(tp)) 5094 tcp_grow_window(sk, skb, false); 5095 skb_condense(skb); 5096 skb_set_owner_r(skb, sk); 5097 } 5098 } 5099 5100 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, 5101 bool *fragstolen) 5102 { 5103 int eaten; 5104 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue); 5105 5106 eaten = (tail && 5107 tcp_try_coalesce(sk, tail, 5108 skb, fragstolen)) ? 1 : 0; 5109 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq); 5110 if (!eaten) { 5111 __skb_queue_tail(&sk->sk_receive_queue, skb); 5112 skb_set_owner_r(skb, sk); 5113 } 5114 return eaten; 5115 } 5116 5117 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size) 5118 { 5119 struct sk_buff *skb; 5120 int err = -ENOMEM; 5121 int data_len = 0; 5122 bool fragstolen; 5123 5124 if (size == 0) 5125 return 0; 5126 5127 if (size > PAGE_SIZE) { 5128 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS); 5129 5130 data_len = npages << PAGE_SHIFT; 5131 size = data_len + (size & ~PAGE_MASK); 5132 } 5133 skb = alloc_skb_with_frags(size - data_len, data_len, 5134 PAGE_ALLOC_COSTLY_ORDER, 5135 &err, sk->sk_allocation); 5136 if (!skb) 5137 goto err; 5138 5139 skb_put(skb, size - data_len); 5140 skb->data_len = data_len; 5141 skb->len = size; 5142 5143 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) { 5144 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP); 5145 goto err_free; 5146 } 5147 5148 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size); 5149 if (err) 5150 goto err_free; 5151 5152 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt; 5153 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size; 5154 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1; 5155 5156 if (tcp_queue_rcv(sk, skb, &fragstolen)) { 5157 WARN_ON_ONCE(fragstolen); /* should not happen */ 5158 __kfree_skb(skb); 5159 } 5160 return size; 5161 5162 err_free: 5163 kfree_skb(skb); 5164 err: 5165 return err; 5166 5167 } 5168 5169 void tcp_data_ready(struct sock *sk) 5170 { 5171 if (tcp_epollin_ready(sk, sk->sk_rcvlowat) || sock_flag(sk, SOCK_DONE)) 5172 sk->sk_data_ready(sk); 5173 } 5174 5175 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb) 5176 { 5177 struct tcp_sock *tp = tcp_sk(sk); 5178 enum skb_drop_reason reason; 5179 bool fragstolen; 5180 int eaten; 5181 5182 /* If a subflow has been reset, the packet should not continue 5183 * to be processed, drop the packet. 5184 */ 5185 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb)) { 5186 __kfree_skb(skb); 5187 return; 5188 } 5189 5190 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) { 5191 __kfree_skb(skb); 5192 return; 5193 } 5194 skb_dst_drop(skb); 5195 __skb_pull(skb, tcp_hdr(skb)->doff * 4); 5196 5197 reason = SKB_DROP_REASON_NOT_SPECIFIED; 5198 tp->rx_opt.dsack = 0; 5199 5200 /* Queue data for delivery to the user. 5201 * Packets in sequence go to the receive queue. 5202 * Out of sequence packets to the out_of_order_queue. 5203 */ 5204 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { 5205 if (tcp_receive_window(tp) == 0) { 5206 /* Some stacks are known to send bare FIN packets 5207 * in a loop even if we send RWIN 0 in our ACK. 5208 * Accepting this FIN does not hurt memory pressure 5209 * because the FIN flag will simply be merged to the 5210 * receive queue tail skb in most cases. 5211 */ 5212 if (!skb->len && 5213 (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 5214 goto queue_and_out; 5215 5216 reason = SKB_DROP_REASON_TCP_ZEROWINDOW; 5217 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP); 5218 goto out_of_window; 5219 } 5220 5221 /* Ok. In sequence. In window. */ 5222 queue_and_out: 5223 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) { 5224 /* TODO: maybe ratelimit these WIN 0 ACK ? */ 5225 inet_csk(sk)->icsk_ack.pending |= 5226 (ICSK_ACK_NOMEM | ICSK_ACK_NOW); 5227 inet_csk_schedule_ack(sk); 5228 sk->sk_data_ready(sk); 5229 5230 if (skb_queue_len(&sk->sk_receive_queue) && skb->len) { 5231 reason = SKB_DROP_REASON_PROTO_MEM; 5232 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP); 5233 goto drop; 5234 } 5235 sk_forced_mem_schedule(sk, skb->truesize); 5236 } 5237 5238 eaten = tcp_queue_rcv(sk, skb, &fragstolen); 5239 if (skb->len) 5240 tcp_event_data_recv(sk, skb); 5241 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 5242 tcp_fin(sk); 5243 5244 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) { 5245 tcp_ofo_queue(sk); 5246 5247 /* RFC5681. 4.2. SHOULD send immediate ACK, when 5248 * gap in queue is filled. 5249 */ 5250 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) 5251 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW; 5252 } 5253 5254 if (tp->rx_opt.num_sacks) 5255 tcp_sack_remove(tp); 5256 5257 tcp_fast_path_check(sk); 5258 5259 if (eaten > 0) 5260 kfree_skb_partial(skb, fragstolen); 5261 if (!sock_flag(sk, SOCK_DEAD)) 5262 tcp_data_ready(sk); 5263 return; 5264 } 5265 5266 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 5267 tcp_rcv_spurious_retrans(sk, skb); 5268 /* A retransmit, 2nd most common case. Force an immediate ack. */ 5269 reason = SKB_DROP_REASON_TCP_OLD_DATA; 5270 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); 5271 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 5272 5273 out_of_window: 5274 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS); 5275 inet_csk_schedule_ack(sk); 5276 drop: 5277 tcp_drop_reason(sk, skb, reason); 5278 return; 5279 } 5280 5281 /* Out of window. F.e. zero window probe. */ 5282 if (!before(TCP_SKB_CB(skb)->seq, 5283 tp->rcv_nxt + tcp_receive_window(tp))) { 5284 reason = SKB_DROP_REASON_TCP_OVERWINDOW; 5285 goto out_of_window; 5286 } 5287 5288 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 5289 /* Partial packet, seq < rcv_next < end_seq */ 5290 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt); 5291 5292 /* If window is closed, drop tail of packet. But after 5293 * remembering D-SACK for its head made in previous line. 5294 */ 5295 if (!tcp_receive_window(tp)) { 5296 reason = SKB_DROP_REASON_TCP_ZEROWINDOW; 5297 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP); 5298 goto out_of_window; 5299 } 5300 goto queue_and_out; 5301 } 5302 5303 tcp_data_queue_ofo(sk, skb); 5304 } 5305 5306 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list) 5307 { 5308 if (list) 5309 return !skb_queue_is_last(list, skb) ? skb->next : NULL; 5310 5311 return skb_rb_next(skb); 5312 } 5313 5314 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb, 5315 struct sk_buff_head *list, 5316 struct rb_root *root) 5317 { 5318 struct sk_buff *next = tcp_skb_next(skb, list); 5319 5320 if (list) 5321 __skb_unlink(skb, list); 5322 else 5323 rb_erase(&skb->rbnode, root); 5324 5325 __kfree_skb(skb); 5326 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED); 5327 5328 return next; 5329 } 5330 5331 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */ 5332 void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb) 5333 { 5334 struct rb_node **p = &root->rb_node; 5335 struct rb_node *parent = NULL; 5336 struct sk_buff *skb1; 5337 5338 while (*p) { 5339 parent = *p; 5340 skb1 = rb_to_skb(parent); 5341 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq)) 5342 p = &parent->rb_left; 5343 else 5344 p = &parent->rb_right; 5345 } 5346 rb_link_node(&skb->rbnode, parent, p); 5347 rb_insert_color(&skb->rbnode, root); 5348 } 5349 5350 /* Collapse contiguous sequence of skbs head..tail with 5351 * sequence numbers start..end. 5352 * 5353 * If tail is NULL, this means until the end of the queue. 5354 * 5355 * Segments with FIN/SYN are not collapsed (only because this 5356 * simplifies code) 5357 */ 5358 static void 5359 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root, 5360 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end) 5361 { 5362 struct sk_buff *skb = head, *n; 5363 struct sk_buff_head tmp; 5364 bool end_of_skbs; 5365 5366 /* First, check that queue is collapsible and find 5367 * the point where collapsing can be useful. 5368 */ 5369 restart: 5370 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) { 5371 n = tcp_skb_next(skb, list); 5372 5373 /* No new bits? It is possible on ofo queue. */ 5374 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 5375 skb = tcp_collapse_one(sk, skb, list, root); 5376 if (!skb) 5377 break; 5378 goto restart; 5379 } 5380 5381 /* The first skb to collapse is: 5382 * - not SYN/FIN and 5383 * - bloated or contains data before "start" or 5384 * overlaps to the next one and mptcp allow collapsing. 5385 */ 5386 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) && 5387 (tcp_win_from_space(sk, skb->truesize) > skb->len || 5388 before(TCP_SKB_CB(skb)->seq, start))) { 5389 end_of_skbs = false; 5390 break; 5391 } 5392 5393 if (n && n != tail && tcp_skb_can_collapse_rx(skb, n) && 5394 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) { 5395 end_of_skbs = false; 5396 break; 5397 } 5398 5399 /* Decided to skip this, advance start seq. */ 5400 start = TCP_SKB_CB(skb)->end_seq; 5401 } 5402 if (end_of_skbs || 5403 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN))) 5404 return; 5405 5406 __skb_queue_head_init(&tmp); 5407 5408 while (before(start, end)) { 5409 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start); 5410 struct sk_buff *nskb; 5411 5412 nskb = alloc_skb(copy, GFP_ATOMIC); 5413 if (!nskb) 5414 break; 5415 5416 memcpy(nskb->cb, skb->cb, sizeof(skb->cb)); 5417 skb_copy_decrypted(nskb, skb); 5418 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start; 5419 if (list) 5420 __skb_queue_before(list, skb, nskb); 5421 else 5422 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */ 5423 skb_set_owner_r(nskb, sk); 5424 mptcp_skb_ext_move(nskb, skb); 5425 5426 /* Copy data, releasing collapsed skbs. */ 5427 while (copy > 0) { 5428 int offset = start - TCP_SKB_CB(skb)->seq; 5429 int size = TCP_SKB_CB(skb)->end_seq - start; 5430 5431 BUG_ON(offset < 0); 5432 if (size > 0) { 5433 size = min(copy, size); 5434 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size)) 5435 BUG(); 5436 TCP_SKB_CB(nskb)->end_seq += size; 5437 copy -= size; 5438 start += size; 5439 } 5440 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 5441 skb = tcp_collapse_one(sk, skb, list, root); 5442 if (!skb || 5443 skb == tail || 5444 !tcp_skb_can_collapse_rx(nskb, skb) || 5445 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN))) 5446 goto end; 5447 } 5448 } 5449 } 5450 end: 5451 skb_queue_walk_safe(&tmp, skb, n) 5452 tcp_rbtree_insert(root, skb); 5453 } 5454 5455 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs 5456 * and tcp_collapse() them until all the queue is collapsed. 5457 */ 5458 static void tcp_collapse_ofo_queue(struct sock *sk) 5459 { 5460 struct tcp_sock *tp = tcp_sk(sk); 5461 u32 range_truesize, sum_tiny = 0; 5462 struct sk_buff *skb, *head; 5463 u32 start, end; 5464 5465 skb = skb_rb_first(&tp->out_of_order_queue); 5466 new_range: 5467 if (!skb) { 5468 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue); 5469 return; 5470 } 5471 start = TCP_SKB_CB(skb)->seq; 5472 end = TCP_SKB_CB(skb)->end_seq; 5473 range_truesize = skb->truesize; 5474 5475 for (head = skb;;) { 5476 skb = skb_rb_next(skb); 5477 5478 /* Range is terminated when we see a gap or when 5479 * we are at the queue end. 5480 */ 5481 if (!skb || 5482 after(TCP_SKB_CB(skb)->seq, end) || 5483 before(TCP_SKB_CB(skb)->end_seq, start)) { 5484 /* Do not attempt collapsing tiny skbs */ 5485 if (range_truesize != head->truesize || 5486 end - start >= SKB_WITH_OVERHEAD(PAGE_SIZE)) { 5487 tcp_collapse(sk, NULL, &tp->out_of_order_queue, 5488 head, skb, start, end); 5489 } else { 5490 sum_tiny += range_truesize; 5491 if (sum_tiny > sk->sk_rcvbuf >> 3) 5492 return; 5493 } 5494 goto new_range; 5495 } 5496 5497 range_truesize += skb->truesize; 5498 if (unlikely(before(TCP_SKB_CB(skb)->seq, start))) 5499 start = TCP_SKB_CB(skb)->seq; 5500 if (after(TCP_SKB_CB(skb)->end_seq, end)) 5501 end = TCP_SKB_CB(skb)->end_seq; 5502 } 5503 } 5504 5505 /* 5506 * Clean the out-of-order queue to make room. 5507 * We drop high sequences packets to : 5508 * 1) Let a chance for holes to be filled. 5509 * This means we do not drop packets from ooo queue if their sequence 5510 * is before incoming packet sequence. 5511 * 2) not add too big latencies if thousands of packets sit there. 5512 * (But if application shrinks SO_RCVBUF, we could still end up 5513 * freeing whole queue here) 5514 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks. 5515 * 5516 * Return true if queue has shrunk. 5517 */ 5518 static bool tcp_prune_ofo_queue(struct sock *sk, const struct sk_buff *in_skb) 5519 { 5520 struct tcp_sock *tp = tcp_sk(sk); 5521 struct rb_node *node, *prev; 5522 bool pruned = false; 5523 int goal; 5524 5525 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) 5526 return false; 5527 5528 goal = sk->sk_rcvbuf >> 3; 5529 node = &tp->ooo_last_skb->rbnode; 5530 5531 do { 5532 struct sk_buff *skb = rb_to_skb(node); 5533 5534 /* If incoming skb would land last in ofo queue, stop pruning. */ 5535 if (after(TCP_SKB_CB(in_skb)->seq, TCP_SKB_CB(skb)->seq)) 5536 break; 5537 pruned = true; 5538 prev = rb_prev(node); 5539 rb_erase(node, &tp->out_of_order_queue); 5540 goal -= skb->truesize; 5541 tcp_drop_reason(sk, skb, SKB_DROP_REASON_TCP_OFO_QUEUE_PRUNE); 5542 tp->ooo_last_skb = rb_to_skb(prev); 5543 if (!prev || goal <= 0) { 5544 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf && 5545 !tcp_under_memory_pressure(sk)) 5546 break; 5547 goal = sk->sk_rcvbuf >> 3; 5548 } 5549 node = prev; 5550 } while (node); 5551 5552 if (pruned) { 5553 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED); 5554 /* Reset SACK state. A conforming SACK implementation will 5555 * do the same at a timeout based retransmit. When a connection 5556 * is in a sad state like this, we care only about integrity 5557 * of the connection not performance. 5558 */ 5559 if (tp->rx_opt.sack_ok) 5560 tcp_sack_reset(&tp->rx_opt); 5561 } 5562 return pruned; 5563 } 5564 5565 /* Reduce allocated memory if we can, trying to get 5566 * the socket within its memory limits again. 5567 * 5568 * Return less than zero if we should start dropping frames 5569 * until the socket owning process reads some of the data 5570 * to stabilize the situation. 5571 */ 5572 static int tcp_prune_queue(struct sock *sk, const struct sk_buff *in_skb) 5573 { 5574 struct tcp_sock *tp = tcp_sk(sk); 5575 5576 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED); 5577 5578 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf) 5579 tcp_clamp_window(sk); 5580 else if (tcp_under_memory_pressure(sk)) 5581 tcp_adjust_rcv_ssthresh(sk); 5582 5583 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 5584 return 0; 5585 5586 tcp_collapse_ofo_queue(sk); 5587 if (!skb_queue_empty(&sk->sk_receive_queue)) 5588 tcp_collapse(sk, &sk->sk_receive_queue, NULL, 5589 skb_peek(&sk->sk_receive_queue), 5590 NULL, 5591 tp->copied_seq, tp->rcv_nxt); 5592 5593 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 5594 return 0; 5595 5596 /* Collapsing did not help, destructive actions follow. 5597 * This must not ever occur. */ 5598 5599 tcp_prune_ofo_queue(sk, in_skb); 5600 5601 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 5602 return 0; 5603 5604 /* If we are really being abused, tell the caller to silently 5605 * drop receive data on the floor. It will get retransmitted 5606 * and hopefully then we'll have sufficient space. 5607 */ 5608 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED); 5609 5610 /* Massive buffer overcommit. */ 5611 tp->pred_flags = 0; 5612 return -1; 5613 } 5614 5615 static bool tcp_should_expand_sndbuf(struct sock *sk) 5616 { 5617 const struct tcp_sock *tp = tcp_sk(sk); 5618 5619 /* If the user specified a specific send buffer setting, do 5620 * not modify it. 5621 */ 5622 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK) 5623 return false; 5624 5625 /* If we are under global TCP memory pressure, do not expand. */ 5626 if (tcp_under_memory_pressure(sk)) { 5627 int unused_mem = sk_unused_reserved_mem(sk); 5628 5629 /* Adjust sndbuf according to reserved mem. But make sure 5630 * it never goes below SOCK_MIN_SNDBUF. 5631 * See sk_stream_moderate_sndbuf() for more details. 5632 */ 5633 if (unused_mem > SOCK_MIN_SNDBUF) 5634 WRITE_ONCE(sk->sk_sndbuf, unused_mem); 5635 5636 return false; 5637 } 5638 5639 /* If we are under soft global TCP memory pressure, do not expand. */ 5640 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0)) 5641 return false; 5642 5643 /* If we filled the congestion window, do not expand. */ 5644 if (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp)) 5645 return false; 5646 5647 return true; 5648 } 5649 5650 static void tcp_new_space(struct sock *sk) 5651 { 5652 struct tcp_sock *tp = tcp_sk(sk); 5653 5654 if (tcp_should_expand_sndbuf(sk)) { 5655 tcp_sndbuf_expand(sk); 5656 tp->snd_cwnd_stamp = tcp_jiffies32; 5657 } 5658 5659 INDIRECT_CALL_1(sk->sk_write_space, sk_stream_write_space, sk); 5660 } 5661 5662 /* Caller made space either from: 5663 * 1) Freeing skbs in rtx queues (after tp->snd_una has advanced) 5664 * 2) Sent skbs from output queue (and thus advancing tp->snd_nxt) 5665 * 5666 * We might be able to generate EPOLLOUT to the application if: 5667 * 1) Space consumed in output/rtx queues is below sk->sk_sndbuf/2 5668 * 2) notsent amount (tp->write_seq - tp->snd_nxt) became 5669 * small enough that tcp_stream_memory_free() decides it 5670 * is time to generate EPOLLOUT. 5671 */ 5672 void tcp_check_space(struct sock *sk) 5673 { 5674 /* pairs with tcp_poll() */ 5675 smp_mb(); 5676 if (sk->sk_socket && 5677 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 5678 tcp_new_space(sk); 5679 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) 5680 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED); 5681 } 5682 } 5683 5684 static inline void tcp_data_snd_check(struct sock *sk) 5685 { 5686 tcp_push_pending_frames(sk); 5687 tcp_check_space(sk); 5688 } 5689 5690 /* 5691 * Check if sending an ack is needed. 5692 */ 5693 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible) 5694 { 5695 struct tcp_sock *tp = tcp_sk(sk); 5696 unsigned long rtt, delay; 5697 5698 /* More than one full frame received... */ 5699 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss && 5700 /* ... and right edge of window advances far enough. 5701 * (tcp_recvmsg() will send ACK otherwise). 5702 * If application uses SO_RCVLOWAT, we want send ack now if 5703 * we have not received enough bytes to satisfy the condition. 5704 */ 5705 (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat || 5706 __tcp_select_window(sk) >= tp->rcv_wnd)) || 5707 /* We ACK each frame or... */ 5708 tcp_in_quickack_mode(sk) || 5709 /* Protocol state mandates a one-time immediate ACK */ 5710 inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) { 5711 /* If we are running from __release_sock() in user context, 5712 * Defer the ack until tcp_release_cb(). 5713 */ 5714 if (sock_owned_by_user_nocheck(sk) && 5715 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_backlog_ack_defer)) { 5716 set_bit(TCP_ACK_DEFERRED, &sk->sk_tsq_flags); 5717 return; 5718 } 5719 send_now: 5720 tcp_send_ack(sk); 5721 return; 5722 } 5723 5724 if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) { 5725 tcp_send_delayed_ack(sk); 5726 return; 5727 } 5728 5729 if (!tcp_is_sack(tp) || 5730 tp->compressed_ack >= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr)) 5731 goto send_now; 5732 5733 if (tp->compressed_ack_rcv_nxt != tp->rcv_nxt) { 5734 tp->compressed_ack_rcv_nxt = tp->rcv_nxt; 5735 tp->dup_ack_counter = 0; 5736 } 5737 if (tp->dup_ack_counter < TCP_FASTRETRANS_THRESH) { 5738 tp->dup_ack_counter++; 5739 goto send_now; 5740 } 5741 tp->compressed_ack++; 5742 if (hrtimer_is_queued(&tp->compressed_ack_timer)) 5743 return; 5744 5745 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */ 5746 5747 rtt = tp->rcv_rtt_est.rtt_us; 5748 if (tp->srtt_us && tp->srtt_us < rtt) 5749 rtt = tp->srtt_us; 5750 5751 delay = min_t(unsigned long, 5752 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns), 5753 rtt * (NSEC_PER_USEC >> 3)/20); 5754 sock_hold(sk); 5755 hrtimer_start_range_ns(&tp->compressed_ack_timer, ns_to_ktime(delay), 5756 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_slack_ns), 5757 HRTIMER_MODE_REL_PINNED_SOFT); 5758 } 5759 5760 static inline void tcp_ack_snd_check(struct sock *sk) 5761 { 5762 if (!inet_csk_ack_scheduled(sk)) { 5763 /* We sent a data segment already. */ 5764 return; 5765 } 5766 __tcp_ack_snd_check(sk, 1); 5767 } 5768 5769 /* 5770 * This routine is only called when we have urgent data 5771 * signaled. Its the 'slow' part of tcp_urg. It could be 5772 * moved inline now as tcp_urg is only called from one 5773 * place. We handle URGent data wrong. We have to - as 5774 * BSD still doesn't use the correction from RFC961. 5775 * For 1003.1g we should support a new option TCP_STDURG to permit 5776 * either form (or just set the sysctl tcp_stdurg). 5777 */ 5778 5779 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th) 5780 { 5781 struct tcp_sock *tp = tcp_sk(sk); 5782 u32 ptr = ntohs(th->urg_ptr); 5783 5784 if (ptr && !READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_stdurg)) 5785 ptr--; 5786 ptr += ntohl(th->seq); 5787 5788 /* Ignore urgent data that we've already seen and read. */ 5789 if (after(tp->copied_seq, ptr)) 5790 return; 5791 5792 /* Do not replay urg ptr. 5793 * 5794 * NOTE: interesting situation not covered by specs. 5795 * Misbehaving sender may send urg ptr, pointing to segment, 5796 * which we already have in ofo queue. We are not able to fetch 5797 * such data and will stay in TCP_URG_NOTYET until will be eaten 5798 * by recvmsg(). Seems, we are not obliged to handle such wicked 5799 * situations. But it is worth to think about possibility of some 5800 * DoSes using some hypothetical application level deadlock. 5801 */ 5802 if (before(ptr, tp->rcv_nxt)) 5803 return; 5804 5805 /* Do we already have a newer (or duplicate) urgent pointer? */ 5806 if (tp->urg_data && !after(ptr, tp->urg_seq)) 5807 return; 5808 5809 /* Tell the world about our new urgent pointer. */ 5810 sk_send_sigurg(sk); 5811 5812 /* We may be adding urgent data when the last byte read was 5813 * urgent. To do this requires some care. We cannot just ignore 5814 * tp->copied_seq since we would read the last urgent byte again 5815 * as data, nor can we alter copied_seq until this data arrives 5816 * or we break the semantics of SIOCATMARK (and thus sockatmark()) 5817 * 5818 * NOTE. Double Dutch. Rendering to plain English: author of comment 5819 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB); 5820 * and expect that both A and B disappear from stream. This is _wrong_. 5821 * Though this happens in BSD with high probability, this is occasional. 5822 * Any application relying on this is buggy. Note also, that fix "works" 5823 * only in this artificial test. Insert some normal data between A and B and we will 5824 * decline of BSD again. Verdict: it is better to remove to trap 5825 * buggy users. 5826 */ 5827 if (tp->urg_seq == tp->copied_seq && tp->urg_data && 5828 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) { 5829 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 5830 tp->copied_seq++; 5831 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) { 5832 __skb_unlink(skb, &sk->sk_receive_queue); 5833 __kfree_skb(skb); 5834 } 5835 } 5836 5837 WRITE_ONCE(tp->urg_data, TCP_URG_NOTYET); 5838 WRITE_ONCE(tp->urg_seq, ptr); 5839 5840 /* Disable header prediction. */ 5841 tp->pred_flags = 0; 5842 } 5843 5844 /* This is the 'fast' part of urgent handling. */ 5845 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th) 5846 { 5847 struct tcp_sock *tp = tcp_sk(sk); 5848 5849 /* Check if we get a new urgent pointer - normally not. */ 5850 if (unlikely(th->urg)) 5851 tcp_check_urg(sk, th); 5852 5853 /* Do we wait for any urgent data? - normally not... */ 5854 if (unlikely(tp->urg_data == TCP_URG_NOTYET)) { 5855 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) - 5856 th->syn; 5857 5858 /* Is the urgent pointer pointing into this packet? */ 5859 if (ptr < skb->len) { 5860 u8 tmp; 5861 if (skb_copy_bits(skb, ptr, &tmp, 1)) 5862 BUG(); 5863 WRITE_ONCE(tp->urg_data, TCP_URG_VALID | tmp); 5864 if (!sock_flag(sk, SOCK_DEAD)) 5865 sk->sk_data_ready(sk); 5866 } 5867 } 5868 } 5869 5870 /* Accept RST for rcv_nxt - 1 after a FIN. 5871 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a 5872 * FIN is sent followed by a RST packet. The RST is sent with the same 5873 * sequence number as the FIN, and thus according to RFC 5961 a challenge 5874 * ACK should be sent. However, Mac OSX rate limits replies to challenge 5875 * ACKs on the closed socket. In addition middleboxes can drop either the 5876 * challenge ACK or a subsequent RST. 5877 */ 5878 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb) 5879 { 5880 const struct tcp_sock *tp = tcp_sk(sk); 5881 5882 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) && 5883 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK | 5884 TCPF_CLOSING)); 5885 } 5886 5887 /* Does PAWS and seqno based validation of an incoming segment, flags will 5888 * play significant role here. 5889 */ 5890 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb, 5891 const struct tcphdr *th, int syn_inerr) 5892 { 5893 struct tcp_sock *tp = tcp_sk(sk); 5894 SKB_DR(reason); 5895 5896 /* RFC1323: H1. Apply PAWS check first. */ 5897 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) && 5898 tp->rx_opt.saw_tstamp && 5899 tcp_paws_discard(sk, skb)) { 5900 if (!th->rst) { 5901 if (unlikely(th->syn)) 5902 goto syn_challenge; 5903 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED); 5904 if (!tcp_oow_rate_limited(sock_net(sk), skb, 5905 LINUX_MIB_TCPACKSKIPPEDPAWS, 5906 &tp->last_oow_ack_time)) 5907 tcp_send_dupack(sk, skb); 5908 SKB_DR_SET(reason, TCP_RFC7323_PAWS); 5909 goto discard; 5910 } 5911 /* Reset is accepted even if it did not pass PAWS. */ 5912 } 5913 5914 /* Step 1: check sequence number */ 5915 reason = tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 5916 if (reason) { 5917 /* RFC793, page 37: "In all states except SYN-SENT, all reset 5918 * (RST) segments are validated by checking their SEQ-fields." 5919 * And page 69: "If an incoming segment is not acceptable, 5920 * an acknowledgment should be sent in reply (unless the RST 5921 * bit is set, if so drop the segment and return)". 5922 */ 5923 if (!th->rst) { 5924 if (th->syn) 5925 goto syn_challenge; 5926 if (!tcp_oow_rate_limited(sock_net(sk), skb, 5927 LINUX_MIB_TCPACKSKIPPEDSEQ, 5928 &tp->last_oow_ack_time)) 5929 tcp_send_dupack(sk, skb); 5930 } else if (tcp_reset_check(sk, skb)) { 5931 goto reset; 5932 } 5933 goto discard; 5934 } 5935 5936 /* Step 2: check RST bit */ 5937 if (th->rst) { 5938 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a 5939 * FIN and SACK too if available): 5940 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or 5941 * the right-most SACK block, 5942 * then 5943 * RESET the connection 5944 * else 5945 * Send a challenge ACK 5946 */ 5947 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt || 5948 tcp_reset_check(sk, skb)) 5949 goto reset; 5950 5951 if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) { 5952 struct tcp_sack_block *sp = &tp->selective_acks[0]; 5953 int max_sack = sp[0].end_seq; 5954 int this_sack; 5955 5956 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; 5957 ++this_sack) { 5958 max_sack = after(sp[this_sack].end_seq, 5959 max_sack) ? 5960 sp[this_sack].end_seq : max_sack; 5961 } 5962 5963 if (TCP_SKB_CB(skb)->seq == max_sack) 5964 goto reset; 5965 } 5966 5967 /* Disable TFO if RST is out-of-order 5968 * and no data has been received 5969 * for current active TFO socket 5970 */ 5971 if (tp->syn_fastopen && !tp->data_segs_in && 5972 sk->sk_state == TCP_ESTABLISHED) 5973 tcp_fastopen_active_disable(sk); 5974 tcp_send_challenge_ack(sk); 5975 SKB_DR_SET(reason, TCP_RESET); 5976 goto discard; 5977 } 5978 5979 /* step 3: check security and precedence [ignored] */ 5980 5981 /* step 4: Check for a SYN 5982 * RFC 5961 4.2 : Send a challenge ack 5983 */ 5984 if (th->syn) { 5985 syn_challenge: 5986 if (syn_inerr) 5987 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS); 5988 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE); 5989 tcp_send_challenge_ack(sk); 5990 SKB_DR_SET(reason, TCP_INVALID_SYN); 5991 goto discard; 5992 } 5993 5994 bpf_skops_parse_hdr(sk, skb); 5995 5996 return true; 5997 5998 discard: 5999 tcp_drop_reason(sk, skb, reason); 6000 return false; 6001 6002 reset: 6003 tcp_reset(sk, skb); 6004 __kfree_skb(skb); 6005 return false; 6006 } 6007 6008 /* 6009 * TCP receive function for the ESTABLISHED state. 6010 * 6011 * It is split into a fast path and a slow path. The fast path is 6012 * disabled when: 6013 * - A zero window was announced from us - zero window probing 6014 * is only handled properly in the slow path. 6015 * - Out of order segments arrived. 6016 * - Urgent data is expected. 6017 * - There is no buffer space left 6018 * - Unexpected TCP flags/window values/header lengths are received 6019 * (detected by checking the TCP header against pred_flags) 6020 * - Data is sent in both directions. Fast path only supports pure senders 6021 * or pure receivers (this means either the sequence number or the ack 6022 * value must stay constant) 6023 * - Unexpected TCP option. 6024 * 6025 * When these conditions are not satisfied it drops into a standard 6026 * receive procedure patterned after RFC793 to handle all cases. 6027 * The first three cases are guaranteed by proper pred_flags setting, 6028 * the rest is checked inline. Fast processing is turned on in 6029 * tcp_data_queue when everything is OK. 6030 */ 6031 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb) 6032 { 6033 enum skb_drop_reason reason = SKB_DROP_REASON_NOT_SPECIFIED; 6034 const struct tcphdr *th = (const struct tcphdr *)skb->data; 6035 struct tcp_sock *tp = tcp_sk(sk); 6036 unsigned int len = skb->len; 6037 6038 /* TCP congestion window tracking */ 6039 trace_tcp_probe(sk, skb); 6040 6041 tcp_mstamp_refresh(tp); 6042 if (unlikely(!rcu_access_pointer(sk->sk_rx_dst))) 6043 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb); 6044 /* 6045 * Header prediction. 6046 * The code loosely follows the one in the famous 6047 * "30 instruction TCP receive" Van Jacobson mail. 6048 * 6049 * Van's trick is to deposit buffers into socket queue 6050 * on a device interrupt, to call tcp_recv function 6051 * on the receive process context and checksum and copy 6052 * the buffer to user space. smart... 6053 * 6054 * Our current scheme is not silly either but we take the 6055 * extra cost of the net_bh soft interrupt processing... 6056 * We do checksum and copy also but from device to kernel. 6057 */ 6058 6059 tp->rx_opt.saw_tstamp = 0; 6060 6061 /* pred_flags is 0xS?10 << 16 + snd_wnd 6062 * if header_prediction is to be made 6063 * 'S' will always be tp->tcp_header_len >> 2 6064 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to 6065 * turn it off (when there are holes in the receive 6066 * space for instance) 6067 * PSH flag is ignored. 6068 */ 6069 6070 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags && 6071 TCP_SKB_CB(skb)->seq == tp->rcv_nxt && 6072 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) { 6073 int tcp_header_len = tp->tcp_header_len; 6074 6075 /* Timestamp header prediction: tcp_header_len 6076 * is automatically equal to th->doff*4 due to pred_flags 6077 * match. 6078 */ 6079 6080 /* Check timestamp */ 6081 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) { 6082 /* No? Slow path! */ 6083 if (!tcp_parse_aligned_timestamp(tp, th)) 6084 goto slow_path; 6085 6086 /* If PAWS failed, check it more carefully in slow path */ 6087 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0) 6088 goto slow_path; 6089 6090 /* DO NOT update ts_recent here, if checksum fails 6091 * and timestamp was corrupted part, it will result 6092 * in a hung connection since we will drop all 6093 * future packets due to the PAWS test. 6094 */ 6095 } 6096 6097 if (len <= tcp_header_len) { 6098 /* Bulk data transfer: sender */ 6099 if (len == tcp_header_len) { 6100 /* Predicted packet is in window by definition. 6101 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 6102 * Hence, check seq<=rcv_wup reduces to: 6103 */ 6104 if (tcp_header_len == 6105 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 6106 tp->rcv_nxt == tp->rcv_wup) 6107 tcp_store_ts_recent(tp); 6108 6109 /* We know that such packets are checksummed 6110 * on entry. 6111 */ 6112 tcp_ack(sk, skb, 0); 6113 __kfree_skb(skb); 6114 tcp_data_snd_check(sk); 6115 /* When receiving pure ack in fast path, update 6116 * last ts ecr directly instead of calling 6117 * tcp_rcv_rtt_measure_ts() 6118 */ 6119 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr; 6120 return; 6121 } else { /* Header too small */ 6122 reason = SKB_DROP_REASON_PKT_TOO_SMALL; 6123 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS); 6124 goto discard; 6125 } 6126 } else { 6127 int eaten = 0; 6128 bool fragstolen = false; 6129 6130 if (tcp_checksum_complete(skb)) 6131 goto csum_error; 6132 6133 if ((int)skb->truesize > sk->sk_forward_alloc) 6134 goto step5; 6135 6136 /* Predicted packet is in window by definition. 6137 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 6138 * Hence, check seq<=rcv_wup reduces to: 6139 */ 6140 if (tcp_header_len == 6141 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 6142 tp->rcv_nxt == tp->rcv_wup) 6143 tcp_store_ts_recent(tp); 6144 6145 tcp_rcv_rtt_measure_ts(sk, skb); 6146 6147 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS); 6148 6149 /* Bulk data transfer: receiver */ 6150 skb_dst_drop(skb); 6151 __skb_pull(skb, tcp_header_len); 6152 eaten = tcp_queue_rcv(sk, skb, &fragstolen); 6153 6154 tcp_event_data_recv(sk, skb); 6155 6156 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) { 6157 /* Well, only one small jumplet in fast path... */ 6158 tcp_ack(sk, skb, FLAG_DATA); 6159 tcp_data_snd_check(sk); 6160 if (!inet_csk_ack_scheduled(sk)) 6161 goto no_ack; 6162 } else { 6163 tcp_update_wl(tp, TCP_SKB_CB(skb)->seq); 6164 } 6165 6166 __tcp_ack_snd_check(sk, 0); 6167 no_ack: 6168 if (eaten) 6169 kfree_skb_partial(skb, fragstolen); 6170 tcp_data_ready(sk); 6171 return; 6172 } 6173 } 6174 6175 slow_path: 6176 if (len < (th->doff << 2) || tcp_checksum_complete(skb)) 6177 goto csum_error; 6178 6179 if (!th->ack && !th->rst && !th->syn) { 6180 reason = SKB_DROP_REASON_TCP_FLAGS; 6181 goto discard; 6182 } 6183 6184 /* 6185 * Standard slow path. 6186 */ 6187 6188 if (!tcp_validate_incoming(sk, skb, th, 1)) 6189 return; 6190 6191 step5: 6192 reason = tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT); 6193 if ((int)reason < 0) { 6194 reason = -reason; 6195 goto discard; 6196 } 6197 tcp_rcv_rtt_measure_ts(sk, skb); 6198 6199 /* Process urgent data. */ 6200 tcp_urg(sk, skb, th); 6201 6202 /* step 7: process the segment text */ 6203 tcp_data_queue(sk, skb); 6204 6205 tcp_data_snd_check(sk); 6206 tcp_ack_snd_check(sk); 6207 return; 6208 6209 csum_error: 6210 reason = SKB_DROP_REASON_TCP_CSUM; 6211 trace_tcp_bad_csum(skb); 6212 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS); 6213 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS); 6214 6215 discard: 6216 tcp_drop_reason(sk, skb, reason); 6217 } 6218 EXPORT_SYMBOL(tcp_rcv_established); 6219 6220 void tcp_init_transfer(struct sock *sk, int bpf_op, struct sk_buff *skb) 6221 { 6222 struct inet_connection_sock *icsk = inet_csk(sk); 6223 struct tcp_sock *tp = tcp_sk(sk); 6224 6225 tcp_mtup_init(sk); 6226 icsk->icsk_af_ops->rebuild_header(sk); 6227 tcp_init_metrics(sk); 6228 6229 /* Initialize the congestion window to start the transfer. 6230 * Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been 6231 * retransmitted. In light of RFC6298 more aggressive 1sec 6232 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK 6233 * retransmission has occurred. 6234 */ 6235 if (tp->total_retrans > 1 && tp->undo_marker) 6236 tcp_snd_cwnd_set(tp, 1); 6237 else 6238 tcp_snd_cwnd_set(tp, tcp_init_cwnd(tp, __sk_dst_get(sk))); 6239 tp->snd_cwnd_stamp = tcp_jiffies32; 6240 6241 bpf_skops_established(sk, bpf_op, skb); 6242 /* Initialize congestion control unless BPF initialized it already: */ 6243 if (!icsk->icsk_ca_initialized) 6244 tcp_init_congestion_control(sk); 6245 tcp_init_buffer_space(sk); 6246 } 6247 6248 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb) 6249 { 6250 struct tcp_sock *tp = tcp_sk(sk); 6251 struct inet_connection_sock *icsk = inet_csk(sk); 6252 6253 tcp_ao_finish_connect(sk, skb); 6254 tcp_set_state(sk, TCP_ESTABLISHED); 6255 icsk->icsk_ack.lrcvtime = tcp_jiffies32; 6256 6257 if (skb) { 6258 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb); 6259 security_inet_conn_established(sk, skb); 6260 sk_mark_napi_id(sk, skb); 6261 } 6262 6263 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB, skb); 6264 6265 /* Prevent spurious tcp_cwnd_restart() on first data 6266 * packet. 6267 */ 6268 tp->lsndtime = tcp_jiffies32; 6269 6270 if (sock_flag(sk, SOCK_KEEPOPEN)) 6271 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp)); 6272 6273 if (!tp->rx_opt.snd_wscale) 6274 __tcp_fast_path_on(tp, tp->snd_wnd); 6275 else 6276 tp->pred_flags = 0; 6277 } 6278 6279 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack, 6280 struct tcp_fastopen_cookie *cookie) 6281 { 6282 struct tcp_sock *tp = tcp_sk(sk); 6283 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL; 6284 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0; 6285 bool syn_drop = false; 6286 6287 if (mss == tp->rx_opt.user_mss) { 6288 struct tcp_options_received opt; 6289 6290 /* Get original SYNACK MSS value if user MSS sets mss_clamp */ 6291 tcp_clear_options(&opt); 6292 opt.user_mss = opt.mss_clamp = 0; 6293 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL); 6294 mss = opt.mss_clamp; 6295 } 6296 6297 if (!tp->syn_fastopen) { 6298 /* Ignore an unsolicited cookie */ 6299 cookie->len = -1; 6300 } else if (tp->total_retrans) { 6301 /* SYN timed out and the SYN-ACK neither has a cookie nor 6302 * acknowledges data. Presumably the remote received only 6303 * the retransmitted (regular) SYNs: either the original 6304 * SYN-data or the corresponding SYN-ACK was dropped. 6305 */ 6306 syn_drop = (cookie->len < 0 && data); 6307 } else if (cookie->len < 0 && !tp->syn_data) { 6308 /* We requested a cookie but didn't get it. If we did not use 6309 * the (old) exp opt format then try so next time (try_exp=1). 6310 * Otherwise we go back to use the RFC7413 opt (try_exp=2). 6311 */ 6312 try_exp = tp->syn_fastopen_exp ? 2 : 1; 6313 } 6314 6315 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp); 6316 6317 if (data) { /* Retransmit unacked data in SYN */ 6318 if (tp->total_retrans) 6319 tp->fastopen_client_fail = TFO_SYN_RETRANSMITTED; 6320 else 6321 tp->fastopen_client_fail = TFO_DATA_NOT_ACKED; 6322 skb_rbtree_walk_from(data) 6323 tcp_mark_skb_lost(sk, data); 6324 tcp_non_congestion_loss_retransmit(sk); 6325 NET_INC_STATS(sock_net(sk), 6326 LINUX_MIB_TCPFASTOPENACTIVEFAIL); 6327 return true; 6328 } 6329 tp->syn_data_acked = tp->syn_data; 6330 if (tp->syn_data_acked) { 6331 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE); 6332 /* SYN-data is counted as two separate packets in tcp_ack() */ 6333 if (tp->delivered > 1) 6334 --tp->delivered; 6335 } 6336 6337 tcp_fastopen_add_skb(sk, synack); 6338 6339 return false; 6340 } 6341 6342 static void smc_check_reset_syn(struct tcp_sock *tp) 6343 { 6344 #if IS_ENABLED(CONFIG_SMC) 6345 if (static_branch_unlikely(&tcp_have_smc)) { 6346 if (tp->syn_smc && !tp->rx_opt.smc_ok) 6347 tp->syn_smc = 0; 6348 } 6349 #endif 6350 } 6351 6352 static void tcp_try_undo_spurious_syn(struct sock *sk) 6353 { 6354 struct tcp_sock *tp = tcp_sk(sk); 6355 u32 syn_stamp; 6356 6357 /* undo_marker is set when SYN or SYNACK times out. The timeout is 6358 * spurious if the ACK's timestamp option echo value matches the 6359 * original SYN timestamp. 6360 */ 6361 syn_stamp = tp->retrans_stamp; 6362 if (tp->undo_marker && syn_stamp && tp->rx_opt.saw_tstamp && 6363 syn_stamp == tp->rx_opt.rcv_tsecr) 6364 tp->undo_marker = 0; 6365 } 6366 6367 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb, 6368 const struct tcphdr *th) 6369 { 6370 struct inet_connection_sock *icsk = inet_csk(sk); 6371 struct tcp_sock *tp = tcp_sk(sk); 6372 struct tcp_fastopen_cookie foc = { .len = -1 }; 6373 int saved_clamp = tp->rx_opt.mss_clamp; 6374 bool fastopen_fail; 6375 SKB_DR(reason); 6376 6377 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc); 6378 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) 6379 tp->rx_opt.rcv_tsecr -= tp->tsoffset; 6380 6381 if (th->ack) { 6382 /* rfc793: 6383 * "If the state is SYN-SENT then 6384 * first check the ACK bit 6385 * If the ACK bit is set 6386 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send 6387 * a reset (unless the RST bit is set, if so drop 6388 * the segment and return)" 6389 */ 6390 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) || 6391 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) { 6392 /* Previous FIN/ACK or RST/ACK might be ignored. */ 6393 if (icsk->icsk_retransmits == 0) 6394 inet_csk_reset_xmit_timer(sk, 6395 ICSK_TIME_RETRANS, 6396 TCP_TIMEOUT_MIN, TCP_RTO_MAX); 6397 SKB_DR_SET(reason, TCP_INVALID_ACK_SEQUENCE); 6398 goto reset_and_undo; 6399 } 6400 6401 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 6402 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp, 6403 tcp_time_stamp_ts(tp))) { 6404 NET_INC_STATS(sock_net(sk), 6405 LINUX_MIB_PAWSACTIVEREJECTED); 6406 SKB_DR_SET(reason, TCP_RFC7323_PAWS); 6407 goto reset_and_undo; 6408 } 6409 6410 /* Now ACK is acceptable. 6411 * 6412 * "If the RST bit is set 6413 * If the ACK was acceptable then signal the user "error: 6414 * connection reset", drop the segment, enter CLOSED state, 6415 * delete TCB, and return." 6416 */ 6417 6418 if (th->rst) { 6419 tcp_reset(sk, skb); 6420 consume: 6421 __kfree_skb(skb); 6422 return 0; 6423 } 6424 6425 /* rfc793: 6426 * "fifth, if neither of the SYN or RST bits is set then 6427 * drop the segment and return." 6428 * 6429 * See note below! 6430 * --ANK(990513) 6431 */ 6432 if (!th->syn) { 6433 SKB_DR_SET(reason, TCP_FLAGS); 6434 goto discard_and_undo; 6435 } 6436 /* rfc793: 6437 * "If the SYN bit is on ... 6438 * are acceptable then ... 6439 * (our SYN has been ACKed), change the connection 6440 * state to ESTABLISHED..." 6441 */ 6442 6443 tcp_ecn_rcv_synack(tp, th); 6444 6445 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); 6446 tcp_try_undo_spurious_syn(sk); 6447 tcp_ack(sk, skb, FLAG_SLOWPATH); 6448 6449 /* Ok.. it's good. Set up sequence numbers and 6450 * move to established. 6451 */ 6452 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1); 6453 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 6454 6455 /* RFC1323: The window in SYN & SYN/ACK segments is 6456 * never scaled. 6457 */ 6458 tp->snd_wnd = ntohs(th->window); 6459 6460 if (!tp->rx_opt.wscale_ok) { 6461 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0; 6462 WRITE_ONCE(tp->window_clamp, 6463 min(tp->window_clamp, 65535U)); 6464 } 6465 6466 if (tp->rx_opt.saw_tstamp) { 6467 tp->rx_opt.tstamp_ok = 1; 6468 tp->tcp_header_len = 6469 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 6470 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 6471 tcp_store_ts_recent(tp); 6472 } else { 6473 tp->tcp_header_len = sizeof(struct tcphdr); 6474 } 6475 6476 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 6477 tcp_initialize_rcv_mss(sk); 6478 6479 /* Remember, tcp_poll() does not lock socket! 6480 * Change state from SYN-SENT only after copied_seq 6481 * is initialized. */ 6482 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 6483 6484 smc_check_reset_syn(tp); 6485 6486 smp_mb(); 6487 6488 tcp_finish_connect(sk, skb); 6489 6490 fastopen_fail = (tp->syn_fastopen || tp->syn_data) && 6491 tcp_rcv_fastopen_synack(sk, skb, &foc); 6492 6493 if (!sock_flag(sk, SOCK_DEAD)) { 6494 sk->sk_state_change(sk); 6495 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT); 6496 } 6497 if (fastopen_fail) 6498 return -1; 6499 if (sk->sk_write_pending || 6500 READ_ONCE(icsk->icsk_accept_queue.rskq_defer_accept) || 6501 inet_csk_in_pingpong_mode(sk)) { 6502 /* Save one ACK. Data will be ready after 6503 * several ticks, if write_pending is set. 6504 * 6505 * It may be deleted, but with this feature tcpdumps 6506 * look so _wonderfully_ clever, that I was not able 6507 * to stand against the temptation 8) --ANK 6508 */ 6509 inet_csk_schedule_ack(sk); 6510 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS); 6511 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 6512 TCP_DELACK_MAX, TCP_RTO_MAX); 6513 goto consume; 6514 } 6515 tcp_send_ack(sk); 6516 return -1; 6517 } 6518 6519 /* No ACK in the segment */ 6520 6521 if (th->rst) { 6522 /* rfc793: 6523 * "If the RST bit is set 6524 * 6525 * Otherwise (no ACK) drop the segment and return." 6526 */ 6527 SKB_DR_SET(reason, TCP_RESET); 6528 goto discard_and_undo; 6529 } 6530 6531 /* PAWS check. */ 6532 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && 6533 tcp_paws_reject(&tp->rx_opt, 0)) { 6534 SKB_DR_SET(reason, TCP_RFC7323_PAWS); 6535 goto discard_and_undo; 6536 } 6537 if (th->syn) { 6538 /* We see SYN without ACK. It is attempt of 6539 * simultaneous connect with crossed SYNs. 6540 * Particularly, it can be connect to self. 6541 */ 6542 #ifdef CONFIG_TCP_AO 6543 struct tcp_ao_info *ao; 6544 6545 ao = rcu_dereference_protected(tp->ao_info, 6546 lockdep_sock_is_held(sk)); 6547 if (ao) { 6548 WRITE_ONCE(ao->risn, th->seq); 6549 ao->rcv_sne = 0; 6550 } 6551 #endif 6552 tcp_set_state(sk, TCP_SYN_RECV); 6553 6554 if (tp->rx_opt.saw_tstamp) { 6555 tp->rx_opt.tstamp_ok = 1; 6556 tcp_store_ts_recent(tp); 6557 tp->tcp_header_len = 6558 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 6559 } else { 6560 tp->tcp_header_len = sizeof(struct tcphdr); 6561 } 6562 6563 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1); 6564 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 6565 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 6566 6567 /* RFC1323: The window in SYN & SYN/ACK segments is 6568 * never scaled. 6569 */ 6570 tp->snd_wnd = ntohs(th->window); 6571 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 6572 tp->max_window = tp->snd_wnd; 6573 6574 tcp_ecn_rcv_syn(tp, th); 6575 6576 tcp_mtup_init(sk); 6577 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 6578 tcp_initialize_rcv_mss(sk); 6579 6580 tcp_send_synack(sk); 6581 #if 0 6582 /* Note, we could accept data and URG from this segment. 6583 * There are no obstacles to make this (except that we must 6584 * either change tcp_recvmsg() to prevent it from returning data 6585 * before 3WHS completes per RFC793, or employ TCP Fast Open). 6586 * 6587 * However, if we ignore data in ACKless segments sometimes, 6588 * we have no reasons to accept it sometimes. 6589 * Also, seems the code doing it in step6 of tcp_rcv_state_process 6590 * is not flawless. So, discard packet for sanity. 6591 * Uncomment this return to process the data. 6592 */ 6593 return -1; 6594 #else 6595 goto consume; 6596 #endif 6597 } 6598 /* "fifth, if neither of the SYN or RST bits is set then 6599 * drop the segment and return." 6600 */ 6601 6602 discard_and_undo: 6603 tcp_clear_options(&tp->rx_opt); 6604 tp->rx_opt.mss_clamp = saved_clamp; 6605 tcp_drop_reason(sk, skb, reason); 6606 return 0; 6607 6608 reset_and_undo: 6609 tcp_clear_options(&tp->rx_opt); 6610 tp->rx_opt.mss_clamp = saved_clamp; 6611 /* we can reuse/return @reason to its caller to handle the exception */ 6612 return reason; 6613 } 6614 6615 static void tcp_rcv_synrecv_state_fastopen(struct sock *sk) 6616 { 6617 struct tcp_sock *tp = tcp_sk(sk); 6618 struct request_sock *req; 6619 6620 /* If we are still handling the SYNACK RTO, see if timestamp ECR allows 6621 * undo. If peer SACKs triggered fast recovery, we can't undo here. 6622 */ 6623 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss && !tp->packets_out) 6624 tcp_try_undo_recovery(sk); 6625 6626 /* Reset rtx states to prevent spurious retransmits_timed_out() */ 6627 tcp_update_rto_time(tp); 6628 tp->retrans_stamp = 0; 6629 inet_csk(sk)->icsk_retransmits = 0; 6630 6631 /* Once we leave TCP_SYN_RECV or TCP_FIN_WAIT_1, 6632 * we no longer need req so release it. 6633 */ 6634 req = rcu_dereference_protected(tp->fastopen_rsk, 6635 lockdep_sock_is_held(sk)); 6636 reqsk_fastopen_remove(sk, req, false); 6637 6638 /* Re-arm the timer because data may have been sent out. 6639 * This is similar to the regular data transmission case 6640 * when new data has just been ack'ed. 6641 * 6642 * (TFO) - we could try to be more aggressive and 6643 * retransmitting any data sooner based on when they 6644 * are sent out. 6645 */ 6646 tcp_rearm_rto(sk); 6647 } 6648 6649 /* 6650 * This function implements the receiving procedure of RFC 793 for 6651 * all states except ESTABLISHED and TIME_WAIT. 6652 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be 6653 * address independent. 6654 */ 6655 6656 enum skb_drop_reason 6657 tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb) 6658 { 6659 struct tcp_sock *tp = tcp_sk(sk); 6660 struct inet_connection_sock *icsk = inet_csk(sk); 6661 const struct tcphdr *th = tcp_hdr(skb); 6662 struct request_sock *req; 6663 int queued = 0; 6664 SKB_DR(reason); 6665 6666 switch (sk->sk_state) { 6667 case TCP_CLOSE: 6668 SKB_DR_SET(reason, TCP_CLOSE); 6669 goto discard; 6670 6671 case TCP_LISTEN: 6672 if (th->ack) 6673 return SKB_DROP_REASON_TCP_FLAGS; 6674 6675 if (th->rst) { 6676 SKB_DR_SET(reason, TCP_RESET); 6677 goto discard; 6678 } 6679 if (th->syn) { 6680 if (th->fin) { 6681 SKB_DR_SET(reason, TCP_FLAGS); 6682 goto discard; 6683 } 6684 /* It is possible that we process SYN packets from backlog, 6685 * so we need to make sure to disable BH and RCU right there. 6686 */ 6687 rcu_read_lock(); 6688 local_bh_disable(); 6689 icsk->icsk_af_ops->conn_request(sk, skb); 6690 local_bh_enable(); 6691 rcu_read_unlock(); 6692 6693 consume_skb(skb); 6694 return 0; 6695 } 6696 SKB_DR_SET(reason, TCP_FLAGS); 6697 goto discard; 6698 6699 case TCP_SYN_SENT: 6700 tp->rx_opt.saw_tstamp = 0; 6701 tcp_mstamp_refresh(tp); 6702 queued = tcp_rcv_synsent_state_process(sk, skb, th); 6703 if (queued >= 0) 6704 return queued; 6705 6706 /* Do step6 onward by hand. */ 6707 tcp_urg(sk, skb, th); 6708 __kfree_skb(skb); 6709 tcp_data_snd_check(sk); 6710 return 0; 6711 } 6712 6713 tcp_mstamp_refresh(tp); 6714 tp->rx_opt.saw_tstamp = 0; 6715 req = rcu_dereference_protected(tp->fastopen_rsk, 6716 lockdep_sock_is_held(sk)); 6717 if (req) { 6718 bool req_stolen; 6719 6720 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV && 6721 sk->sk_state != TCP_FIN_WAIT1); 6722 6723 if (!tcp_check_req(sk, skb, req, true, &req_stolen)) { 6724 SKB_DR_SET(reason, TCP_FASTOPEN); 6725 goto discard; 6726 } 6727 } 6728 6729 if (!th->ack && !th->rst && !th->syn) { 6730 SKB_DR_SET(reason, TCP_FLAGS); 6731 goto discard; 6732 } 6733 if (!tcp_validate_incoming(sk, skb, th, 0)) 6734 return 0; 6735 6736 /* step 5: check the ACK field */ 6737 reason = tcp_ack(sk, skb, FLAG_SLOWPATH | 6738 FLAG_UPDATE_TS_RECENT | 6739 FLAG_NO_CHALLENGE_ACK); 6740 6741 if ((int)reason <= 0) { 6742 if (sk->sk_state == TCP_SYN_RECV) { 6743 /* send one RST */ 6744 if (!reason) 6745 return SKB_DROP_REASON_TCP_OLD_ACK; 6746 return -reason; 6747 } 6748 /* accept old ack during closing */ 6749 if ((int)reason < 0) { 6750 tcp_send_challenge_ack(sk); 6751 reason = -reason; 6752 goto discard; 6753 } 6754 } 6755 SKB_DR_SET(reason, NOT_SPECIFIED); 6756 switch (sk->sk_state) { 6757 case TCP_SYN_RECV: 6758 tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */ 6759 if (!tp->srtt_us) 6760 tcp_synack_rtt_meas(sk, req); 6761 6762 if (req) { 6763 tcp_rcv_synrecv_state_fastopen(sk); 6764 } else { 6765 tcp_try_undo_spurious_syn(sk); 6766 tp->retrans_stamp = 0; 6767 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB, 6768 skb); 6769 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 6770 } 6771 tcp_ao_established(sk); 6772 smp_mb(); 6773 tcp_set_state(sk, TCP_ESTABLISHED); 6774 sk->sk_state_change(sk); 6775 6776 /* Note, that this wakeup is only for marginal crossed SYN case. 6777 * Passively open sockets are not waked up, because 6778 * sk->sk_sleep == NULL and sk->sk_socket == NULL. 6779 */ 6780 if (sk->sk_socket) 6781 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT); 6782 6783 tp->snd_una = TCP_SKB_CB(skb)->ack_seq; 6784 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale; 6785 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); 6786 6787 if (tp->rx_opt.tstamp_ok) 6788 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 6789 6790 if (!inet_csk(sk)->icsk_ca_ops->cong_control) 6791 tcp_update_pacing_rate(sk); 6792 6793 /* Prevent spurious tcp_cwnd_restart() on first data packet */ 6794 tp->lsndtime = tcp_jiffies32; 6795 6796 tcp_initialize_rcv_mss(sk); 6797 tcp_fast_path_on(tp); 6798 if (sk->sk_shutdown & SEND_SHUTDOWN) 6799 tcp_shutdown(sk, SEND_SHUTDOWN); 6800 break; 6801 6802 case TCP_FIN_WAIT1: { 6803 int tmo; 6804 6805 if (req) 6806 tcp_rcv_synrecv_state_fastopen(sk); 6807 6808 if (tp->snd_una != tp->write_seq) 6809 break; 6810 6811 tcp_set_state(sk, TCP_FIN_WAIT2); 6812 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | SEND_SHUTDOWN); 6813 6814 sk_dst_confirm(sk); 6815 6816 if (!sock_flag(sk, SOCK_DEAD)) { 6817 /* Wake up lingering close() */ 6818 sk->sk_state_change(sk); 6819 break; 6820 } 6821 6822 if (READ_ONCE(tp->linger2) < 0) { 6823 tcp_done(sk); 6824 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 6825 return SKB_DROP_REASON_TCP_ABORT_ON_DATA; 6826 } 6827 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 6828 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { 6829 /* Receive out of order FIN after close() */ 6830 if (tp->syn_fastopen && th->fin) 6831 tcp_fastopen_active_disable(sk); 6832 tcp_done(sk); 6833 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 6834 return SKB_DROP_REASON_TCP_ABORT_ON_DATA; 6835 } 6836 6837 tmo = tcp_fin_time(sk); 6838 if (tmo > TCP_TIMEWAIT_LEN) { 6839 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN); 6840 } else if (th->fin || sock_owned_by_user(sk)) { 6841 /* Bad case. We could lose such FIN otherwise. 6842 * It is not a big problem, but it looks confusing 6843 * and not so rare event. We still can lose it now, 6844 * if it spins in bh_lock_sock(), but it is really 6845 * marginal case. 6846 */ 6847 inet_csk_reset_keepalive_timer(sk, tmo); 6848 } else { 6849 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 6850 goto consume; 6851 } 6852 break; 6853 } 6854 6855 case TCP_CLOSING: 6856 if (tp->snd_una == tp->write_seq) { 6857 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 6858 goto consume; 6859 } 6860 break; 6861 6862 case TCP_LAST_ACK: 6863 if (tp->snd_una == tp->write_seq) { 6864 tcp_update_metrics(sk); 6865 tcp_done(sk); 6866 goto consume; 6867 } 6868 break; 6869 } 6870 6871 /* step 6: check the URG bit */ 6872 tcp_urg(sk, skb, th); 6873 6874 /* step 7: process the segment text */ 6875 switch (sk->sk_state) { 6876 case TCP_CLOSE_WAIT: 6877 case TCP_CLOSING: 6878 case TCP_LAST_ACK: 6879 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 6880 /* If a subflow has been reset, the packet should not 6881 * continue to be processed, drop the packet. 6882 */ 6883 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb)) 6884 goto discard; 6885 break; 6886 } 6887 fallthrough; 6888 case TCP_FIN_WAIT1: 6889 case TCP_FIN_WAIT2: 6890 /* RFC 793 says to queue data in these states, 6891 * RFC 1122 says we MUST send a reset. 6892 * BSD 4.4 also does reset. 6893 */ 6894 if (sk->sk_shutdown & RCV_SHUTDOWN) { 6895 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 6896 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { 6897 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 6898 tcp_reset(sk, skb); 6899 return SKB_DROP_REASON_TCP_ABORT_ON_DATA; 6900 } 6901 } 6902 fallthrough; 6903 case TCP_ESTABLISHED: 6904 tcp_data_queue(sk, skb); 6905 queued = 1; 6906 break; 6907 } 6908 6909 /* tcp_data could move socket to TIME-WAIT */ 6910 if (sk->sk_state != TCP_CLOSE) { 6911 tcp_data_snd_check(sk); 6912 tcp_ack_snd_check(sk); 6913 } 6914 6915 if (!queued) { 6916 discard: 6917 tcp_drop_reason(sk, skb, reason); 6918 } 6919 return 0; 6920 6921 consume: 6922 __kfree_skb(skb); 6923 return 0; 6924 } 6925 EXPORT_SYMBOL(tcp_rcv_state_process); 6926 6927 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family) 6928 { 6929 struct inet_request_sock *ireq = inet_rsk(req); 6930 6931 if (family == AF_INET) 6932 net_dbg_ratelimited("drop open request from %pI4/%u\n", 6933 &ireq->ir_rmt_addr, port); 6934 #if IS_ENABLED(CONFIG_IPV6) 6935 else if (family == AF_INET6) 6936 net_dbg_ratelimited("drop open request from %pI6/%u\n", 6937 &ireq->ir_v6_rmt_addr, port); 6938 #endif 6939 } 6940 6941 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set 6942 * 6943 * If we receive a SYN packet with these bits set, it means a 6944 * network is playing bad games with TOS bits. In order to 6945 * avoid possible false congestion notifications, we disable 6946 * TCP ECN negotiation. 6947 * 6948 * Exception: tcp_ca wants ECN. This is required for DCTCP 6949 * congestion control: Linux DCTCP asserts ECT on all packets, 6950 * including SYN, which is most optimal solution; however, 6951 * others, such as FreeBSD do not. 6952 * 6953 * Exception: At least one of the reserved bits of the TCP header (th->res1) is 6954 * set, indicating the use of a future TCP extension (such as AccECN). See 6955 * RFC8311 §4.3 which updates RFC3168 to allow the development of such 6956 * extensions. 6957 */ 6958 static void tcp_ecn_create_request(struct request_sock *req, 6959 const struct sk_buff *skb, 6960 const struct sock *listen_sk, 6961 const struct dst_entry *dst) 6962 { 6963 const struct tcphdr *th = tcp_hdr(skb); 6964 const struct net *net = sock_net(listen_sk); 6965 bool th_ecn = th->ece && th->cwr; 6966 bool ect, ecn_ok; 6967 u32 ecn_ok_dst; 6968 6969 if (!th_ecn) 6970 return; 6971 6972 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield); 6973 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK); 6974 ecn_ok = READ_ONCE(net->ipv4.sysctl_tcp_ecn) || ecn_ok_dst; 6975 6976 if (((!ect || th->res1) && ecn_ok) || tcp_ca_needs_ecn(listen_sk) || 6977 (ecn_ok_dst & DST_FEATURE_ECN_CA) || 6978 tcp_bpf_ca_needs_ecn((struct sock *)req)) 6979 inet_rsk(req)->ecn_ok = 1; 6980 } 6981 6982 static void tcp_openreq_init(struct request_sock *req, 6983 const struct tcp_options_received *rx_opt, 6984 struct sk_buff *skb, const struct sock *sk) 6985 { 6986 struct inet_request_sock *ireq = inet_rsk(req); 6987 6988 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */ 6989 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq; 6990 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 6991 tcp_rsk(req)->snt_synack = 0; 6992 tcp_rsk(req)->last_oow_ack_time = 0; 6993 req->mss = rx_opt->mss_clamp; 6994 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0; 6995 ireq->tstamp_ok = rx_opt->tstamp_ok; 6996 ireq->sack_ok = rx_opt->sack_ok; 6997 ireq->snd_wscale = rx_opt->snd_wscale; 6998 ireq->wscale_ok = rx_opt->wscale_ok; 6999 ireq->acked = 0; 7000 ireq->ecn_ok = 0; 7001 ireq->ir_rmt_port = tcp_hdr(skb)->source; 7002 ireq->ir_num = ntohs(tcp_hdr(skb)->dest); 7003 ireq->ir_mark = inet_request_mark(sk, skb); 7004 #if IS_ENABLED(CONFIG_SMC) 7005 ireq->smc_ok = rx_opt->smc_ok && !(tcp_sk(sk)->smc_hs_congested && 7006 tcp_sk(sk)->smc_hs_congested(sk)); 7007 #endif 7008 } 7009 7010 /* 7011 * Return true if a syncookie should be sent 7012 */ 7013 static bool tcp_syn_flood_action(struct sock *sk, const char *proto) 7014 { 7015 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue; 7016 const char *msg = "Dropping request"; 7017 struct net *net = sock_net(sk); 7018 bool want_cookie = false; 7019 u8 syncookies; 7020 7021 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies); 7022 7023 #ifdef CONFIG_SYN_COOKIES 7024 if (syncookies) { 7025 msg = "Sending cookies"; 7026 want_cookie = true; 7027 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES); 7028 } else 7029 #endif 7030 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP); 7031 7032 if (!READ_ONCE(queue->synflood_warned) && syncookies != 2 && 7033 xchg(&queue->synflood_warned, 1) == 0) { 7034 if (IS_ENABLED(CONFIG_IPV6) && sk->sk_family == AF_INET6) { 7035 net_info_ratelimited("%s: Possible SYN flooding on port [%pI6c]:%u. %s.\n", 7036 proto, inet6_rcv_saddr(sk), 7037 sk->sk_num, msg); 7038 } else { 7039 net_info_ratelimited("%s: Possible SYN flooding on port %pI4:%u. %s.\n", 7040 proto, &sk->sk_rcv_saddr, 7041 sk->sk_num, msg); 7042 } 7043 } 7044 7045 return want_cookie; 7046 } 7047 7048 static void tcp_reqsk_record_syn(const struct sock *sk, 7049 struct request_sock *req, 7050 const struct sk_buff *skb) 7051 { 7052 if (tcp_sk(sk)->save_syn) { 7053 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb); 7054 struct saved_syn *saved_syn; 7055 u32 mac_hdrlen; 7056 void *base; 7057 7058 if (tcp_sk(sk)->save_syn == 2) { /* Save full header. */ 7059 base = skb_mac_header(skb); 7060 mac_hdrlen = skb_mac_header_len(skb); 7061 len += mac_hdrlen; 7062 } else { 7063 base = skb_network_header(skb); 7064 mac_hdrlen = 0; 7065 } 7066 7067 saved_syn = kmalloc(struct_size(saved_syn, data, len), 7068 GFP_ATOMIC); 7069 if (saved_syn) { 7070 saved_syn->mac_hdrlen = mac_hdrlen; 7071 saved_syn->network_hdrlen = skb_network_header_len(skb); 7072 saved_syn->tcp_hdrlen = tcp_hdrlen(skb); 7073 memcpy(saved_syn->data, base, len); 7074 req->saved_syn = saved_syn; 7075 } 7076 } 7077 } 7078 7079 /* If a SYN cookie is required and supported, returns a clamped MSS value to be 7080 * used for SYN cookie generation. 7081 */ 7082 u16 tcp_get_syncookie_mss(struct request_sock_ops *rsk_ops, 7083 const struct tcp_request_sock_ops *af_ops, 7084 struct sock *sk, struct tcphdr *th) 7085 { 7086 struct tcp_sock *tp = tcp_sk(sk); 7087 u16 mss; 7088 7089 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_syncookies) != 2 && 7090 !inet_csk_reqsk_queue_is_full(sk)) 7091 return 0; 7092 7093 if (!tcp_syn_flood_action(sk, rsk_ops->slab_name)) 7094 return 0; 7095 7096 if (sk_acceptq_is_full(sk)) { 7097 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS); 7098 return 0; 7099 } 7100 7101 mss = tcp_parse_mss_option(th, tp->rx_opt.user_mss); 7102 if (!mss) 7103 mss = af_ops->mss_clamp; 7104 7105 return mss; 7106 } 7107 EXPORT_SYMBOL_GPL(tcp_get_syncookie_mss); 7108 7109 int tcp_conn_request(struct request_sock_ops *rsk_ops, 7110 const struct tcp_request_sock_ops *af_ops, 7111 struct sock *sk, struct sk_buff *skb) 7112 { 7113 struct tcp_fastopen_cookie foc = { .len = -1 }; 7114 struct tcp_options_received tmp_opt; 7115 struct tcp_sock *tp = tcp_sk(sk); 7116 struct net *net = sock_net(sk); 7117 struct sock *fastopen_sk = NULL; 7118 struct request_sock *req; 7119 bool want_cookie = false; 7120 struct dst_entry *dst; 7121 struct flowi fl; 7122 u8 syncookies; 7123 u32 isn; 7124 7125 #ifdef CONFIG_TCP_AO 7126 const struct tcp_ao_hdr *aoh; 7127 #endif 7128 7129 isn = __this_cpu_read(tcp_tw_isn); 7130 if (isn) { 7131 /* TW buckets are converted to open requests without 7132 * limitations, they conserve resources and peer is 7133 * evidently real one. 7134 */ 7135 __this_cpu_write(tcp_tw_isn, 0); 7136 } else { 7137 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies); 7138 7139 if (syncookies == 2 || inet_csk_reqsk_queue_is_full(sk)) { 7140 want_cookie = tcp_syn_flood_action(sk, 7141 rsk_ops->slab_name); 7142 if (!want_cookie) 7143 goto drop; 7144 } 7145 } 7146 7147 if (sk_acceptq_is_full(sk)) { 7148 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS); 7149 goto drop; 7150 } 7151 7152 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie); 7153 if (!req) 7154 goto drop; 7155 7156 req->syncookie = want_cookie; 7157 tcp_rsk(req)->af_specific = af_ops; 7158 tcp_rsk(req)->ts_off = 0; 7159 tcp_rsk(req)->req_usec_ts = false; 7160 #if IS_ENABLED(CONFIG_MPTCP) 7161 tcp_rsk(req)->is_mptcp = 0; 7162 #endif 7163 7164 tcp_clear_options(&tmp_opt); 7165 tmp_opt.mss_clamp = af_ops->mss_clamp; 7166 tmp_opt.user_mss = tp->rx_opt.user_mss; 7167 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0, 7168 want_cookie ? NULL : &foc); 7169 7170 if (want_cookie && !tmp_opt.saw_tstamp) 7171 tcp_clear_options(&tmp_opt); 7172 7173 if (IS_ENABLED(CONFIG_SMC) && want_cookie) 7174 tmp_opt.smc_ok = 0; 7175 7176 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp; 7177 tcp_openreq_init(req, &tmp_opt, skb, sk); 7178 inet_rsk(req)->no_srccheck = inet_test_bit(TRANSPARENT, sk); 7179 7180 /* Note: tcp_v6_init_req() might override ir_iif for link locals */ 7181 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb); 7182 7183 dst = af_ops->route_req(sk, skb, &fl, req, isn); 7184 if (!dst) 7185 goto drop_and_free; 7186 7187 if (tmp_opt.tstamp_ok) { 7188 tcp_rsk(req)->req_usec_ts = dst_tcp_usec_ts(dst); 7189 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb); 7190 } 7191 if (!want_cookie && !isn) { 7192 int max_syn_backlog = READ_ONCE(net->ipv4.sysctl_max_syn_backlog); 7193 7194 /* Kill the following clause, if you dislike this way. */ 7195 if (!syncookies && 7196 (max_syn_backlog - inet_csk_reqsk_queue_len(sk) < 7197 (max_syn_backlog >> 2)) && 7198 !tcp_peer_is_proven(req, dst)) { 7199 /* Without syncookies last quarter of 7200 * backlog is filled with destinations, 7201 * proven to be alive. 7202 * It means that we continue to communicate 7203 * to destinations, already remembered 7204 * to the moment of synflood. 7205 */ 7206 pr_drop_req(req, ntohs(tcp_hdr(skb)->source), 7207 rsk_ops->family); 7208 goto drop_and_release; 7209 } 7210 7211 isn = af_ops->init_seq(skb); 7212 } 7213 7214 tcp_ecn_create_request(req, skb, sk, dst); 7215 7216 if (want_cookie) { 7217 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss); 7218 if (!tmp_opt.tstamp_ok) 7219 inet_rsk(req)->ecn_ok = 0; 7220 } 7221 7222 #ifdef CONFIG_TCP_AO 7223 if (tcp_parse_auth_options(tcp_hdr(skb), NULL, &aoh)) 7224 goto drop_and_release; /* Invalid TCP options */ 7225 if (aoh) { 7226 tcp_rsk(req)->used_tcp_ao = true; 7227 tcp_rsk(req)->ao_rcv_next = aoh->keyid; 7228 tcp_rsk(req)->ao_keyid = aoh->rnext_keyid; 7229 7230 } else { 7231 tcp_rsk(req)->used_tcp_ao = false; 7232 } 7233 #endif 7234 tcp_rsk(req)->snt_isn = isn; 7235 tcp_rsk(req)->txhash = net_tx_rndhash(); 7236 tcp_rsk(req)->syn_tos = TCP_SKB_CB(skb)->ip_dsfield; 7237 tcp_openreq_init_rwin(req, sk, dst); 7238 sk_rx_queue_set(req_to_sk(req), skb); 7239 if (!want_cookie) { 7240 tcp_reqsk_record_syn(sk, req, skb); 7241 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst); 7242 } 7243 if (fastopen_sk) { 7244 af_ops->send_synack(fastopen_sk, dst, &fl, req, 7245 &foc, TCP_SYNACK_FASTOPEN, skb); 7246 /* Add the child socket directly into the accept queue */ 7247 if (!inet_csk_reqsk_queue_add(sk, req, fastopen_sk)) { 7248 reqsk_fastopen_remove(fastopen_sk, req, false); 7249 bh_unlock_sock(fastopen_sk); 7250 sock_put(fastopen_sk); 7251 goto drop_and_free; 7252 } 7253 sk->sk_data_ready(sk); 7254 bh_unlock_sock(fastopen_sk); 7255 sock_put(fastopen_sk); 7256 } else { 7257 tcp_rsk(req)->tfo_listener = false; 7258 if (!want_cookie) { 7259 req->timeout = tcp_timeout_init((struct sock *)req); 7260 if (unlikely(!inet_csk_reqsk_queue_hash_add(sk, req, 7261 req->timeout))) { 7262 reqsk_free(req); 7263 return 0; 7264 } 7265 7266 } 7267 af_ops->send_synack(sk, dst, &fl, req, &foc, 7268 !want_cookie ? TCP_SYNACK_NORMAL : 7269 TCP_SYNACK_COOKIE, 7270 skb); 7271 if (want_cookie) { 7272 reqsk_free(req); 7273 return 0; 7274 } 7275 } 7276 reqsk_put(req); 7277 return 0; 7278 7279 drop_and_release: 7280 dst_release(dst); 7281 drop_and_free: 7282 __reqsk_free(req); 7283 drop: 7284 tcp_listendrop(sk); 7285 return 0; 7286 } 7287 EXPORT_SYMBOL(tcp_conn_request); 7288